├── .github ├── ISSUE_TEMPLATE.md ├── ISSUE_TEMPLATE │ ├── bug_report.md │ ├── documentation.md │ ├── feature_request.md │ └── how-to-question.md ├── PULL_REQUEST_TEMPLATE.md ├── stale.yml └── workflows │ ├── build.yml │ └── build_wheels.yml ├── .gitignore ├── .gitmodules ├── CODE_OF_CONDUCT.md ├── CONTRIBUTING.md ├── LICENSE ├── README.md ├── data └── scripts │ └── apply_spm.py ├── docs ├── Makefile ├── _static │ └── theme_overrides.css ├── command_line_tools.rst ├── conf.py ├── criterions.rst ├── data.rst ├── docutils.conf ├── fairseq.gif ├── getting_started.rst ├── hydra_integration.md ├── index.rst ├── lr_scheduler.rst ├── make.bat ├── models.rst ├── modules.rst ├── optim.rst ├── overview.rst ├── requirements.txt ├── tasks.rst ├── tutorial_classifying_names.rst └── tutorial_simple_lstm.rst ├── examples ├── .gitignore ├── __init__.py ├── adaptive_span │ ├── README.md │ ├── __init__.py │ ├── adagrad_with_grad_clip.py │ ├── adaptive_span_attention.py │ ├── adaptive_span_loss.py │ ├── adaptive_span_model.py │ ├── adaptive_span_model_wrapper.py │ └── truncated_bptt_lm_task.py ├── backtranslation │ ├── README.md │ ├── deduplicate_lines.py │ ├── extract_bt_data.py │ ├── prepare-de-monolingual.sh │ ├── prepare-wmt18en2de.sh │ ├── sacrebleu.sh │ └── tokenized_bleu.sh ├── bart │ ├── README.glue.md │ ├── README.md │ ├── README.summarization.md │ └── summarize.py ├── byte_level_bpe │ ├── README.md │ ├── get_bitext.py │ ├── get_data.sh │ └── gru_transformer.py ├── camembert │ └── README.md ├── constrained_decoding │ ├── README.md │ ├── normalize.py │ └── tok.py ├── conv_seq2seq │ └── README.md ├── criss │ ├── README.md │ ├── download_and_preprocess_flores_test.sh │ ├── download_and_preprocess_tatoeba.sh │ ├── mining │ │ ├── mine.py │ │ └── mine_example.sh │ ├── save_encoder.py │ ├── sentence_retrieval │ │ ├── encoder_analysis.py │ │ └── sentence_retrieval_tatoeba.sh │ └── unsupervised_mt │ │ └── eval.sh ├── cross_lingual_language_model │ └── README.md ├── discriminative_reranking_nmt │ ├── README.md │ ├── __init__.py │ ├── config │ │ └── deen.yaml │ ├── criterions │ │ ├── __init__.py │ │ └── discriminative_reranking_criterion.py │ ├── drnmt_rerank.py │ ├── models │ │ ├── __init__.py │ │ └── discriminative_reranking_model.py │ ├── scripts │ │ └── prep_data.py │ └── tasks │ │ ├── __init__.py │ │ └── discriminative_reranking_task.py ├── fast_noisy_channel │ ├── README.md │ ├── __init__.py │ ├── noisy_channel_beam_search.py │ ├── noisy_channel_sequence_generator.py │ └── noisy_channel_translation.py ├── flores101 │ └── README.md ├── fully_sharded_data_parallel │ └── README.md ├── gottbert │ └── README.md ├── hubert │ ├── README.md │ ├── config │ │ ├── decode │ │ │ ├── ax_sweep │ │ │ │ ├── ngram.yaml │ │ │ │ └── transformer.yaml │ │ │ ├── infer_fsqlm.yaml │ │ │ ├── infer_kenlm.yaml │ │ │ ├── infer_viterbi.yaml │ │ │ └── run │ │ │ │ ├── submitit_slurm.yaml │ │ │ │ └── submitit_slurm_8gpu.yaml │ │ ├── finetune │ │ │ ├── base_10h.yaml │ │ │ ├── ckpt │ │ │ │ └── it1.yaml │ │ │ ├── lm │ │ │ │ └── ls_4gram.yaml │ │ │ └── run │ │ │ │ └── submitit_reg.yaml │ │ └── pretrain │ │ │ ├── data │ │ │ ├── iter1.yaml │ │ │ └── iter2.yaml │ │ │ ├── hubert_base_librispeech.yaml │ │ │ ├── hubert_large_librivox.yaml │ │ │ ├── hubert_xlarge_librivox.yaml │ │ │ └── run │ │ │ └── submitit_reg.yaml │ ├── measure_teacher_quality.py │ ├── simple_kmeans │ │ ├── README.md │ │ ├── dump_hubert_feature.py │ │ ├── dump_hubert_feature_s2t.py │ │ ├── dump_km_label.py │ │ ├── dump_mfcc_feature.py │ │ ├── dump_w2v2_feature.py │ │ ├── feature_utils.py │ │ └── learn_kmeans.py │ └── update_ckpt.py ├── joint_alignment_translation │ ├── README.md │ └── prepare-wmt18en2de_no_norm_no_escape_no_agressive.sh ├── language_model │ ├── README.adaptive_inputs.md │ ├── README.conv.md │ ├── README.md │ └── prepare-wikitext-103.sh ├── laser │ ├── README.md │ └── laser_src │ │ ├── __init__.py │ │ ├── laser_lstm.py │ │ ├── laser_task.py │ │ ├── laser_transformer.py │ │ └── multitask_data_utils.py ├── latent_depth │ ├── README.md │ └── latent_depth_src │ │ ├── __init__.py │ │ ├── loss │ │ ├── __init__.py │ │ └── latent_depth.py │ │ ├── models │ │ ├── __init__.py │ │ ├── latent_multilingual_transformer.py │ │ └── latent_transformer.py │ │ ├── modules │ │ ├── __init__.py │ │ └── latent_layers.py │ │ └── multilingual_translation_latent_depth.py ├── layerdrop │ └── README.md ├── linformer │ ├── README.md │ └── linformer_src │ │ ├── __init__.py │ │ ├── models │ │ ├── __init__.py │ │ └── linformer_roberta.py │ │ └── modules │ │ ├── __init__.py │ │ ├── linformer_sentence_encoder.py │ │ ├── linformer_sentence_encoder_layer.py │ │ └── multihead_linear_attention.py ├── m2m_100 │ ├── README.md │ ├── install_dependecies.sh │ ├── process_data │ │ ├── clean_histogram.py │ │ ├── dedup_data.py │ │ └── remove_too_much_punc.py │ ├── tok.sh │ └── tokenizers │ │ ├── README.md │ │ ├── seg_ja.sh │ │ ├── seg_ko.sh │ │ ├── thirdparty │ │ └── .gitignore │ │ ├── tokenize_indic.py │ │ ├── tokenize_thai.py │ │ ├── tokenize_zh.py │ │ └── tokenizer_ar.sh ├── mbart │ └── README.md ├── megatron_11b │ ├── README.md │ └── detok.py ├── multilingual │ ├── ML50_langs.txt │ ├── README.md │ ├── data_scripts │ │ ├── README.md │ │ ├── binarize.py │ │ ├── check_iswlt_test_data.py │ │ ├── check_self_overlaps.py │ │ ├── check_valid_test_overlaps.py │ │ ├── dedup_all.py │ │ ├── download_ML50_v1.sh │ │ ├── download_af_xh.sh │ │ ├── download_flores_data.sh │ │ ├── download_iitb.sh │ │ ├── download_iwslt_and_extract.sh │ │ ├── download_lotus.sh │ │ ├── download_ted_and_extract.py │ │ ├── download_wat19_my.sh │ │ ├── download_wmt19_and_before.py │ │ ├── download_wmt20.sh │ │ ├── preprocess_ML50_v1.sh │ │ ├── remove_valid_test_in_train.py │ │ ├── requirement.txt │ │ └── utils │ │ │ ├── dedup.py │ │ │ ├── fasttext_multi_filter.py │ │ │ └── strip_sgm.sh │ ├── finetune_multilingual_model.sh │ ├── multilingual_fairseq_gen.sh │ └── train_multilingual_model.sh ├── noisychannel │ ├── README.md │ ├── __init__.py │ ├── rerank.py │ ├── rerank_generate.py │ ├── rerank_options.py │ ├── rerank_score_bw.py │ ├── rerank_score_lm.py │ ├── rerank_tune.py │ └── rerank_utils.py ├── nonautoregressive_translation │ ├── README.md │ └── scripts.md ├── paraphraser │ ├── README.md │ └── paraphrase.py ├── pay_less_attention_paper │ └── README.md ├── pointer_generator │ ├── README.md │ ├── README.xsum.md │ ├── pointer_generator_src │ │ ├── __init__.py │ │ └── transformer_pg.py │ ├── postprocess.py │ └── preprocess.py ├── quant_noise │ ├── README.md │ └── transformer_quantization_config.yaml ├── roberta │ ├── README.custom_classification.md │ ├── README.glue.md │ ├── README.md │ ├── README.pretraining.md │ ├── README.race.md │ ├── commonsense_qa │ │ ├── README.md │ │ ├── __init__.py │ │ ├── commonsense_qa_task.py │ │ └── download_cqa_data.sh │ ├── config │ │ ├── finetuning │ │ │ ├── cola.yaml │ │ │ ├── mnli.yaml │ │ │ ├── mrpc.yaml │ │ │ ├── qnli.yaml │ │ │ ├── qqp.yaml │ │ │ ├── rte.yaml │ │ │ ├── sst_2.yaml │ │ │ └── sts_b.yaml │ │ └── pretraining │ │ │ └── base.yaml │ ├── multiprocessing_bpe_encoder.py │ ├── preprocess_GLUE_tasks.sh │ ├── preprocess_RACE.py │ ├── preprocess_RACE.sh │ └── wsc │ │ ├── README.md │ │ ├── __init__.py │ │ ├── wsc_criterion.py │ │ ├── wsc_task.py │ │ └── wsc_utils.py ├── rxf │ ├── README.md │ ├── __init__.py │ └── rxf_src │ │ ├── __init__.py │ │ ├── label_smoothed_cross_entropy_r3f.py │ │ └── sentence_prediction_r3f.py ├── scaling_nmt │ └── README.md ├── simultaneous_translation │ ├── README.md │ ├── __init__.py │ ├── docs │ │ ├── ende-mma.md │ │ └── enja-waitk.md │ ├── eval │ │ └── agents │ │ │ └── simul_t2t_enja.py │ ├── models │ │ ├── __init__.py │ │ ├── convtransformer_simul_trans.py │ │ └── transformer_monotonic_attention.py │ ├── modules │ │ ├── __init__.py │ │ ├── fixed_pre_decision.py │ │ ├── monotonic_multihead_attention.py │ │ └── monotonic_transformer_layer.py │ └── utils │ │ ├── __init__.py │ │ ├── data_utils.py │ │ ├── functions.py │ │ ├── latency.py │ │ └── p_choose_strategy.py ├── speech_recognition │ ├── README.md │ ├── __init__.py │ ├── criterions │ │ ├── ASG_loss.py │ │ ├── __init__.py │ │ └── cross_entropy_acc.py │ ├── data │ │ ├── __init__.py │ │ ├── asr_dataset.py │ │ ├── collaters.py │ │ ├── data_utils.py │ │ └── replabels.py │ ├── datasets │ │ ├── asr_prep_json.py │ │ └── prepare-librispeech.sh │ ├── infer.py │ ├── kaldi │ │ ├── __init__.py │ │ ├── add-self-loop-simple.cc │ │ ├── config │ │ │ └── kaldi_initializer.yaml │ │ ├── kaldi_decoder.py │ │ └── kaldi_initializer.py │ ├── models │ │ ├── __init__.py │ │ ├── vggtransformer.py │ │ └── w2l_conv_glu_enc.py │ ├── new │ │ ├── README.md │ │ ├── __init__.py │ │ ├── conf │ │ │ ├── hydra │ │ │ │ └── sweeper │ │ │ │ │ └── ax.yaml │ │ │ └── infer.yaml │ │ ├── decoders │ │ │ ├── __init__.py │ │ │ ├── base_decoder.py │ │ │ ├── decoder.py │ │ │ ├── decoder_config.py │ │ │ ├── flashlight_decoder.py │ │ │ └── viterbi_decoder.py │ │ └── infer.py │ ├── tasks │ │ ├── __init__.py │ │ └── speech_recognition.py │ ├── utils │ │ └── wer_utils.py │ └── w2l_decoder.py ├── speech_to_text │ ├── README.md │ ├── data_utils.py │ ├── docs │ │ ├── covost_example.md │ │ ├── librispeech_example.md │ │ ├── mtedx_example.md │ │ ├── mustc_example.md │ │ └── simulst_mustc_example.md │ ├── hdfs_utils.py │ ├── learn_dict.py │ ├── learn_dict_raw.py │ ├── prep_covost_data.py │ ├── prep_librispeech_data.py │ ├── prep_librispeech_data_raw.py │ ├── prep_mtedx_data.py │ ├── prep_mustc_data.py │ ├── prep_mustc_data_raw.py │ ├── seg_mustc_data.py │ ├── simultaneous_translation │ │ └── agents │ │ │ └── fairseq_simul_st_agent.py │ └── tmp.py ├── stories │ └── README.md ├── translation │ ├── README.md │ ├── prepare-iwslt14.sh │ ├── prepare-iwslt17-multilingual.sh │ ├── prepare-wmt14en2de.sh │ └── prepare-wmt14en2fr.sh ├── translation_moe │ ├── README.md │ ├── score.py │ └── translation_moe_src │ │ ├── __init__.py │ │ ├── logsumexp_moe.py │ │ ├── mean_pool_gating_network.py │ │ └── translation_moe.py ├── truncated_bptt │ ├── README.md │ ├── __init__.py │ ├── transformer_xl_model.py │ └── truncated_bptt_lm_task.py ├── unsupervised_quality_estimation │ ├── README.md │ ├── aggregate_scores.py │ ├── meteor.py │ └── repeat_lines.py ├── wav2vec │ ├── README.md │ ├── __init__.py │ ├── config │ │ ├── finetuning │ │ │ ├── base_100h.yaml │ │ │ ├── base_10h.yaml │ │ │ ├── base_10m.yaml │ │ │ ├── base_1h.yaml │ │ │ ├── base_960h.yaml │ │ │ ├── vox_100h.yaml │ │ │ ├── vox_10h.yaml │ │ │ ├── vox_10m.yaml │ │ │ ├── vox_1h.yaml │ │ │ └── vox_960h.yaml │ │ └── pretraining │ │ │ ├── wav2vec2_base_librispeech.yaml │ │ │ ├── wav2vec2_large_librivox.yaml │ │ │ ├── wav2vec2_large_librivox_tpu-pod.yaml │ │ │ └── wav2vec2_large_librivox_tpu.yaml │ ├── libri_labels.py │ ├── scripts │ │ └── binarize_manifest.sh │ ├── unsupervised │ │ ├── README.md │ │ ├── __init__.py │ │ ├── config │ │ │ ├── finetuning │ │ │ │ └── w2v_finetune.yaml │ │ │ ├── gan │ │ │ │ └── w2vu.yaml │ │ │ ├── generate │ │ │ │ └── viterbi.yaml │ │ │ ├── timit_matched │ │ │ │ ├── test.uid │ │ │ │ ├── train.uid │ │ │ │ ├── train_text.uid │ │ │ │ └── valid.uid │ │ │ └── timit_unmatched │ │ │ │ ├── test.uid │ │ │ │ ├── train.uid │ │ │ │ ├── train_text.uid │ │ │ │ └── valid.uid │ │ ├── data │ │ │ ├── __init__.py │ │ │ ├── extracted_features_dataset.py │ │ │ └── random_input_dataset.py │ │ ├── kaldi_self_train │ │ │ ├── README.md │ │ │ └── st │ │ │ │ ├── cmd.sh │ │ │ │ ├── decode_phone.sh │ │ │ │ ├── decode_word_step1.sh │ │ │ │ ├── decode_word_step2.sh │ │ │ │ ├── local │ │ │ │ ├── copy_aligned_text.py │ │ │ │ ├── decode.sh │ │ │ │ ├── prepare_data_from_w2v.py │ │ │ │ ├── prepare_lang.sh │ │ │ │ ├── prepare_lang_word.sh │ │ │ │ ├── prepare_lm.sh │ │ │ │ ├── score.sh │ │ │ │ ├── show_wer.sh │ │ │ │ ├── train_subset_lgbeam.sh │ │ │ │ ├── unsup_select.py │ │ │ │ ├── unsup_select_decode.sh │ │ │ │ └── unsup_select_decode_word.sh │ │ │ │ ├── path.sh │ │ │ │ ├── steps │ │ │ │ ├── steps_gan │ │ │ │ ├── train_deltas.sh │ │ │ │ ├── train_lda_mllt.sh │ │ │ │ └── train_sat.sh │ │ │ │ ├── train.sh │ │ │ │ └── utils │ │ ├── models │ │ │ ├── __init__.py │ │ │ └── wav2vec_u.py │ │ ├── scripts │ │ │ ├── apply_pca.py │ │ │ ├── copy_labels.py │ │ │ ├── filter_lexicon.py │ │ │ ├── filter_tsv.py │ │ │ ├── g2p_wrd_to_phn.py │ │ │ ├── ltr_to_wrd.py │ │ │ ├── mean_pool.py │ │ │ ├── merge_clusters.py │ │ │ ├── normalize_and_filter_text.py │ │ │ ├── normalize_text.py │ │ │ ├── pca.py │ │ │ ├── phonemize_with_sil.py │ │ │ ├── prepare_audio.sh │ │ │ ├── prepare_text.sh │ │ │ ├── prepare_timit.sh │ │ │ ├── remove_silence.py │ │ │ ├── vads.py │ │ │ ├── wav2vec_apply_cluster_faiss.py │ │ │ ├── wav2vec_cluster_faiss.py │ │ │ ├── wav2vec_extract_features.py │ │ │ ├── wer.py │ │ │ └── wrd_to_ltr.py │ │ ├── tasks │ │ │ ├── __init__.py │ │ │ └── unpaired_audio_text.py │ │ └── w2vu_generate.py │ ├── vq-wav2vec_featurize.py │ ├── wav2vec_featurize.py │ └── wav2vec_manifest.py ├── wmt19 │ └── README.md ├── wmt20 │ └── README.md └── xlmr │ └── README.md ├── fairseq ├── __init__.py ├── benchmark │ ├── __init__.py │ ├── dummy_dataset.py │ ├── dummy_lm.py │ ├── dummy_masked_lm.py │ ├── dummy_model.py │ └── dummy_mt.py ├── binarizer.py ├── checkpoint_utils.py ├── clib │ ├── cuda │ │ ├── ngram_repeat_block_cuda.cpp │ │ └── ngram_repeat_block_cuda_kernel.cu │ ├── libbase │ │ └── balanced_assignment.cpp │ ├── libbleu │ │ ├── libbleu.cpp │ │ └── module.cpp │ ├── libnat │ │ └── edit_dist.cpp │ └── libnat_cuda │ │ ├── binding.cpp │ │ ├── edit_dist.cu │ │ └── edit_dist.h ├── config │ ├── __init__.py │ ├── config.yaml │ └── model │ │ ├── transformer_lm │ │ ├── transformer_lm_baevski_gbw.yaml │ │ ├── transformer_lm_baevski_wiki103.yaml │ │ ├── transformer_lm_big.yaml │ │ ├── transformer_lm_gbw.yaml │ │ ├── transformer_lm_gpt.yaml │ │ ├── transformer_lm_gpt2_big.yaml │ │ ├── transformer_lm_gpt2_medium.yaml │ │ ├── transformer_lm_gpt2_small.yaml │ │ └── transformer_lm_wiki103.yaml │ │ ├── wav2vec │ │ └── vq_wav2vec_gumbel.yaml │ │ └── wav2vec2 │ │ ├── wav2vec2_base.yaml │ │ └── wav2vec2_large.yaml ├── criterions │ ├── __init__.py │ ├── adaptive_loss.py │ ├── composite_loss.py │ ├── cross_entropy.py │ ├── ctc.py │ ├── fairseq_criterion.py │ ├── hubert_criterion.py │ ├── label_smoothed_cross_entropy.py │ ├── label_smoothed_cross_entropy_with_stmm_self_learning.py │ ├── legacy_masked_lm.py │ ├── masked_lm.py │ ├── model_criterion.py │ ├── nat_loss.py │ ├── sentence_prediction.py │ ├── sentence_ranking.py │ └── wav2vec_criterion.py ├── data │ ├── __init__.py │ ├── add_target_dataset.py │ ├── append_token_dataset.py │ ├── audio │ │ ├── __init__.py │ │ ├── audio_utils.py │ │ ├── feature_transforms │ │ │ ├── __init__.py │ │ │ ├── global_cmvn.py │ │ │ ├── specaugment.py │ │ │ └── utterance_cmvn.py │ │ ├── hubert_dataset.py │ │ ├── raw_audio_dataset.py │ │ └── speech_to_text_dataset.py │ ├── backtranslation_dataset.py │ ├── base_wrapper_dataset.py │ ├── bucket_pad_length_dataset.py │ ├── colorize_dataset.py │ ├── concat_dataset.py │ ├── concat_sentences_dataset.py │ ├── data_utils.py │ ├── data_utils_fast.pyx │ ├── denoising_dataset.py │ ├── dictionary.py │ ├── encoders │ │ ├── __init__.py │ │ ├── byte_bpe.py │ │ ├── byte_utils.py │ │ ├── bytes.py │ │ ├── characters.py │ │ ├── fastbpe.py │ │ ├── gpt2_bpe.py │ │ ├── gpt2_bpe_utils.py │ │ ├── hf_bert_bpe.py │ │ ├── hf_byte_bpe.py │ │ ├── moses_tokenizer.py │ │ ├── nltk_tokenizer.py │ │ ├── sentencepiece_bpe.py │ │ ├── space_tokenizer.py │ │ ├── subword_nmt_bpe.py │ │ └── utils.py │ ├── fairseq_dataset.py │ ├── fasta_dataset.py │ ├── id_dataset.py │ ├── indexed_dataset.py │ ├── iterators.py │ ├── language_pair_dataset.py │ ├── legacy │ │ ├── __init__.py │ │ ├── block_pair_dataset.py │ │ ├── masked_lm_dataset.py │ │ └── masked_lm_dictionary.py │ ├── list_dataset.py │ ├── lm_context_window_dataset.py │ ├── lru_cache_dataset.py │ ├── mask_tokens_dataset.py │ ├── monolingual_dataset.py │ ├── multi_corpus_dataset.py │ ├── multi_corpus_sampled_dataset.py │ ├── multilingual │ │ ├── __init__.py │ │ ├── multilingual_data_manager.py │ │ ├── multilingual_utils.py │ │ ├── sampled_multi_dataset.py │ │ ├── sampled_multi_epoch_dataset.py │ │ └── sampling_method.py │ ├── nested_dictionary_dataset.py │ ├── noising.py │ ├── num_samples_dataset.py │ ├── numel_dataset.py │ ├── offset_tokens_dataset.py │ ├── pad_dataset.py │ ├── plasma_utils.py │ ├── prepend_dataset.py │ ├── prepend_token_dataset.py │ ├── raw_label_dataset.py │ ├── replace_dataset.py │ ├── resampling_dataset.py │ ├── roll_dataset.py │ ├── round_robin_zip_datasets.py │ ├── shorten_dataset.py │ ├── sort_dataset.py │ ├── strip_token_dataset.py │ ├── subsample_dataset.py │ ├── token_block_dataset.py │ ├── token_block_utils_fast.pyx │ ├── transform_eos_dataset.py │ └── transform_eos_lang_pair_dataset.py ├── dataclass │ ├── __init__.py │ ├── configs.py │ ├── constants.py │ ├── initialize.py │ └── utils.py ├── distributed │ ├── __init__.py │ ├── distributed_timeout_wrapper.py │ ├── fully_sharded_data_parallel.py │ ├── legacy_distributed_data_parallel.py │ ├── module_proxy_wrapper.py │ ├── tpu_distributed_data_parallel.py │ └── utils.py ├── file_chunker_utils.py ├── file_io.py ├── file_utils.py ├── hub_utils.py ├── incremental_decoding_utils.py ├── iterative_refinement_generator.py ├── logging │ ├── __init__.py │ ├── meters.py │ ├── metrics.py │ └── progress_bar.py ├── model_parallel │ ├── __init__.py │ ├── criterions │ │ ├── __init__.py │ │ └── vocab_parallel_cross_entropy.py │ ├── megatron_trainer.py │ ├── models │ │ ├── __init__.py │ │ ├── pipeline_parallel_transformer │ │ │ ├── __init__.py │ │ │ ├── layers.py │ │ │ └── model.py │ │ ├── roberta │ │ │ ├── __init__.py │ │ │ └── model.py │ │ ├── transformer.py │ │ └── transformer_lm.py │ └── modules │ │ ├── __init__.py │ │ ├── multihead_attention.py │ │ └── transformer_layer.py ├── models │ ├── __init__.py │ ├── bart │ │ ├── __init__.py │ │ ├── hub_interface.py │ │ └── model.py │ ├── composite_encoder.py │ ├── distributed_fairseq_model.py │ ├── fairseq_decoder.py │ ├── fairseq_encoder.py │ ├── fairseq_incremental_decoder.py │ ├── fairseq_model.py │ ├── fconv.py │ ├── fconv_lm.py │ ├── fconv_self_att.py │ ├── hubert │ │ ├── __init__.py │ │ ├── hubert.py │ │ └── hubert_asr.py │ ├── huggingface │ │ ├── __init__.py │ │ └── hf_gpt2.py │ ├── lightconv.py │ ├── lightconv_lm.py │ ├── lstm.py │ ├── lstm_lm.py │ ├── masked_lm.py │ ├── model_utils.py │ ├── multilingual_transformer.py │ ├── nat │ │ ├── __init__.py │ │ ├── cmlm_transformer.py │ │ ├── fairseq_nat_model.py │ │ ├── insertion_transformer.py │ │ ├── iterative_nonautoregressive_transformer.py │ │ ├── levenshtein_transformer.py │ │ ├── levenshtein_utils.py │ │ ├── nat_crf_transformer.py │ │ ├── nonautoregressive_ensembles.py │ │ └── nonautoregressive_transformer.py │ ├── roberta │ │ ├── __init__.py │ │ ├── alignment_utils.py │ │ ├── enc_dec.py │ │ ├── hub_interface.py │ │ ├── model.py │ │ ├── model_camembert.py │ │ ├── model_gottbert.py │ │ └── model_xlmr.py │ ├── speech_to_text │ │ ├── __init__.py │ │ ├── berard.py │ │ ├── convtransformer.py │ │ ├── modules │ │ │ ├── augmented_memory_attention.py │ │ │ └── emformer.py │ │ ├── s2t_transformer.py │ │ ├── s2t_transformer_w2v2.py │ │ └── utils.py │ ├── transformer.py │ ├── transformer_align.py │ ├── transformer_from_pretrained_xlm.py │ ├── transformer_lm.py │ └── wav2vec │ │ ├── __init__.py │ │ ├── wav2vec.py │ │ ├── wav2vec2.py │ │ └── wav2vec2_asr.py ├── modules │ ├── __init__.py │ ├── adaptive_input.py │ ├── adaptive_softmax.py │ ├── base_layer.py │ ├── beamable_mm.py │ ├── character_token_embedder.py │ ├── checkpoint_activations.py │ ├── conv_tbc.py │ ├── cross_entropy.py │ ├── cuda_utils.cu │ ├── downsampled_multihead_attention.py │ ├── dynamic_convolution.py │ ├── dynamic_crf_layer.py │ ├── dynamicconv_layer │ │ ├── __init__.py │ │ ├── cuda_function_gen.py │ │ ├── dynamicconv_cuda.cpp │ │ ├── dynamicconv_cuda.cuh │ │ ├── dynamicconv_cuda_kernel.cu │ │ ├── dynamicconv_layer.py │ │ ├── dynamiconv_cpu.cpp │ │ └── setup.py │ ├── fairseq_dropout.py │ ├── fp32_group_norm.py │ ├── gelu.py │ ├── grad_multiply.py │ ├── gumbel_vector_quantizer.py │ ├── kmeans_attention.py │ ├── kmeans_vector_quantizer.py │ ├── layer_drop.py │ ├── layer_norm.py │ ├── learned_positional_embedding.py │ ├── lightconv_layer │ │ ├── __init__.py │ │ ├── cuda_function_gen.py │ │ ├── lightconv_cuda.cpp │ │ ├── lightconv_cuda.cuh │ │ ├── lightconv_cuda_kernel.cu │ │ ├── lightconv_layer.py │ │ └── setup.py │ ├── lightweight_convolution.py │ ├── linearized_convolution.py │ ├── multihead_attention.py │ ├── positional_embedding.py │ ├── quant_noise.py │ ├── quantization │ │ ├── __init__.py │ │ ├── pq │ │ │ ├── __init__.py │ │ │ ├── em.py │ │ │ ├── modules │ │ │ │ ├── __init__.py │ │ │ │ ├── qconv.py │ │ │ │ ├── qemb.py │ │ │ │ └── qlinear.py │ │ │ ├── pq.py │ │ │ └── utils.py │ │ ├── quantization_options.py │ │ └── scalar │ │ │ ├── __init__.py │ │ │ ├── modules │ │ │ ├── __init__.py │ │ │ ├── qact.py │ │ │ ├── qconv.py │ │ │ ├── qemb.py │ │ │ └── qlinear.py │ │ │ ├── ops.py │ │ │ └── utils.py │ ├── same_pad.py │ ├── scalar_bias.py │ ├── sinusoidal_positional_embedding.py │ ├── sparse_multihead_attention.py │ ├── sparse_transformer_sentence_encoder.py │ ├── sparse_transformer_sentence_encoder_layer.py │ ├── transformer_layer.py │ ├── transformer_sentence_encoder.py │ ├── transformer_sentence_encoder_layer.py │ ├── transpose_last.py │ ├── unfold.py │ └── vggblock.py ├── nan_detector.py ├── ngram_repeat_block.py ├── optim │ ├── __init__.py │ ├── adadelta.py │ ├── adafactor.py │ ├── adagrad.py │ ├── adam.py │ ├── adamax.py │ ├── amp_optimizer.py │ ├── bmuf.py │ ├── composite.py │ ├── cpu_adam.py │ ├── dynamic_loss_scaler.py │ ├── fairseq_optimizer.py │ ├── fp16_optimizer.py │ ├── fused_adam.py │ ├── fused_lamb.py │ ├── lr_scheduler │ │ ├── __init__.py │ │ ├── cosine_lr_scheduler.py │ │ ├── fairseq_lr_scheduler.py │ │ ├── fixed_schedule.py │ │ ├── inverse_square_root_schedule.py │ │ ├── manual_lr_scheduler.py │ │ ├── pass_through.py │ │ ├── polynomial_decay_schedule.py │ │ ├── reduce_lr_on_plateau.py │ │ ├── tri_stage_lr_scheduler.py │ │ └── triangular_lr_scheduler.py │ ├── nag.py │ ├── sgd.py │ └── shard.py ├── options.py ├── pdb.py ├── quantization_utils.py ├── registry.py ├── scoring │ ├── __init__.py │ ├── bleu.py │ ├── chrf.py │ ├── tokenizer.py │ └── wer.py ├── search.py ├── sequence_generator.py ├── sequence_scorer.py ├── tasks │ ├── __init__.py │ ├── audio_pretraining.py │ ├── cross_lingual_lm.py │ ├── denoising.py │ ├── fairseq_task.py │ ├── hubert_pretraining.py │ ├── language_modeling.py │ ├── legacy_masked_lm.py │ ├── masked_lm.py │ ├── multilingual_denoising.py │ ├── multilingual_masked_lm.py │ ├── multilingual_translation.py │ ├── online_backtranslation.py │ ├── semisupervised_translation.py │ ├── sentence_prediction.py │ ├── sentence_ranking.py │ ├── simultaneous_translation.py │ ├── speech_to_text.py │ ├── translation.py │ ├── translation_from_pretrained_bart.py │ ├── translation_from_pretrained_xlm.py │ ├── translation_lev.py │ └── translation_multi_simple_epoch.py ├── token_generation_constraints.py ├── tokenizer.py ├── trainer.py ├── utils.py └── version.txt ├── fairseq_cli ├── __init__.py ├── eval_lm.py ├── generate.py ├── hydra_train.py ├── interactive.py ├── preprocess.py ├── score.py ├── train.py └── validate.py ├── fbank_scripts ├── pretrain_asr.sh ├── pretrain_asr_base.sh ├── pretrain_mt.sh ├── train_baseline.sh ├── train_baseline_base.sh ├── train_stmm.sh ├── train_stmm_base.sh └── train_stmm_base_static.sh ├── hubconf.py ├── preprocess.sh ├── preprocess_scripts ├── clean_mustc.py ├── convert_format.py ├── group.py ├── postprocess_raw.py └── word_align_info_raw.py ├── pretrain_mt.sh ├── pretrain_mt_ext.sh ├── pyproject.toml ├── scripts ├── __init__.py ├── average_checkpoints.py ├── build_sym_alignment.py ├── compare_namespaces.py ├── compound_split_bleu.sh ├── constraints │ ├── extract.py │ └── validate.py ├── convert_dictionary.lua ├── convert_model.lua ├── count_docs.py ├── read_binarized.py ├── rm_pt.py ├── sacrebleu.sh ├── shard_docs.py ├── split_train_valid_docs.py ├── spm_decode.py ├── spm_encode.py ├── spm_train.py └── test_fsdp.sh ├── setup.py ├── test.sh ├── tests ├── __init__.py ├── distributed │ ├── __init__.py │ ├── test_bmuf.py │ ├── test_distributed_timeout_wrapper.py │ ├── test_module_proxy_wrapper.py │ ├── test_utils.py │ └── utils.py ├── gpu │ ├── __init__.py │ ├── test_binaries_gpu.py │ └── transformer_quantization_config.yaml ├── speech_recognition │ ├── __init__.py │ ├── asr_test_base.py │ ├── test_collaters.py │ ├── test_cross_entropy.py │ ├── test_data_utils.py │ └── test_vggtransformer.py ├── test_activation_checkpointing.py ├── test_amp_optimizer.py ├── test_average_checkpoints.py ├── test_backtranslation_dataset.py ├── test_binaries.py ├── test_character_token_embedder.py ├── test_checkpoint_utils.py ├── test_concat_dataset.py ├── test_constraints.py ├── test_convtbc.py ├── test_data_utils.py ├── test_dataset.py ├── test_dictionary.py ├── test_export.py ├── test_file_chunker_utils.py ├── test_file_io.py ├── test_fp16_optimizer.py ├── test_inference_dropout.py ├── test_iopath.py ├── test_iterators.py ├── test_label_smoothing.py ├── test_lm_context_window.py ├── test_lstm_jitable.py ├── test_memory_efficient_fp16.py ├── test_metrics.py ├── test_multi_corpus_dataset.py ├── test_multi_corpus_sampled_dataset.py ├── test_multihead_attention.py ├── test_noising.py ├── test_online_backtranslation.py ├── test_plasma_utils.py ├── test_reproducibility.py ├── test_resampling_dataset.py ├── test_roberta.py ├── test_sequence_generator.py ├── test_sequence_scorer.py ├── test_sparse_multihead_attention.py ├── test_token_block_dataset.py ├── test_train.py ├── test_transformer.py ├── test_utils.py ├── test_valid_subset_checks.py └── utils.py ├── train.py └── train.sh /.github/ISSUE_TEMPLATE.md: -------------------------------------------------------------------------------- 1 | ## 👉 [Please follow one of these issue templates](https://github.com/pytorch/fairseq/issues/new/choose) 👈 2 | 3 | Note: to keep the backlog clean and actionable, issues may be immediately closed if they do not follow one of the above issue templates. 4 | -------------------------------------------------------------------------------- /.github/ISSUE_TEMPLATE/bug_report.md: -------------------------------------------------------------------------------- 1 | --- 2 | name: 🐛 Bug Report 3 | about: Submit a bug report to help us improve 4 | labels: 'bug, needs triage' 5 | --- 6 | 7 | ## 🐛 Bug 8 | 9 | 10 | 11 | ### To Reproduce 12 | 13 | Steps to reproduce the behavior (**always include the command you ran**): 14 | 15 | 1. Run cmd '....' 16 | 2. See error 17 | 18 | 19 | 20 | 21 | #### Code sample 22 | 24 | 25 | ### Expected behavior 26 | 27 | 28 | 29 | ### Environment 30 | 31 | - fairseq Version (e.g., 1.0 or master): 32 | - PyTorch Version (e.g., 1.0) 33 | - OS (e.g., Linux): 34 | - How you installed fairseq (`pip`, source): 35 | - Build command you used (if compiling from source): 36 | - Python version: 37 | - CUDA/cuDNN version: 38 | - GPU models and configuration: 39 | - Any other relevant information: 40 | 41 | ### Additional context 42 | 43 | 44 | -------------------------------------------------------------------------------- /.github/ISSUE_TEMPLATE/documentation.md: -------------------------------------------------------------------------------- 1 | --- 2 | name: 📚 Documentation/Typos 3 | about: Report an issue related to documentation or a typo 4 | labels: 'documentation, needs triage' 5 | --- 6 | 7 | ## 📚 Documentation 8 | 9 | For typos and doc fixes, please go ahead and: 10 | 11 | 1. Create an issue. 12 | 2. Fix the typo. 13 | 3. Submit a PR. 14 | 15 | Thanks! 16 | -------------------------------------------------------------------------------- /.github/ISSUE_TEMPLATE/feature_request.md: -------------------------------------------------------------------------------- 1 | --- 2 | name: 🚀 Feature Request 3 | about: Submit a proposal/request for a new feature 4 | labels: 'enhancement, help wanted, needs triage' 5 | --- 6 | 7 | ## 🚀 Feature Request 8 | 9 | 10 | ### Motivation 11 | 12 | 13 | 14 | ### Pitch 15 | 16 | 17 | 18 | ### Alternatives 19 | 20 | 21 | 22 | ### Additional context 23 | 24 | 25 | -------------------------------------------------------------------------------- /.github/ISSUE_TEMPLATE/how-to-question.md: -------------------------------------------------------------------------------- 1 | --- 2 | name: ❓ Questions/Help 3 | about: If you have questions, please first search existing issues and docs 4 | labels: 'question, needs triage' 5 | --- 6 | 7 | ## ❓ Questions and Help 8 | 9 | ### Before asking: 10 | 1. search the issues. 11 | 2. search the docs. 12 | 13 | 14 | 15 | #### What is your question? 16 | 17 | #### Code 18 | 19 | 20 | 21 | #### What have you tried? 22 | 23 | #### What's your environment? 24 | 25 | - fairseq Version (e.g., 1.0 or master): 26 | - PyTorch Version (e.g., 1.0) 27 | - OS (e.g., Linux): 28 | - How you installed fairseq (`pip`, source): 29 | - Build command you used (if compiling from source): 30 | - Python version: 31 | - CUDA/cuDNN version: 32 | - GPU models and configuration: 33 | - Any other relevant information: 34 | -------------------------------------------------------------------------------- /.github/PULL_REQUEST_TEMPLATE.md: -------------------------------------------------------------------------------- 1 | # Before submitting 2 | 3 | - [ ] Was this discussed/approved via a Github issue? (no need for typos, doc improvements) 4 | - [ ] Did you read the [contributor guideline](https://github.com/pytorch/fairseq/blob/master/CONTRIBUTING.md)? 5 | - [ ] Did you make sure to update the docs? 6 | - [ ] Did you write any new necessary tests? 7 | 8 | ## What does this PR do? 9 | Fixes # (issue). 10 | 11 | ## PR review 12 | Anyone in the community is free to review the PR once the tests have passed. 13 | If we didn't discuss your PR in Github issues there's a high chance it will not be merged. 14 | 15 | ## Did you have fun? 16 | Make sure you had fun coding 🙃 17 | -------------------------------------------------------------------------------- /.github/workflows/build_wheels.yml: -------------------------------------------------------------------------------- 1 | name: build_wheels 2 | 3 | on: 4 | push: 5 | branches: 6 | - v[0-9]+.[0-9]+.[x0-9]+ 7 | tags: 8 | - v* 9 | 10 | jobs: 11 | build_wheels: 12 | name: Build wheels on ${{ matrix.os }} 13 | runs-on: ${{ matrix.os }} 14 | strategy: 15 | matrix: 16 | os: [ubuntu-latest, macos-latest] 17 | 18 | steps: 19 | - uses: actions/checkout@v2 20 | 21 | - name: Install Python 22 | uses: actions/setup-python@v2 23 | with: 24 | python-version: '3.7' 25 | 26 | - name: Install cibuildwheel 27 | run: | 28 | python -m pip install cibuildwheel 29 | 30 | - name: Build wheels for CPython 31 | run: | 32 | python -m cibuildwheel --output-dir dist 33 | env: 34 | CIBW_BUILD: "cp36-*64 cp37-*64 cp38-*64" 35 | CIBW_MANYLINUX_X86_64_IMAGE: manylinux1 36 | CIBW_BEFORE_BUILD: git submodule update --init --recursive && pip install . 37 | 38 | - uses: actions/upload-artifact@v2 39 | with: 40 | name: wheels 41 | path: ./dist/*.whl 42 | -------------------------------------------------------------------------------- /.gitmodules: -------------------------------------------------------------------------------- 1 | [submodule "fairseq/model_parallel/megatron"] 2 | path = fairseq/model_parallel/megatron 3 | url = https://github.com/ngoyal2707/Megatron-LM 4 | branch = fairseq 5 | -------------------------------------------------------------------------------- /CONTRIBUTING.md: -------------------------------------------------------------------------------- 1 | # Contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq) 2 | We want to make contributing to this project as easy and transparent as 3 | possible. 4 | 5 | ## Pull Requests 6 | We actively welcome your pull requests. 7 | 8 | 1. Fork the repo and create your branch from `master`. 9 | 2. If you've added code that should be tested, add tests. 10 | 3. If you've changed APIs, update the documentation. 11 | 4. Ensure the test suite passes. 12 | 5. Make sure your code lints. 13 | 6. If you haven't already, complete the Contributor License Agreement ("CLA"). 14 | 15 | ## Contributor License Agreement ("CLA") 16 | In order to accept your pull request, we need you to submit a CLA. You only need 17 | to do this once to work on any of Facebook's open source projects. 18 | 19 | Complete your CLA here: 20 | 21 | ## Issues 22 | We use GitHub issues to track public bugs. Please ensure your description is 23 | clear and has sufficient instructions to be able to reproduce the issue. 24 | 25 | ## License 26 | By contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq), 27 | you agree that your contributions will be licensed under the LICENSE file in 28 | the root directory of this source tree. 29 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) Facebook, Inc. and its affiliates. 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /data/scripts/apply_spm.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | from argparse import Namespace 3 | from fairseq.data import encoders 4 | import tqdm 5 | 6 | parser = argparse.ArgumentParser() 7 | parser.add_argument("--input-file", type=str, required=True) 8 | parser.add_argument("--output-file", type=str, required=True) 9 | parser.add_argument("--model", type=str, required=True) 10 | 11 | args = parser.parse_args() 12 | 13 | bpe_tokenizer = encoders.build_bpe( 14 | Namespace( 15 | bpe='sentencepiece', 16 | sentencepiece_model=args.model, 17 | ) 18 | ) 19 | 20 | with open(args.input_file, 'r') as input_file: 21 | input_lines = input_file.readlines() 22 | 23 | with open(args.output_file, 'w') as output_file: 24 | for line in tqdm.tqdm(input_lines): 25 | encoded_line = bpe_tokenizer.encode(line) 26 | output_file.write(encoded_line + '\n') -------------------------------------------------------------------------------- /docs/Makefile: -------------------------------------------------------------------------------- 1 | # Minimal makefile for Sphinx documentation 2 | # 3 | 4 | # You can set these variables from the command line. 5 | SPHINXOPTS = 6 | SPHINXBUILD = python -msphinx 7 | SPHINXPROJ = fairseq 8 | SOURCEDIR = . 9 | BUILDDIR = _build 10 | 11 | # Put it first so that "make" without argument is like "make help". 12 | help: 13 | @$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) 14 | 15 | .PHONY: help Makefile 16 | 17 | # Catch-all target: route all unknown targets to Sphinx using the new 18 | # "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS). 19 | %: Makefile 20 | @$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) -------------------------------------------------------------------------------- /docs/_static/theme_overrides.css: -------------------------------------------------------------------------------- 1 | .wy-table-responsive table td kbd { 2 | white-space: nowrap; 3 | } 4 | .wy-table-responsive table td { 5 | white-space: normal !important; 6 | } 7 | .wy-table-responsive { 8 | overflow: visible !important; 9 | } 10 | -------------------------------------------------------------------------------- /docs/criterions.rst: -------------------------------------------------------------------------------- 1 | .. role:: hidden 2 | :class: hidden-section 3 | 4 | .. _Criterions: 5 | 6 | Criterions 7 | ========== 8 | 9 | Criterions compute the loss function given the model and batch, roughly:: 10 | 11 | loss = criterion(model, batch) 12 | 13 | .. automodule:: fairseq.criterions 14 | :members: 15 | 16 | .. autoclass:: fairseq.criterions.FairseqCriterion 17 | :members: 18 | :undoc-members: 19 | 20 | .. autoclass:: fairseq.criterions.adaptive_loss.AdaptiveLoss 21 | :members: 22 | :undoc-members: 23 | .. autoclass:: fairseq.criterions.composite_loss.CompositeLoss 24 | :members: 25 | :undoc-members: 26 | .. autoclass:: fairseq.criterions.cross_entropy.CrossEntropyCriterion 27 | :members: 28 | :undoc-members: 29 | .. autoclass:: fairseq.criterions.label_smoothed_cross_entropy.LabelSmoothedCrossEntropyCriterion 30 | :members: 31 | :undoc-members: 32 | -------------------------------------------------------------------------------- /docs/docutils.conf: -------------------------------------------------------------------------------- 1 | [writers] 2 | option-limit=0 3 | -------------------------------------------------------------------------------- /docs/fairseq.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/docs/fairseq.gif -------------------------------------------------------------------------------- /docs/index.rst: -------------------------------------------------------------------------------- 1 | .. fairseq documentation master file, created by 2 | sphinx-quickstart on Fri Aug 17 21:45:30 2018. 3 | You can adapt this file completely to your liking, but it should at least 4 | contain the root `toctree` directive. 5 | 6 | :github_url: https://github.com/pytorch/fairseq 7 | 8 | 9 | fairseq documentation 10 | ===================== 11 | 12 | Fairseq is a sequence modeling toolkit written in `PyTorch 13 | `_ that allows researchers and developers to 14 | train custom models for translation, summarization, language modeling and other 15 | text generation tasks. 16 | 17 | .. toctree:: 18 | :maxdepth: 1 19 | :caption: Getting Started 20 | 21 | getting_started 22 | command_line_tools 23 | 24 | .. toctree:: 25 | :maxdepth: 1 26 | :caption: Extending Fairseq 27 | 28 | overview 29 | tutorial_simple_lstm 30 | tutorial_classifying_names 31 | 32 | .. toctree:: 33 | :maxdepth: 2 34 | :caption: Library Reference 35 | 36 | tasks 37 | models 38 | criterions 39 | optim 40 | lr_scheduler 41 | data 42 | modules 43 | 44 | 45 | Indices and tables 46 | ================== 47 | 48 | * :ref:`genindex` 49 | * :ref:`search` 50 | -------------------------------------------------------------------------------- /docs/lr_scheduler.rst: -------------------------------------------------------------------------------- 1 | .. role:: hidden 2 | :class: hidden-section 3 | 4 | .. _Learning Rate Schedulers: 5 | 6 | Learning Rate Schedulers 7 | ======================== 8 | 9 | Learning Rate Schedulers update the learning rate over the course of training. 10 | Learning rates can be updated after each update via :func:`step_update` or at 11 | epoch boundaries via :func:`step`. 12 | 13 | .. automodule:: fairseq.optim.lr_scheduler 14 | :members: 15 | 16 | .. autoclass:: fairseq.optim.lr_scheduler.FairseqLRScheduler 17 | :members: 18 | :undoc-members: 19 | 20 | .. autoclass:: fairseq.optim.lr_scheduler.cosine_lr_scheduler.CosineSchedule 21 | :members: 22 | :undoc-members: 23 | .. autoclass:: fairseq.optim.lr_scheduler.fixed_schedule.FixedSchedule 24 | :members: 25 | :undoc-members: 26 | .. autoclass:: fairseq.optim.lr_scheduler.inverse_square_root_schedule.InverseSquareRootSchedule 27 | :members: 28 | :undoc-members: 29 | .. autoclass:: fairseq.optim.lr_scheduler.reduce_lr_on_plateau.ReduceLROnPlateau 30 | :members: 31 | :undoc-members: 32 | .. autoclass:: fairseq.optim.lr_scheduler.triangular_lr_scheduler.TriangularSchedule 33 | :members: 34 | :undoc-members: 35 | -------------------------------------------------------------------------------- /docs/make.bat: -------------------------------------------------------------------------------- 1 | @ECHO OFF 2 | 3 | pushd %~dp0 4 | 5 | REM Command file for Sphinx documentation 6 | 7 | if "%SPHINXBUILD%" == "" ( 8 | set SPHINXBUILD=python -msphinx 9 | ) 10 | set SOURCEDIR=. 11 | set BUILDDIR=_build 12 | set SPHINXPROJ=fairseq 13 | 14 | if "%1" == "" goto help 15 | 16 | %SPHINXBUILD% >NUL 2>NUL 17 | if errorlevel 9009 ( 18 | echo. 19 | echo.The Sphinx module was not found. Make sure you have Sphinx installed, 20 | echo.then set the SPHINXBUILD environment variable to point to the full 21 | echo.path of the 'sphinx-build' executable. Alternatively you may add the 22 | echo.Sphinx directory to PATH. 23 | echo. 24 | echo.If you don't have Sphinx installed, grab it from 25 | echo.http://sphinx-doc.org/ 26 | exit /b 1 27 | ) 28 | 29 | %SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% 30 | goto end 31 | 32 | :help 33 | %SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% 34 | 35 | :end 36 | popd 37 | -------------------------------------------------------------------------------- /docs/modules.rst: -------------------------------------------------------------------------------- 1 | Modules 2 | ======= 3 | 4 | Fairseq provides several stand-alone :class:`torch.nn.Module` classes that may 5 | be helpful when implementing a new :class:`~fairseq.models.BaseFairseqModel`. 6 | 7 | .. automodule:: fairseq.modules 8 | :members: 9 | :undoc-members: 10 | -------------------------------------------------------------------------------- /docs/optim.rst: -------------------------------------------------------------------------------- 1 | .. role:: hidden 2 | :class: hidden-section 3 | 4 | .. _optimizers: 5 | 6 | Optimizers 7 | ========== 8 | 9 | Optimizers update the Model parameters based on the gradients. 10 | 11 | .. automodule:: fairseq.optim 12 | :members: 13 | 14 | .. autoclass:: fairseq.optim.FairseqOptimizer 15 | :members: 16 | :undoc-members: 17 | 18 | .. autoclass:: fairseq.optim.adadelta.Adadelta 19 | :members: 20 | :undoc-members: 21 | .. autoclass:: fairseq.optim.adagrad.Adagrad 22 | :members: 23 | :undoc-members: 24 | .. autoclass:: fairseq.optim.adafactor.FairseqAdafactor 25 | :members: 26 | :undoc-members: 27 | .. autoclass:: fairseq.optim.adam.FairseqAdam 28 | :members: 29 | :undoc-members: 30 | .. autoclass:: fairseq.optim.fp16_optimizer.FP16Optimizer 31 | :members: 32 | :undoc-members: 33 | .. autoclass:: fairseq.optim.nag.FairseqNAG 34 | :members: 35 | :undoc-members: 36 | .. autoclass:: fairseq.optim.sgd.SGD 37 | :members: 38 | :undoc-members: 39 | -------------------------------------------------------------------------------- /docs/requirements.txt: -------------------------------------------------------------------------------- 1 | sphinx<2.0 2 | sphinx-argparse 3 | -------------------------------------------------------------------------------- /examples/.gitignore: -------------------------------------------------------------------------------- 1 | !*/*.sh 2 | !*/*.md 3 | -------------------------------------------------------------------------------- /examples/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | try: 7 | from fairseq.version import __version__ # noqa 8 | except ImportError: 9 | pass 10 | -------------------------------------------------------------------------------- /examples/adaptive_span/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import importlib 7 | import os 8 | 9 | # automatically import any Python files in the current directory 10 | cur_dir = os.path.dirname(__file__) 11 | for file in os.listdir(cur_dir): 12 | path = os.path.join(cur_dir, file) 13 | if ( 14 | not file.startswith("_") 15 | and not file.startswith(".") 16 | and (file.endswith(".py") or os.path.isdir(path)) 17 | ): 18 | mod_name = file[: file.find(".py")] if file.endswith(".py") else file 19 | module = importlib.import_module(__name__ + "." + mod_name) 20 | -------------------------------------------------------------------------------- /examples/adaptive_span/truncated_bptt_lm_task.py: -------------------------------------------------------------------------------- 1 | ../truncated_bptt/truncated_bptt_lm_task.py -------------------------------------------------------------------------------- /examples/backtranslation/deduplicate_lines.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/python3 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import argparse 8 | import fileinput 9 | import hashlib 10 | import sys 11 | from multiprocessing import Pool 12 | 13 | 14 | def get_hashes_and_lines(raw_line): 15 | hash = hashlib.md5(raw_line).hexdigest() 16 | return hash, raw_line 17 | 18 | 19 | def main(): 20 | parser = argparse.ArgumentParser() 21 | parser.add_argument("--workers", type=int, default=10) 22 | parser.add_argument("files", nargs="*", help="input files") 23 | args = parser.parse_args() 24 | 25 | seen = set() 26 | with fileinput.input(args.files, mode="rb") as h: 27 | pool = Pool(args.workers) 28 | results = pool.imap_unordered(get_hashes_and_lines, h, 1000) 29 | for i, (hash, raw_line) in enumerate(results): 30 | if hash not in seen: 31 | seen.add(hash) 32 | sys.stdout.buffer.write(raw_line) 33 | if i % 1000000 == 0: 34 | print(i, file=sys.stderr, end="", flush=True) 35 | elif i % 100000 == 0: 36 | print(".", file=sys.stderr, end="", flush=True) 37 | print(file=sys.stderr, flush=True) 38 | 39 | 40 | if __name__ == "__main__": 41 | main() 42 | -------------------------------------------------------------------------------- /examples/backtranslation/sacrebleu.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | if [ $# -ne 5 ]; then 4 | echo "usage: $0 [dataset=wmt14/full] [langpair=en-de] [databin] [bpecode] [model]" 5 | exit 6 | fi 7 | 8 | 9 | DATASET=$1 10 | LANGPAIR=$2 11 | DATABIN=$3 12 | BPECODE=$4 13 | MODEL=$5 14 | 15 | SRCLANG=$(echo $LANGPAIR | cut -d '-' -f 1) 16 | TGTLANG=$(echo $LANGPAIR | cut -d '-' -f 2) 17 | 18 | 19 | BPEROOT=examples/backtranslation/subword-nmt/subword_nmt 20 | if [ ! -e $BPEROOT ]; then 21 | BPEROOT=subword-nmt/subword_nmt 22 | if [ ! -e $BPEROOT ]; then 23 | echo 'Cloning Subword NMT repository (for BPE pre-processing)...' 24 | git clone https://github.com/rsennrich/subword-nmt.git 25 | fi 26 | fi 27 | 28 | 29 | sacrebleu -t $DATASET -l $LANGPAIR --echo src \ 30 | | sacremoses tokenize -a -l $SRCLANG -q \ 31 | | python $BPEROOT/apply_bpe.py -c $BPECODE \ 32 | | fairseq-interactive $DATABIN --path $MODEL \ 33 | -s $SRCLANG -t $TGTLANG \ 34 | --beam 5 --remove-bpe --buffer-size 1024 --max-tokens 8000 \ 35 | | grep ^H- | cut -f 3- \ 36 | | sacremoses detokenize -l $TGTLANG -q \ 37 | | sacrebleu -t $DATASET -l $LANGPAIR 38 | -------------------------------------------------------------------------------- /examples/backtranslation/tokenized_bleu.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | if [ $# -ne 5 ]; then 4 | echo "usage: $0 [dataset=wmt14/full] [langpair=en-de] [databin] [bpecode] [model]" 5 | exit 6 | fi 7 | 8 | 9 | DATASET=$1 10 | LANGPAIR=$2 11 | DATABIN=$3 12 | BPECODE=$4 13 | MODEL=$5 14 | 15 | SRCLANG=$(echo $LANGPAIR | cut -d '-' -f 1) 16 | TGTLANG=$(echo $LANGPAIR | cut -d '-' -f 2) 17 | 18 | 19 | BPEROOT=examples/backtranslation/subword-nmt/subword_nmt 20 | if [ ! -e $BPEROOT ]; then 21 | BPEROOT=subword-nmt/subword_nmt 22 | if [ ! -e $BPEROOT ]; then 23 | echo 'Cloning Subword NMT repository (for BPE pre-processing)...' 24 | git clone https://github.com/rsennrich/subword-nmt.git 25 | fi 26 | fi 27 | 28 | 29 | TMP_REF=$(mktemp) 30 | 31 | sacrebleu -t $DATASET -l $LANGPAIR --echo ref -q \ 32 | | sacremoses normalize -l $TGTLANG -q \ 33 | | sacremoses tokenize -a -l $TGTLANG -q \ 34 | > $TMP_REF 35 | 36 | sacrebleu -t $DATASET -l $LANGPAIR --echo src -q \ 37 | | sacremoses normalize -l $SRCLANG -q \ 38 | | sacremoses tokenize -a -l $SRCLANG -q \ 39 | | python $BPEROOT/apply_bpe.py -c $BPECODE \ 40 | | fairseq-interactive $DATABIN --path $MODEL \ 41 | -s $SRCLANG -t $TGTLANG \ 42 | --beam 5 --remove-bpe --buffer-size 1024 --max-tokens 8000 \ 43 | | grep ^H- | cut -f 3- \ 44 | | fairseq-score --ref $TMP_REF 45 | 46 | rm -f $TMP_REF 47 | -------------------------------------------------------------------------------- /examples/constrained_decoding/normalize.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # 3 | # Copyright (c) Facebook, Inc. and its affiliates. 4 | # 5 | # This source code is licensed under the MIT license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | import sys 9 | 10 | from sacremoses.normalize import MosesPunctNormalizer 11 | 12 | 13 | def main(args): 14 | normalizer = MosesPunctNormalizer(lang=args.lang, penn=args.penn) 15 | for line in sys.stdin: 16 | print(normalizer.normalize(line.rstrip()), flush=True) 17 | 18 | 19 | if __name__ == "__main__": 20 | import argparse 21 | 22 | parser = argparse.ArgumentParser() 23 | parser.add_argument("--lang", "-l", default="en") 24 | parser.add_argument("--penn", "-p", action="store_true") 25 | args = parser.parse_args() 26 | 27 | main(args) 28 | -------------------------------------------------------------------------------- /examples/constrained_decoding/tok.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # 3 | # Copyright (c) Facebook, Inc. and its affiliates. 4 | # 5 | # This source code is licensed under the MIT license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | import sys 9 | 10 | import sacremoses 11 | 12 | 13 | def main(args): 14 | """Tokenizes, preserving tabs""" 15 | mt = sacremoses.MosesTokenizer(lang=args.lang) 16 | 17 | def tok(s): 18 | return mt.tokenize(s, return_str=True) 19 | 20 | for line in sys.stdin: 21 | parts = list(map(tok, line.split("\t"))) 22 | print(*parts, sep="\t", flush=True) 23 | 24 | 25 | if __name__ == "__main__": 26 | import argparse 27 | 28 | parser = argparse.ArgumentParser() 29 | parser.add_argument("--lang", "-l", default="en") 30 | parser.add_argument("--penn", "-p", action="store_true") 31 | parser.add_argument("--fields", "-f", help="fields to tokenize") 32 | args = parser.parse_args() 33 | 34 | main(args) 35 | -------------------------------------------------------------------------------- /examples/discriminative_reranking_nmt/__init__.py: -------------------------------------------------------------------------------- 1 | from . import criterions, models, tasks # noqa 2 | -------------------------------------------------------------------------------- /examples/discriminative_reranking_nmt/config/deen.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 50 7 | seed: 2 8 | 9 | checkpoint: 10 | no_epoch_checkpoints: true 11 | best_checkpoint_metric: bleu 12 | maximize_best_checkpoint_metric: true 13 | 14 | task: 15 | _name: discriminative_reranking_nmt 16 | data: ??? 17 | num_data_splits: ??? 18 | include_src: true 19 | mt_beam: 50 20 | eval_target_metric: true 21 | target_metric: bleu 22 | 23 | dataset: 24 | batch_size: 50 25 | num_workers: 6 26 | required_batch_size_multiple: 50 27 | valid_subset: ??? 28 | 29 | criterion: 30 | _name: kl_divergence_rereanking 31 | target_dist_norm: minmax 32 | temperature: 0.5 33 | 34 | optimization: 35 | max_epoch: 200 36 | lr: [0.00005] 37 | update_freq: [32] 38 | 39 | optimizer: 40 | _name: adam 41 | adam_betas: (0.9,0.98) 42 | adam_eps: 1e-06 43 | 44 | lr_scheduler: 45 | _name: polynomial_decay 46 | warmup_updates: 8000 47 | total_num_update: 320000 48 | 49 | model: 50 | _name: discriminative_nmt_reranker 51 | pretrained_model: ??? 52 | classifier_dropout: 0.2 53 | 54 | distributed_training: 55 | ddp_backend: no_c10d 56 | distributed_world_size: 16 57 | -------------------------------------------------------------------------------- /examples/discriminative_reranking_nmt/criterions/__init__.py: -------------------------------------------------------------------------------- 1 | from .discriminative_reranking_criterion import KLDivergenceRerankingCriterion 2 | 3 | 4 | __all__ = [ 5 | "KLDivergenceRerankingCriterion", 6 | ] 7 | -------------------------------------------------------------------------------- /examples/discriminative_reranking_nmt/models/__init__.py: -------------------------------------------------------------------------------- 1 | from .discriminative_reranking_model import DiscriminativeNMTReranker 2 | 3 | 4 | __all__ = [ 5 | "DiscriminativeNMTReranker", 6 | ] 7 | -------------------------------------------------------------------------------- /examples/discriminative_reranking_nmt/tasks/__init__.py: -------------------------------------------------------------------------------- 1 | from .discriminative_reranking_task import DiscriminativeRerankingNMTTask 2 | 3 | 4 | __all__ = [ 5 | "DiscriminativeRerankingNMTTask", 6 | ] 7 | -------------------------------------------------------------------------------- /examples/fast_noisy_channel/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import noisy_channel_translation # noqa 7 | from . import noisy_channel_sequence_generator # noqa 8 | from . import noisy_channel_beam_search # noqa 9 | -------------------------------------------------------------------------------- /examples/hubert/config/decode/ax_sweep/ngram.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | 3 | common_eval: 4 | results_path: ${decoding.exp_dir}/decode/${decoding.decoder.name}_ax/${dataset.gen_subset} 5 | 6 | hydra: 7 | sweeper: 8 | ax_config: 9 | max_trials: 60 10 | early_stop: 11 | minimize: true 12 | max_epochs_without_improvement: 10 13 | epsilon: 0.025 14 | experiment: 15 | name: ${dataset.gen_subset} 16 | objective_name: wer 17 | minimize: true 18 | parameter_constraints: null 19 | outcome_constraints: null 20 | status_quo: null 21 | client: 22 | verbose_logging: false 23 | random_seed: null 24 | params: 25 | decoding.decoder.lmweight: 26 | type: range 27 | bounds: [0.0, 8.0] 28 | decoding.decoder.wordscore: 29 | type: range 30 | bounds: [-5.0, 5.0] 31 | decoding.decoder.silweight: 32 | type: range 33 | bounds: [-10.0, 0.0] 34 | -------------------------------------------------------------------------------- /examples/hubert/config/decode/ax_sweep/transformer.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | 3 | common_eval: 4 | results_path: ${decoding.exp_dir}/decode/${decoding.decoder.name}_ax/${dataset.gen_subset} 5 | 6 | hydra: 7 | sweeper: 8 | ax_config: 9 | max_trials: 60 10 | early_stop: 11 | minimize: true 12 | max_epochs_without_improvement: 10 13 | epsilon: 0.025 14 | experiment: 15 | name: ${dataset.gen_subset} 16 | objective_name: wer 17 | minimize: true 18 | parameter_constraints: null 19 | outcome_constraints: null 20 | status_quo: null 21 | client: 22 | verbose_logging: false 23 | random_seed: null 24 | params: 25 | decoding.decoder.lmweight: 26 | type: range 27 | bounds: [0.0, 4.0] 28 | decoding.decoder.wordscore: 29 | type: range 30 | bounds: [-5.0, 5.0] 31 | decoding.decoder.silweight: 32 | type: range 33 | bounds: [-8.0, 0.0] 34 | -------------------------------------------------------------------------------- /examples/hubert/config/decode/infer_fsqlm.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | defaults: 4 | - model: null 5 | 6 | hydra: 7 | run: 8 | dir: ${common_eval.results_path}/beam${decoding.decoder.beam}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} 9 | sweep: 10 | dir: ${common_eval.results_path} 11 | subdir: beam${decoding.decoder.beam}_th${decoding.decoder.beamthreshold}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} 12 | 13 | task: 14 | _name: hubert_pretraining 15 | single_target: true 16 | data: ??? 17 | normalize: ??? 18 | 19 | decoding: 20 | type: fairseqlm 21 | lexicon: ??? 22 | lmpath: ??? 23 | beamthreshold: 25 # 100 24 | beam: 500 25 | lmweight: 2 26 | wordscore: -1 27 | silweight: 0 28 | unique_wer_file: true 29 | beam: 500 30 | common_eval: 31 | results_path: ??? 32 | path: ??? 33 | post_process: letter 34 | dataset: 35 | max_tokens: 1100000 36 | gen_subset: ??? 37 | -------------------------------------------------------------------------------- /examples/hubert/config/decode/infer_kenlm.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | defaults: 4 | - model: null 5 | 6 | hydra: 7 | run: 8 | dir: ${common_eval.results_path}/beam${decoding.decoder.beam}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} 9 | sweep: 10 | dir: ${common_eval.results_path} 11 | subdir: beam${decoding.decoder.beam}_th${decoding.decoder.beamthreshold}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} 12 | 13 | task: 14 | _name: hubert_pretraining 15 | single_target: true 16 | data: ??? 17 | normalize: ??? 18 | 19 | decoding: 20 | type: kenlm 21 | lexicon: ??? 22 | lmpath: ??? 23 | beamthreshold: 100 24 | beam: 500 25 | lmweight: 2 26 | wordscore: -1 27 | silweight: 0 28 | unique_wer_file: true 29 | beam: 500 30 | common_eval: 31 | results_path: ??? 32 | path: ??? 33 | post_process: letter 34 | dataset: 35 | max_tokens: 1100000 36 | gen_subset: ??? 37 | -------------------------------------------------------------------------------- /examples/hubert/config/decode/infer_viterbi.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | defaults: 4 | - model: null 5 | 6 | hydra: 7 | run: 8 | dir: ${common_eval.results_path}/beam${decoding.decoder.beam}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} 9 | sweep: 10 | dir: ${common_eval.results_path} 11 | subdir: beam${decoding.decoder.beam}_th${decoding.decoder.beamthreshold}_lmw${decoding.decoder.lmweight}_wrd${decoding.decoder.wordscore}_sil${decoding.decoder.silweight} 12 | 13 | task: 14 | _name: hubert_pretraining 15 | single_target: true 16 | data: ??? 17 | normalize: ??? 18 | 19 | decoding: 20 | type: viterbi 21 | unique_wer_file: true 22 | common_eval: 23 | results_path: ??? 24 | path: ??? 25 | post_process: letter 26 | generation: 27 | nbest: 1 28 | beam: 500 29 | dataset: 30 | max_tokens: 1100000 31 | gen_subset: ??? 32 | -------------------------------------------------------------------------------- /examples/hubert/config/decode/run/submitit_slurm.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | hydra: 3 | launcher: 4 | cpus_per_task: ${distributed_training.distributed_world_size} 5 | gpus_per_node: ${distributed_training.distributed_world_size} 6 | tasks_per_node: ${hydra.launcher.gpus_per_node} 7 | nodes: 1 8 | mem_gb: 200 9 | timeout_min: 4320 10 | max_num_timeout: 50 11 | name: ${hydra.job.config_name} 12 | submitit_folder: ${hydra.sweep.dir}/submitit 13 | 14 | distributed_training: 15 | distributed_world_size: 1 16 | distributed_no_spawn: true 17 | distributed_port: 29761 18 | -------------------------------------------------------------------------------- /examples/hubert/config/decode/run/submitit_slurm_8gpu.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | hydra: 3 | launcher: 4 | cpus_per_task: ${distributed_training.distributed_world_size} 5 | gpus_per_node: ${distributed_training.distributed_world_size} 6 | tasks_per_node: ${hydra.launcher.gpus_per_node} 7 | nodes: 1 8 | mem_gb: 200 9 | timeout_min: 4320 10 | max_num_timeout: 50 11 | name: ${hydra.job.config_name} 12 | submitit_folder: ${hydra.sweep.dir}/submitit 13 | 14 | distributed_training: 15 | distributed_world_size: 8 16 | distributed_no_spawn: true 17 | distributed_port: 29761 18 | -------------------------------------------------------------------------------- /examples/hubert/config/finetune/ckpt/it1.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | 3 | task: 4 | normalize: false 5 | 6 | model: 7 | w2v_path: /checkpoint/wnhsu/w2v/hubert_final/iter1/hubert.km.randcrop.pmw1_0.puw0_0.grpnorm.ml10.mp0_8.untie.mxsz250000.ufreq1.maxtok1400000.MU400k.s1337.ngpu32/checkpoint_last.pt 8 | -------------------------------------------------------------------------------- /examples/hubert/config/finetune/lm/ls_4gram.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | 3 | criterion: 4 | wer_kenlm_model: /checkpoint/abdo/old_checkpoint02/datasets/librispeech/4-gram.bin 5 | wer_lexicon: /checkpoint/abdo/old_checkpoint02/datasets/librispeech/10h/raw/lexicon_ltr.lst 6 | wer_lm_weight: 2.0 7 | wer_word_score: -1.0 8 | -------------------------------------------------------------------------------- /examples/hubert/config/finetune/run/submitit_reg.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | 3 | hydra: 4 | launcher: 5 | cpus_per_task: 8 6 | gpus_per_node: 8 7 | tasks_per_node: ${hydra.launcher.gpus_per_node} 8 | nodes: 1 9 | comment: null 10 | mem_gb: 384 11 | timeout_min: 4320 12 | max_num_timeout: 100 13 | constraint: volta32gb 14 | name: ${hydra.job.config_name}/${hydra.job.override_dirname} 15 | submitit_folder: ${hydra.sweep.dir}/submitit/%j 16 | 17 | distributed_training: 18 | distributed_world_size: 8 19 | distributed_port: 29671 20 | nprocs_per_node: 8 21 | -------------------------------------------------------------------------------- /examples/hubert/config/pretrain/data/iter1.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | 3 | task: 4 | label_dir: ??? 5 | labels: ["km"] 6 | 7 | model: 8 | label_rate: 100 9 | -------------------------------------------------------------------------------- /examples/hubert/config/pretrain/data/iter2.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | 3 | task: 4 | label_dir: ??? 5 | labels: ["km"] 6 | 7 | model: 8 | label_rate: 50 9 | -------------------------------------------------------------------------------- /examples/hubert/config/pretrain/run/submitit_reg.yaml: -------------------------------------------------------------------------------- 1 | # @package _global_ 2 | 3 | hydra: 4 | launcher: 5 | cpus_per_task: 8 6 | gpus_per_node: 8 7 | tasks_per_node: ${hydra.launcher.gpus_per_node} 8 | nodes: 4 9 | comment: null 10 | mem_gb: 384 11 | timeout_min: 4320 12 | max_num_timeout: 100 13 | constraint: volta32gb 14 | name: ${hydra.job.config_name}/${hydra.job.override_dirname} 15 | submitit_folder: ${hydra.sweep.dir}/submitit/%j 16 | 17 | distributed_training: 18 | distributed_world_size: 32 19 | distributed_port: 29671 20 | nprocs_per_node: 8 21 | -------------------------------------------------------------------------------- /examples/hubert/update_ckpt.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import torch 7 | 8 | src_ckpt = "/checkpoint/wnhsu/w2v/archived/hubert_base_ls960_it2.pt" 9 | ref_ckpt = "/checkpoint/wnhsu/w2v/hubert_icassp_oss_v3/iter2_km100-400k-grp-L6/oss.km500_p0_1_s334.pmw1_0.puw0_0.grpnorm.ml10.mp0_8.untie.mxsz250000.ufreq1.maxtok1400000.MU100k.s1337.ngpu32/checkpoint_last.pt" 10 | new_ckpt = "/checkpoint/wnhsu/w2v/archived/hubert_base_ls960_it2_updated.pt" 11 | 12 | 13 | def update_state(state): 14 | state["model"]["label_embs_concat"] = state["model"].pop("label_embs") 15 | state["args"].task = "hubert_pretraining" 16 | state["args"].labels = f"['{state['args'].labels}']" 17 | return state 18 | 19 | 20 | src_state = torch.load(src_ckpt) 21 | src_state = update_state(src_state) 22 | torch.save(src_state, new_ckpt) 23 | -------------------------------------------------------------------------------- /examples/language_model/prepare-wikitext-103.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Adapted from https://github.com/facebookresearch/MIXER/blob/master/prepareData.sh 3 | 4 | URLS=( 5 | "https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-v1.zip" 6 | ) 7 | FILES=( 8 | "wikitext-103-v1.zip" 9 | ) 10 | 11 | for ((i=0;i<${#URLS[@]};++i)); do 12 | file=${FILES[i]} 13 | if [ -f $file ]; then 14 | echo "$file already exists, skipping download" 15 | else 16 | url=${URLS[i]} 17 | wget "$url" 18 | if [ -f $file ]; then 19 | echo "$url successfully downloaded." 20 | else 21 | echo "$url not successfully downloaded." 22 | exit -1 23 | fi 24 | if [ ${file: -4} == ".tgz" ]; then 25 | tar zxvf $file 26 | elif [ ${file: -4} == ".tar" ]; then 27 | tar xvf $file 28 | elif [ ${file: -4} == ".zip" ]; then 29 | unzip $file 30 | fi 31 | fi 32 | done 33 | cd .. 34 | -------------------------------------------------------------------------------- /examples/laser/laser_src/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .laser_task import * # noqa 7 | from .laser_lstm import * # noqa 8 | from .laser_transformer import * # noqa 9 | -------------------------------------------------------------------------------- /examples/latent_depth/latent_depth_src/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import multilingual_translation_latent_depth # noqa 7 | from .loss import latent_depth # noqa 8 | from .models import latent_multilingual_transformer # noqa 9 | from .modules import latent_layers # noqa 10 | -------------------------------------------------------------------------------- /examples/latent_depth/latent_depth_src/loss/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/latent_depth/latent_depth_src/loss/__init__.py -------------------------------------------------------------------------------- /examples/latent_depth/latent_depth_src/models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/latent_depth/latent_depth_src/models/__init__.py -------------------------------------------------------------------------------- /examples/latent_depth/latent_depth_src/modules/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/latent_depth/latent_depth_src/modules/__init__.py -------------------------------------------------------------------------------- /examples/linformer/README.md: -------------------------------------------------------------------------------- 1 | # Linformer: Self-Attention with Linear Complexity (Wang et al., 2020) 2 | 3 | This example contains code to train Linformer models as described in our paper 4 | [Linformer: Self-Attention with Linear Complexity](https://arxiv.org/abs/2006.04768). 5 | 6 | ## Training a new Linformer RoBERTa model 7 | 8 | You can mostly follow the [RoBERTa pretraining README](/examples/roberta/README.pretraining.md), 9 | updating your training command with `--user-dir examples/linformer/linformer_src --arch linformer_roberta_base`. 10 | 11 | ## Citation 12 | 13 | If you use our work, please cite: 14 | 15 | ```bibtex 16 | @article{wang2020linformer, 17 | title={Linformer: Self-Attention with Linear Complexity}, 18 | author={Wang, Sinong and Li, Belinda and Khabsa, Madian and Fang, Han and Ma, Hao}, 19 | journal={arXiv preprint arXiv:2006.04768}, 20 | year={2020} 21 | } 22 | ``` 23 | -------------------------------------------------------------------------------- /examples/linformer/linformer_src/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .models import linformer_roberta # noqa 7 | -------------------------------------------------------------------------------- /examples/linformer/linformer_src/models/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/linformer/linformer_src/models/__init__.py -------------------------------------------------------------------------------- /examples/linformer/linformer_src/modules/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/linformer/linformer_src/modules/__init__.py -------------------------------------------------------------------------------- /examples/m2m_100/tokenizers/README.md: -------------------------------------------------------------------------------- 1 | # M2M-100 Tokenization 2 | 3 | We apply different tokenization strategies for different languages following the existing literature. Here we provide tok.sh a tokenizer that can be used to reproduce our results. 4 | 5 | To reproduce the results, follow these steps: 6 | 7 | ``` 8 | tgt_lang=... 9 | reference_translation=... 10 | cat generation_output | grep -P "^H" | sort -V | cut -f 3- | sh tok.sh $tgt_lang > hyp 11 | cat $reference_translation |sh tok.sh $tgt_lang > ref 12 | sacrebleu -tok 'none' ref < hyp 13 | ``` 14 | 15 | ## Installation 16 | 17 | Tools needed for all the languages except Arabic can be installed by running install_dependencies.sh 18 | If you want to evaluate Arabic models, please follow the instructions provided here: http://alt.qcri.org/tools/arabic-normalizer/ to install 19 | -------------------------------------------------------------------------------- /examples/m2m_100/tokenizers/seg_ja.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | SCRIPT=`realpath $0` 7 | KYTEA=`dirname $SCRIPT`/thirdparty/kytea 8 | export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$KYTEA/lib:/usr/local/lib 9 | export PATH=$PATH:"$KYTEA/bin" 10 | 11 | cat - | tr -d "[:blank:]" | kytea -notags 12 | -------------------------------------------------------------------------------- /examples/m2m_100/tokenizers/seg_ko.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | SCRIPT=`realpath $0` 7 | MECAB=`dirname $SCRIPT`/thirdparty/mecab-0.996-ko-0.9.2 8 | 9 | export PATH=$PATH:"$MECAB/bin":"$MECAB/lib" 10 | export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:"$MECAB/lib" 11 | 12 | cat - | mecab -O wakati 13 | -------------------------------------------------------------------------------- /examples/m2m_100/tokenizers/thirdparty/.gitignore: -------------------------------------------------------------------------------- 1 | seg_my.py 2 | indic_nlp_library/ 3 | indic_nlp_resources/ 4 | kytea/ 5 | mecab-0.996-ko-0.9.2.tar.gz 6 | mecab-0.996-ko-0.9.2/ 7 | mosesdecoder/ 8 | wat2020.my-en.zip 9 | wat2020.my-en/ 10 | wmt16-scripts/ 11 | mecab-ko-dic-2.1.1-20180720/ 12 | mecab-ko-dic-2.1.1-20180720.tar.gz -------------------------------------------------------------------------------- /examples/m2m_100/tokenizers/tokenize_indic.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | # Use: echo {text} | python tokenize_indic.py {language} 8 | 9 | import sys 10 | 11 | from indicnlp.normalize.indic_normalize import IndicNormalizerFactory 12 | from indicnlp.tokenize.indic_tokenize import trivial_tokenize 13 | 14 | 15 | factory = IndicNormalizerFactory() 16 | normalizer = factory.get_normalizer( 17 | sys.argv[1], remove_nuktas=False, nasals_mode="do_nothing" 18 | ) 19 | 20 | for line in sys.stdin: 21 | normalized_line = normalizer.normalize(line.strip()) 22 | tokenized_line = " ".join(trivial_tokenize(normalized_line, sys.argv[1])) 23 | print(tokenized_line) 24 | -------------------------------------------------------------------------------- /examples/m2m_100/tokenizers/tokenize_thai.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import sys 8 | 9 | from pythainlp import word_tokenize 10 | 11 | 12 | for line in sys.stdin: 13 | print(" ".join(word_tokenize(line.strip()))) 14 | -------------------------------------------------------------------------------- /examples/m2m_100/tokenizers/tokenize_zh.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | 8 | import fileinput 9 | 10 | import sacrebleu 11 | 12 | 13 | for line in fileinput.input(): 14 | print(sacrebleu.tokenize_zh(line)) 15 | -------------------------------------------------------------------------------- /examples/m2m_100/tokenizers/tokenizer_ar.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env sh 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | # 7 | # Please follow the instructions here http://alt.qcri.org/tools/arabic-normalizer/ 8 | # to install tools needed for Arabic 9 | 10 | echo "Please install Arabic tools: http://alt.qcri.org/tools/arabic-normalizer/" 11 | echo "Then update environment variables in tokenizer_ar.sh" 12 | exit 1 13 | 14 | SVMTOOL=... 15 | GOMOSESGO=... 16 | QCRI_ARABIC_NORMALIZER=... 17 | 18 | export PERL5LIB="$SVMTOOL/lib":"$GOMOSESGO/bin/MADA-3.2":$PERL5LIB 19 | 20 | 21 | tempfile=$(mktemp) 22 | cat - > $tempfile 23 | 24 | cd $QCRI_ARABIC_NORMALIZER 25 | 26 | bash qcri_normalizer_mada3.2_aramorph1.2.1.sh $tempfile 27 | cat $tempfile.mada_norm-aramorph.europarl_tok 28 | -------------------------------------------------------------------------------- /examples/megatron_11b/detok.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 -u 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import argparse 8 | import fileinput 9 | 10 | import sacremoses 11 | 12 | 13 | def main(): 14 | parser = argparse.ArgumentParser(description="") 15 | parser.add_argument("files", nargs="*", help="input files") 16 | args = parser.parse_args() 17 | 18 | detok = sacremoses.MosesDetokenizer() 19 | 20 | for line in fileinput.input(args.files, openhook=fileinput.hook_compressed): 21 | print( 22 | detok.detokenize(line.strip().split(" ")) 23 | .replace(" @", "") 24 | .replace("@ ", "") 25 | .replace(" =", "=") 26 | .replace("= ", "=") 27 | .replace(" – ", "–") 28 | ) 29 | 30 | 31 | if __name__ == "__main__": 32 | main() 33 | -------------------------------------------------------------------------------- /examples/multilingual/ML50_langs.txt: -------------------------------------------------------------------------------- 1 | ar_AR 2 | cs_CZ 3 | de_DE 4 | en_XX 5 | es_XX 6 | et_EE 7 | fi_FI 8 | fr_XX 9 | gu_IN 10 | hi_IN 11 | it_IT 12 | ja_XX 13 | kk_KZ 14 | ko_KR 15 | lt_LT 16 | lv_LV 17 | my_MM 18 | ne_NP 19 | nl_XX 20 | ro_RO 21 | ru_RU 22 | si_LK 23 | tr_TR 24 | vi_VN 25 | zh_CN 26 | af_ZA 27 | az_AZ 28 | bn_IN 29 | fa_IR 30 | he_IL 31 | hr_HR 32 | id_ID 33 | ka_GE 34 | km_KH 35 | mk_MK 36 | ml_IN 37 | mn_MN 38 | mr_IN 39 | pl_PL 40 | ps_AF 41 | pt_XX 42 | sv_SE 43 | sw_KE 44 | ta_IN 45 | te_IN 46 | th_TH 47 | tl_XX 48 | uk_UA 49 | ur_PK 50 | xh_ZA 51 | gl_ES 52 | sl_SI -------------------------------------------------------------------------------- /examples/multilingual/data_scripts/README.md: -------------------------------------------------------------------------------- 1 | 2 | # Install dependency 3 | ```bash 4 | pip install -r requirement.txt 5 | ``` 6 | 7 | # Download the data set 8 | ```bash 9 | export WORKDIR_ROOT= 10 | 11 | ``` 12 | The downloaded data will be at $WORKDIR_ROOT/ML50 13 | 14 | # preprocess the data 15 | Install SPM [here](https://github.com/google/sentencepiece) 16 | ```bash 17 | export WORKDIR_ROOT= 18 | export SPM_PATH= 19 | ``` 20 | * $WORKDIR_ROOT/ML50/raw: extracted raw data 21 | * $WORKDIR_ROOT/ML50/dedup: dedup data 22 | * $WORKDIR_ROOT/ML50/clean: data with valid and test sentences removed from the dedup data 23 | 24 | 25 | -------------------------------------------------------------------------------- /examples/multilingual/data_scripts/download_ML50_v1.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # All rights reserved. 4 | # 5 | # This source code is licensed under the license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | if [ -z $WORKDIR_ROOT ] ; 9 | then 10 | echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." 11 | exit 12 | fi 13 | 14 | # first run download_wmt20.sh; it will install a few useful tools for other scripts 15 | # TODO: need to print out instructions on downloading a few files which requires manually authentication from the websites 16 | bash ./download_wmt20.sh 17 | 18 | python ./download_wmt19_and_before.py 19 | bash ./download_wat19_my.sh 20 | python ./download_ted_and_extract.py 21 | bash ./download_lotus.sh 22 | bash ./download_iitb.sh 23 | bash ./download_af_xh.sh 24 | 25 | 26 | # IWSLT downloading URLs have changed in between; TODO: fix them: 27 | bash ./download_iwslt_and_extract.sh 28 | 29 | # TODO: globalvoices URLs changed; need to be fixed 30 | bash ./download_flores_data.sh 31 | -------------------------------------------------------------------------------- /examples/multilingual/data_scripts/download_iitb.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # All rights reserved. 4 | # 5 | # This source code is licensed under the license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | 9 | if [ -z $WORKDIR_ROOT ] ; 10 | then 11 | echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." 12 | exit 13 | fi 14 | 15 | IITB=$WORKDIR_ROOT/IITB 16 | mkdir -p $IITB 17 | pushd $IITB 18 | 19 | wget http://www.cfilt.iitb.ac.in/~moses/iitb_en_hi_parallel/iitb_corpus_download/parallel.tgz 20 | tar -xvzf parallel.tgz 21 | 22 | wget http://www.cfilt.iitb.ac.in/~moses/iitb_en_hi_parallel/iitb_corpus_download/dev_test.tgz 23 | tar -xvzf dev_test.tgz 24 | 25 | DESTDIR=${WORKDIR_ROOT}/ML50/raw/ 26 | 27 | cp parallel/IITB.en-hi.en $DESTDIR/train.hi_IN-en_XX.en_XX 28 | cp parallel/IITB.en-hi.hi $DESTDIR/train.hi_IN-en_XX.hi_IN 29 | 30 | cp dev_test/dev.en $DESTDIR/valid.hi_IN-en_XX.en_XX 31 | cp dev_test/dev.hi $DESTDIR/valid.hi_IN-en_XX.hi_IN 32 | 33 | cp dev_test/test.en $DESTDIR/test.hi_IN-en_XX.en_XX 34 | cp dev_test/test.hi $DESTDIR/test.hi_IN-en_XX.hi_IN 35 | popd -------------------------------------------------------------------------------- /examples/multilingual/data_scripts/download_wat19_my.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # All rights reserved. 4 | # 5 | # This source code is licensed under the license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | 9 | if [ -z $WORKDIR_ROOT ] ; 10 | then 11 | echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." 12 | exit 13 | fi 14 | 15 | 16 | SRCDIR=$WORKDIR_ROOT/indic_languages_corpus 17 | DESTDIR=$WORKDIR_ROOT/ML50/raw 18 | mkdir -p $SRCDIR 19 | mkdir -p $DESTDIR 20 | 21 | WAT_MY_EN=wat2020.my-en.zip 22 | cd $SRCDIR 23 | # please refer to http://lotus.kuee.kyoto-u.ac.jp/WAT/my-en-data/ for latest URL if the following url expired 24 | #- The data used for WAT2020 are identical to those used in WAT2019. 25 | wget http://lotus.kuee.kyoto-u.ac.jp/WAT/my-en-data/$WAT_MY_EN 26 | unzip $WAT_MY_EN 27 | 28 | 29 | SRC_EXTRACT_DIR=$SRCDIR/wat2020.my-en/alt 30 | 31 | cp $SRC_EXTRACT_DIR/train.alt.en $DESTDIR/train.my_MM-en_XX.en_XX 32 | cp $SRC_EXTRACT_DIR/train.alt.my $DESTDIR/train.my_MM-en_XX.my_MM 33 | cp $SRC_EXTRACT_DIR/dev.alt.en $DESTDIR/valid.my_MM-en_XX.en_XX 34 | cp $SRC_EXTRACT_DIR/dev.alt.my $DESTDIR/valid.my_MM-en_XX.my_MM 35 | cp $SRC_EXTRACT_DIR/test.alt.en $DESTDIR/test.my_MM-en_XX.en_XX 36 | cp $SRC_EXTRACT_DIR/test.alt.my $DESTDIR/test.my_MM-en_XX.my_MM 37 | -------------------------------------------------------------------------------- /examples/multilingual/data_scripts/preprocess_ML50_v1.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # All rights reserved. 4 | # 5 | # This source code is licensed under the license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | if [ -z $WORKDIR_ROOT ] ; 9 | then 10 | echo "please specify your working directory root in environment variable WORKDIR_ROOT. Exitting..." 11 | exit 12 | fi 13 | 14 | if [ -z $SPM_PATH ] ; 15 | then 16 | echo "Please install sentence piecence from https://github.com/google/sentencepiece and set SPM_PATH pointing to the installed spm_encode.py. Exitting..." 17 | exit 18 | fi 19 | 20 | ML50=${WORKDIR_ROOT}/ML50 21 | 22 | mkdir -p $ML50/dedup 23 | mkdir -p $ML50/cleaned_dedup 24 | 25 | python ./dedup_all.py --from-folder $ML50/raw --to-folder $ML50/dedup 26 | python ./remove_valid_test_in_train.py --from-folder $ML50/dedup --to-folder $ML50/clean 27 | python ./binarize.py --raw-folder $ML50/clean -------------------------------------------------------------------------------- /examples/multilingual/data_scripts/requirement.txt: -------------------------------------------------------------------------------- 1 | wget 2 | pandas -------------------------------------------------------------------------------- /examples/multilingual/data_scripts/utils/strip_sgm.sh: -------------------------------------------------------------------------------- 1 | grep "seg id" | sed 's///g' | sed 's/<\/seg>//g' 2 | -------------------------------------------------------------------------------- /examples/multilingual/multilingual_fairseq_gen.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # All rights reserved. 4 | # 5 | # This source code is licensed under the license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | lang_pairs="en-fr,en-cs,fr-en,cs-en" 9 | path_2_data=$1 # 10 | lang_list=$2 # 11 | model=$3 # 12 | source_lang=cs 13 | target_lang=en 14 | 15 | fairseq-generate "$path_2_data" \ 16 | --path "$model" \ 17 | --task translation_multi_simple_epoch \ 18 | --gen-subset test \ 19 | --source-lang "$source_lang" \ 20 | --target-lang "$target_lang" \ 21 | --sacrebleu --remove-bpe 'sentencepiece'\ 22 | --batch-size 32 \ 23 | --encoder-langtok "src" \ 24 | --decoder-langtok \ 25 | --lang-dict "$lang_list" \ 26 | --lang-pairs "$lang_pairs" 27 | -------------------------------------------------------------------------------- /examples/noisychannel/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .rerank_options import * # noqa 7 | -------------------------------------------------------------------------------- /examples/pointer_generator/pointer_generator_src/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import transformer_pg # noqa 7 | -------------------------------------------------------------------------------- /examples/quant_noise/transformer_quantization_config.yaml: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | # This file defines example configuration arguments for quantizing 7 | # a transformer model with product quantization 8 | 9 | # Number of Centroids for Product Quantization, by default 256 (byte-aligned) 10 | n_centroids: 11 | Linear: 12 | key: in_features 13 | value: {"*": 256} 14 | Embedding: 15 | key: embedding_dim 16 | value: {"*": 256} 17 | 18 | # Block Sizes for Product Quantization 19 | # We suggest: 8 for FFN, 4 for ATTN, 4 for embedding projections, 8 for embeddings 20 | block_sizes: 21 | Linear: 22 | key: fuzzy_name 23 | value: {fc: 8, attn: 4, emb: 4} 24 | Embedding: 25 | key: fuzzy_name 26 | value: {emb: 8} 27 | 28 | # Layers to Quantize Sequentially 29 | # We suggest: first FFN, then EMB, then ATTN 30 | layers_to_quantize: 31 | - decoder\\.layers\\.\d+\\.fc[12] 32 | - decoder\\.embed_tokens\\.embeddings\\.[012]\\.[01] 33 | - decoder\\.layers\\.\d+\\.self_attn\\.(k_proj|v_proj|q_proj|out_proj) 34 | -------------------------------------------------------------------------------- /examples/roberta/commonsense_qa/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import commonsense_qa_task # noqa 7 | -------------------------------------------------------------------------------- /examples/roberta/commonsense_qa/download_cqa_data.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | OUTDIR=data/CommonsenseQA 8 | 9 | mkdir -p $OUTDIR 10 | 11 | wget -O $OUTDIR/train.jsonl https://s3.amazonaws.com/commensenseqa/train_rand_split.jsonl 12 | wget -O $OUTDIR/valid.jsonl https://s3.amazonaws.com/commensenseqa/dev_rand_split.jsonl 13 | wget -O $OUTDIR/test.jsonl https://s3.amazonaws.com/commensenseqa/test_rand_split_no_answers.jsonl 14 | wget -O $OUTDIR/dict.txt https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt 15 | -------------------------------------------------------------------------------- /examples/roberta/config/finetuning/cola.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | fp16_init_scale: 4 6 | threshold_loss_scale: 1 7 | fp16_scale_window: 128 8 | log_format: json 9 | log_interval: 200 10 | 11 | task: 12 | _name: sentence_prediction 13 | data: ??? 14 | init_token: 0 15 | separator_token: 2 16 | num_classes: 2 17 | max_positions: 512 18 | 19 | checkpoint: 20 | restore_file: ??? 21 | reset_optimizer: true 22 | reset_dataloader: true 23 | reset_meters: true 24 | best_checkpoint_metric: accuracy 25 | maximize_best_checkpoint_metric: true 26 | no_epoch_checkpoints: true 27 | 28 | distributed_training: 29 | find_unused_parameters: true 30 | distributed_world_size: 1 31 | 32 | criterion: 33 | _name: sentence_prediction 34 | 35 | dataset: 36 | batch_size: 16 37 | required_batch_size_multiple: 1 38 | max_tokens: 4400 39 | 40 | optimizer: 41 | _name: adam 42 | weight_decay: 0.1 43 | adam_betas: (0.9,0.98) 44 | adam_eps: 1e-06 45 | 46 | lr_scheduler: 47 | _name: polynomial_decay 48 | warmup_updates: 320 49 | 50 | optimization: 51 | clip_norm: 0.0 52 | lr: [1e-05] 53 | max_update: 5336 54 | max_epoch: 10 55 | 56 | model: 57 | _name: roberta 58 | dropout: 0.1 59 | attention_dropout: 0.1 60 | -------------------------------------------------------------------------------- /examples/roberta/config/finetuning/mnli.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | fp16_init_scale: 4 6 | threshold_loss_scale: 1 7 | fp16_scale_window: 128 8 | log_format: json 9 | log_interval: 200 10 | 11 | task: 12 | _name: sentence_prediction 13 | data: ??? 14 | init_token: 0 15 | separator_token: 2 16 | num_classes: 3 17 | max_positions: 512 18 | 19 | checkpoint: 20 | restore_file: ??? 21 | reset_optimizer: true 22 | reset_dataloader: true 23 | reset_meters: true 24 | best_checkpoint_metric: accuracy 25 | maximize_best_checkpoint_metric: true 26 | no_epoch_checkpoints: true 27 | 28 | distributed_training: 29 | find_unused_parameters: true 30 | distributed_world_size: 1 31 | 32 | criterion: 33 | _name: sentence_prediction 34 | 35 | dataset: 36 | batch_size: 32 37 | required_batch_size_multiple: 1 38 | max_tokens: 4400 39 | 40 | optimizer: 41 | _name: adam 42 | weight_decay: 0.1 43 | adam_betas: (0.9,0.98) 44 | adam_eps: 1e-06 45 | 46 | lr_scheduler: 47 | _name: polynomial_decay 48 | warmup_updates: 7432 49 | 50 | optimization: 51 | clip_norm: 0.0 52 | lr: [1e-05] 53 | max_update: 123873 54 | max_epoch: 10 55 | 56 | model: 57 | _name: roberta 58 | dropout: 0.1 59 | attention_dropout: 0.1 60 | -------------------------------------------------------------------------------- /examples/roberta/config/finetuning/mrpc.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | fp16_init_scale: 4 6 | threshold_loss_scale: 1 7 | fp16_scale_window: 128 8 | log_format: json 9 | log_interval: 200 10 | 11 | task: 12 | _name: sentence_prediction 13 | data: ??? 14 | init_token: 0 15 | separator_token: 2 16 | num_classes: 2 17 | max_positions: 512 18 | 19 | checkpoint: 20 | restore_file: ??? 21 | reset_optimizer: true 22 | reset_dataloader: true 23 | reset_meters: true 24 | best_checkpoint_metric: accuracy 25 | maximize_best_checkpoint_metric: true 26 | no_epoch_checkpoints: true 27 | 28 | distributed_training: 29 | find_unused_parameters: true 30 | distributed_world_size: 1 31 | 32 | criterion: 33 | _name: sentence_prediction 34 | 35 | dataset: 36 | batch_size: 16 37 | required_batch_size_multiple: 1 38 | max_tokens: 4400 39 | 40 | optimizer: 41 | _name: adam 42 | weight_decay: 0.1 43 | adam_betas: (0.9,0.98) 44 | adam_eps: 1e-06 45 | 46 | lr_scheduler: 47 | _name: polynomial_decay 48 | warmup_updates: 137 49 | 50 | optimization: 51 | clip_norm: 0.0 52 | lr: [1e-05] 53 | max_update: 2296 54 | max_epoch: 10 55 | 56 | model: 57 | _name: roberta 58 | dropout: 0.1 59 | attention_dropout: 0.1 60 | -------------------------------------------------------------------------------- /examples/roberta/config/finetuning/qnli.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | fp16_init_scale: 4 6 | threshold_loss_scale: 1 7 | fp16_scale_window: 128 8 | log_format: json 9 | log_interval: 200 10 | 11 | task: 12 | _name: sentence_prediction 13 | data: ??? 14 | init_token: 0 15 | separator_token: 2 16 | num_classes: 2 17 | max_positions: 512 18 | 19 | checkpoint: 20 | restore_file: ??? 21 | reset_optimizer: true 22 | reset_dataloader: true 23 | reset_meters: true 24 | best_checkpoint_metric: accuracy 25 | maximize_best_checkpoint_metric: true 26 | no_epoch_checkpoints: true 27 | 28 | distributed_training: 29 | find_unused_parameters: true 30 | distributed_world_size: 1 31 | 32 | criterion: 33 | _name: sentence_prediction 34 | 35 | dataset: 36 | batch_size: 32 37 | required_batch_size_multiple: 1 38 | max_tokens: 4400 39 | 40 | optimizer: 41 | _name: adam 42 | weight_decay: 0.1 43 | adam_betas: (0.9,0.98) 44 | adam_eps: 1e-06 45 | 46 | lr_scheduler: 47 | _name: polynomial_decay 48 | warmup_updates: 1986 49 | 50 | optimization: 51 | clip_norm: 0.0 52 | lr: [1e-05] 53 | max_update: 33112 54 | max_epoch: 10 55 | 56 | model: 57 | _name: roberta 58 | dropout: 0.1 59 | attention_dropout: 0.1 60 | -------------------------------------------------------------------------------- /examples/roberta/config/finetuning/qqp.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | fp16_init_scale: 4 6 | threshold_loss_scale: 1 7 | fp16_scale_window: 128 8 | log_format: json 9 | log_interval: 200 10 | 11 | task: 12 | _name: sentence_prediction 13 | data: ??? 14 | init_token: 0 15 | separator_token: 2 16 | num_classes: 2 17 | max_positions: 512 18 | 19 | checkpoint: 20 | restore_file: ??? 21 | reset_optimizer: true 22 | reset_dataloader: true 23 | reset_meters: true 24 | best_checkpoint_metric: accuracy 25 | maximize_best_checkpoint_metric: true 26 | no_epoch_checkpoints: true 27 | 28 | distributed_training: 29 | find_unused_parameters: true 30 | distributed_world_size: 1 31 | 32 | criterion: 33 | _name: sentence_prediction 34 | 35 | dataset: 36 | batch_size: 32 37 | required_batch_size_multiple: 1 38 | max_tokens: 4400 39 | 40 | optimizer: 41 | _name: adam 42 | weight_decay: 0.1 43 | adam_betas: (0.9,0.98) 44 | adam_eps: 1e-06 45 | 46 | lr_scheduler: 47 | _name: polynomial_decay 48 | warmup_updates: 28318 49 | 50 | optimization: 51 | clip_norm: 0.0 52 | lr: [1e-05] 53 | max_update: 113272 54 | max_epoch: 10 55 | 56 | model: 57 | _name: roberta 58 | dropout: 0.1 59 | attention_dropout: 0.1 60 | -------------------------------------------------------------------------------- /examples/roberta/config/finetuning/rte.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | fp16_init_scale: 4 6 | threshold_loss_scale: 1 7 | fp16_scale_window: 128 8 | log_format: json 9 | log_interval: 200 10 | 11 | task: 12 | _name: sentence_prediction 13 | data: ??? 14 | init_token: 0 15 | separator_token: 2 16 | num_classes: 2 17 | max_positions: 512 18 | 19 | checkpoint: 20 | restore_file: ??? 21 | reset_optimizer: true 22 | reset_dataloader: true 23 | reset_meters: true 24 | best_checkpoint_metric: accuracy 25 | maximize_best_checkpoint_metric: true 26 | no_epoch_checkpoints: true 27 | 28 | distributed_training: 29 | find_unused_parameters: true 30 | distributed_world_size: 1 31 | 32 | criterion: 33 | _name: sentence_prediction 34 | 35 | dataset: 36 | batch_size: 16 37 | required_batch_size_multiple: 1 38 | max_tokens: 4400 39 | 40 | optimizer: 41 | _name: adam 42 | weight_decay: 0.1 43 | adam_betas: (0.9,0.98) 44 | adam_eps: 1e-06 45 | 46 | lr_scheduler: 47 | _name: polynomial_decay 48 | warmup_updates: 122 49 | 50 | optimization: 51 | clip_norm: 0.0 52 | lr: [2e-05] 53 | max_update: 2036 54 | max_epoch: 10 55 | 56 | model: 57 | _name: roberta 58 | dropout: 0.1 59 | attention_dropout: 0.1 60 | -------------------------------------------------------------------------------- /examples/roberta/config/finetuning/sst_2.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | fp16_init_scale: 4 6 | threshold_loss_scale: 1 7 | fp16_scale_window: 128 8 | log_format: json 9 | log_interval: 200 10 | 11 | task: 12 | _name: sentence_prediction 13 | data: ??? 14 | init_token: 0 15 | separator_token: 2 16 | num_classes: 2 17 | max_positions: 512 18 | 19 | checkpoint: 20 | restore_file: ??? 21 | reset_optimizer: true 22 | reset_dataloader: true 23 | reset_meters: true 24 | best_checkpoint_metric: accuracy 25 | maximize_best_checkpoint_metric: true 26 | no_epoch_checkpoints: true 27 | 28 | distributed_training: 29 | find_unused_parameters: true 30 | distributed_world_size: 1 31 | 32 | criterion: 33 | _name: sentence_prediction 34 | 35 | dataset: 36 | batch_size: 32 37 | required_batch_size_multiple: 1 38 | max_tokens: 4400 39 | 40 | optimizer: 41 | _name: adam 42 | weight_decay: 0.1 43 | adam_betas: (0.9,0.98) 44 | adam_eps: 1e-06 45 | 46 | lr_scheduler: 47 | _name: polynomial_decay 48 | warmup_updates: 1256 49 | 50 | optimization: 51 | clip_norm: 0.0 52 | lr: [1e-05] 53 | max_update: 20935 54 | max_epoch: 10 55 | 56 | model: 57 | _name: roberta 58 | dropout: 0.1 59 | attention_dropout: 0.1 60 | -------------------------------------------------------------------------------- /examples/roberta/config/finetuning/sts_b.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | fp16_init_scale: 4 6 | threshold_loss_scale: 1 7 | fp16_scale_window: 128 8 | log_format: json 9 | log_interval: 200 10 | 11 | task: 12 | _name: sentence_prediction 13 | data: ??? 14 | init_token: 0 15 | separator_token: 2 16 | num_classes: 1 17 | max_positions: 512 18 | 19 | checkpoint: 20 | restore_file: ??? 21 | reset_optimizer: true 22 | reset_dataloader: true 23 | reset_meters: true 24 | no_epoch_checkpoints: true 25 | 26 | distributed_training: 27 | find_unused_parameters: true 28 | distributed_world_size: 1 29 | 30 | criterion: 31 | _name: sentence_prediction 32 | regression_target: true 33 | 34 | dataset: 35 | batch_size: 16 36 | required_batch_size_multiple: 1 37 | max_tokens: 4400 38 | 39 | optimizer: 40 | _name: adam 41 | weight_decay: 0.1 42 | adam_betas: (0.9,0.98) 43 | adam_eps: 1e-06 44 | 45 | lr_scheduler: 46 | _name: polynomial_decay 47 | warmup_updates: 214 48 | 49 | optimization: 50 | clip_norm: 0.0 51 | lr: [2e-05] 52 | max_update: 3598 53 | max_epoch: 10 54 | 55 | model: 56 | _name: roberta 57 | dropout: 0.1 58 | attention_dropout: 0.1 59 | -------------------------------------------------------------------------------- /examples/roberta/config/pretraining/base.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | common: 3 | fp16: true 4 | log_format: json 5 | log_interval: 200 6 | 7 | checkpoint: 8 | no_epoch_checkpoints: true 9 | 10 | task: 11 | _name: masked_lm 12 | data: ??? 13 | sample_break_mode: complete 14 | tokens_per_sample: 512 15 | 16 | criterion: masked_lm 17 | 18 | dataset: 19 | batch_size: 16 20 | ignore_unused_valid_subsets: true 21 | 22 | optimizer: 23 | _name: adam 24 | weight_decay: 0.01 25 | adam_betas: (0.9,0.98) 26 | adam_eps: 1e-06 27 | 28 | lr_scheduler: 29 | _name: polynomial_decay 30 | warmup_updates: 10000 31 | 32 | optimization: 33 | clip_norm: 0 34 | lr: [0.0005] 35 | max_update: 125000 36 | update_freq: [16] 37 | 38 | model: 39 | _name: roberta 40 | max_positions: 512 41 | dropout: 0.1 42 | attention_dropout: 0.1 43 | -------------------------------------------------------------------------------- /examples/roberta/wsc/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import wsc_criterion # noqa 7 | from . import wsc_task # noqa 8 | -------------------------------------------------------------------------------- /examples/rxf/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import rxf_src # noqa 7 | -------------------------------------------------------------------------------- /examples/rxf/rxf_src/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import label_smoothed_cross_entropy_r3f, sentence_prediction_r3f # noqa 7 | -------------------------------------------------------------------------------- /examples/simultaneous_translation/README.md: -------------------------------------------------------------------------------- 1 | # Simultaneous Translation 2 | Examples of simultaneous translation in fairseq 3 | - [English-to-Japanese text-to-text wait-k model](docs/enja-waitk.md) 4 | - [English-to-Germen text-to-text monotonic multihead attention model](docs/ende-mma.md) 5 | - [English-to-Germen speech-to-text simultaneous translation model](../speech_to_text/docs/simulst_mustc_example.md) 6 | -------------------------------------------------------------------------------- /examples/simultaneous_translation/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import models # noqa 7 | -------------------------------------------------------------------------------- /examples/simultaneous_translation/models/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import importlib 7 | import os 8 | 9 | 10 | for file in sorted(os.listdir(os.path.dirname(__file__))): 11 | if file.endswith(".py") and not file.startswith("_"): 12 | model_name = file[: file.find(".py")] 13 | importlib.import_module( 14 | "examples.simultaneous_translation.models." + model_name 15 | ) 16 | -------------------------------------------------------------------------------- /examples/simultaneous_translation/modules/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import importlib 7 | import os 8 | 9 | from fairseq import registry 10 | 11 | 12 | ( 13 | build_monotonic_attention, 14 | register_monotonic_attention, 15 | MONOTONIC_ATTENTION_REGISTRY, 16 | _, 17 | ) = registry.setup_registry("--simul-type") 18 | 19 | for file in sorted(os.listdir(os.path.dirname(__file__))): 20 | if file.endswith(".py") and not file.startswith("_"): 21 | model_name = file[: file.find(".py")] 22 | importlib.import_module( 23 | "examples.simultaneous_translation.modules." + model_name 24 | ) 25 | -------------------------------------------------------------------------------- /examples/simultaneous_translation/utils/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import importlib 7 | import os 8 | 9 | 10 | # automatically import any Python files in the criterions/ directory 11 | for file in sorted(os.listdir(os.path.dirname(__file__))): 12 | if file.endswith(".py") and not file.startswith("_"): 13 | module = file[: file.find(".py")] 14 | importlib.import_module("examples.simultaneous_translation.utils." + module) 15 | -------------------------------------------------------------------------------- /examples/speech_recognition/__init__.py: -------------------------------------------------------------------------------- 1 | from . import criterions, models, tasks # noqa 2 | -------------------------------------------------------------------------------- /examples/speech_recognition/criterions/__init__.py: -------------------------------------------------------------------------------- 1 | import importlib 2 | import os 3 | 4 | 5 | # ASG loss requires flashlight bindings 6 | files_to_skip = set() 7 | try: 8 | import flashlight.lib.sequence.criterion 9 | except ImportError: 10 | files_to_skip.add("ASG_loss.py") 11 | 12 | for file in sorted(os.listdir(os.path.dirname(__file__))): 13 | if file.endswith(".py") and not file.startswith("_") and file not in files_to_skip: 14 | criterion_name = file[: file.find(".py")] 15 | importlib.import_module( 16 | "examples.speech_recognition.criterions." + criterion_name 17 | ) 18 | -------------------------------------------------------------------------------- /examples/speech_recognition/data/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .asr_dataset import AsrDataset 7 | 8 | 9 | __all__ = [ 10 | "AsrDataset", 11 | ] 12 | -------------------------------------------------------------------------------- /examples/speech_recognition/kaldi/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/speech_recognition/kaldi/__init__.py -------------------------------------------------------------------------------- /examples/speech_recognition/kaldi/config/kaldi_initializer.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | data_dir: ??? 4 | fst_dir: ??? 5 | in_labels: ??? 6 | kaldi_root: ??? 7 | lm_arpa: ??? 8 | blank_symbol: 9 | -------------------------------------------------------------------------------- /examples/speech_recognition/models/__init__.py: -------------------------------------------------------------------------------- 1 | import importlib 2 | import os 3 | 4 | 5 | for file in sorted(os.listdir(os.path.dirname(__file__))): 6 | if file.endswith(".py") and not file.startswith("_"): 7 | model_name = file[: file.find(".py")] 8 | importlib.import_module("examples.speech_recognition.models." + model_name) 9 | -------------------------------------------------------------------------------- /examples/speech_recognition/new/README.md: -------------------------------------------------------------------------------- 1 | # Flashlight Decoder 2 | 3 | This script runs decoding for pre-trained speech recognition models. 4 | 5 | ## Usage 6 | 7 | Assuming a few variables: 8 | 9 | ```bash 10 | checkpoint= 11 | data= 12 | lm_model= 13 | lexicon= 14 | ``` 15 | 16 | Example usage for decoding a fine-tuned Wav2Vec model: 17 | 18 | ```bash 19 | python $FAIRSEQ_ROOT/examples/speech_recognition/new/infer.py --multirun \ 20 | task=audio_pretraining \ 21 | task.data=$data \ 22 | task.labels=ltr \ 23 | common_eval.path=$checkpoint \ 24 | decoding.type=kenlm \ 25 | decoding.lexicon=$lexicon \ 26 | decoding.lmpath=$lm_model \ 27 | dataset.gen_subset=dev_clean,dev_other,test_clean,test_other 28 | ``` 29 | 30 | Example usage for using Ax to sweep WER parameters (requires `pip install hydra-ax-sweeper`): 31 | 32 | ```bash 33 | python $FAIRSEQ_ROOT/examples/speech_recognition/new/infer.py --multirun \ 34 | hydra/sweeper=ax \ 35 | task=audio_pretraining \ 36 | task.data=$data \ 37 | task.labels=ltr \ 38 | common_eval.path=$checkpoint \ 39 | decoding.type=kenlm \ 40 | decoding.lexicon=$lexicon \ 41 | decoding.lmpath=$lm_model \ 42 | dataset.gen_subset=dev_other 43 | ``` 44 | -------------------------------------------------------------------------------- /examples/speech_recognition/new/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/speech_recognition/new/__init__.py -------------------------------------------------------------------------------- /examples/speech_recognition/new/conf/hydra/sweeper/ax.yaml: -------------------------------------------------------------------------------- 1 | # @package hydra.sweeper 2 | _target_: hydra_plugins.hydra_ax_sweeper.ax_sweeper.AxSweeper 3 | max_batch_size: null 4 | ax_config: 5 | max_trials: 128 6 | early_stop: 7 | minimize: true 8 | max_epochs_without_improvement: 32 9 | epsilon: 1.0e-05 10 | experiment: 11 | name: ${dataset.gen_subset} 12 | objective_name: wer 13 | minimize: true 14 | parameter_constraints: null 15 | outcome_constraints: null 16 | status_quo: null 17 | client: 18 | verbose_logging: false 19 | random_seed: null 20 | params: 21 | decoding.lmweight: 22 | type: range 23 | bounds: [0.0, 5.0] 24 | decoding.wordscore: 25 | type: range 26 | bounds: [-5.0, 5.0] 27 | -------------------------------------------------------------------------------- /examples/speech_recognition/new/conf/infer.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | defaults: 4 | - task: null 5 | - model: null 6 | 7 | hydra: 8 | run: 9 | dir: ${common_eval.results_path}/${dataset.gen_subset} 10 | sweep: 11 | dir: ${common_eval.results_path} 12 | subdir: ${dataset.gen_subset} 13 | common_eval: 14 | results_path: null 15 | path: null 16 | post_process: letter 17 | quiet: true 18 | dataset: 19 | max_tokens: 1000000 20 | gen_subset: test 21 | distributed_training: 22 | distributed_world_size: 1 23 | decoding: 24 | beam: 5 25 | type: viterbi 26 | -------------------------------------------------------------------------------- /examples/speech_recognition/new/decoders/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/speech_recognition/new/decoders/__init__.py -------------------------------------------------------------------------------- /examples/speech_recognition/new/decoders/decoder.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | 3 | # Copyright (c) Facebook, Inc. and its affiliates. 4 | # 5 | # This source code is licensed under the MIT license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | from typing import Union 9 | 10 | from fairseq.data.dictionary import Dictionary 11 | 12 | from .decoder_config import DecoderConfig, FlashlightDecoderConfig 13 | from .base_decoder import BaseDecoder 14 | 15 | 16 | def Decoder( 17 | cfg: Union[DecoderConfig, FlashlightDecoderConfig], tgt_dict: Dictionary 18 | ) -> BaseDecoder: 19 | 20 | if cfg.type == "viterbi": 21 | from .viterbi_decoder import ViterbiDecoder 22 | 23 | return ViterbiDecoder(tgt_dict) 24 | if cfg.type == "kenlm": 25 | from .flashlight_decoder import KenLMDecoder 26 | 27 | return KenLMDecoder(cfg, tgt_dict) 28 | if cfg.type == "fairseqlm": 29 | from .flashlight_decoder import FairseqLMDecoder 30 | 31 | return FairseqLMDecoder(cfg, tgt_dict) 32 | raise NotImplementedError(f"Invalid decoder name: {cfg.name}") 33 | -------------------------------------------------------------------------------- /examples/speech_recognition/new/decoders/viterbi_decoder.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | 3 | # Copyright (c) Facebook, Inc. and its affiliates. 4 | # 5 | # This source code is licensed under the MIT license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | import torch 9 | 10 | from typing import List, Dict 11 | 12 | from .base_decoder import BaseDecoder 13 | 14 | 15 | class ViterbiDecoder(BaseDecoder): 16 | def decode( 17 | self, 18 | emissions: torch.FloatTensor, 19 | ) -> List[List[Dict[str, torch.LongTensor]]]: 20 | def get_pred(e): 21 | toks = e.argmax(dim=-1).unique_consecutive() 22 | return toks[toks != self.blank] 23 | 24 | return [[{"tokens": get_pred(x), "score": 0}] for x in emissions] 25 | -------------------------------------------------------------------------------- /examples/speech_recognition/tasks/__init__.py: -------------------------------------------------------------------------------- 1 | import importlib 2 | import os 3 | 4 | 5 | for file in sorted(os.listdir(os.path.dirname(__file__))): 6 | if file.endswith(".py") and not file.startswith("_"): 7 | task_name = file[: file.find(".py")] 8 | importlib.import_module("examples.speech_recognition.tasks." + task_name) 9 | -------------------------------------------------------------------------------- /examples/translation_moe/translation_moe_src/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import translation_moe # noqa 7 | -------------------------------------------------------------------------------- /examples/translation_moe/translation_moe_src/logsumexp_moe.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import torch 7 | 8 | 9 | class LogSumExpMoE(torch.autograd.Function): 10 | """Standard LogSumExp forward pass, but use *posterior* for the backward. 11 | 12 | See `"Mixture Models for Diverse Machine Translation: Tricks of the Trade" 13 | (Shen et al., 2019) `_. 14 | """ 15 | 16 | @staticmethod 17 | def forward(ctx, logp, posterior, dim=-1): 18 | ctx.save_for_backward(posterior) 19 | ctx.dim = dim 20 | return torch.logsumexp(logp, dim=dim) 21 | 22 | @staticmethod 23 | def backward(ctx, grad_output): 24 | (posterior,) = ctx.saved_tensors 25 | grad_logp = grad_output.unsqueeze(ctx.dim) * posterior 26 | return grad_logp, None, None 27 | -------------------------------------------------------------------------------- /examples/truncated_bptt/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import transformer_xl_model, truncated_bptt_lm_task # noqa 7 | -------------------------------------------------------------------------------- /examples/unsupervised_quality_estimation/aggregate_scores.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import argparse 7 | import sys 8 | 9 | import numpy as np 10 | 11 | 12 | aggregate_funcs = { 13 | "std": np.std, 14 | "var": np.var, 15 | "median": np.median, 16 | "mean": np.mean, 17 | "min": np.min, 18 | "max": np.max, 19 | } 20 | 21 | 22 | def main(): 23 | parser = argparse.ArgumentParser() 24 | parser.add_argument("-i", "--input_file", required=True, type=str) 25 | parser.add_argument("-n", "--repeat_times", required=True, type=int) 26 | parser.add_argument("-o", "--output_file", required=False) 27 | parser.add_argument("-f", "--func", required=False, default="mean") 28 | args = parser.parse_args() 29 | 30 | stream = open(args.output_file, "w") if args.output_file else sys.stdout 31 | 32 | segment_scores = [] 33 | for line in open(args.input_file): 34 | segment_scores.append(float(line.strip())) 35 | if len(segment_scores) == args.repeat_times: 36 | stream.write("{}\n".format(aggregate_funcs[args.func](segment_scores))) 37 | segment_scores = [] 38 | 39 | 40 | if __name__ == "__main__": 41 | main() 42 | -------------------------------------------------------------------------------- /examples/unsupervised_quality_estimation/repeat_lines.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import argparse 7 | import sys 8 | 9 | 10 | def _normalize_spaces(line): 11 | return " ".join(line.split()) 12 | 13 | 14 | def main(): 15 | parser = argparse.ArgumentParser() 16 | parser.add_argument("-i", "--input_file", required=True, type=str) 17 | parser.add_argument("-n", "--repeat_times", required=True, type=int) 18 | parser.add_argument("-o", "--output_file", required=False, type=str) 19 | args = parser.parse_args() 20 | stream = open(args.output_file, "w") if args.output_file else sys.stdout 21 | 22 | for line in open(args.input_file): 23 | for _ in range(args.repeat_times): 24 | stream.write(_normalize_spaces(line) + "\n") 25 | 26 | 27 | if __name__ == "__main__": 28 | main() 29 | -------------------------------------------------------------------------------- /examples/wav2vec/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/wav2vec/__init__.py -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/base_100h.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | no_epoch_checkpoints: true 10 | best_checkpoint_metric: wer 11 | 12 | task: 13 | _name: audio_pretraining 14 | data: ??? 15 | normalize: false 16 | labels: ltr 17 | 18 | dataset: 19 | num_workers: 6 20 | max_tokens: 3200000 21 | skip_invalid_size_inputs_valid_test: true 22 | valid_subset: dev_other 23 | 24 | distributed_training: 25 | ddp_backend: legacy_ddp 26 | distributed_world_size: 2 27 | 28 | criterion: 29 | _name: ctc 30 | zero_infinity: true 31 | 32 | optimization: 33 | max_update: 80000 34 | lr: [0.00003] 35 | sentence_avg: true 36 | update_freq: [4] 37 | 38 | optimizer: 39 | _name: adam 40 | adam_betas: (0.9,0.98) 41 | adam_eps: 1e-08 42 | 43 | lr_scheduler: 44 | _name: tri_stage 45 | phase_ratio: [0.1, 0.4, 0.5] 46 | final_lr_scale: 0.05 47 | 48 | model: 49 | _name: wav2vec_ctc 50 | w2v_path: ??? 51 | apply_mask: true 52 | mask_prob: 0.65 53 | mask_channel_prob: 0.5 54 | mask_channel_length: 64 55 | layerdrop: 0.1 56 | activation_dropout: 0.1 57 | feature_grad_mult: 0.0 58 | freeze_finetune_updates: 0 59 | 60 | -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/base_10h.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | save_interval: 50 10 | save_interval_updates: 10000 11 | keep_interval_updates: 1 12 | no_epoch_checkpoints: true 13 | best_checkpoint_metric: wer 14 | 15 | task: 16 | _name: audio_pretraining 17 | data: ??? 18 | normalize: false 19 | labels: ltr 20 | 21 | dataset: 22 | num_workers: 6 23 | max_tokens: 3200000 24 | skip_invalid_size_inputs_valid_test: true 25 | validate_after_updates: 10000 26 | validate_interval: 50 27 | valid_subset: dev_other 28 | 29 | distributed_training: 30 | ddp_backend: legacy_ddp 31 | distributed_world_size: 2 32 | 33 | criterion: 34 | _name: ctc 35 | zero_infinity: true 36 | 37 | optimization: 38 | max_update: 20000 39 | lr: [0.00005] 40 | sentence_avg: true 41 | update_freq: [4] 42 | 43 | optimizer: 44 | _name: adam 45 | adam_betas: (0.9,0.98) 46 | adam_eps: 1e-08 47 | 48 | lr_scheduler: 49 | _name: tri_stage 50 | phase_ratio: [0.1, 0.4, 0.5] 51 | final_lr_scale: 0.05 52 | 53 | model: 54 | _name: wav2vec_ctc 55 | w2v_path: ??? 56 | apply_mask: true 57 | mask_prob: 0.65 58 | mask_channel_prob: 0.5 59 | mask_channel_length: 64 60 | layerdrop: 0.05 61 | activation_dropout: 0.1 62 | feature_grad_mult: 0.0 63 | freeze_finetune_updates: 10000 64 | 65 | -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/base_10m.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | save_interval: 1000 10 | save_interval_updates: 50 11 | keep_interval_updates: 1 12 | no_epoch_checkpoints: true 13 | best_checkpoint_metric: wer 14 | 15 | task: 16 | _name: audio_pretraining 17 | data: ??? 18 | normalize: false 19 | labels: ltr 20 | 21 | dataset: 22 | num_workers: 6 23 | max_tokens: 3200000 24 | skip_invalid_size_inputs_valid_test: true 25 | validate_after_updates: 10000 26 | validate_interval: 1000 27 | valid_subset: dev_other 28 | 29 | distributed_training: 30 | ddp_backend: legacy_ddp 31 | distributed_world_size: 2 32 | 33 | criterion: 34 | _name: ctc 35 | zero_infinity: true 36 | 37 | optimization: 38 | max_update: 13000 39 | lr: [0.00005] 40 | sentence_avg: true 41 | update_freq: [4] 42 | 43 | optimizer: 44 | _name: adam 45 | adam_betas: (0.9,0.98) 46 | adam_eps: 1e-08 47 | 48 | lr_scheduler: 49 | _name: tri_stage 50 | phase_ratio: [0.1, 0.4, 0.5] 51 | final_lr_scale: 0.05 52 | 53 | model: 54 | _name: wav2vec_ctc 55 | w2v_path: ??? 56 | apply_mask: true 57 | mask_prob: 0.65 58 | mask_channel_prob: 0.25 59 | mask_channel_length: 64 60 | layerdrop: 0.1 61 | activation_dropout: 0.1 62 | feature_grad_mult: 0.0 63 | freeze_finetune_updates: 10000 64 | 65 | -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/base_1h.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | save_interval: 1000 10 | save_interval_updates: 50 11 | keep_interval_updates: 1 12 | no_epoch_checkpoints: true 13 | best_checkpoint_metric: wer 14 | 15 | task: 16 | _name: audio_pretraining 17 | data: ??? 18 | normalize: false 19 | labels: ltr 20 | 21 | dataset: 22 | num_workers: 6 23 | max_tokens: 3200000 24 | skip_invalid_size_inputs_valid_test: true 25 | validate_after_updates: 10000 26 | validate_interval: 1000 27 | valid_subset: dev_other 28 | 29 | distributed_training: 30 | ddp_backend: legacy_ddp 31 | distributed_world_size: 2 32 | 33 | criterion: 34 | _name: ctc 35 | zero_infinity: true 36 | 37 | optimization: 38 | max_update: 13000 39 | lr: [0.00005] 40 | sentence_avg: true 41 | update_freq: [4] 42 | 43 | optimizer: 44 | _name: adam 45 | adam_betas: (0.9,0.98) 46 | adam_eps: 1e-08 47 | 48 | lr_scheduler: 49 | _name: tri_stage 50 | phase_ratio: [0.1, 0.4, 0.5] 51 | final_lr_scale: 0.05 52 | 53 | model: 54 | _name: wav2vec_ctc 55 | w2v_path: ??? 56 | apply_mask: true 57 | mask_prob: 0.65 58 | mask_channel_prob: 0.25 59 | mask_channel_length: 64 60 | layerdrop: 0.1 61 | activation_dropout: 0.1 62 | feature_grad_mult: 0.0 63 | freeze_finetune_updates: 10000 64 | 65 | -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/base_960h.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | no_epoch_checkpoints: true 10 | best_checkpoint_metric: wer 11 | 12 | task: 13 | _name: audio_pretraining 14 | data: ??? 15 | normalize: false 16 | labels: ltr 17 | 18 | dataset: 19 | num_workers: 6 20 | max_tokens: 3200000 21 | skip_invalid_size_inputs_valid_test: true 22 | valid_subset: dev_other 23 | 24 | distributed_training: 25 | ddp_backend: legacy_ddp 26 | distributed_world_size: 8 27 | 28 | criterion: 29 | _name: ctc 30 | zero_infinity: true 31 | 32 | optimization: 33 | max_update: 320000 34 | lr: [0.0001] 35 | sentence_avg: true 36 | 37 | optimizer: 38 | _name: adam 39 | adam_betas: (0.9,0.98) 40 | adam_eps: 1e-08 41 | 42 | lr_scheduler: 43 | _name: tri_stage 44 | phase_ratio: [0.1, 0.4, 0.5] 45 | final_lr_scale: 0.05 46 | 47 | model: 48 | _name: wav2vec_ctc 49 | w2v_path: ??? 50 | apply_mask: true 51 | mask_prob: 0.5 52 | mask_channel_prob: 0.1 53 | mask_channel_length: 64 54 | layerdrop: 0.1 55 | activation_dropout: 0.1 56 | feature_grad_mult: 0.0 57 | freeze_finetune_updates: 0 58 | 59 | -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/vox_100h.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | no_epoch_checkpoints: true 10 | best_checkpoint_metric: wer 11 | 12 | task: 13 | _name: audio_pretraining 14 | data: ??? 15 | normalize: true 16 | labels: ltr 17 | 18 | dataset: 19 | num_workers: 6 20 | max_tokens: 1280000 21 | skip_invalid_size_inputs_valid_test: true 22 | valid_subset: dev_other 23 | 24 | distributed_training: 25 | ddp_backend: legacy_ddp 26 | distributed_world_size: 4 27 | 28 | criterion: 29 | _name: ctc 30 | zero_infinity: true 31 | 32 | optimization: 33 | max_update: 80000 34 | lr: [0.00003] 35 | sentence_avg: true 36 | update_freq: [5] 37 | 38 | optimizer: 39 | _name: adam 40 | adam_betas: (0.9,0.98) 41 | adam_eps: 1e-08 42 | 43 | lr_scheduler: 44 | _name: tri_stage 45 | phase_ratio: [0.1, 0.4, 0.5] 46 | final_lr_scale: 0.05 47 | 48 | model: 49 | _name: wav2vec_ctc 50 | w2v_path: ??? 51 | apply_mask: true 52 | mask_prob: 0.5 53 | mask_channel_prob: 0.5 54 | mask_channel_length: 64 55 | layerdrop: 0.1 56 | activation_dropout: 0.1 57 | feature_grad_mult: 0.0 58 | freeze_finetune_updates: 10000 59 | 60 | -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/vox_10h.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | save_interval: 50 10 | save_interval_updates: 10000 11 | keep_interval_updates: 1 12 | no_epoch_checkpoints: true 13 | best_checkpoint_metric: wer 14 | 15 | task: 16 | _name: audio_pretraining 17 | data: ??? 18 | normalize: true 19 | labels: ltr 20 | 21 | dataset: 22 | num_workers: 6 23 | max_tokens: 1280000 24 | skip_invalid_size_inputs_valid_test: true 25 | validate_after_updates: 10000 26 | validate_interval: 50 27 | valid_subset: dev_other 28 | 29 | distributed_training: 30 | ddp_backend: legacy_ddp 31 | distributed_world_size: 4 32 | 33 | criterion: 34 | _name: ctc 35 | zero_infinity: true 36 | 37 | optimization: 38 | max_update: 20000 39 | lr: [0.0001] 40 | sentence_avg: true 41 | update_freq: [5] 42 | 43 | optimizer: 44 | _name: adam 45 | adam_betas: (0.9,0.98) 46 | adam_eps: 1e-08 47 | 48 | lr_scheduler: 49 | _name: tri_stage 50 | phase_ratio: [0.1, 0.4, 0.5] 51 | final_lr_scale: 0.05 52 | 53 | model: 54 | _name: wav2vec_ctc 55 | w2v_path: ??? 56 | apply_mask: true 57 | mask_prob: 0.75 58 | mask_channel_prob: 0.25 59 | mask_channel_length: 64 60 | layerdrop: 0.1 61 | activation_dropout: 0.1 62 | feature_grad_mult: 0.0 63 | freeze_finetune_updates: 10000 64 | 65 | -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/vox_10m.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | save_interval: 1000 10 | save_interval_updates: 50 11 | keep_interval_updates: 1 12 | no_epoch_checkpoints: true 13 | best_checkpoint_metric: wer 14 | 15 | task: 16 | _name: audio_pretraining 17 | data: ??? 18 | normalize: true 19 | labels: ltr 20 | 21 | dataset: 22 | num_workers: 6 23 | max_tokens: 1280000 24 | skip_invalid_size_inputs_valid_test: true 25 | validate_after_updates: 10000 26 | validate_interval: 1000 27 | valid_subset: dev_other 28 | 29 | distributed_training: 30 | ddp_backend: legacy_ddp 31 | distributed_world_size: 4 32 | 33 | criterion: 34 | _name: ctc 35 | zero_infinity: true 36 | 37 | optimization: 38 | max_update: 13000 39 | lr: [0.0001] 40 | sentence_avg: true 41 | update_freq: [5] 42 | 43 | optimizer: 44 | _name: adam 45 | adam_betas: (0.9,0.98) 46 | adam_eps: 1e-08 47 | 48 | lr_scheduler: 49 | _name: tri_stage 50 | phase_ratio: [0.1, 0.4, 0.5] 51 | final_lr_scale: 0.05 52 | 53 | model: 54 | _name: wav2vec_ctc 55 | w2v_path: ??? 56 | apply_mask: true 57 | mask_prob: 0.65 58 | mask_channel_prob: 0.25 59 | mask_channel_length: 64 60 | layerdrop: 0.1 61 | activation_dropout: 0.1 62 | feature_grad_mult: 0.0 63 | freeze_finetune_updates: 10000 64 | 65 | -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/vox_1h.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | save_interval: 1000 10 | save_interval_updates: 50 11 | keep_interval_updates: 1 12 | no_epoch_checkpoints: true 13 | best_checkpoint_metric: wer 14 | 15 | task: 16 | _name: audio_pretraining 17 | data: ??? 18 | normalize: true 19 | labels: ltr 20 | 21 | dataset: 22 | num_workers: 6 23 | max_tokens: 1280000 24 | skip_invalid_size_inputs_valid_test: true 25 | validate_after_updates: 10000 26 | validate_interval: 1000 27 | valid_subset: dev_other 28 | 29 | distributed_training: 30 | ddp_backend: legacy_ddp 31 | distributed_world_size: 4 32 | 33 | criterion: 34 | _name: ctc 35 | zero_infinity: true 36 | 37 | optimization: 38 | max_update: 13000 39 | lr: [0.0003] 40 | sentence_avg: true 41 | update_freq: [5] 42 | 43 | optimizer: 44 | _name: adam 45 | adam_betas: (0.9,0.98) 46 | adam_eps: 1e-08 47 | 48 | lr_scheduler: 49 | _name: tri_stage 50 | phase_ratio: [0.1, 0.4, 0.5] 51 | final_lr_scale: 0.05 52 | 53 | model: 54 | _name: wav2vec_ctc 55 | w2v_path: ??? 56 | apply_mask: true 57 | mask_prob: 0.75 58 | mask_channel_prob: 0.25 59 | mask_channel_length: 64 60 | layerdrop: 0.1 61 | activation_dropout: 0.1 62 | feature_grad_mult: 0.0 63 | freeze_finetune_updates: 10000 64 | 65 | -------------------------------------------------------------------------------- /examples/wav2vec/config/finetuning/vox_960h.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | no_epoch_checkpoints: true 10 | best_checkpoint_metric: wer 11 | 12 | task: 13 | _name: audio_pretraining 14 | data: ??? 15 | normalize: true 16 | labels: ltr 17 | 18 | dataset: 19 | num_workers: 6 20 | max_tokens: 1280000 21 | skip_invalid_size_inputs_valid_test: true 22 | valid_subset: dev_other 23 | 24 | distributed_training: 25 | ddp_backend: legacy_ddp 26 | distributed_world_size: 24 27 | 28 | criterion: 29 | _name: ctc 30 | zero_infinity: true 31 | 32 | optimization: 33 | max_update: 320000 34 | lr: [0.00003] 35 | sentence_avg: true 36 | 37 | optimizer: 38 | _name: adam 39 | adam_betas: (0.9,0.98) 40 | adam_eps: 1e-08 41 | 42 | lr_scheduler: 43 | _name: tri_stage 44 | phase_ratio: [0.1, 0.4, 0.5] 45 | final_lr_scale: 0.05 46 | 47 | model: 48 | _name: wav2vec_ctc 49 | w2v_path: ??? 50 | apply_mask: true 51 | mask_prob: 0.5 52 | mask_channel_prob: 0.25 53 | mask_channel_length: 64 54 | layerdrop: 0.1 55 | activation_dropout: 0.1 56 | feature_grad_mult: 0.0 57 | freeze_finetune_updates: 10000 58 | 59 | -------------------------------------------------------------------------------- /examples/wav2vec/config/pretraining/wav2vec2_base_librispeech.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | 8 | checkpoint: 9 | save_interval_updates: 25000 10 | keep_interval_updates: 1 11 | no_epoch_checkpoints: true 12 | 13 | task: 14 | _name: audio_pretraining 15 | data: ??? 16 | max_sample_size: 250000 17 | min_sample_size: 32000 18 | normalize: false 19 | 20 | dataset: 21 | num_workers: 6 22 | max_tokens: 1400000 23 | skip_invalid_size_inputs_valid_test: true 24 | 25 | distributed_training: 26 | distributed_world_size: 64 27 | ddp_backend: legacy_ddp 28 | 29 | criterion: 30 | _name: wav2vec 31 | infonce: true 32 | log_keys: ["prob_perplexity","code_perplexity","temp"] 33 | loss_weights: [0.1, 10] 34 | 35 | optimization: 36 | max_update: 400000 37 | lr: [0.0005] 38 | 39 | optimizer: 40 | _name: adam 41 | adam_betas: (0.9,0.98) 42 | adam_eps: 1e-06 43 | weight_decay: 0.01 44 | 45 | lr_scheduler: 46 | _name: polynomial_decay 47 | warmup_updates: 32000 48 | 49 | model: 50 | _name: wav2vec2 51 | quantize_targets: true 52 | final_dim: 256 53 | encoder_layerdrop: 0.05 54 | dropout_input: 0.1 55 | dropout_features: 0.1 56 | feature_grad_mult: 0.1 57 | encoder_embed_dim: 768 58 | -------------------------------------------------------------------------------- /examples/wav2vec/scripts/binarize_manifest.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | 3 | # usage: bash binarize_manifest 4 | 5 | DEST_DIR=$1 6 | TRAIN_SPLIT=$2 7 | VALID_SPLIT=$3 8 | FAIRSEQ_ROOT=$4 9 | 10 | mkdir -p $DEST_DIR 11 | 12 | # split file path and lengths into separate files 13 | cut -f1 $TRAIN_SPLIT.tsv > $DEST_DIR/train_fnames.txt 14 | cut -f1 $VALID_SPLIT.tsv > $DEST_DIR/valid_fnames.txt 15 | cut -f2 $TRAIN_SPLIT.tsv > $DEST_DIR/train.lengths 16 | cut -f2 $VALID_SPLIT.tsv > $DEST_DIR/valid.lengths 17 | 18 | # copy root directory 19 | head -1 $TRAIN_SPLIT.tsv > $DEST_DIR/train.root 20 | head -1 $VALID_SPLIT.tsv > $DEST_DIR/valid.root 21 | 22 | # remove root directory 23 | sed -i '1d' $DEST_DIR/train_fnames.txt 24 | sed -i '1d' $DEST_DIR/valid_fnames.txt 25 | sed -i '1d' $DEST_DIR/train.lengths 26 | sed -i '1d' $DEST_DIR/valid.lengths 27 | 28 | # insert spaces between characters 29 | sed -i -e 's/\(.\)/\1 /g' $DEST_DIR/train_fnames.txt 30 | sed -i -e 's/\(.\)/\1 /g' $DEST_DIR/valid_fnames.txt 31 | 32 | # run preprocessor 33 | PYTHONPATH=$FAIRSEQ_ROOT python $FAIRSEQ_ROOT/fairseq_cli/preprocess.py --dataset-impl mmap --trainpref $DEST_DIR/train_fnames.txt --validpref $DEST_DIR/valid_fnames.txt --workers 60 --only-source --destdir $DEST_DIR 34 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/examples/wav2vec/unsupervised/__init__.py -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/config/finetuning/w2v_finetune.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | common: 4 | fp16: true 5 | log_format: json 6 | log_interval: 200 7 | tensorboard_logdir: tb 8 | 9 | checkpoint: 10 | no_epoch_checkpoints: true 11 | save_interval_updates: 20000 12 | 13 | task: 14 | _name: audio_pretraining 15 | data: ??? 16 | normalize: true 17 | labels: ltr 18 | 19 | dataset: 20 | num_workers: 6 21 | max_tokens: 800000 22 | skip_invalid_size_inputs_valid_test: true 23 | train_subset: train 24 | valid_subset: valid 25 | 26 | distributed_training: 27 | ddp_backend: legacy_ddp 28 | distributed_world_size: 8 29 | find_unused_parameters: True 30 | 31 | criterion: 32 | _name: ctc 33 | zero_infinity: true 34 | post_process: letter 35 | 36 | optimization: 37 | max_update: 80000 38 | lr: [0.00003] 39 | sentence_avg: true 40 | update_freq: [1] 41 | 42 | optimizer: 43 | _name: adam 44 | adam_betas: (0.9,0.98) 45 | adam_eps: 1e-08 46 | 47 | lr_scheduler: 48 | _name: tri_stage 49 | phase_ratio: [0.1, 0.4, 0.5] 50 | final_lr_scale: 0.05 51 | 52 | model: 53 | _name: wav2vec_ctc 54 | w2v_path: ??? 55 | apply_mask: true 56 | mask_prob: 0.25 57 | mask_channel_prob: 0.1 58 | mask_channel_length: 64 59 | layerdrop: 0.1 60 | activation_dropout: 0.1 61 | feature_grad_mult: 0.0 62 | freeze_finetune_updates: 0 63 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/config/generate/viterbi.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | fairseq: 4 | task: 5 | _name: unpaired_audio_text 6 | labels: phn 7 | data: ??? 8 | sort_by_length: false 9 | shuffle: false 10 | text_data: '' 11 | 12 | common_eval: 13 | path: ??? 14 | quiet: true 15 | 16 | dataset: 17 | gen_subset: valid 18 | batch_size: 1 19 | 20 | w2l_decoder: VITERBI 21 | post_process: silence 22 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/data/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .extracted_features_dataset import ExtractedFeaturesDataset 7 | from .random_input_dataset import RandomInputDataset 8 | 9 | 10 | __all__ = [ 11 | "ExtractedFeaturesDataset", 12 | "RandomInputDataset", 13 | ] 14 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/cmd.sh: -------------------------------------------------------------------------------- 1 | # you can change cmd.sh depending on what type of queue you are using. 2 | # If you have no queueing system and want to run on a local machine, you 3 | # can change all instances 'queue.pl' to run.pl (but be careful and run 4 | # commands one by one: most recipes will exhaust the memory on your 5 | # machine). queue.pl works with GridEngine (qsub). slurm.pl works 6 | # with slurm. Different queues are configured differently, with different 7 | # queue names and different ways of specifying things like memory; 8 | # to account for these differences you can create and edit the file 9 | # conf/queue.conf to match your queue's configuration. Search for 10 | # conf/queue.conf in http://kaldi-asr.org/doc/queue.html for more information, 11 | # or search for the string 'default_config' in utils/queue.pl or utils/slurm.pl. 12 | 13 | export train_cmd="run.pl --mem 2G" 14 | export decode_cmd="run.pl --mem 4G" 15 | export mkgraph_cmd="run.pl --mem 8G" 16 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/decode_phone.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | # decode into phones (and prepare a new data directory for HMM outputs) 4 | 5 | . ./path.sh 6 | 7 | set -eu 8 | 9 | out_dir= # same as in train.sh 10 | dec_lmparam= # LM hyperparameters (e.g., 7.0.0) 11 | dec_exp= 12 | dec_script= 13 | dec_splits="train valid" 14 | dec_data_dir=$out_dir/dec_data # where to write HMM output 15 | 16 | data_dir=${out_dir}/data 17 | 18 | local/decode.sh --nj 40 --graph_name graph \ 19 | --val_sets "$dec_splits" --decode_script $dec_script \ 20 | $out_dir/exp/$dec_exp $data_dir $data_dir/lang_test 21 | 22 | if [ ! -z $dec_lmparam ]; then 23 | for x in $dec_splits; do 24 | mkdir -p $dec_data_dir/$x 25 | cp $data_dir/$x/{feats.scp,cmvn.scp,utt2spk,spk2utt} $dec_data_dir/$x/ 26 | 27 | tra=$out_dir/exp/$dec_exp/decode_${x}/scoring/${dec_lmparam}.tra 28 | cat $tra | utils/int2sym.pl -f 2- $data_dir/lang/words.txt | \ 29 | sed 's:::g' | sed 's:::g' > $dec_data_dir/${x}/text 30 | utils/fix_data_dir.sh $dec_data_dir/${x} 31 | echo "WER on ${x} is" $(compute-wer ark:$data_dir/${x}_gt/text ark:$dec_data_dir/$x/text | cut -d" " -f2-) 32 | done 33 | fi 34 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/decode_word_step2.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | # prepare a new data directory of HMM word output 4 | 5 | . ./path.sh 6 | 7 | set -eu 8 | 9 | out_dir= # same as in train.sh 10 | dec_lmparam= # LM hyperparameters (e.g., 7.0.0) 11 | 12 | dec_exp=tri3b # what HMM stage to decode (e.g., tri3b) 13 | dec_suffix=word 14 | dec_splits="train valid" 15 | dec_data_dir=$out_dir/dec_data_word # where to write HMM output 16 | 17 | data_dir=$out_dir/data 18 | wrd_data_dir=$out_dir/data_word 19 | 20 | for x in $dec_splits; do 21 | mkdir -p $dec_data_dir/$x 22 | cp $data_dir/$x/{feats.scp,cmvn.scp,utt2spk,spk2utt} $dec_data_dir/$x/ 23 | 24 | tra=$out_dir/exp/$dec_exp/decode${dec_suffix}_${x}/scoring/${dec_lmparam}.tra 25 | cat $tra | utils/int2sym.pl -f 2- $data_dir/lang_word/words.txt | \ 26 | sed 's:::g' | sed 's:::g' > $dec_data_dir/$x/text 27 | utils/fix_data_dir.sh $dec_data_dir/$x 28 | echo "WER on $x is" $(compute-wer ark:$wrd_data_dir/${x}_gt/text ark:$dec_data_dir/$x/text | cut -d" " -f2-) 29 | done 30 | 31 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/local/copy_aligned_text.py: -------------------------------------------------------------------------------- 1 | import sys 2 | 3 | for idx, line in enumerate(sys.stdin): 4 | print(f"utt{idx:010d} {line}", end='') -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/local/decode.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | set -u 4 | 5 | val_sets="dev_other" 6 | graph_name=graph 7 | decode_suffix="" 8 | decode_script="steps/decode_fmllr.sh" 9 | decode_args="" 10 | nj=60 11 | 12 | . ./cmd.sh 13 | . ./path.sh 14 | . parse_options.sh 15 | 16 | set -x 17 | exp_dir=$1 18 | data_root=$2 19 | lang_test=$3 20 | 21 | graph=$exp_dir/$graph_name 22 | 23 | if [ ! -d $graph ]; then 24 | utils/mkgraph.sh $lang_test $exp_dir $graph 25 | fi 26 | 27 | for part in $val_sets; do 28 | dec_dir=$exp_dir/decode${decode_suffix}_${part} 29 | if [ ! -d $dec_dir ]; then 30 | echo "decoding $part for $exp_dir" 31 | $decode_script --nj $nj --cmd "$decode_cmd" $decode_args \ 32 | $graph $data_root/$part $dec_dir & 33 | else 34 | echo "$dec_dir exists. skip" 35 | fi 36 | done 37 | 38 | wait 39 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lang.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | sil_prob=0.5 4 | num_sil_states=3 5 | num_nonsil_states=1 6 | 7 | . ./cmd.sh 8 | . ./path.sh 9 | . parse_options.sh 10 | 11 | set -eux 12 | 13 | dict=$1 14 | data_dir=$2 15 | 16 | dict_dir=$data_dir/local/dict 17 | tmplm_dir=$data_dir/local/lang_tmp 18 | lm_dir=$data_dir/lang 19 | 20 | mkdir -p $dict_dir $tmplm_dir $lm_dir 21 | 22 | # prepare dict 23 | echo "SIL" > $dict_dir/silence_phones.txt 24 | echo "SIL" > $dict_dir/optional_silence.txt 25 | awk '{print $1}' $dict > $dict_dir/nonsilence_phones.txt 26 | 27 | echo "SIL SIL" > $dict_dir/lexicon.txt 28 | echo " SIL" >> $dict_dir/lexicon.txt 29 | awk '{print $1" "$1}' $dict >> $dict_dir/lexicon.txt 30 | 31 | echo "SIL" > $dict_dir/extra_questions.txt 32 | awk '{printf $1" "} END {printf "\n"}' $dict >> $dict_dir/extra_questions.txt 33 | 34 | # prepare lang 35 | utils/prepare_lang.sh --sil-prob $sil_prob --position-dependent-phones false \ 36 | --num_sil_states $num_sil_states --num_nonsil_states $num_nonsil_states \ 37 | $dict_dir "" $tmplm_dir $lm_dir 38 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lang_word.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | num_sil_states=3 4 | num_nonsil_states=1 5 | 6 | . ./cmd.sh 7 | . ./path.sh 8 | . parse_options.sh 9 | 10 | set -eux 11 | 12 | dict=$1 13 | data_dir=$2 14 | lexicon=$3 15 | 16 | dict_dir=$data_dir/local/dict_word 17 | tmplm_dir=$data_dir/local/lang_tmp_word 18 | lm_dir=$data_dir/lang_word 19 | 20 | mkdir -p $dict_dir $tmplm_dir $lm_dir 21 | 22 | # prepare dict 23 | echo "SIL" > $dict_dir/silence_phones.txt 24 | echo "SIL" > $dict_dir/optional_silence.txt 25 | awk '{print $1}' $dict > $dict_dir/nonsilence_phones.txt 26 | 27 | (echo "!SIL SIL"; echo " SIL";) | cat - $lexicon > $dict_dir/lexicon.txt 28 | 29 | echo "SIL" > $dict_dir/extra_questions.txt 30 | awk '{printf $1" "} END {printf "\n"}' $dict >> $dict_dir/extra_questions.txt 31 | 32 | # prepare lang 33 | utils/prepare_lang.sh --position-dependent-phones false \ 34 | --num_sil_states $num_sil_states --num_nonsil_states $num_nonsil_states \ 35 | $dict_dir "" $tmplm_dir $lm_dir 36 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/local/prepare_lm.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | 3 | langdir="" 4 | lmdir="" 5 | 6 | . ./cmd.sh 7 | . ./path.sh 8 | . parse_options.sh 9 | 10 | arpa_lm=$1 11 | data=$2 12 | 13 | if [ -z $langdir ]; then 14 | langdir=$data/lang 15 | fi 16 | if [ -z $lmdir ]; then 17 | lmdir=$data/lang_test 18 | fi 19 | 20 | if [ ! -d $langdir ]; then 21 | echo "$langdir not found. run local/prepare_lang.sh first" && exit 1 22 | fi 23 | 24 | mkdir -p $lmdir 25 | cp -r $langdir/* $lmdir 26 | 27 | if [[ "$arpa_lm" == *.gz ]]; then 28 | gunzip -c $arpa_lm | arpa2fst --disambig-symbol=#0 --read-symbol-table=$lmdir/words.txt - $lmdir/G.fst 29 | else 30 | arpa2fst --disambig-symbol=#0 --read-symbol-table=$lmdir/words.txt $arpa_lm $lmdir/G.fst 31 | fi 32 | fstisstochastic $lmdir/G.fst 33 | utils/validate_lang.pl $lmdir || exit 1 34 | 35 | echo "done preparing lm ($lmdir)" 36 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select_decode.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | split="dev_other" 4 | ref_txt="" # ground truth transcript path 5 | psd_txt="" # pseudo transcript path 6 | get_best_wer=true 7 | dec_name="decode" 8 | graph_name="graph" 9 | kenlm_path=/checkpoint/abaevski/data/speech/libri/librispeech_lm_novox.phnc_o6.bin 10 | 11 | . ./cmd.sh 12 | . ./path.sh 13 | . parse_options.sh 14 | 15 | exp_root=$1 16 | unsup_args="" 17 | if [ $# -ge 2 ]; then 18 | unsup_args=$2 19 | fi 20 | 21 | set -eu 22 | 23 | if [ ! -z $ref_txt ] && $get_best_wer; then 24 | echo "==== WER w.r.t. real transcript (select based on unsupervised metric)" 25 | for x in $exp_root/*/${dec_name}_${split}*; do 26 | lang=$(dirname $x)/$graph_name 27 | 28 | ( 29 | for tra in $x/scoring/*.tra; do 30 | cat $tra | utils/int2sym.pl -f 2- $lang/words.txt | sed 's:::g' | sed 's:::g' > $tra.txt 31 | python local/unsup_select.py $psd_txt $tra.txt --kenlm_path $kenlm_path --gt_tra $ref_txt $unsup_args 32 | done 2>/dev/null | grep "score=" | sed 's/=/ /g' | sed 's/;//g' | sort -k3n | head -n1 33 | ) & 34 | done 35 | fi 36 | wait 37 | 38 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/local/unsup_select_decode_word.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | split="dev_other" 4 | ref_txt="" # ground truth transcript path 5 | psd_txt="" # pseudo transcript path 6 | get_best_wer=true 7 | dec_name="decode" 8 | graph_name="graph" 9 | kenlm_path=/checkpoint/abaevski/data/speech/libri/librispeech_lm_novox.phnc_o6.bin 10 | phonemize_lexicon="" 11 | 12 | . ./cmd.sh 13 | . ./path.sh 14 | . parse_options.sh 15 | . /private/home/wnhsu/unsup_asr/fairseq-py-unsup/env.sh 16 | 17 | exp_root=$1 18 | 19 | set -eu 20 | 21 | if [ ! -z $ref_txt ] && $get_best_wer; then 22 | echo "==== WER w.r.t. real transcript (select based on unsupervised metric)" 23 | for x in $exp_root/*/${dec_name}_${split}*; do 24 | lang=$(dirname $x)/$graph_name 25 | 26 | for tra in $x/scoring/*.tra; do 27 | cat $tra | utils/int2sym.pl -f 2- $lang/words.txt | sed 's:\::g' > $tra.txt 28 | python local/unsup_select.py $psd_txt $tra.txt \ 29 | --kenlm_path $kenlm_path --gt_tra $ref_txt --phonemize \ 30 | --phonemize_lexicon "$phonemize_lexicon" 31 | done | grep "score=" | sed 's/=/ /g' | sed 's/;//g' | sort -k3n | head -n1 32 | done 33 | fi 34 | 35 | 36 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/path.sh: -------------------------------------------------------------------------------- 1 | export KALDI_ROOT=`pwd`/../../.. 2 | export PATH=$PWD/utils/:$KALDI_ROOT/tools/openfst/bin:$PWD:$PATH 3 | [ ! -f $KALDI_ROOT/tools/config/common_path.sh ] && echo >&2 "The standard file $KALDI_ROOT/tools/config/common_path.sh is not present -> Exit!" && exit 1 4 | . $KALDI_ROOT/tools/config/common_path.sh 5 | export LC_ALL=C 6 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/steps: -------------------------------------------------------------------------------- 1 | ../../wsj/s5/steps -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/utils: -------------------------------------------------------------------------------- 1 | ../../wsj/s5/utils -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/models/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .wav2vec_u import Wav2vec_U 7 | 8 | 9 | __all__ = [ 10 | "Wav2vec_U", 11 | ] 12 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/scripts/copy_labels.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 -u 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import sys 8 | 9 | for idx, line in enumerate(sys.stdin): 10 | print(f"utt{idx:010d} {line}", end="") 11 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/scripts/filter_lexicon.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 -u 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import argparse 8 | import sys 9 | 10 | from fairseq.data import Dictionary 11 | 12 | 13 | def get_parser(): 14 | parser = argparse.ArgumentParser( 15 | description="filters a lexicon given a unit dictionary" 16 | ) 17 | parser.add_argument("-d", "--unit-dict", help="unit dictionary", required=True) 18 | return parser 19 | 20 | 21 | def main(): 22 | parser = get_parser() 23 | args = parser.parse_args() 24 | 25 | d = Dictionary.load(args.unit_dict) 26 | symbols = set(d.symbols) 27 | 28 | for line in sys.stdin: 29 | items = line.rstrip().split() 30 | skip = len(items) < 2 31 | for x in items[1:]: 32 | if x not in symbols: 33 | skip = True 34 | break 35 | if not skip: 36 | print(line, end="") 37 | 38 | 39 | if __name__ == "__main__": 40 | main() 41 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/scripts/filter_tsv.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 -u 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import os 8 | import argparse 9 | import sys 10 | 11 | 12 | parser = argparse.ArgumentParser() 13 | parser.add_argument("--tsv", required=True, type=str) 14 | parser.add_argument("--no-skip", action="store_true") 15 | parser.add_argument("--keep", action="store_true") 16 | params = parser.parse_args() 17 | 18 | 19 | def get_fname(line): 20 | p = os.path.basename(line.split("\t")[0]) 21 | p = os.path.splitext(p)[0] 22 | return p 23 | 24 | 25 | # filenames to exclude 26 | seen = set() 27 | with open(params.tsv) as f: 28 | if not params.no_skip: 29 | root = next(f).rstrip() 30 | for line in f: 31 | seen.add(get_fname(line)) 32 | 33 | for i, line in enumerate(sys.stdin): 34 | exists = get_fname(line) in seen 35 | keep = (exists and params.keep) or (not exists and not params.keep) 36 | if i == 0 or keep: 37 | print(line, end="") 38 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/scripts/g2p_wrd_to_phn.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 -u 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import argparse 8 | import sys 9 | 10 | from g2p_en import G2p 11 | 12 | 13 | def main(): 14 | parser = argparse.ArgumentParser() 15 | parser.add_argument( 16 | "--compact", 17 | action="store_true", 18 | help="if set, compacts phones", 19 | ) 20 | args = parser.parse_args() 21 | 22 | compact = args.compact 23 | 24 | wrd_to_phn = {} 25 | g2p = G2p() 26 | for line in sys.stdin: 27 | words = line.strip().split() 28 | phones = [] 29 | for w in words: 30 | if w not in wrd_to_phn: 31 | wrd_to_phn[w] = g2p(w) 32 | if compact: 33 | wrd_to_phn[w] = [ 34 | p[:-1] if p[-1].isnumeric() else p for p in wrd_to_phn[w] 35 | ] 36 | phones.extend(wrd_to_phn[w]) 37 | try: 38 | print(" ".join(phones)) 39 | except: 40 | print(wrd_to_phn, words, phones, file=sys.stderr) 41 | raise 42 | 43 | 44 | if __name__ == "__main__": 45 | main() 46 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/scripts/ltr_to_wrd.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 -u 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import sys 8 | 9 | 10 | def main(): 11 | for line in sys.stdin: 12 | print(line.replace(" ", "").replace("|", " ").strip()) 13 | 14 | 15 | if __name__ == "__main__": 16 | main() 17 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/scripts/normalize_text.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import regex 8 | import sys 9 | 10 | 11 | def main(): 12 | filter_r = regex.compile(r"[^\p{L}\p{N}\p{M}\' \-]") 13 | 14 | for line in sys.stdin: 15 | line = line.strip() 16 | line = filter_r.sub(" ", line) 17 | line = " ".join(line.split()) 18 | print(line) 19 | 20 | 21 | if __name__ == "__main__": 22 | main() 23 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/scripts/wrd_to_ltr.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 -u 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | import sys 8 | 9 | 10 | def main(): 11 | for line in sys.stdin: 12 | print(" ".join(list(line.strip().replace(" ", "|"))) + " |") 13 | 14 | 15 | if __name__ == "__main__": 16 | main() 17 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/tasks/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .unpaired_audio_text import UnpairedAudioText 7 | 8 | 9 | __all__ = [ 10 | "UnpairedAudioText", 11 | ] 12 | -------------------------------------------------------------------------------- /fairseq/benchmark/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | # import models/tasks to register them 7 | from . import dummy_dataset, dummy_lm, dummy_masked_lm, dummy_model, dummy_mt # noqa 8 | -------------------------------------------------------------------------------- /fairseq/benchmark/dummy_dataset.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from fairseq.data import FairseqDataset 3 | 4 | 5 | class DummyDataset(FairseqDataset): 6 | def __init__(self, batch, num_items, item_size): 7 | super().__init__() 8 | self.batch = batch 9 | self.num_items = num_items 10 | self.item_size = item_size 11 | 12 | def __getitem__(self, index): 13 | return index 14 | 15 | def __len__(self): 16 | return self.num_items 17 | 18 | def collater(self, samples): 19 | return self.batch 20 | 21 | @property 22 | def sizes(self): 23 | return np.array([self.item_size] * self.num_items) 24 | 25 | def num_tokens(self, index): 26 | return self.item_size 27 | 28 | def size(self, index): 29 | return self.item_size 30 | 31 | def ordered_indices(self): 32 | return np.arange(self.num_items) 33 | 34 | @property 35 | def supports_prefetch(self): 36 | return False 37 | -------------------------------------------------------------------------------- /fairseq/clib/libbleu/module.cpp: -------------------------------------------------------------------------------- 1 | /** 2 | * Copyright 2017-present, Facebook, Inc. 3 | * All rights reserved. 4 | * 5 | * This source code is licensed under the license found in the 6 | * LICENSE file in the root directory of this source tree. 7 | */ 8 | 9 | #include 10 | 11 | 12 | static PyMethodDef method_def[] = { 13 | {NULL, NULL, 0, NULL} 14 | }; 15 | 16 | static struct PyModuleDef module_def = { 17 | PyModuleDef_HEAD_INIT, 18 | "libbleu", /* name of module */ 19 | NULL, /* module documentation, may be NULL */ 20 | -1, /* size of per-interpreter state of the module, 21 | or -1 if the module keeps state in global variables. */ 22 | method_def 23 | }; 24 | 25 | 26 | #if PY_MAJOR_VERSION == 2 27 | PyMODINIT_FUNC init_libbleu() 28 | #else 29 | PyMODINIT_FUNC PyInit_libbleu() 30 | #endif 31 | { 32 | PyObject *m = PyModule_Create(&module_def); 33 | if (!m) { 34 | return NULL; 35 | } 36 | return m; 37 | } 38 | -------------------------------------------------------------------------------- /fairseq/clib/libnat_cuda/edit_dist.h: -------------------------------------------------------------------------------- 1 | /** 2 | * Copyright 2017-present, Facebook, Inc. 3 | * All rights reserved. 4 | * 5 | * This source code is licensed under the license found in the 6 | * LICENSE file in the root directory of this source tree. 7 | */ 8 | 9 | #pragma once 10 | 11 | #include 12 | 13 | torch::Tensor LevenshteinDistanceCuda( 14 | torch::Tensor source, 15 | torch::Tensor target, 16 | torch::Tensor source_length, 17 | torch::Tensor target_length); 18 | 19 | torch::Tensor GenerateDeletionLabelCuda( 20 | torch::Tensor source, 21 | torch::Tensor operations); 22 | 23 | std::pair GenerateInsertionLabelCuda( 24 | torch::Tensor source, 25 | torch::Tensor operations); 26 | -------------------------------------------------------------------------------- /fairseq/config/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | -------------------------------------------------------------------------------- /fairseq/config/config.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | hydra: 4 | run: 5 | dir: . 6 | 7 | defaults: 8 | - task: null 9 | - model: null 10 | - criterion: cross_entropy 11 | - optimizer: null 12 | - lr_scheduler: fixed 13 | - bpe: null 14 | - tokenizer: null 15 | - scoring: null 16 | - generation: null 17 | - common_eval: null 18 | - eval_lm: null 19 | -------------------------------------------------------------------------------- /fairseq/config/model/transformer_lm/transformer_lm_baevski_gbw.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation_fn: "relu" 3 | dropout: 0.1 4 | attention_dropout: 0.1 5 | activation_dropout: 0.0 6 | relu_dropout: 0.0 7 | decoder_embed_dim: 512 8 | decoder_output_dim: 512 9 | decoder_input_dim: 512 10 | decoder_ffn_embed_dim: 4096 11 | decoder_layers: 12 12 | decoder_attention_heads: 16 13 | decoder_normalize_before: true 14 | no_decoder_final_norm: true 15 | adaptive_softmax_cutoff: null 16 | adaptive_softmax_dropout: 0 17 | adaptive_softmax_factor: 4 18 | no_token_positional_embeddings: false 19 | share_decoder_input_output_embed: false 20 | character_embeddings: false 21 | character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" 22 | character_embedding_dim: 4 23 | char_embedder_highway_layers: 2 24 | adaptive_input: false 25 | adaptive_input_factor: 4 26 | adaptive_input_cutoff: null 27 | tie_adaptive_weights: false 28 | tie_adaptive_proj: false 29 | decoder_learned_pos: false 30 | decoder_layerdrop: 0 31 | decoder_layers_to_keep: null 32 | layernorm_embedding: false 33 | no_scale_embedding: false 34 | quant_noise_pq: 0 35 | quant_noise_pq_block_size: 8 36 | quant_noise_scalar: 0 37 | -------------------------------------------------------------------------------- /fairseq/config/model/transformer_lm/transformer_lm_baevski_wiki103.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation_fn: "relu" 3 | dropout: 0.3 4 | attention_dropout: 0.1 5 | activation_dropout: 0.1 6 | relu_dropout: 0.1 7 | decoder_embed_dim: 1024 8 | decoder_output_dim: 1024 9 | decoder_input_dim: 1024 10 | decoder_ffn_embed_dim: 4096 11 | decoder_layers: 16 12 | decoder_attention_heads: 8 13 | decoder_normalize_before: true 14 | no_decoder_final_norm: true 15 | adaptive_softmax_cutoff: "20000,60000" 16 | adaptive_softmax_dropout: 0.2 17 | adaptive_softmax_factor: 4 18 | no_token_positional_embeddings: false 19 | share_decoder_input_output_embed: false 20 | character_embeddings: false 21 | character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" 22 | character_embedding_dim: 4 23 | char_embedder_highway_layers: 2 24 | adaptive_input: true 25 | adaptive_input_factor: 4 26 | adaptive_input_cutoff: "20000,60000" 27 | tie_adaptive_weights: true 28 | tie_adaptive_proj: true 29 | decoder_learned_pos: false 30 | decoder_layerdrop: 0 31 | decoder_layers_to_keep: null 32 | layernorm_embedding: false 33 | no_scale_embedding: false 34 | quant_noise_pq: 0 35 | quant_noise_pq_block_size: 8 36 | quant_noise_scalar: 0 37 | -------------------------------------------------------------------------------- /fairseq/config/model/transformer_lm/transformer_lm_big.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation_fn: "relu" 3 | dropout: 0.1 4 | attention_dropout: 0.0 5 | activation_dropout: 0.0 6 | relu_dropout: 0.0 7 | decoder_embed_dim: 1024 8 | decoder_output_dim: 1024 9 | decoder_input_dim: 1024 10 | decoder_ffn_embed_dim: 4096 11 | decoder_layers: 12 12 | decoder_attention_heads: 16 13 | decoder_normalize_before: true 14 | no_decoder_final_norm: false 15 | adaptive_softmax_cutoff: null 16 | adaptive_softmax_dropout: 0 17 | adaptive_softmax_factor: 4 18 | no_token_positional_embeddings: false 19 | share_decoder_input_output_embed: false 20 | character_embeddings: false 21 | character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" 22 | character_embedding_dim: 4 23 | char_embedder_highway_layers: 2 24 | adaptive_input: false 25 | adaptive_input_factor: 4 26 | adaptive_input_cutoff: null 27 | tie_adaptive_weights: false 28 | tie_adaptive_proj: false 29 | decoder_learned_pos: false 30 | decoder_layerdrop: 0 31 | decoder_layers_to_keep: null 32 | layernorm_embedding: false 33 | no_scale_embedding: false 34 | quant_noise_pq: 0 35 | quant_noise_pq_block_size: 8 36 | quant_noise_scalar: 0 37 | -------------------------------------------------------------------------------- /fairseq/config/model/transformer_lm/transformer_lm_gbw.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation_fn: "relu" 3 | dropout: 0.1 4 | attention_dropout: 0.1 5 | activation_dropout: 0.0 6 | relu_dropout: 0.0 7 | decoder_embed_dim: 512 8 | decoder_output_dim: 512 9 | decoder_input_dim: 512 10 | decoder_ffn_embed_dim: 4096 11 | decoder_layers: 12 12 | decoder_attention_heads: 16 13 | decoder_normalize_before: true 14 | no_decoder_final_norm: true 15 | adaptive_softmax_cutoff: null 16 | adaptive_softmax_dropout: 0 17 | adaptive_softmax_factor: 4 18 | no_token_positional_embeddings: false 19 | share_decoder_input_output_embed: false 20 | character_embeddings: false 21 | character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" 22 | character_embedding_dim: 4 23 | char_embedder_highway_layers: 2 24 | adaptive_input: false 25 | adaptive_input_factor: 4 26 | adaptive_input_cutoff: null 27 | tie_adaptive_weights: false 28 | tie_adaptive_proj: false 29 | decoder_learned_pos: false 30 | decoder_layerdrop: 0 31 | decoder_layers_to_keep: null 32 | layernorm_embedding: false 33 | no_scale_embedding: false 34 | quant_noise_pq: 0 35 | quant_noise_pq_block_size: 8 36 | quant_noise_scalar: 0 37 | -------------------------------------------------------------------------------- /fairseq/config/model/transformer_lm/transformer_lm_gpt.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation_fn: "gelu" 3 | dropout: 0.1 4 | attention_dropout: 0.1 5 | activation_dropout: 0.0 6 | relu_dropout: 0.0 7 | decoder_embed_dim: 768 8 | decoder_output_dim: 768 9 | decoder_input_dim: 768 10 | decoder_ffn_embed_dim: 3072 11 | decoder_layers: 12 12 | decoder_attention_heads: 12 13 | decoder_normalize_before: true 14 | no_decoder_final_norm: false 15 | adaptive_softmax_cutoff: null 16 | adaptive_softmax_dropout: 0 17 | adaptive_softmax_factor: 4 18 | no_token_positional_embeddings: false 19 | share_decoder_input_output_embed: false 20 | character_embeddings: false 21 | character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" 22 | character_embedding_dim: 4 23 | char_embedder_highway_layers: 2 24 | adaptive_input: false 25 | adaptive_input_factor: 4 26 | adaptive_input_cutoff: null 27 | tie_adaptive_weights: false 28 | tie_adaptive_proj: false 29 | decoder_learned_pos: false 30 | decoder_layerdrop: 0 31 | decoder_layers_to_keep: null 32 | layernorm_embedding: false 33 | no_scale_embedding: false 34 | quant_noise_pq: 0 35 | quant_noise_pq_block_size: 8 36 | quant_noise_scalar: 0 37 | -------------------------------------------------------------------------------- /fairseq/config/model/transformer_lm/transformer_lm_gpt2_big.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation_fn: "gelu" 3 | dropout: 0.1 4 | attention_dropout: 0.1 5 | activation_dropout: 0.0 6 | relu_dropout: 0.0 7 | decoder_embed_dim: 1600 8 | decoder_output_dim: 1600 9 | decoder_input_dim: 1600 10 | decoder_ffn_embed_dim: 6400 11 | decoder_layers: 48 12 | decoder_attention_heads: 25 13 | decoder_normalize_before: true 14 | no_decoder_final_norm: false 15 | adaptive_softmax_cutoff: null 16 | adaptive_softmax_dropout: 0 17 | adaptive_softmax_factor: 4 18 | no_token_positional_embeddings: false 19 | share_decoder_input_output_embed: false 20 | character_embeddings: false 21 | character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" 22 | character_embedding_dim: 4 23 | char_embedder_highway_layers: 2 24 | adaptive_input: false 25 | adaptive_input_factor: 4 26 | adaptive_input_cutoff: null 27 | tie_adaptive_weights: false 28 | tie_adaptive_proj: false 29 | decoder_learned_pos: false 30 | decoder_layerdrop: 0 31 | decoder_layers_to_keep: null 32 | layernorm_embedding: false 33 | no_scale_embedding: false 34 | quant_noise_pq: 0 35 | quant_noise_pq_block_size: 8 36 | quant_noise_scalar: 0 37 | -------------------------------------------------------------------------------- /fairseq/config/model/transformer_lm/transformer_lm_gpt2_medium.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation_fn: "gelu" 3 | dropout: 0.1 4 | attention_dropout: 0.1 5 | activation_dropout: 0.0 6 | relu_dropout: 0.0 7 | decoder_embed_dim: 1280 8 | decoder_output_dim: 1280 9 | decoder_input_dim: 1280 10 | decoder_ffn_embed_dim: 5120 11 | decoder_layers: 36 12 | decoder_attention_heads: 20 13 | decoder_normalize_before: true 14 | no_decoder_final_norm: false 15 | adaptive_softmax_cutoff: null 16 | adaptive_softmax_dropout: 0 17 | adaptive_softmax_factor: 4 18 | no_token_positional_embeddings: false 19 | share_decoder_input_output_embed: false 20 | character_embeddings: false 21 | character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" 22 | character_embedding_dim: 4 23 | char_embedder_highway_layers: 2 24 | adaptive_input: false 25 | adaptive_input_factor: 4 26 | adaptive_input_cutoff: null 27 | tie_adaptive_weights: false 28 | tie_adaptive_proj: false 29 | decoder_learned_pos: false 30 | decoder_layerdrop: 0 31 | decoder_layers_to_keep: null 32 | layernorm_embedding: false 33 | no_scale_embedding: false 34 | quant_noise_pq: 0 35 | quant_noise_pq_block_size: 8 36 | quant_noise_scalar: 0 37 | -------------------------------------------------------------------------------- /fairseq/config/model/transformer_lm/transformer_lm_gpt2_small.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation_fn: "gelu" 3 | dropout: 0.1 4 | attention_dropout: 0.1 5 | activation_dropout: 0.0 6 | relu_dropout: 0.0 7 | decoder_embed_dim: 1024 8 | decoder_output_dim: 1024 9 | decoder_input_dim: 1024 10 | decoder_ffn_embed_dim: 4096 11 | decoder_layers: 24 12 | decoder_attention_heads: 16 13 | decoder_normalize_before: true 14 | no_decoder_final_norm: false 15 | adaptive_softmax_cutoff: null 16 | adaptive_softmax_dropout: 0 17 | adaptive_softmax_factor: 4 18 | no_token_positional_embeddings: false 19 | share_decoder_input_output_embed: false 20 | character_embeddings: false 21 | character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" 22 | character_embedding_dim: 4 23 | char_embedder_highway_layers: 2 24 | adaptive_input: false 25 | adaptive_input_factor: 4 26 | adaptive_input_cutoff: null 27 | tie_adaptive_weights: false 28 | tie_adaptive_proj: false 29 | decoder_learned_pos: false 30 | decoder_layerdrop: 0 31 | decoder_layers_to_keep: null 32 | layernorm_embedding: false 33 | no_scale_embedding: false 34 | quant_noise_pq: 0 35 | quant_noise_pq_block_size: 8 36 | quant_noise_scalar: 0 37 | -------------------------------------------------------------------------------- /fairseq/config/model/transformer_lm/transformer_lm_wiki103.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation_fn: "relu" 3 | dropout: 0.3 4 | attention_dropout: 0.1 5 | activation_dropout: 0.1 6 | relu_dropout: 0.1 7 | decoder_embed_dim: 1024 8 | decoder_output_dim: 1024 9 | decoder_input_dim: 1024 10 | decoder_ffn_embed_dim: 4096 11 | decoder_layers: 16 12 | decoder_attention_heads: 8 13 | decoder_normalize_before: true 14 | no_decoder_final_norm: true 15 | adaptive_softmax_cutoff: "20000,60000" 16 | adaptive_softmax_dropout: 0.2 17 | adaptive_softmax_factor: 4 18 | no_token_positional_embeddings: false 19 | share_decoder_input_output_embed: false 20 | character_embeddings: false 21 | character_filters: "[(1, 64), (2, 128), (3, 192), (4, 256), (5, 256), (6, 256), (7, 256)]" 22 | character_embedding_dim: 4 23 | char_embedder_highway_layers: 2 24 | adaptive_input: true 25 | adaptive_input_factor: 4 26 | adaptive_input_cutoff: "20000,60000" 27 | tie_adaptive_weights: true 28 | tie_adaptive_proj: true 29 | decoder_learned_pos: false 30 | decoder_layerdrop: 0 31 | decoder_layers_to_keep: null 32 | layernorm_embedding: false 33 | no_scale_embedding: false 34 | quant_noise_pq: 0 35 | quant_noise_pq_block_size: 8 36 | quant_noise_scalar: 0 37 | -------------------------------------------------------------------------------- /fairseq/config/model/wav2vec/vq_wav2vec_gumbel.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | activation: gelu 3 | vq_type: gumbel 4 | vq_depth: 2 5 | combine_groups: true 6 | -------------------------------------------------------------------------------- /fairseq/config/model/wav2vec2/wav2vec2_base.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | quantize_targets: true 4 | final_dim: 256 5 | encoder_layerdrop: 0.05 6 | dropout_input: 0.1 7 | dropout_features: 0.1 8 | feature_grad_mult: 0.1 9 | -------------------------------------------------------------------------------- /fairseq/config/model/wav2vec2/wav2vec2_large.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | quantize_targets: true 4 | extractor_mode: layer_norm 5 | layer_norm_first: true 6 | final_dim: 768 7 | latent_temp: [2.0,0.1,0.999995] 8 | encoder_layerdrop: 0.0 9 | dropout_input: 0.0 10 | dropout_features: 0.0 11 | dropout: 0.0 12 | attention_dropout: 0.0 13 | conv_bias: true 14 | 15 | encoder_layers: 24 16 | encoder_embed_dim: 1024 17 | encoder_ffn_embed_dim: 4096 18 | encoder_attention_heads: 16 19 | 20 | feature_grad_mult: 1.0 21 | -------------------------------------------------------------------------------- /fairseq/criterions/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | """isort:skip_file""" 6 | 7 | import importlib 8 | import os 9 | 10 | from fairseq import registry 11 | from fairseq.criterions.fairseq_criterion import ( # noqa 12 | FairseqCriterion, 13 | LegacyFairseqCriterion, 14 | ) 15 | from omegaconf import DictConfig 16 | 17 | 18 | ( 19 | build_criterion_, 20 | register_criterion, 21 | CRITERION_REGISTRY, 22 | CRITERION_DATACLASS_REGISTRY, 23 | ) = registry.setup_registry( 24 | "--criterion", base_class=FairseqCriterion, default="cross_entropy" 25 | ) 26 | 27 | 28 | def build_criterion(cfg: DictConfig, task): 29 | return build_criterion_(cfg, task) 30 | 31 | 32 | # automatically import any Python files in the criterions/ directory 33 | for file in sorted(os.listdir(os.path.dirname(__file__))): 34 | if file.endswith(".py") and not file.startswith("_"): 35 | file_name = file[: file.find(".py")] 36 | importlib.import_module("fairseq.criterions." + file_name) 37 | -------------------------------------------------------------------------------- /fairseq/data/append_token_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import numpy as np 7 | import torch 8 | 9 | from . import BaseWrapperDataset 10 | 11 | 12 | class AppendTokenDataset(BaseWrapperDataset): 13 | def __init__(self, dataset, token=None): 14 | super().__init__(dataset) 15 | self.token = token 16 | if token is not None: 17 | self._sizes = np.array(dataset.sizes) + 1 18 | else: 19 | self._sizes = dataset.sizes 20 | 21 | def __getitem__(self, idx): 22 | item = self.dataset[idx] 23 | if self.token is not None: 24 | item = torch.cat([item, item.new([self.token])]) 25 | return item 26 | 27 | @property 28 | def sizes(self): 29 | return self._sizes 30 | 31 | def num_tokens(self, index): 32 | n = self.dataset.num_tokens(index) 33 | if self.token is not None: 34 | n += 1 35 | return n 36 | 37 | def size(self, index): 38 | n = self.dataset.size(index) 39 | if self.token is not None: 40 | n += 1 41 | return n 42 | -------------------------------------------------------------------------------- /fairseq/data/audio/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/fairseq/data/audio/__init__.py -------------------------------------------------------------------------------- /fairseq/data/audio/feature_transforms/global_cmvn.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from fairseq.data.audio.feature_transforms import ( 3 | AudioFeatureTransform, 4 | register_audio_feature_transform, 5 | ) 6 | 7 | 8 | @register_audio_feature_transform("global_cmvn") 9 | class GlobalCMVN(AudioFeatureTransform): 10 | """Global CMVN (cepstral mean and variance normalization). The global mean 11 | and variance need to be pre-computed and stored in NumPy format (.npz).""" 12 | 13 | @classmethod 14 | def from_config_dict(cls, config=None): 15 | _config = {} if config is None else config 16 | return GlobalCMVN(_config.get("stats_npz_path")) 17 | 18 | def __init__(self, stats_npz_path): 19 | self.stats_npz_path = stats_npz_path 20 | stats = np.load(stats_npz_path) 21 | self.mean, self.std = stats["mean"], stats["std"] 22 | 23 | def __repr__(self): 24 | return self.__class__.__name__ + f'(stats_npz_path="{self.stats_npz_path}")' 25 | 26 | def __call__(self, x): 27 | x = np.subtract(x, self.mean) 28 | x = np.divide(x, self.std) 29 | return x 30 | -------------------------------------------------------------------------------- /fairseq/data/audio/feature_transforms/utterance_cmvn.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from fairseq.data.audio.feature_transforms import ( 3 | AudioFeatureTransform, 4 | register_audio_feature_transform, 5 | ) 6 | 7 | 8 | @register_audio_feature_transform("utterance_cmvn") 9 | class UtteranceCMVN(AudioFeatureTransform): 10 | """Utterance-level CMVN (cepstral mean and variance normalization)""" 11 | 12 | @classmethod 13 | def from_config_dict(cls, config=None): 14 | _config = {} if config is None else config 15 | return UtteranceCMVN( 16 | _config.get("norm_means", True), 17 | _config.get("norm_vars", True), 18 | ) 19 | 20 | def __init__(self, norm_means=True, norm_vars=True): 21 | self.norm_means, self.norm_vars = norm_means, norm_vars 22 | 23 | def __repr__(self): 24 | return ( 25 | self.__class__.__name__ 26 | + f"(norm_means={self.norm_means}, norm_vars={self.norm_vars})" 27 | ) 28 | 29 | def __call__(self, x): 30 | mean = x.mean(axis=0) 31 | square_sums = (x ** 2).sum(axis=0) 32 | 33 | if self.norm_means: 34 | x = np.subtract(x, mean) 35 | if self.norm_vars: 36 | var = square_sums / x.shape[0] - mean ** 2 37 | std = np.sqrt(np.maximum(var, 1e-10)) 38 | x = np.divide(x, std) 39 | 40 | return x 41 | -------------------------------------------------------------------------------- /fairseq/data/colorize_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import torch 7 | 8 | from . import BaseWrapperDataset 9 | 10 | 11 | class ColorizeDataset(BaseWrapperDataset): 12 | """ Adds 'colors' property to net input that is obtained from the provided color getter for use by models """ 13 | 14 | def __init__(self, dataset, color_getter): 15 | super().__init__(dataset) 16 | self.color_getter = color_getter 17 | 18 | def collater(self, samples): 19 | base_collate = super().collater(samples) 20 | if len(base_collate) > 0: 21 | base_collate["net_input"]["colors"] = torch.tensor( 22 | list(self.color_getter(self.dataset, s["id"]) for s in samples), 23 | dtype=torch.long, 24 | ) 25 | return base_collate 26 | -------------------------------------------------------------------------------- /fairseq/data/encoders/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | 7 | import importlib 8 | import os 9 | 10 | from fairseq import registry 11 | 12 | 13 | build_tokenizer, register_tokenizer, TOKENIZER_REGISTRY, _ = registry.setup_registry( 14 | "--tokenizer", 15 | default=None, 16 | ) 17 | 18 | 19 | build_bpe, register_bpe, BPE_REGISTRY, _ = registry.setup_registry( 20 | "--bpe", 21 | default=None, 22 | ) 23 | 24 | 25 | # automatically import any Python files in the encoders/ directory 26 | for file in sorted(os.listdir(os.path.dirname(__file__))): 27 | if file.endswith(".py") and not file.startswith("_"): 28 | module = file[: file.find(".py")] 29 | importlib.import_module("fairseq.data.encoders." + module) 30 | -------------------------------------------------------------------------------- /fairseq/data/encoders/bytes.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | 7 | from fairseq.data.encoders import register_bpe 8 | from fairseq.data.encoders.byte_utils import ( 9 | SPACE, 10 | SPACE_ESCAPE, 11 | byte_encode, 12 | smart_byte_decode, 13 | ) 14 | 15 | 16 | @register_bpe("bytes") 17 | class Bytes(object): 18 | def __init__(self, *unused): 19 | pass 20 | 21 | @staticmethod 22 | def add_args(parser): 23 | pass 24 | 25 | @staticmethod 26 | def encode(x: str) -> str: 27 | encoded = byte_encode(x) 28 | escaped = encoded.replace(SPACE, SPACE_ESCAPE) 29 | return SPACE.join(list(escaped)) 30 | 31 | @staticmethod 32 | def decode(x: str) -> str: 33 | unescaped = x.replace(SPACE, "").replace(SPACE_ESCAPE, SPACE) 34 | return smart_byte_decode(unescaped) 35 | -------------------------------------------------------------------------------- /fairseq/data/encoders/characters.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | 7 | from fairseq.data.encoders import register_bpe 8 | 9 | 10 | SPACE = chr(32) 11 | SPACE_ESCAPE = chr(9601) 12 | 13 | 14 | @register_bpe("characters") 15 | class Characters(object): 16 | def __init__(self, *unused): 17 | pass 18 | 19 | @staticmethod 20 | def add_args(parser): 21 | pass 22 | 23 | @staticmethod 24 | def encode(x: str) -> str: 25 | escaped = x.replace(SPACE, SPACE_ESCAPE) 26 | return SPACE.join(list(escaped)) 27 | 28 | @staticmethod 29 | def decode(x: str) -> str: 30 | return x.replace(SPACE, "").replace(SPACE_ESCAPE, SPACE) 31 | -------------------------------------------------------------------------------- /fairseq/data/encoders/fastbpe.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from dataclasses import dataclass, field 7 | 8 | from fairseq import file_utils 9 | from fairseq.data.encoders import register_bpe 10 | from fairseq.dataclass import FairseqDataclass 11 | 12 | 13 | @dataclass 14 | class fastBPEConfig(FairseqDataclass): 15 | bpe_codes: str = field(default="???", metadata={"help": "path to fastBPE BPE"}) 16 | 17 | 18 | @register_bpe("fastbpe", dataclass=fastBPEConfig) 19 | class fastBPE(object): 20 | def __init__(self, cfg): 21 | if cfg.bpe_codes is None: 22 | raise ValueError("--bpe-codes is required for --bpe=fastbpe") 23 | codes = file_utils.cached_path(cfg.bpe_codes) 24 | try: 25 | import fastBPE 26 | 27 | self.bpe = fastBPE.fastBPE(codes) 28 | self.bpe_symbol = "@@ " 29 | except ImportError: 30 | raise ImportError("Please install fastBPE with: pip install fastBPE") 31 | 32 | def encode(self, x: str) -> str: 33 | return self.bpe.apply([x])[0] 34 | 35 | def decode(self, x: str) -> str: 36 | return (x + " ").replace(self.bpe_symbol, "").rstrip() 37 | -------------------------------------------------------------------------------- /fairseq/data/encoders/nltk_tokenizer.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from fairseq.data.encoders import register_tokenizer 7 | from fairseq.dataclass import FairseqDataclass 8 | 9 | 10 | @register_tokenizer("nltk", dataclass=FairseqDataclass) 11 | class NLTKTokenizer(object): 12 | def __init__(self, *unused): 13 | try: 14 | from nltk.tokenize import word_tokenize 15 | 16 | self.word_tokenize = word_tokenize 17 | except ImportError: 18 | raise ImportError("Please install nltk with: pip install nltk") 19 | 20 | def encode(self, x: str) -> str: 21 | return " ".join(self.word_tokenize(x)) 22 | 23 | def decode(self, x: str) -> str: 24 | return x 25 | -------------------------------------------------------------------------------- /fairseq/data/encoders/space_tokenizer.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import re 7 | 8 | from fairseq.data.encoders import register_tokenizer 9 | from fairseq.dataclass import FairseqDataclass 10 | 11 | 12 | @register_tokenizer("space", dataclass=FairseqDataclass) 13 | class SpaceTokenizer(object): 14 | def __init__(self, *unused): 15 | self.space_tok = re.compile(r"\s+") 16 | 17 | def encode(self, x: str) -> str: 18 | return self.space_tok.sub(" ", x) 19 | 20 | def decode(self, x: str) -> str: 21 | return x 22 | -------------------------------------------------------------------------------- /fairseq/data/encoders/utils.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import torch 7 | from fairseq.data import encoders 8 | 9 | 10 | def get_whole_word_mask(args, dictionary): 11 | bpe = encoders.build_bpe(args) 12 | if bpe is not None: 13 | 14 | def is_beginning_of_word(i): 15 | if i < dictionary.nspecial: 16 | # special elements are always considered beginnings 17 | return True 18 | tok = dictionary[i] 19 | if tok.startswith("madeupword"): 20 | return True 21 | try: 22 | return bpe.is_beginning_of_word(tok) 23 | except ValueError: 24 | return True 25 | 26 | mask_whole_words = torch.ByteTensor( 27 | list(map(is_beginning_of_word, range(len(dictionary)))) 28 | ) 29 | return mask_whole_words 30 | return None 31 | -------------------------------------------------------------------------------- /fairseq/data/id_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import torch 7 | 8 | from . import FairseqDataset 9 | 10 | 11 | class IdDataset(FairseqDataset): 12 | def __getitem__(self, index): 13 | return index 14 | 15 | def __len__(self): 16 | return 0 17 | 18 | def collater(self, samples): 19 | return torch.tensor(samples) 20 | -------------------------------------------------------------------------------- /fairseq/data/legacy/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .block_pair_dataset import BlockPairDataset 7 | from .masked_lm_dataset import MaskedLMDataset 8 | from .masked_lm_dictionary import BertDictionary, MaskedLMDictionary 9 | 10 | 11 | __all__ = [ 12 | "BertDictionary", 13 | "BlockPairDataset", 14 | "MaskedLMDataset", 15 | "MaskedLMDictionary", 16 | ] 17 | -------------------------------------------------------------------------------- /fairseq/data/list_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import BaseWrapperDataset 7 | 8 | 9 | class ListDataset(BaseWrapperDataset): 10 | def __init__(self, dataset, sizes=None): 11 | super().__init__(dataset) 12 | self._sizes = sizes 13 | 14 | def __iter__(self): 15 | for x in self.dataset: 16 | yield x 17 | 18 | def collater(self, samples): 19 | return samples 20 | 21 | @property 22 | def sizes(self): 23 | return self._sizes 24 | 25 | def num_tokens(self, index): 26 | return self.sizes[index] 27 | 28 | def size(self, index): 29 | return self.sizes[index] 30 | 31 | def set_epoch(self, epoch): 32 | pass 33 | -------------------------------------------------------------------------------- /fairseq/data/lru_cache_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from functools import lru_cache 7 | 8 | from . import BaseWrapperDataset 9 | 10 | 11 | class LRUCacheDataset(BaseWrapperDataset): 12 | def __init__(self, dataset, token=None): 13 | super().__init__(dataset) 14 | 15 | @lru_cache(maxsize=8) 16 | def __getitem__(self, index): 17 | return self.dataset[index] 18 | 19 | @lru_cache(maxsize=8) 20 | def collater(self, samples): 21 | return self.dataset.collater(samples) 22 | -------------------------------------------------------------------------------- /fairseq/data/multilingual/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | -------------------------------------------------------------------------------- /fairseq/data/num_samples_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import FairseqDataset 7 | 8 | 9 | class NumSamplesDataset(FairseqDataset): 10 | def __getitem__(self, index): 11 | return 1 12 | 13 | def __len__(self): 14 | return 0 15 | 16 | def collater(self, samples): 17 | return sum(samples) 18 | -------------------------------------------------------------------------------- /fairseq/data/numel_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import numpy as np 7 | import torch 8 | 9 | from . import BaseWrapperDataset 10 | 11 | 12 | class NumelDataset(BaseWrapperDataset): 13 | def __init__(self, dataset, reduce=False): 14 | super().__init__(dataset) 15 | self.reduce = reduce 16 | 17 | def __getitem__(self, index): 18 | item = self.dataset[index] 19 | if torch.is_tensor(item): 20 | return torch.numel(item) 21 | else: 22 | return np.size(item) 23 | 24 | def __len__(self): 25 | return len(self.dataset) 26 | 27 | def collater(self, samples): 28 | if self.reduce: 29 | return sum(samples) 30 | else: 31 | return torch.tensor(samples) 32 | -------------------------------------------------------------------------------- /fairseq/data/offset_tokens_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import BaseWrapperDataset 7 | 8 | 9 | class OffsetTokensDataset(BaseWrapperDataset): 10 | def __init__(self, dataset, offset): 11 | super().__init__(dataset) 12 | self.offset = offset 13 | 14 | def __getitem__(self, idx): 15 | return self.dataset[idx] + self.offset 16 | -------------------------------------------------------------------------------- /fairseq/data/pad_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from fairseq.data import data_utils 7 | 8 | from . import BaseWrapperDataset 9 | 10 | 11 | class PadDataset(BaseWrapperDataset): 12 | def __init__(self, dataset, pad_idx, left_pad): 13 | super().__init__(dataset) 14 | self.pad_idx = pad_idx 15 | self.left_pad = left_pad 16 | 17 | def collater(self, samples): 18 | return data_utils.collate_tokens(samples, self.pad_idx, left_pad=self.left_pad) 19 | 20 | 21 | class LeftPadDataset(PadDataset): 22 | def __init__(self, dataset, pad_idx): 23 | super().__init__(dataset, pad_idx, left_pad=True) 24 | 25 | 26 | class RightPadDataset(PadDataset): 27 | def __init__(self, dataset, pad_idx): 28 | super().__init__(dataset, pad_idx, left_pad=False) 29 | -------------------------------------------------------------------------------- /fairseq/data/prepend_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import numpy as np 7 | import torch 8 | 9 | from . import BaseWrapperDataset 10 | 11 | 12 | class PrependDataset(BaseWrapperDataset): 13 | def __init__(self, dataset, prepend_getter, ensure_first_token_is=None): 14 | super().__init__(dataset) 15 | self.prepend_getter = prepend_getter 16 | self.ensure_first_token = ensure_first_token_is 17 | 18 | def __getitem__(self, idx): 19 | item = self.dataset[idx] 20 | is_tuple = isinstance(item, tuple) 21 | src = item[0] if is_tuple else item 22 | 23 | assert self.ensure_first_token is None or src[0] == self.ensure_first_token 24 | prepend_idx = self.prepend_getter(self.dataset, idx) 25 | assert isinstance(prepend_idx, int) 26 | src[0] = prepend_idx 27 | item = tuple((src,) + item[1:]) if is_tuple else src 28 | return item 29 | -------------------------------------------------------------------------------- /fairseq/data/prepend_token_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import numpy as np 7 | import torch 8 | 9 | from . import BaseWrapperDataset 10 | 11 | 12 | class PrependTokenDataset(BaseWrapperDataset): 13 | def __init__(self, dataset, token=None): 14 | super().__init__(dataset) 15 | self.token = token 16 | if token is not None: 17 | self._sizes = np.array(dataset.sizes) + 1 18 | else: 19 | self._sizes = dataset.sizes 20 | 21 | def __getitem__(self, idx): 22 | item = self.dataset[idx] 23 | if self.token is not None: 24 | item = torch.cat([item.new([self.token]), item]) 25 | return item 26 | 27 | @property 28 | def sizes(self): 29 | return self._sizes 30 | 31 | def num_tokens(self, index): 32 | n = self.dataset.num_tokens(index) 33 | if self.token is not None: 34 | n += 1 35 | return n 36 | 37 | def size(self, index): 38 | n = self.dataset.size(index) 39 | if self.token is not None: 40 | n += 1 41 | return n 42 | -------------------------------------------------------------------------------- /fairseq/data/raw_label_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import torch 7 | 8 | from . import FairseqDataset 9 | 10 | 11 | class RawLabelDataset(FairseqDataset): 12 | def __init__(self, labels): 13 | super().__init__() 14 | self.labels = labels 15 | 16 | def __getitem__(self, index): 17 | return self.labels[index] 18 | 19 | def __len__(self): 20 | return len(self.labels) 21 | 22 | def collater(self, samples): 23 | return torch.tensor(samples) 24 | -------------------------------------------------------------------------------- /fairseq/data/roll_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import torch 7 | 8 | from . import BaseWrapperDataset 9 | 10 | 11 | class RollDataset(BaseWrapperDataset): 12 | def __init__(self, dataset, shifts): 13 | super().__init__(dataset) 14 | self.shifts = shifts 15 | 16 | def __getitem__(self, index): 17 | item = self.dataset[index] 18 | return torch.roll(item, self.shifts) 19 | -------------------------------------------------------------------------------- /fairseq/data/sort_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import numpy as np 7 | 8 | from . import BaseWrapperDataset 9 | 10 | 11 | class SortDataset(BaseWrapperDataset): 12 | def __init__(self, dataset, sort_order): 13 | super().__init__(dataset) 14 | if not isinstance(sort_order, (list, tuple)): 15 | sort_order = [sort_order] 16 | self.sort_order = sort_order 17 | 18 | assert all(len(so) == len(dataset) for so in sort_order) 19 | 20 | def ordered_indices(self): 21 | return np.lexsort(self.sort_order) 22 | -------------------------------------------------------------------------------- /fairseq/data/strip_token_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import BaseWrapperDataset 7 | 8 | 9 | class StripTokenDataset(BaseWrapperDataset): 10 | def __init__(self, dataset, id_to_strip): 11 | super().__init__(dataset) 12 | self.id_to_strip = id_to_strip 13 | 14 | def __getitem__(self, index): 15 | item = self.dataset[index] 16 | while len(item) > 0 and item[-1] == self.id_to_strip: 17 | item = item[:-1] 18 | while len(item) > 0 and item[0] == self.id_to_strip: 19 | item = item[1:] 20 | return item 21 | -------------------------------------------------------------------------------- /fairseq/dataclass/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .configs import FairseqDataclass 7 | from .constants import ChoiceEnum 8 | 9 | 10 | __all__ = [ 11 | "FairseqDataclass", 12 | "ChoiceEnum", 13 | ] 14 | -------------------------------------------------------------------------------- /fairseq/distributed/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .distributed_timeout_wrapper import DistributedTimeoutWrapper 7 | from .fully_sharded_data_parallel import fsdp_enable_wrap, fsdp_wrap, FullyShardedDataParallel 8 | from .legacy_distributed_data_parallel import LegacyDistributedDataParallel 9 | from .module_proxy_wrapper import ModuleProxyWrapper 10 | from .tpu_distributed_data_parallel import TPUDistributedDataParallel 11 | 12 | 13 | __all__ = [ 14 | "DistributedTimeoutWrapper", 15 | "fsdp_enable_wrap", 16 | "fsdp_wrap", 17 | "FullyShardedDataParallel", 18 | "LegacyDistributedDataParallel", 19 | "ModuleProxyWrapper", 20 | "TPUDistributedDataParallel", 21 | ] 22 | -------------------------------------------------------------------------------- /fairseq/logging/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/fairseq/logging/__init__.py -------------------------------------------------------------------------------- /fairseq/model_parallel/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from . import criterions, models, modules # noqa 7 | -------------------------------------------------------------------------------- /fairseq/model_parallel/criterions/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import importlib 7 | import os 8 | 9 | 10 | # automatically import any Python files in the criterions/ directory 11 | for file in sorted(os.listdir(os.path.dirname(__file__))): 12 | if file.endswith(".py") and not file.startswith("_"): 13 | module = file[: file.find(".py")] 14 | importlib.import_module("fairseq.model_parallel.criterions." + module) 15 | -------------------------------------------------------------------------------- /fairseq/model_parallel/models/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import importlib 7 | import os 8 | 9 | 10 | # automatically import any Python files in the models/ directory 11 | models_dir = os.path.dirname(__file__) 12 | for file in os.listdir(models_dir): 13 | path = os.path.join(models_dir, file) 14 | if ( 15 | not file.startswith("_") 16 | and not file.startswith(".") 17 | and (file.endswith(".py") or os.path.isdir(path)) 18 | ): 19 | model_name = file[: file.find(".py")] if file.endswith(".py") else file 20 | module = importlib.import_module("fairseq.model_parallel.models." + model_name) 21 | -------------------------------------------------------------------------------- /fairseq/model_parallel/models/pipeline_parallel_transformer/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .model import * # noqa 7 | -------------------------------------------------------------------------------- /fairseq/model_parallel/models/roberta/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .model import * # noqa 7 | -------------------------------------------------------------------------------- /fairseq/model_parallel/modules/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | """isort:skip_file""" 6 | 7 | from .multihead_attention import ModelParallelMultiheadAttention 8 | from .transformer_layer import ( 9 | ModelParallelTransformerEncoderLayer, 10 | ModelParallelTransformerDecoderLayer, 11 | ) 12 | 13 | __all__ = [ 14 | "ModelParallelMultiheadAttention", 15 | "ModelParallelTransformerEncoderLayer", 16 | "ModelParallelTransformerDecoderLayer", 17 | ] 18 | -------------------------------------------------------------------------------- /fairseq/models/bart/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .hub_interface import * # noqa 7 | from .model import * # noqa 8 | -------------------------------------------------------------------------------- /fairseq/models/hubert/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .hubert import * # noqa 7 | from .hubert_asr import * # noqa 8 | -------------------------------------------------------------------------------- /fairseq/models/huggingface/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import importlib 7 | import os 8 | 9 | 10 | # automatically import any Python files in the models/huggingface/ directory 11 | models_dir = os.path.dirname(__file__) 12 | for file in os.listdir(models_dir): 13 | path = os.path.join(models_dir, file) 14 | if ( 15 | not file.startswith("_") 16 | and not file.startswith(".") 17 | and (file.endswith(".py") or os.path.isdir(path)) 18 | ): 19 | model_name = file[: file.find(".py")] if file.endswith(".py") else file 20 | module = importlib.import_module("fairseq.models.huggingface." + model_name) 21 | -------------------------------------------------------------------------------- /fairseq/models/nat/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | """isort:skip_file""" 6 | 7 | from .fairseq_nat_model import * 8 | from .nonautoregressive_transformer import * 9 | from .nat_crf_transformer import * 10 | from .iterative_nonautoregressive_transformer import * 11 | from .cmlm_transformer import * 12 | from .levenshtein_transformer import * 13 | from .insertion_transformer import * 14 | -------------------------------------------------------------------------------- /fairseq/models/roberta/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .hub_interface import * # noqa 7 | from .model import * # noqa 8 | from .enc_dec import * # noqa 9 | from .model_camembert import * # noqa 10 | from .model_gottbert import * # noqa 11 | from .model_xlmr import * # noqa 12 | -------------------------------------------------------------------------------- /fairseq/models/speech_to_text/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .berard import * # noqa 7 | from .convtransformer import * # noqa 8 | from .s2t_transformer import * # noqa 9 | from .s2t_transformer_w2v2 import * -------------------------------------------------------------------------------- /fairseq/models/wav2vec/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .wav2vec import * # noqa 7 | from .wav2vec2 import * # noqa 8 | from .wav2vec2_asr import * # noqa 9 | -------------------------------------------------------------------------------- /fairseq/modules/dynamicconv_layer/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .dynamicconv_layer import DynamicconvLayer # noqa 7 | -------------------------------------------------------------------------------- /fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp: -------------------------------------------------------------------------------- 1 | #include 2 | #include 3 | 4 | std::vector dynamicconv_cpu_forward( 5 | float* input, 6 | float* filters, 7 | int padding_l); 8 | 9 | std::vector dynamicconv_cpu_backward( 10 | float* gradOutput, 11 | int padding_l, 12 | float* input, 13 | float* filters); 14 | 15 | std::vector dynamicconv_forward( 16 | float* input, 17 | float* filters, 18 | int padding_l) { 19 | 20 | return dynamicconv_cpu_forward(input, filters, padding_l); 21 | } 22 | 23 | std::vector dynamicconv_backward( 24 | float* gradOutput, 25 | int padding_l, 26 | float* input, 27 | float* filters) { 28 | 29 | return dynamicconv_cpu_backward(gradOutput, padding_l, input, filters); 30 | } 31 | 32 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 33 | m.def("forward", &dynamicconv_forward, "dynamicconv forward (CPU)"); 34 | m.def("backward", &dynamicconv_backward, "dynamicconv backward (CPU)"); 35 | } 36 | -------------------------------------------------------------------------------- /fairseq/modules/dynamicconv_layer/setup.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | from setuptools import setup 8 | from torch.utils.cpp_extension import BuildExtension, CUDAExtension 9 | 10 | 11 | setup( 12 | name="dynamicconv_layer", 13 | ext_modules=[ 14 | CUDAExtension( 15 | name="dynamicconv_cuda", 16 | sources=[ 17 | "dynamicconv_cuda.cpp", 18 | "dynamicconv_cuda_kernel.cu", 19 | ], 20 | ), 21 | ], 22 | cmdclass={"build_ext": BuildExtension}, 23 | ) 24 | -------------------------------------------------------------------------------- /fairseq/modules/fp32_group_norm.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | """ 6 | Layer norm done in fp32 (for fp16 training) 7 | """ 8 | 9 | import torch.nn as nn 10 | import torch.nn.functional as F 11 | 12 | 13 | class Fp32GroupNorm(nn.GroupNorm): 14 | def __init__(self, *args, **kwargs): 15 | super().__init__(*args, **kwargs) 16 | 17 | def forward(self, input): 18 | output = F.group_norm( 19 | input.float(), 20 | self.num_groups, 21 | self.weight.float() if self.weight is not None else None, 22 | self.bias.float() if self.bias is not None else None, 23 | self.eps, 24 | ) 25 | return output.type_as(input) 26 | -------------------------------------------------------------------------------- /fairseq/modules/gelu.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | """ 6 | See "Gaussian Error Linear Units (GELUs)" by Dan Hendrycks and Kevin Gimpel with 7 | the corresponding GitHub repo: https://github.com/hendrycks/GELUs 8 | """ 9 | 10 | import math 11 | 12 | import torch 13 | import torch.nn as nn 14 | 15 | 16 | def gelu_accurate(x): 17 | if not hasattr(gelu_accurate, "_a"): 18 | gelu_accurate._a = math.sqrt(2 / math.pi) 19 | return ( 20 | 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3)))) 21 | ) 22 | 23 | 24 | def gelu(x: torch.Tensor) -> torch.Tensor: 25 | return torch.nn.functional.gelu(x.float()).type_as(x) 26 | -------------------------------------------------------------------------------- /fairseq/modules/grad_multiply.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import torch 7 | 8 | 9 | class GradMultiply(torch.autograd.Function): 10 | @staticmethod 11 | def forward(ctx, x, scale): 12 | ctx.scale = scale 13 | res = x.new(x) 14 | return res 15 | 16 | @staticmethod 17 | def backward(ctx, grad): 18 | return grad * ctx.scale, None 19 | -------------------------------------------------------------------------------- /fairseq/modules/lightconv_layer/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .lightconv_layer import LightconvLayer # noqa 7 | -------------------------------------------------------------------------------- /fairseq/modules/lightconv_layer/setup.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | 7 | from setuptools import setup 8 | from torch.utils.cpp_extension import BuildExtension, CUDAExtension 9 | 10 | 11 | setup( 12 | name="lightconv_layer", 13 | ext_modules=[ 14 | CUDAExtension( 15 | "lightconv_cuda", 16 | [ 17 | "lightconv_cuda.cpp", 18 | "lightconv_cuda_kernel.cu", 19 | ], 20 | ), 21 | ], 22 | cmdclass={"build_ext": BuildExtension}, 23 | ) 24 | -------------------------------------------------------------------------------- /fairseq/modules/quantization/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/fairseq/modules/quantization/__init__.py -------------------------------------------------------------------------------- /fairseq/modules/quantization/pq/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .utils import SizeTracker, quantize_model_ # NOQA 7 | -------------------------------------------------------------------------------- /fairseq/modules/quantization/pq/modules/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .qconv import PQConv2d # NOQA 7 | from .qemb import PQEmbedding # NOQA 8 | from .qlinear import PQLinear # NOQA 9 | -------------------------------------------------------------------------------- /fairseq/modules/quantization/scalar/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .utils import quantize_model_ # NOQA 7 | -------------------------------------------------------------------------------- /fairseq/modules/quantization/scalar/modules/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .qact import ActivationQuantizer # NOQA 7 | from .qconv import IntConv2d # NOQA 8 | from .qemb import IntEmbedding # NOQA 9 | from .qlinear import IntLinear # NOQA 10 | -------------------------------------------------------------------------------- /fairseq/modules/same_pad.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | 7 | from torch import nn 8 | 9 | 10 | class SamePad(nn.Module): 11 | def __init__(self, kernel_size, causal=False): 12 | super().__init__() 13 | if causal: 14 | self.remove = kernel_size - 1 15 | else: 16 | self.remove = 1 if kernel_size % 2 == 0 else 0 17 | 18 | def forward(self, x): 19 | if self.remove > 0: 20 | x = x[:, :, : -self.remove] 21 | return x 22 | -------------------------------------------------------------------------------- /fairseq/modules/scalar_bias.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | # 6 | 7 | import torch 8 | 9 | 10 | class ScalarBias(torch.autograd.Function): 11 | """ 12 | Adds a vector of scalars, used in self-attention mechanism to allow 13 | the model to optionally attend to this vector instead of the past 14 | """ 15 | 16 | @staticmethod 17 | def forward(ctx, input, dim, bias_init): 18 | size = list(input.size()) 19 | size[dim] += 1 20 | output = input.new(*size).fill_(bias_init) 21 | output.narrow(dim, 1, size[dim] - 1).copy_(input) 22 | ctx.dim = dim 23 | return output 24 | 25 | @staticmethod 26 | def backward(ctx, grad): 27 | return grad.narrow(ctx.dim, 1, grad.size(ctx.dim) - 1), None, None 28 | 29 | 30 | def scalar_bias(input, dim, bias_init=0): 31 | return ScalarBias.apply(input, dim, bias_init) 32 | -------------------------------------------------------------------------------- /fairseq/modules/transpose_last.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | """ 6 | transpose last 2 dimensions of the input 7 | """ 8 | 9 | import torch.nn as nn 10 | 11 | 12 | class TransposeLast(nn.Module): 13 | def __init__(self, deconstruct_idx=None): 14 | super().__init__() 15 | self.deconstruct_idx = deconstruct_idx 16 | 17 | def forward(self, x): 18 | if self.deconstruct_idx is not None: 19 | x = x[self.deconstruct_idx] 20 | return x.transpose(-2, -1) 21 | -------------------------------------------------------------------------------- /fairseq/modules/unfold.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import torch.nn.functional as F 7 | 8 | 9 | def unfold1d(x, kernel_size, padding_l, pad_value=0): 10 | """unfold T x B x C to T x B x C x K""" 11 | if kernel_size > 1: 12 | T, B, C = x.size() 13 | x = F.pad( 14 | x, (0, 0, 0, 0, padding_l, kernel_size - 1 - padding_l), value=pad_value 15 | ) 16 | x = x.as_strided((T, B, C, kernel_size), (B * C, C, 1, B * C)) 17 | else: 18 | x = x.unsqueeze(3) 19 | return x 20 | -------------------------------------------------------------------------------- /fairseq/optim/lr_scheduler/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | """isort:skip_file""" 6 | 7 | import importlib 8 | import os 9 | 10 | from fairseq import registry 11 | from fairseq.optim.lr_scheduler.fairseq_lr_scheduler import ( # noqa 12 | FairseqLRScheduler, 13 | LegacyFairseqLRScheduler, 14 | ) 15 | from omegaconf import DictConfig 16 | 17 | 18 | ( 19 | build_lr_scheduler_, 20 | register_lr_scheduler, 21 | LR_SCHEDULER_REGISTRY, 22 | LR_SCHEDULER_DATACLASS_REGISTRY, 23 | ) = registry.setup_registry( 24 | "--lr-scheduler", base_class=FairseqLRScheduler, default="fixed" 25 | ) 26 | 27 | 28 | def build_lr_scheduler(cfg: DictConfig, optimizer): 29 | return build_lr_scheduler_(cfg, optimizer) 30 | 31 | 32 | # automatically import any Python files in the optim/lr_scheduler/ directory 33 | for file in sorted(os.listdir(os.path.dirname(__file__))): 34 | if file.endswith(".py") and not file.startswith("_"): 35 | file_name = file[: file.find(".py")] 36 | importlib.import_module("fairseq.optim.lr_scheduler." + file_name) 37 | -------------------------------------------------------------------------------- /fairseq/pdb.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import multiprocessing 7 | import os 8 | import pdb 9 | import sys 10 | 11 | 12 | __all__ = ["set_trace"] 13 | 14 | 15 | _stdin = [None] 16 | _stdin_lock = multiprocessing.Lock() 17 | try: 18 | _stdin_fd = sys.stdin.fileno() 19 | except Exception: 20 | _stdin_fd = None 21 | 22 | 23 | class MultiprocessingPdb(pdb.Pdb): 24 | """A Pdb wrapper that works in a multiprocessing environment. 25 | 26 | Usage: `from fairseq import pdb; pdb.set_trace()` 27 | """ 28 | 29 | def __init__(self): 30 | pdb.Pdb.__init__(self, nosigint=True) 31 | 32 | def _cmdloop(self): 33 | stdin_bak = sys.stdin 34 | with _stdin_lock: 35 | try: 36 | if _stdin_fd is not None: 37 | if not _stdin[0]: 38 | _stdin[0] = os.fdopen(_stdin_fd) 39 | sys.stdin = _stdin[0] 40 | self.cmdloop() 41 | finally: 42 | sys.stdin = stdin_bak 43 | 44 | 45 | def set_trace(): 46 | pdb = MultiprocessingPdb() 47 | pdb.set_trace(sys._getframe().f_back) 48 | -------------------------------------------------------------------------------- /fairseq/scoring/chrf.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from fairseq.scoring import BaseScorer, register_scorer 7 | 8 | 9 | @register_scorer("chrf") 10 | class ChrFScorer(BaseScorer): 11 | def __init__(self, args): 12 | super(ChrFScorer, self).__init__(args) 13 | import sacrebleu 14 | 15 | self.sacrebleu = sacrebleu 16 | 17 | def add_string(self, ref, pred): 18 | self.ref.append(ref) 19 | self.pred.append(pred) 20 | 21 | def score(self, order=4): 22 | return self.result_string(order).score 23 | 24 | def result_string(self, order=4): 25 | if order != 4: 26 | raise NotImplementedError 27 | return self.sacrebleu.corpus_chrf(self.pred, [self.ref]).format() 28 | -------------------------------------------------------------------------------- /fairseq/tokenizer.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import re 7 | 8 | 9 | SPACE_NORMALIZER = re.compile(r"\s+") 10 | 11 | 12 | def tokenize_line(line): 13 | line = SPACE_NORMALIZER.sub(" ", line) 14 | line = line.strip() 15 | return line.split() 16 | -------------------------------------------------------------------------------- /fairseq/version.txt: -------------------------------------------------------------------------------- 1 | 1.0.0a0 2 | -------------------------------------------------------------------------------- /fairseq_cli/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/fairseq_cli/__init__.py -------------------------------------------------------------------------------- /fbank_scripts/pretrain_asr.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_asr 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config.yaml --train-subset train --valid-subset dev \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 40000 --max-update 100000 \ 6 | --task speech_to_text --s2t-task asr --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ 7 | --arch s2t_transformer_s_12aenc_0tenc_6dec --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 2 \ 11 | --patience 10 | tee logs/${exp}.txt -------------------------------------------------------------------------------- /fbank_scripts/pretrain_asr_base.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_asr_base 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config.yaml --train-subset train --valid-subset dev \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 40000 --max-update 100000 \ 6 | --task speech_to_text --s2t-task asr --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ 7 | --arch s2t_transformer_b_12aenc_0tenc_6dec --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 1 \ 11 | --patience 10 | tee logs/${exp}.txt -------------------------------------------------------------------------------- /fbank_scripts/pretrain_mt.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_mt 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config.yaml --train-subset train --valid-subset dev \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 4096 --max-update 100000 \ 6 | --task speech_to_text --s2t-task mt --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ 7 | --arch s2t_transformer_s_0aenc_6tenc_6dec --optimizer adam --lr 1e-3 --lr-scheduler inverse_sqrt \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 8000 --clip-norm 10.0 --seed 1 --update-freq 2 \ 11 | --layernorm-embedding \ 12 | --patience 10 | tee logs/${exp}.txt -------------------------------------------------------------------------------- /fbank_scripts/train_baseline.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_fbank 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config.yaml --train-subset train --valid-subset dev \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 40000 --max-update 100000 \ 6 | --task speech_to_text --s2t-task stack --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ 7 | --arch s2t_transformer_s_12aenc_6tenc_6dec --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 2 \ 11 | --patience 10 \ 12 | --max-epoch 25 \ 13 | --load-pretrained-asr-encoder-from checkpoints/mustc_en${lang}_asr.pt \ 14 | --load-pretrained-mt-encoder-decoder-from checkpoints/mustc_en${lang}_mt.pt | tee logs/${exp}.txt -------------------------------------------------------------------------------- /fbank_scripts/train_baseline_base.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_fbank_base 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config.yaml --train-subset train --valid-subset dev \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 40000 --max-update 100000 \ 6 | --task speech_to_text --s2t-task stack --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ 7 | --arch s2t_transformer_b_12aenc_6tenc_6dec --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 2 \ 11 | --patience 10 \ 12 | --max-epoch 25 \ 13 | --load-pretrained-asr-encoder-from checkpoints/mustc_en${lang}_asr_base.pt \ 14 | --load-pretrained-mt-encoder-decoder-from checkpoints/mustc_en${lang}_mt_base.pt | tee logs/${exp}.txt -------------------------------------------------------------------------------- /fbank_scripts/train_stmm.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_stmm_self_learning_fbank 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config.yaml --train-subset train --valid-subset dev \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 40000 --max-update 100000 \ 6 | --task speech_to_text --s2t-task stack --criterion label_smoothed_cross_entropy_with_stmm_self_learning --label-smoothing 0.1 --report-accuracy \ 7 | --arch s2t_transformer_s_12aenc_6tenc_6dec --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 2 \ 11 | --patience 10 \ 12 | --max-epoch 25 \ 13 | --load-pretrained-asr-encoder-from checkpoints/mustc_en${lang}_asr.pt \ 14 | --load-pretrained-mt-encoder-decoder-from checkpoints/mustc_en${lang}_mt.pt \ 15 | --mixup --mixup-arguments fix,100000,1.0 | tee logs/${exp}.txt -------------------------------------------------------------------------------- /fbank_scripts/train_stmm_base.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_stmm_self_learning_fbank_base 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config.yaml --train-subset train --valid-subset dev \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 20000 --max-update 100000 \ 6 | --task speech_to_text --s2t-task stack --criterion label_smoothed_cross_entropy_with_stmm_self_learning --label-smoothing 0.1 --report-accuracy \ 7 | --arch s2t_transformer_b_12aenc_6tenc_6dec --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 4 \ 11 | --patience 10 \ 12 | --max-epoch 25 \ 13 | --load-pretrained-asr-encoder-from checkpoints/mustc_en${lang}_asr_base.pt \ 14 | --load-pretrained-mt-encoder-decoder-from checkpoints/mustc_en${lang}_mt_base.pt \ 15 | --mixup --mixup-arguments fix,100000,1.0 | tee logs/${exp}.txt -------------------------------------------------------------------------------- /fbank_scripts/train_stmm_base_static.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_stmm_self_learning_fbank_base_static 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config.yaml --train-subset train --valid-subset dev \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 20000 --max-update 100000 \ 6 | --task speech_to_text --s2t-task stack --criterion label_smoothed_cross_entropy_with_stmm_self_learning --label-smoothing 0.1 --report-accuracy \ 7 | --arch s2t_transformer_b_12aenc_6tenc_6dec --optimizer adam --lr 2e-3 --lr-scheduler inverse_sqrt \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 10000 --clip-norm 10.0 --seed 1 --update-freq 2 \ 11 | --patience 10 \ 12 | --max-epoch 25 \ 13 | --load-pretrained-asr-encoder-from checkpoints/mustc_en${lang}_asr_base.pt \ 14 | --load-pretrained-mt-encoder-decoder-from checkpoints/mustc_en${lang}_mt_base.pt \ 15 | --mixup --mixup-arguments fix,100000,1.0 | tee logs/${exp}.txt -------------------------------------------------------------------------------- /preprocess_scripts/group.py: -------------------------------------------------------------------------------- 1 | import os 2 | import shutil 3 | import argparse 4 | 5 | parser = argparse.ArgumentParser() 6 | parser.add_argument("--lang", help="target language") 7 | args = parser.parse_args() 8 | 9 | splits = ['dev', 'tst-COMMON', 'tst-HE', 'train'] 10 | root = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) 11 | seg_path = os.path.join(root, 'data', 'mustc', f'en-{args.lang}', 'segment') 12 | 13 | for split in splits: 14 | split_path = os.path.join(seg_path, split) 15 | for f in os.listdir(split_path): 16 | if f.startswith('ted'): 17 | speaker = f.split('_')[1] 18 | speaker_dir = os.path.join(split_path, speaker) 19 | os.makedirs(speaker_dir, exist_ok=True) 20 | shutil.move(os.path.join(split_path, f), speaker_dir) -------------------------------------------------------------------------------- /pretrain_mt.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_mt 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config_raw.yaml --train-subset train_raw --valid-subset dev_raw \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 4096 --max-update 100000 \ 6 | --task speech_to_text --s2t-task mt --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --report-accuracy \ 7 | --arch s2t_transformer_b_0aenc_6tenc_6dec --optimizer adam --lr 1e-3 --lr-scheduler inverse_sqrt \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 8000 --clip-norm 10.0 --seed 1 --update-freq 1 \ 11 | --layernorm-embedding \ 12 | --patience 10 \ 13 | --load-pretrained-mt-encoder-decoder-from checkpoints/en${lang}_mt.pt | tee logs/${exp}.txt -------------------------------------------------------------------------------- /pretrain_mt_ext.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=en${lang}_mt 3 | fairseq-train data/ext_en${lang}/binary --task translation \ 4 | --arch transformer_wmt_en_de --share-all-embeddings --dropout 0.1 \ 5 | --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \ 6 | --lr-scheduler inverse_sqrt --warmup-init-lr 1e-07 --warmup-updates 4000 \ 7 | --lr 0.0007 --stop-min-lr 1e-09 \ 8 | --criterion label_smoothed_cross_entropy --label-smoothing 0.1 --weight-decay 0.0 \ 9 | --max-tokens 4096 \ 10 | --update-freq 1 --no-progress-bar --log-format json --log-interval 100 \ 11 | --layernorm-embedding \ 12 | --keep-last-epochs 20 \ 13 | --save-dir checkpoints/$exp \ 14 | --ddp-backend=no_c10d \ 15 | --max-update 250000 | tee logs/$exp.txt -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [build-system] 2 | requires = ["setuptools", "wheel", "cython"] 3 | build-backend = "setuptools.build_meta" 4 | -------------------------------------------------------------------------------- /scripts/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/scripts/__init__.py -------------------------------------------------------------------------------- /scripts/compare_namespaces.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | """Helper script to compare two argparse.Namespace objects.""" 3 | 4 | from argparse import Namespace # noqa 5 | 6 | 7 | def main(): 8 | 9 | ns1 = eval(input("Namespace 1: ")) 10 | ns2 = eval(input("Namespace 2: ")) 11 | 12 | def keys(ns): 13 | ks = set() 14 | for k in dir(ns): 15 | if not k.startswith("_"): 16 | ks.add(k) 17 | return ks 18 | 19 | k1 = keys(ns1) 20 | k2 = keys(ns2) 21 | 22 | def print_keys(ks, ns1, ns2=None): 23 | for k in ks: 24 | if ns2 is None: 25 | print("{}\t{}".format(k, getattr(ns1, k, None))) 26 | else: 27 | print( 28 | "{}\t{}\t{}".format(k, getattr(ns1, k, None), getattr(ns2, k, None)) 29 | ) 30 | 31 | print("Keys unique to namespace 1:") 32 | print_keys(k1 - k2, ns1) 33 | print() 34 | 35 | print("Keys unique to namespace 2:") 36 | print_keys(k2 - k1, ns2) 37 | print() 38 | 39 | print("Overlapping keys with different values:") 40 | ks = [k for k in k1 & k2 if getattr(ns1, k, "None") != getattr(ns2, k, "None")] 41 | print_keys(ks, ns1, ns2) 42 | print() 43 | 44 | 45 | if __name__ == "__main__": 46 | main() 47 | -------------------------------------------------------------------------------- /scripts/compound_split_bleu.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | if [ $# -ne 1 ]; then 4 | echo "usage: $0 GENERATE_PY_OUTPUT" 5 | exit 1 6 | fi 7 | 8 | GEN=$1 9 | 10 | SYS=$GEN.sys 11 | REF=$GEN.ref 12 | 13 | if [ $(tail -n 1 $GEN | grep BLEU | wc -l) -ne 1 ]; then 14 | echo "not done generating" 15 | exit 16 | fi 17 | 18 | grep ^H $GEN | awk -F '\t' '{print $NF}' | perl -ple 's{(\S)-(\S)}{$1 ##AT##-##AT## $2}g' > $SYS 19 | grep ^T $GEN | cut -f2- | perl -ple 's{(\S)-(\S)}{$1 ##AT##-##AT## $2}g' > $REF 20 | fairseq-score --sys $SYS --ref $REF 21 | -------------------------------------------------------------------------------- /scripts/constraints/validate.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 2 | # 3 | # Copyright (c) Facebook, Inc. and its affiliates. 4 | # 5 | # This source code is licensed under the MIT license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | import sys 9 | 10 | 11 | """Reads in a fairseq output file, and verifies that the constraints 12 | (C- lines) are present in the output (the first H- line). Assumes that 13 | constraints are listed prior to the first hypothesis. 14 | """ 15 | 16 | constraints = [] 17 | found = 0 18 | total = 0 19 | for line in sys.stdin: 20 | if line.startswith("C-"): 21 | constraints.append(line.rstrip().split("\t")[1]) 22 | elif line.startswith("H-"): 23 | text = line.split("\t")[2] 24 | 25 | for constraint in constraints: 26 | total += 1 27 | if constraint in text: 28 | found += 1 29 | else: 30 | print(f"No {constraint} in {text}", file=sys.stderr) 31 | 32 | constraints = [] 33 | 34 | print(f"Found {found} / {total} = {100 * found / total:.1f}%") 35 | -------------------------------------------------------------------------------- /scripts/convert_dictionary.lua: -------------------------------------------------------------------------------- 1 | -- Copyright (c) Facebook, Inc. and its affiliates. 2 | -- 3 | -- This source code is licensed under the MIT license found in the 4 | -- LICENSE file in the root directory of this source tree. 5 | -- 6 | -- Usage: convert_dictionary.lua 7 | require 'fairseq' 8 | require 'torch' 9 | require 'paths' 10 | 11 | if #arg < 1 then 12 | print('usage: convert_dictionary.lua ') 13 | os.exit(1) 14 | end 15 | if not paths.filep(arg[1]) then 16 | print('error: file does not exit: ' .. arg[1]) 17 | os.exit(1) 18 | end 19 | 20 | dict = torch.load(arg[1]) 21 | dst = paths.basename(arg[1]):gsub('.th7', '.txt') 22 | assert(dst:match('.txt$')) 23 | 24 | f = io.open(dst, 'w') 25 | for idx, symbol in ipairs(dict.index_to_symbol) do 26 | if idx > dict.cutoff then 27 | break 28 | end 29 | f:write(symbol) 30 | f:write(' ') 31 | f:write(dict.index_to_freq[idx]) 32 | f:write('\n') 33 | end 34 | f:close() 35 | -------------------------------------------------------------------------------- /scripts/sacrebleu.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | if [ $# -ne 4 ]; then 4 | echo "usage: $0 TESTSET SRCLANG TGTLANG GEN" 5 | exit 1 6 | fi 7 | 8 | TESTSET=$1 9 | SRCLANG=$2 10 | TGTLANG=$3 11 | 12 | GEN=$4 13 | 14 | if ! command -v sacremoses &> /dev/null 15 | then 16 | echo "sacremoses could not be found, please install with: pip install sacremoses" 17 | exit 18 | fi 19 | 20 | grep ^H $GEN \ 21 | | sed 's/^H\-//' \ 22 | | sort -n -k 1 \ 23 | | cut -f 3 \ 24 | | sacremoses detokenize \ 25 | > $GEN.sorted.detok 26 | 27 | sacrebleu --test-set $TESTSET --language-pair "${SRCLANG}-${TGTLANG}" < $GEN.sorted.detok 28 | -------------------------------------------------------------------------------- /scripts/spm_train.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # All rights reserved. 4 | # 5 | # This source code is licensed under the license found in the 6 | # LICENSE file in the root directory of this source tree. 7 | 8 | from __future__ import absolute_import, division, print_function, unicode_literals 9 | 10 | import sys 11 | 12 | import sentencepiece as spm 13 | 14 | 15 | if __name__ == "__main__": 16 | spm.SentencePieceTrainer.Train(" ".join(sys.argv[1:])) 17 | -------------------------------------------------------------------------------- /scripts/test_fsdp.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | rm -rf fsdp_dummy 3 | mkdir -p fsdp_dummy 4 | CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train /private/home/sshleifer/data-bin/stories_mmap \ 5 | --ddp-backend fully_sharded --fp16 --fp16-init-scale 4 \ 6 | --cpu-offload --checkpoint-activations \ 7 | --task language_modeling --tokens-per-sample 256 --batch-size 8 \ 8 | --arch transformer_lm_gpt2_tiny \ 9 | --optimizer cpu_adam --adam-betas "(0.9,0.98)" \ 10 | --lr 0.0001 --lr-scheduler polynomial_decay --warmup-updates 5 --total-num-update 10 \ 11 | --max-update 5 --log-format json --log-interval 1 \ 12 | --save-interval-updates 5 --save-dir fsdp_dummy --disable-validation \ 13 | --restore-file x.pt "$@" 14 | 15 | # Now we try to load the checkpoint 16 | CUDA_VISIBLE_DEVICES=0,1 fairseq-train /private/home/sshleifer/data-bin/stories_mmap \ 17 | --ddp-backend fully_sharded --fp16 --fp16-init-scale 4 \ 18 | --cpu-offload --checkpoint-activations \ 19 | --task language_modeling --tokens-per-sample 256 --batch-size 8 \ 20 | --arch transformer_lm_gpt2_tiny \ 21 | --optimizer cpu_adam --adam-betas "(0.9,0.98)" \ 22 | --lr 0.0001 --lr-scheduler polynomial_decay --warmup-updates 5 --total-num-update 10 \ 23 | --max-update 2 --log-format json --log-interval 1 \ 24 | --save-interval-updates 2 --save-dir fsdp_dummy 25 | -------------------------------------------------------------------------------- /test.sh: -------------------------------------------------------------------------------- 1 | exp=$1 2 | lang=$2 3 | CHECKPOINT_FILENAME=avg_last_10_checkpoint.pt 4 | python3 scripts/average_checkpoints.py \ 5 | --inputs checkpoints/${exp} --num-epoch-checkpoints 10 \ 6 | --output "checkpoints/${exp}/${CHECKPOINT_FILENAME}" 7 | fairseq-generate data/mustc/en-${lang} \ 8 | --config-yaml config_raw.yaml --gen-subset tst-COMMON_raw --task speech_to_text --s2t-task stack \ 9 | --path checkpoints/${exp}/${CHECKPOINT_FILENAME} \ 10 | --max-audio-positions 900000 \ 11 | --max-tokens 2000000 --beam 5 --scoring sacrebleu | tee result/${exp}.txt -------------------------------------------------------------------------------- /tests/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/tests/__init__.py -------------------------------------------------------------------------------- /tests/distributed/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/tests/distributed/__init__.py -------------------------------------------------------------------------------- /tests/gpu/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/tests/gpu/__init__.py -------------------------------------------------------------------------------- /tests/gpu/transformer_quantization_config.yaml: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | # This file defines example configuration arguments for quantizing 7 | # a transformer model with product quantization 8 | 9 | n_centroids: 10 | Linear: 11 | key: in_features 12 | value: {"*": 8} 13 | Embedding: 14 | key: embedding_dim 15 | value: {"*": 8} 16 | 17 | block_sizes: 18 | Linear: 19 | key: fuzzy_name 20 | value: {fc: 8, attn: 4, emb: 4} 21 | Embedding: 22 | key: fuzzy_name 23 | value: {emb: 8} 24 | 25 | layers_to_quantize: 26 | - decoder\\.layers\\.\d+\\.fc[12] 27 | - decoder\\.embed_tokens\\.embeddings\\.[012]\\.[01] 28 | - decoder\\.layers\\.\d+\\.self_attn\\.(k_proj|v_proj|q_proj|out_proj) 29 | -------------------------------------------------------------------------------- /tests/speech_recognition/__init__.py: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ictnlp/STEMM/f6583ec5e85bb005b0873fd011f7b0dcc06da649/tests/speech_recognition/__init__.py -------------------------------------------------------------------------------- /tests/test_iopath.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import unittest 7 | from unittest import mock 8 | 9 | 10 | class TestIOPath(unittest.TestCase): 11 | 12 | def test_no_iopath(self): 13 | from .test_reproducibility import TestReproducibility 14 | 15 | with mock.patch.dict("sys.modules", {"iopath": None}): 16 | # reuse reproducibility tests, which are e2e tests that should cover 17 | # most checkpoint related functionality 18 | TestReproducibility._test_reproducibility(self, "test_reproducibility") 19 | 20 | def test_no_supports_rename(self): 21 | from .test_reproducibility import TestReproducibility 22 | 23 | with mock.patch("fairseq.file_io.PathManager.supports_rename") as mock_fn: 24 | mock_fn.return_value = False 25 | TestReproducibility._test_reproducibility(self, "test_reproducibility") 26 | 27 | 28 | if __name__ == "__main__": 29 | unittest.main() 30 | -------------------------------------------------------------------------------- /train.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python3 -u 2 | # Copyright (c) Facebook, Inc. and its affiliates. 3 | # 4 | # This source code is licensed under the MIT license found in the 5 | # LICENSE file in the root directory of this source tree. 6 | """ 7 | Legacy entry point. Use fairseq_cli/train.py or fairseq-train instead. 8 | """ 9 | 10 | from fairseq_cli.train import cli_main 11 | 12 | 13 | if __name__ == "__main__": 14 | cli_main() 15 | -------------------------------------------------------------------------------- /train.sh: -------------------------------------------------------------------------------- 1 | lang=$1 2 | exp=mustc_en${lang}_stmm_self_learning 3 | fairseq-train data/mustc/en-${lang} \ 4 | --config-yaml config_raw.yaml --train-subset train_raw --valid-subset dev_raw \ 5 | --save-dir checkpoints/${exp} --num-workers 4 --max-tokens 2000000 --max-update 100000 \ 6 | --task speech_to_text --s2t-task stack --criterion label_smoothed_cross_entropy_with_stmm_self_learning --label-smoothing 0.1 \ 7 | --arch s2t_transformer_b_w2v_6tenc_6dec --optimizer adam --adam-betas '(0.9, 0.98)' --lr 1e-4 --lr-scheduler inverse_sqrt --weight-decay 0.0001 \ 8 | --no-progress-bar --log-format json --log-interval 100 \ 9 | --ddp-backend=legacy_ddp \ 10 | --warmup-updates 4000 --clip-norm 0.0 --seed 1 --update-freq 1 \ 11 | --layernorm-embedding \ 12 | --patience 10 \ 13 | --max-epoch 25 \ 14 | --max-audio-positions 900000 \ 15 | --fp16 \ 16 | --w2v2-model-path checkpoints/wav2vec_small.pt \ 17 | --load-pretrained-mt-encoder-decoder-from checkpoints/mustc_en${lang}_mt.pt \ 18 | --mixup --mixup-arguments fix,100000,1.0 | tee logs/${exp}.txt --------------------------------------------------------------------------------