├── tests ├── __init__.py ├── gpu │ ├── __init__.py │ └── transformer_quantization_config.yaml ├── distributed │ └── __init__.py ├── speech_recognition │ └── __init__.py ├── speech │ ├── test_s2t_transformer.py │ ├── test_s2t_conformer.py │ └── test_xm_transformer.py ├── test_hf_hub.py └── test_iopath.py ├── fairseq_cli └── __init__.py ├── scripts ├── __init__.py ├── spm_train.py ├── compound_split_bleu.sh ├── sacrebleu.sh ├── convert_dictionary.lua └── constraints │ └── validate.py ├── examples ├── wav2vec │ ├── __init__.py │ └── unsupervised │ │ ├── __init__.py │ │ ├── kaldi_self_train │ │ └── st │ │ │ ├── steps │ │ │ ├── utils │ │ │ ├── local │ │ │ ├── copy_aligned_text.py │ │ │ ├── decode.sh │ │ │ ├── prepare_lm.sh │ │ │ ├── prepare_lang_word.sh │ │ │ ├── prepare_lang.sh │ │ │ ├── unsup_select_decode.sh │ │ │ └── unsup_select_decode_word.sh │ │ │ ├── path.sh │ │ │ ├── cmd.sh │ │ │ └── decode_word_step2.sh │ │ ├── models │ │ └── __init__.py │ │ ├── tasks │ │ └── __init__.py │ │ ├── scripts │ │ ├── copy_labels.py │ │ ├── ltr_to_wrd.py │ │ ├── wrd_to_ltr.py │ │ ├── normalize_text.py │ │ ├── filter_tsv.py │ │ └── filter_lexicon.py │ │ ├── config │ │ └── generate │ │ │ └── viterbi.yaml │ │ └── data │ │ └── __init__.py ├── speech_matrix │ ├── __init__.py │ ├── data_helper │ │ ├── __init__.py │ │ ├── config_multitask.yaml │ │ └── config.yaml │ ├── mined_train_sets │ │ ├── __init__.py │ │ └── download_speech_to_unit.py │ ├── speech_to_speech │ │ └── __init__.py │ └── valid_test_sets │ │ └── __init__.py ├── .gitignore ├── speech_recognition │ ├── new │ │ ├── __init__.py │ │ ├── decoders │ │ │ ├── __init__.py │ │ │ ├── viterbi_decoder.py │ │ │ └── decoder.py │ │ └── conf │ │ │ ├── infer.yaml │ │ │ └── hydra │ │ │ └── sweeper │ │ │ └── ax.yaml │ ├── kaldi │ │ ├── __init__.py │ │ └── config │ │ │ └── kaldi_initializer.yaml │ ├── __init__.py │ ├── data │ │ └── __init__.py │ ├── tasks │ │ └── __init__.py │ ├── models │ │ └── __init__.py │ └── criterions │ │ └── __init__.py ├── hubert │ ├── tests │ │ ├── sample.base.L9.len │ │ ├── sample.large.L20.len │ │ ├── sample.xlarge.L30.len │ │ ├── 6313-76958-0021.flac │ │ ├── sample.base.L9.npy │ │ ├── sample.large.L20.npy │ │ ├── sample.xlarge.L30.npy │ │ ├── sample.large.hypo.word │ │ └── sample.xlarge.hypo.word │ ├── config │ │ ├── pretrain │ │ │ ├── data │ │ │ │ ├── iter1.yaml │ │ │ │ └── iter2.yaml │ │ │ └── run │ │ │ │ └── submitit_reg.yaml │ │ ├── finetune │ │ │ ├── ckpt │ │ │ │ └── it1.yaml │ │ │ ├── lm │ │ │ │ └── ls_4gram.yaml │ │ │ └── run │ │ │ │ └── submitit_reg.yaml │ │ └── decode │ │ │ ├── run │ │ │ ├── submitit_slurm.yaml │ │ │ └── submitit_slurm_8gpu.yaml │ │ │ ├── infer_viterbi.yaml │ │ │ ├── infer_kenlm.yaml │ │ │ ├── infer_fsqlm.yaml │ │ │ └── ax_sweep │ │ │ ├── ngram.yaml │ │ │ └── transformer.yaml │ └── update_ckpt.py ├── speech_to_speech │ ├── asr_bleu │ │ ├── __init__.py │ │ └── requirements.txt │ ├── __init__.py │ ├── preprocessing │ │ └── __init__.py │ ├── enhance │ │ └── config.yaml │ ├── benchmarking │ │ └── configs │ │ │ ├── S2T.yaml │ │ │ ├── 2StageS2ST.yaml │ │ │ ├── DirectS2U.yaml │ │ │ └── 3StageS2ST.yaml │ └── unity │ │ └── __init__.py ├── attention_head_selection │ └── src │ │ ├── __init__.py │ │ ├── data │ │ └── __init__.py │ │ ├── loss │ │ ├── __init__.py │ │ └── attention_head_selection.py │ │ ├── models │ │ └── __init__.py │ │ └── modules │ │ └── __init__.py ├── linformer │ ├── linformer_src │ │ ├── models │ │ │ └── __init__.py │ │ ├── modules │ │ │ └── __init__.py │ │ └── __init__.py │ └── README.md ├── textless_nlp │ ├── gslm │ │ ├── speech2unit │ │ │ ├── __init__.py │ │ │ ├── clustering │ │ │ │ ├── __init__.py │ │ │ │ └── utils.py │ │ │ └── pretrained │ │ │ │ └── logmel_feature_reader.py │ │ ├── unit2speech │ │ │ ├── tacotron2 │ │ │ │ ├── __init__.py │ │ │ │ └── symbols.py │ │ │ └── multiproc.py │ │ └── metrics │ │ │ ├── asr_metrics │ │ │ └── misc │ │ │ │ └── dict.ltr.txt │ │ │ └── README.md │ └── speech-resynth │ │ └── img │ │ └── fig.png ├── latent_depth │ └── latent_depth_src │ │ ├── loss │ │ └── __init__.py │ │ ├── models │ │ └── __init__.py │ │ ├── modules │ │ └── __init__.py │ │ └── __init__.py ├── multilingual │ ├── data_scripts │ │ ├── requirement.txt │ │ ├── utils │ │ │ └── strip_sgm.sh │ │ ├── README.md │ │ ├── preprocess_ML50_v1.sh │ │ └── download_ML50_v1.sh │ ├── ML50_langs.txt │ └── multilingual_fairseq_gen.sh ├── adaptive_span │ ├── truncated_bptt_lm_task.py │ └── __init__.py ├── MMPT │ ├── vlm.png │ ├── videoclip.png │ ├── projects │ │ ├── mtm │ │ │ ├── vlm.yaml │ │ │ ├── mmfusionmtm.yaml │ │ │ └── vlm │ │ │ │ ├── test_vtt.yaml │ │ │ │ ├── test_vttqa.yaml │ │ │ │ ├── test_youcook.yaml │ │ │ │ ├── test_coin.yaml │ │ │ │ └── test_youcookcap.yaml │ │ ├── task │ │ │ ├── test_vtt_videoclip.yaml │ │ │ ├── test_vttqa_videoclip.yaml │ │ │ ├── test_youcook_videoclip.yaml │ │ │ ├── coin_videoclip.yaml │ │ │ ├── test_coin_videoclip.yaml │ │ │ ├── test_crosstask_videoclip.yaml │ │ │ ├── test_crosstask_zs_videoclip.yaml │ │ │ ├── youcook_videoclip.yaml │ │ │ ├── vttqa_videoclip.yaml │ │ │ ├── test.yaml │ │ │ ├── vtt_videoclip.yaml │ │ │ ├── crosstask_videoclip.yaml │ │ │ ├── test_coin_zs.yaml │ │ │ ├── test_vtt_zs.yaml │ │ │ ├── test_vttqa_zs.yaml │ │ │ ├── test_youcook_zs.yaml │ │ │ ├── ft.yaml │ │ │ ├── test_vtt.yaml │ │ │ ├── test_vttqa.yaml │ │ │ ├── vttqa.yaml │ │ │ ├── default.yaml │ │ │ ├── youcookcap.yaml │ │ │ ├── how2.yaml │ │ │ ├── test_didemo_zs.yaml │ │ │ ├── coin.yaml │ │ │ ├── vtt.yaml │ │ │ ├── test_youcookcap.yaml │ │ │ ├── test_coin.yaml │ │ │ ├── youcook.yaml │ │ │ └── test_youcook.yaml │ │ └── retri │ │ │ ├── videoclip.yaml │ │ │ └── videoclip │ │ │ ├── test_vttqa_zs.yaml │ │ │ ├── test_didemo_zs.yaml │ │ │ ├── test_vtt_zs.yaml │ │ │ ├── test_vttqa_videoclip.yaml │ │ │ ├── test_vtt_videoclip.yaml │ │ │ ├── test_coin_zs.yaml │ │ │ ├── test_youcook_zs.yaml │ │ │ ├── test_youcook_videoclip.yaml │ │ │ └── test_coin_videoclip.yaml │ ├── scripts │ │ ├── text_token_extractor │ │ │ └── configs │ │ │ │ └── bert-base-uncased.yaml │ │ └── video_feature_extractor │ │ │ ├── how2 │ │ │ └── s3d.sh │ │ │ └── random_sequence_shuffler.py │ ├── mmpt │ │ ├── modules │ │ │ └── __init__.py │ │ ├── datasets │ │ │ └── __init__.py │ │ ├── evaluators │ │ │ └── __init__.py │ │ ├── losses │ │ │ └── __init__.py │ │ ├── __init__.py │ │ ├── models │ │ │ └── __init__.py │ │ ├── tasks │ │ │ ├── __init__.py │ │ │ ├── vlmtask.py │ │ │ └── milncetask.py │ │ └── processors │ │ │ └── __init__.py │ └── setup.py ├── discriminative_reranking_nmt │ ├── __init__.py │ ├── models │ │ └── __init__.py │ ├── tasks │ │ └── __init__.py │ └── criterions │ │ └── __init__.py ├── flores101 │ └── flores_logo.png ├── speech_synthesis │ ├── __init__.py │ ├── evaluation │ │ └── __init__.py │ └── preprocessing │ │ ├── __init__.py │ │ └── denoiser │ │ └── __init__.py ├── rxf │ ├── __init__.py │ └── rxf_src │ │ └── __init__.py ├── noisychannel │ └── __init__.py ├── simultaneous_translation │ ├── __init__.py │ ├── README.md │ ├── models │ │ └── __init__.py │ ├── utils │ │ └── __init__.py │ └── modules │ │ └── __init__.py ├── roberta │ ├── commonsense_qa │ │ ├── __init__.py │ │ └── download_cqa_data.sh │ ├── wsc │ │ └── __init__.py │ └── config │ │ └── pretraining │ │ └── base.yaml ├── speech_text_joint_to_text │ ├── models │ │ └── __init__.py │ ├── tasks │ │ └── __init__.py │ ├── __init__.py │ ├── criterions │ │ └── __init__.py │ └── configs │ │ └── mustc_noise.list ├── translation_moe │ └── translation_moe_src │ │ ├── __init__.py │ │ └── logsumexp_moe.py ├── pointer_generator │ └── pointer_generator_src │ │ └── __init__.py ├── truncated_bptt │ └── __init__.py ├── __init__.py ├── m2m_100 │ └── tokenizers │ │ ├── thirdparty │ │ └── .gitignore │ │ ├── tokenize_zh.py │ │ ├── tokenize_thai.py │ │ ├── seg_ja.sh │ │ ├── seg_ko.sh │ │ ├── README.md │ │ ├── tokenize_indic.py │ │ └── tokenizer_ar.sh ├── laser │ └── laser_src │ │ └── __init__.py ├── fast_noisy_channel │ └── __init__.py ├── operators │ ├── alignment_train_cuda.h │ ├── utils.h │ └── alignment_train_cuda.cpp ├── constrained_decoding │ ├── normalize.py │ └── tok.py ├── wmt21 │ ├── README.md │ └── scripts │ │ └── replace-unicode-punctuation.perl ├── megatron_11b │ └── detok.py ├── unsupervised_quality_estimation │ └── repeat_lines.py ├── language_model │ └── prepare-wikitext-103.sh └── backtranslation │ └── sacrebleu.sh ├── fairseq ├── logging │ └── __init__.py ├── version.txt ├── modules │ ├── quantization │ │ ├── __init__.py │ │ ├── scalar │ │ │ ├── __init__.py │ │ │ └── modules │ │ │ │ └── __init__.py │ │ └── pq │ │ │ ├── __init__.py │ │ │ └── modules │ │ │ └── __init__.py │ ├── lightconv_layer │ │ ├── __init__.py │ │ └── setup.py │ ├── dynamicconv_layer │ │ ├── __init__.py │ │ ├── setup.py │ │ └── dynamiconv_cpu.cpp │ ├── grad_multiply.py │ ├── same_pad.py │ ├── transpose_last.py │ ├── unfold.py │ ├── gelu.py │ ├── fp32_group_norm.py │ └── scalar_bias.py ├── models │ ├── speech_to_speech │ │ ├── modules │ │ │ ├── __init__.py │ │ │ └── ctc_decoder.py │ │ └── __init__.py │ ├── speech_to_text │ │ ├── modules │ │ │ └── __init__.py │ │ └── __init__.py │ ├── bart │ │ └── __init__.py │ ├── hubert │ │ └── __init__.py │ ├── wav2vec │ │ ├── __init__.py │ │ └── utils.py │ ├── text_to_speech │ │ └── __init__.py │ ├── roberta │ │ └── __init__.py │ ├── nat │ │ └── __init__.py │ ├── ema │ │ └── __init__.py │ └── huggingface │ │ └── __init__.py ├── config │ ├── model │ │ ├── wav2vec │ │ │ └── vq_wav2vec_gumbel.yaml │ │ ├── wav2vec2 │ │ │ ├── wav2vec2_base.yaml │ │ │ └── wav2vec2_large.yaml │ │ └── transformer_lm │ │ │ ├── transformer_lm_gbw.yaml │ │ │ ├── transformer_lm_gpt.yaml │ │ │ └── transformer_lm_big.yaml │ ├── __init__.py │ └── config.yaml ├── data │ ├── multilingual │ │ └── __init__.py │ ├── setup.py │ ├── num_samples_dataset.py │ ├── id_dataset.py │ ├── offset_tokens_dataset.py │ ├── legacy │ │ └── __init__.py │ ├── roll_dataset.py │ ├── raw_label_dataset.py │ ├── lru_cache_dataset.py │ ├── huffman │ │ └── __init__.py │ ├── encoders │ │ ├── space_tokenizer.py │ │ ├── characters.py │ │ ├── nltk_tokenizer.py │ │ ├── __init__.py │ │ ├── bytes.py │ │ └── utils.py │ ├── sort_dataset.py │ ├── strip_token_dataset.py │ ├── list_dataset.py │ ├── numel_dataset.py │ ├── colorize_dataset.py │ ├── pad_dataset.py │ ├── prepend_dataset.py │ └── audio │ │ └── feature_transforms │ │ └── global_cmvn.py ├── model_parallel │ ├── models │ │ ├── roberta │ │ │ └── __init__.py │ │ ├── pipeline_parallel_transformer │ │ │ └── __init__.py │ │ └── __init__.py │ ├── __init__.py │ ├── criterions │ │ └── __init__.py │ └── modules │ │ └── __init__.py ├── benchmark │ ├── __init__.py │ └── dummy_dataset.py ├── dataclass │ └── __init__.py ├── tokenizer.py ├── clib │ ├── libnat_cuda │ │ └── edit_dist.h │ └── libbleu │ │ └── module.cpp ├── distributed │ └── __init__.py ├── scoring │ └── chrf.py └── criterions │ └── __init__.py ├── research ├── TranSpeech │ ├── __init__.py │ ├── asr_bleu │ │ ├── __init__.py │ │ └── requirements.txt │ ├── hifigan │ │ ├── __init__.py │ │ └── env.py │ ├── hubertCTC │ │ ├── __init__.py │ │ └── functions │ │ │ └── __init__.py │ └── tfcompat │ │ └── __init__.py └── __init__.py ├── docs ├── docutils.conf ├── requirements.txt ├── fairseq.gif ├── fairseq_logo.png ├── _static │ └── theme_overrides.css ├── modules.rst ├── Makefile ├── criterions.rst ├── make.bat └── optim.rst ├── pyproject.toml ├── setup.cfg ├── .gitmodules ├── train.py └── LICENSE /tests/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /fairseq_cli/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /scripts/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /tests/gpu/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/wav2vec/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /fairseq/logging/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /fairseq/version.txt: -------------------------------------------------------------------------------- 1 | 1.0.0a0 2 | -------------------------------------------------------------------------------- /research/TranSpeech/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /tests/distributed/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/speech_matrix/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /tests/speech_recognition/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/.gitignore: -------------------------------------------------------------------------------- 1 | !*/*.sh 2 | !*/*.md 3 | -------------------------------------------------------------------------------- /examples/speech_recognition/new/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /fairseq/modules/quantization/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /research/TranSpeech/asr_bleu/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /research/TranSpeech/hifigan/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /research/TranSpeech/hubertCTC/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /research/TranSpeech/tfcompat/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/hubert/tests/sample.base.L9.len: -------------------------------------------------------------------------------- 1 | 596 2 | -------------------------------------------------------------------------------- /examples/speech_matrix/data_helper/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/speech_recognition/kaldi/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/speech_to_speech/asr_bleu/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /docs/docutils.conf: -------------------------------------------------------------------------------- 1 | [writers] 2 | option-limit=0 3 | -------------------------------------------------------------------------------- /examples/attention_head_selection/src/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/hubert/tests/sample.large.L20.len: -------------------------------------------------------------------------------- 1 | 596 2 | -------------------------------------------------------------------------------- /examples/hubert/tests/sample.xlarge.L30.len: -------------------------------------------------------------------------------- 1 | 596 2 | -------------------------------------------------------------------------------- /examples/linformer/linformer_src/models/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/linformer/linformer_src/modules/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/speech_matrix/mined_train_sets/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/speech_matrix/speech_to_speech/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/speech_matrix/valid_test_sets/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/speech_recognition/new/decoders/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/textless_nlp/gslm/speech2unit/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /fairseq/models/speech_to_speech/modules/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /fairseq/models/speech_to_text/modules/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /research/TranSpeech/hubertCTC/functions/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /docs/requirements.txt: -------------------------------------------------------------------------------- 1 | sphinx<2.0 2 | sphinx-argparse 3 | -------------------------------------------------------------------------------- /examples/attention_head_selection/src/data/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/attention_head_selection/src/loss/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/attention_head_selection/src/models/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/attention_head_selection/src/modules/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/latent_depth/latent_depth_src/loss/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/latent_depth/latent_depth_src/models/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/latent_depth/latent_depth_src/modules/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/textless_nlp/gslm/speech2unit/clustering/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/textless_nlp/gslm/unit2speech/tacotron2/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /examples/multilingual/data_scripts/requirement.txt: -------------------------------------------------------------------------------- 1 | wget 2 | pandas -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/steps: -------------------------------------------------------------------------------- 1 | ../../wsj/s5/steps -------------------------------------------------------------------------------- /examples/wav2vec/unsupervised/kaldi_self_train/st/utils: -------------------------------------------------------------------------------- 1 | ../../wsj/s5/utils -------------------------------------------------------------------------------- /docs/fairseq.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/docs/fairseq.gif -------------------------------------------------------------------------------- /examples/adaptive_span/truncated_bptt_lm_task.py: -------------------------------------------------------------------------------- 1 | ../truncated_bptt/truncated_bptt_lm_task.py -------------------------------------------------------------------------------- /examples/speech_recognition/__init__.py: -------------------------------------------------------------------------------- 1 | from . import criterions, models, tasks # noqa 2 | -------------------------------------------------------------------------------- /docs/fairseq_logo.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/docs/fairseq_logo.png -------------------------------------------------------------------------------- /examples/MMPT/vlm.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/examples/MMPT/vlm.png -------------------------------------------------------------------------------- /examples/discriminative_reranking_nmt/__init__.py: -------------------------------------------------------------------------------- 1 | from . import criterions, models, tasks # noqa 2 | -------------------------------------------------------------------------------- /examples/MMPT/videoclip.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/examples/MMPT/videoclip.png -------------------------------------------------------------------------------- /examples/flores101/flores_logo.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/examples/flores101/flores_logo.png -------------------------------------------------------------------------------- /examples/multilingual/data_scripts/utils/strip_sgm.sh: -------------------------------------------------------------------------------- 1 | grep "seg id" | sed 's///g' | sed 's/<\/seg>//g' 2 | -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [build-system] 2 | requires = ["setuptools", "wheel", "cython"] 3 | build-backend = "setuptools.build_meta" 4 | -------------------------------------------------------------------------------- /setup.cfg: -------------------------------------------------------------------------------- 1 | [flake8] 2 | max-line-length = 127 3 | extend-ignore = E203, W503 4 | extend-exclude = fairseq/model_parallel/megatron 5 | -------------------------------------------------------------------------------- /examples/hubert/tests/6313-76958-0021.flac: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/examples/hubert/tests/6313-76958-0021.flac -------------------------------------------------------------------------------- /examples/hubert/tests/sample.base.L9.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/examples/hubert/tests/sample.base.L9.npy -------------------------------------------------------------------------------- /examples/hubert/tests/sample.large.L20.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/examples/hubert/tests/sample.large.L20.npy -------------------------------------------------------------------------------- /examples/hubert/tests/sample.xlarge.L30.npy: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/examples/hubert/tests/sample.xlarge.L30.npy -------------------------------------------------------------------------------- /examples/textless_nlp/speech-resynth/img/fig.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/Rongjiehuang/TranSpeech/HEAD/examples/textless_nlp/speech-resynth/img/fig.png -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/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='') -------------------------------------------------------------------------------- /research/TranSpeech/asr_bleu/requirements.txt: -------------------------------------------------------------------------------- 1 | fairseq==0.12.2 2 | pandas==1.4.3 3 | sacrebleu==2.2.0 4 | torch==1.12.1 5 | torchaudio==0.12.1 6 | tqdm==4.64.0 7 | transformers==4.21.1 8 | -------------------------------------------------------------------------------- /.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 | -------------------------------------------------------------------------------- /examples/speech_to_speech/asr_bleu/requirements.txt: -------------------------------------------------------------------------------- 1 | fairseq==0.12.2 2 | pandas==1.4.3 3 | sacrebleu==2.2.0 4 | torch==1.12.1 5 | torchaudio==0.12.1 6 | tqdm==4.64.0 7 | transformers==4.21.1 8 | -------------------------------------------------------------------------------- /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/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/discriminative_reranking_nmt/criterions/__init__.py: -------------------------------------------------------------------------------- 1 | from .discriminative_reranking_criterion import KLDivergenceRerankingCriterion 2 | 3 | 4 | __all__ = [ 5 | "KLDivergenceRerankingCriterion", 6 | ] 7 | -------------------------------------------------------------------------------- /examples/MMPT/projects/mtm/vlm.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/mtm/mmfusionmtm.yaml 2 | project_dir: mtm/vlm 3 | task_group: 4 | pretrain: 5 | dataset: 6 | sampled_min_len: 8 7 | loss: 8 | loss_cls: MTM 9 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /examples/speech_matrix/data_helper/config_multitask.yaml: -------------------------------------------------------------------------------- 1 | source_unit: 2 | data: source_unit 3 | decoder_type: transformer 4 | dict: source_unit/dict.txt 5 | encoder_layer: 6 6 | loss_weight: 8.0 7 | target_type: text 8 | -------------------------------------------------------------------------------- /examples/speech_synthesis/__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 | -------------------------------------------------------------------------------- /examples/speech_to_speech/__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/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 | -------------------------------------------------------------------------------- /examples/speech_synthesis/evaluation/__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 | -------------------------------------------------------------------------------- /examples/speech_synthesis/preprocessing/__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 | -------------------------------------------------------------------------------- /examples/speech_to_speech/preprocessing/__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 | -------------------------------------------------------------------------------- /examples/MMPT/scripts/text_token_extractor/configs/bert-base-uncased.yaml: -------------------------------------------------------------------------------- 1 | dataset: 2 | bert_name: bert-base-uncased 3 | caption_pkl_path: data/how2/raw_caption_dedup.pkl 4 | use_fast: true 5 | target_dir: data/feat/feat_how2_s3d_shard_small 6 | -------------------------------------------------------------------------------- /examples/speech_synthesis/preprocessing/denoiser/__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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /fairseq/data/setup.py: -------------------------------------------------------------------------------- 1 | from distutils.core import setup 2 | from Cython.Build import cythonize 3 | import numpy 4 | 5 | setup( 6 | name = 'data_utils_fast', 7 | ext_modules = cythonize("data_utils_fast.pyx"), 8 | include_dirs=[numpy.get_include()] 9 | ) -------------------------------------------------------------------------------- /research/__init__.py: -------------------------------------------------------------------------------- 1 | from research.TranSpeech.task import * 2 | from research.TranSpeech.dataset import * 3 | from research.TranSpeech.nar_transformer import * 4 | from research.TranSpeech.nar_conformer import * 5 | from research.TranSpeech.criterion import * 6 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/MMPT/projects/task/test_vtt_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_vtt.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | 9 | -------------------------------------------------------------------------------- /examples/hubert/tests/sample.large.hypo.word: -------------------------------------------------------------------------------- 1 | KEEP A GOING AN IF YOU'RE LUCKY YOU'LL RUN PLUMB INTO THEM WAS THE JEERING ANSWER AS THE SLEEPY COWMEN SPURRED THEIR PONIES ON TOWARD CAMP MUTTERING THEIR DISAPPROVAL OF TAKING ALONG A BUNCH OF BOYS ON A CATTLE DRIVE (None-0) 2 | -------------------------------------------------------------------------------- /examples/hubert/tests/sample.xlarge.hypo.word: -------------------------------------------------------------------------------- 1 | KEEP A GOIN AND IF YOU'RE LUCKY YOU'LL RUN PLUMB INTO THEM WAS THE JEERING ANSWER AS THE SLEEPY COWMEN SPURRED THEIR PONIES ON TOWARD CAMP MUTTERING THEIR DISAPPROVAL OF TAKING ALONG A BUNCH OF BOYS ON A CATTLE DRIVE (None-0) 2 | -------------------------------------------------------------------------------- /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/MMPT/projects/task/test_vttqa_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_vttqa.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | 9 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_youcook_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_youcook.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | 9 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/speech_text_joint_to_text/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 | -------------------------------------------------------------------------------- /examples/speech_text_joint_to_text/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 | import importlib 7 | import os 8 | 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 | -------------------------------------------------------------------------------- /examples/speech_text_joint_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 . import tasks, criterions, models # noqa 7 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/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/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 | -------------------------------------------------------------------------------- /examples/MMPT/scripts/video_feature_extractor/how2/s3d.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | 3 | 4 | python scripts/video_feature_extractor/extract.py \ 5 | --vdir \ 6 | --fdir data/feat/feat_how2_s3d \ 7 | --type=s3d --num_decoding_thread=4 \ 8 | --batch_size 32 --half_precision 1 9 | -------------------------------------------------------------------------------- /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/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/MMPT/projects/task/coin_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/coin.yaml 2 | model: 3 | model_cls: MMFusionSeparateActionSegmentation 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForTokenClassification 6 | text_encoder_cls: BertModel # dummy, not used. 7 | num_hidden_video_layers: 6 8 | -------------------------------------------------------------------------------- /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/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/MMPT/projects/task/test_coin_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_coin.yaml 2 | model: 3 | model_cls: MMFusionSeparateActionSegmentation 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForTokenClassification 6 | text_encoder_cls: BertModel # dummy, not used. 7 | num_hidden_video_layers: 6 8 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_crosstask_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_crosstask.yaml 2 | model: 3 | model_cls: MMFusionSeparateActionLocalization 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel # dummy, not used. 7 | num_hidden_video_layers: 6 8 | -------------------------------------------------------------------------------- /examples/speech_to_speech/enhance/config.yaml: -------------------------------------------------------------------------------- 1 | audio_root: /large_experiments/ust/spopuri/datasets/s2u/es_en/it3_layer11.km1000 2 | standardize_audio: true 3 | use_audio_input: true 4 | vocab_filename: dict_1003_unitmbart.txt 5 | prepend_bos_and_append_tgt_lang_tag: true 6 | tgt_lang: 7 | eos_token: 8 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_crosstask_zs_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_crosstask_zs.yaml 2 | model: 3 | model_cls: MMFusionSeparateActionLocalization 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel # dummy, not used. 7 | num_hidden_video_layers: 6 8 | -------------------------------------------------------------------------------- /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, get_param, attrsetter, quantize_model_ # NOQA 7 | -------------------------------------------------------------------------------- /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/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/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/speech_to_speech/benchmarking/configs/S2T.yaml: -------------------------------------------------------------------------------- 1 | general: 2 | dataset_path: $npy_dataset 3 | cpu: True 4 | model_type: S2T 5 | dataset_size: 1 6 | 7 | stage1: 8 | data: $data_bin 9 | task: speech_to_text 10 | path: $checkpoint 11 | config_yaml: config.yaml 12 | max_len_a: 2 13 | max_len_b: 500 14 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/mmpt/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 | from .mm import * 6 | 7 | try: 8 | from .expmm import * 9 | except ImportError: 10 | pass 11 | -------------------------------------------------------------------------------- /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_to_speech/unity/__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 sequence_generator # noqa 7 | from . import sequence_generator_multi_decoder # noqa 8 | -------------------------------------------------------------------------------- /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/MMPT/projects/task/youcook_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/youcook.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | # model_cls: MMFusionShare 9 | # mm_encoder_cls: MMBertForEncoder 10 | -------------------------------------------------------------------------------- /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/MMPT/projects/task/vttqa_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/vttqa.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | 9 | # model_cls: MMFusionShare 10 | # mm_encoder_cls: MMBertForEncoder 11 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /fairseq/models/text_to_speech/__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 .tacotron2 import * # noqa 7 | from .tts_transformer import * # noqa 8 | from .fastspeech2 import * # noqa 9 | -------------------------------------------------------------------------------- /examples/MMPT/mmpt/datasets/__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 | from .mmdataset import * 6 | 7 | try: 8 | from .fairseqmmdataset import * 9 | except ImportError: 10 | pass 11 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/retri/videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/retri/videoretri.yaml 2 | project_dir: retri/videoclip 3 | task_group: 4 | pretrain: 5 | model: 6 | model_cls: MMFusionSeparate 7 | mm_encoder_cls: 8 | video_encoder_cls: MMBertForEncoder 9 | text_encoder_cls: BertModel 10 | num_hidden_video_layers: 6 11 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /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 | . $KALDI_ROOT/tools/config/common_path.sh 5 | export LC_ALL=C 6 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test.yaml: -------------------------------------------------------------------------------- 1 | # this yaml cannot be run alone: implement a test_${dataset}.yaml 2 | slurm_config: big 3 | task_type: local_predict 4 | dataset: 5 | split: test 6 | video_processor: VideoProcessor 7 | aligner: DSAligner 8 | bert_name: bert-base-uncased 9 | fairseq: 10 | dataset: 11 | batch_size: 256 12 | valid_subset: test 13 | num_workers: 2 14 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/vtt_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/vtt.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | fairseq: 9 | dataset: 10 | batch_size: 224 11 | # model_cls: MMFusionShare 12 | # mm_encoder_cls: MMBertForEncoder 13 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/mmpt/evaluators/__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 | from .metric import * 6 | from .evaluator import * 7 | 8 | 9 | # experimental. 10 | try: 11 | from .expmetric import * 12 | except ImportError: 13 | pass 14 | -------------------------------------------------------------------------------- /examples/textless_nlp/gslm/metrics/asr_metrics/misc/dict.ltr.txt: -------------------------------------------------------------------------------- 1 | | 94802 2 | E 51860 3 | T 38431 4 | A 33152 5 | O 31495 6 | N 28855 7 | I 28794 8 | H 27187 9 | S 26071 10 | R 23546 11 | D 18289 12 | L 16308 13 | U 12400 14 | M 10685 15 | W 10317 16 | C 9844 17 | F 9062 18 | G 8924 19 | Y 8226 20 | P 6890 21 | B 6339 22 | V 3936 23 | K 3456 24 | ' 1023 25 | X 636 26 | J 598 27 | Q 437 28 | Z 213 29 | -------------------------------------------------------------------------------- /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/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/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 | -------------------------------------------------------------------------------- /fairseq/config/config.yaml: -------------------------------------------------------------------------------- 1 | # @package _group_ 2 | 3 | hydra: 4 | run: 5 | dir: . 6 | 7 | defaults: 8 | - _self_ 9 | - task: null 10 | - model: null 11 | - criterion: cross_entropy 12 | - optimizer: null 13 | - lr_scheduler: fixed 14 | - bpe: null 15 | - tokenizer: null 16 | - scoring: null 17 | - generation: null 18 | - common_eval: null 19 | - eval_lm: null 20 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/crosstask_videoclip.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/crosstask.yaml 2 | model: 3 | model_cls: MMFusionSeparateActionLocalization 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel # dummy, not used. 7 | num_hidden_video_layers: 6 8 | fairseq: 9 | checkpoint: 10 | restore_file: runs/task/checkpoint_best.pt # overwrite the default of VLM. 11 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_coin_zs.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_coin.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | eval: 9 | save_path: runs/task/coin_zs/eval 10 | fairseq: 11 | common_eval: 12 | path: runs/task/checkpoint_best.pt 13 | predictor: COINZSPredictor 14 | -------------------------------------------------------------------------------- /fairseq/models/speech_to_speech/__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 .s2s_conformer import * # noqa 7 | from .s2s_conformer_translatotron2 import * # noqa 8 | from .s2s_conformer_unity import * # noqa 9 | from .s2s_transformer import * # noqa 10 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_vtt_zs.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_vtt.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | eval: 9 | save_path: runs/task/vtt_zs/eval 10 | fairseq: 11 | # read code and find what is the checkpoint arg. 12 | common_eval: 13 | path: runs/task/checkpoint_best.pt 14 | -------------------------------------------------------------------------------- /examples/MMPT/mmpt/losses/__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 | from .loss import * 6 | from .nce import * 7 | 8 | try: 9 | from .fairseqmmloss import * 10 | except ImportError: 11 | pass 12 | 13 | try: 14 | from .expnce import * 15 | except ImportError: 16 | pass 17 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_vttqa_zs.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_vttqa.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | eval: 9 | save_path: runs/task/vttqa_zs/eval 10 | fairseq: 11 | # read code and find what is the checkpoint arg. 12 | common_eval: 13 | path: runs/task/checkpoint_best.pt 14 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_youcook_zs.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test_youcook.yaml 2 | model: 3 | model_cls: MMFusionSeparate 4 | mm_encoder_cls: 5 | video_encoder_cls: MMBertForEncoder 6 | text_encoder_cls: BertModel 7 | num_hidden_video_layers: 6 8 | eval: 9 | save_path: runs/task/youcook_zs/eval 10 | fairseq: 11 | # read code and find what is the checkpoint arg. 12 | common_eval: 13 | path: runs/task/checkpoint_best.pt 14 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/MMPT/mmpt/__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 | try: 6 | # fairseq user dir 7 | from .datasets import FairseqMMDataset 8 | from .losses import FairseqCriterion 9 | from .models import FairseqMMModel 10 | from .tasks import FairseqMMTask 11 | except ImportError: 12 | pass 13 | -------------------------------------------------------------------------------- /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/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/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/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/speech_to_speech/benchmarking/configs/2StageS2ST.yaml: -------------------------------------------------------------------------------- 1 | general: 2 | dataset_path: $npy_dataset 3 | cpu: True 4 | model_type: 2StageS2ST 5 | dataset_size: 1 6 | 7 | stage1: 8 | data: $data_bin_stage1 9 | task: speech_to_text 10 | path: $checkpoint_stage1 11 | config_yaml: config.yaml 12 | max_len_a: 2 13 | max_len_b: 500 14 | 15 | stage2: 16 | data: $data_bin_stage2 17 | task: text_to_speech 18 | path: $checkpoint_stage2 19 | config_yaml: config.yaml 20 | -------------------------------------------------------------------------------- /research/TranSpeech/hifigan/env.py: -------------------------------------------------------------------------------- 1 | import os 2 | import shutil 3 | 4 | 5 | class AttrDict(dict): 6 | def __init__(self, *args, **kwargs): 7 | super(AttrDict, self).__init__(*args, **kwargs) 8 | self.__dict__ = self 9 | 10 | 11 | def build_env(config, config_name, path): 12 | t_path = os.path.join(path, config_name) 13 | if config != t_path: 14 | os.makedirs(path, exist_ok=True) 15 | shutil.copyfile(config, os.path.join(path, config_name)) 16 | -------------------------------------------------------------------------------- /examples/operators/alignment_train_cuda.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 // @manual=//caffe2:torch_extension 12 | 13 | void alignmentTrainCUDAWrapper( 14 | const torch::Tensor& p_choose, 15 | torch::Tensor& alpha, 16 | float eps); 17 | -------------------------------------------------------------------------------- /examples/MMPT/mmpt/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 | from .mmfusion import * 6 | from .transformermodel import * 7 | from .mmfusionnlg import * 8 | 9 | try: 10 | from .fairseqmmmodel import * 11 | except ImportError: 12 | pass 13 | 14 | try: 15 | from .expmmfusion import * 16 | except ImportError: 17 | pass 18 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/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/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/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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/mtm/mmfusionmtm.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/mfmmlm.yaml 2 | project_dir: mtm/mmfusionmtm 3 | task_group: 4 | pretrain: 5 | task: VLMTask # reproducible 6 | dataset: 7 | aligner: MFMMLMAligner 8 | model: 9 | use_seg_emb: True # reproducible 10 | model_cls: MMFusionMTM 11 | mm_encoder_cls: MMBertForMFMMLM 12 | loss: 13 | loss_cls: MTM 14 | finetune: 15 | model: 16 | use_seg_emb: True # reproducible 17 | test: 18 | model: 19 | use_seg_emb: True # reproducible 20 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/ft.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/default.yaml 2 | # all derived config will be run by fairseq-train. 3 | task_type: sweep_small 4 | fairseq: 5 | optimization: 6 | warmup_updates: 122 # copied from roberta glue: https://github.com/pytorch/fairseq/blob/master/examples/roberta/README.glue.md 7 | checkpoint: 8 | # save_interval_updates: 512 9 | # borrowed from Roberta script. 10 | restore_file: runs/task/checkpoint_best.pt 11 | reset_optimizer: True 12 | reset_dataloader: True 13 | reset_meters: True 14 | -------------------------------------------------------------------------------- /examples/MMPT/mmpt/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 | from .task import * 6 | from .vlmtask import * 7 | from .retritask import * 8 | 9 | try: 10 | from .fairseqmmtask import * 11 | except ImportError: 12 | pass 13 | 14 | try: 15 | from .milncetask import * 16 | except ImportError: 17 | pass 18 | 19 | try: 20 | from .expretritask import * 21 | except ImportError: 22 | pass 23 | -------------------------------------------------------------------------------- /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 -------------------------------------------------------------------------------- /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 .multi_modality_model import * # noqa 9 | from .s2t_conformer import * # noqa 10 | from .s2t_transformer import * # noqa 11 | from .s2t_wav_transformer import * # noqa 12 | from .xm_transformer import * # noqa 13 | from .xm_transformer_unity import * # noqa 14 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/speech_to_speech/benchmarking/configs/DirectS2U.yaml: -------------------------------------------------------------------------------- 1 | general: 2 | dataset_path: $npy_dataset_path 3 | cpu: True 4 | model_type: S2UT 5 | dataset_size: 5 6 | dump_speech_waveforms_dir: $dump_waveforms_dir_path 7 | 8 | stage1: 9 | data: $data_bin 10 | task: speech_to_speech 11 | path: $checkpoint 12 | config_yaml: config.yaml 13 | max_len_b: 100000 14 | beam: 10 15 | target_is_code: True 16 | max_target_positions: 3000 17 | target_code_size: 100 18 | 19 | stage2: 20 | vocoder: $vocoder_path 21 | vocoder_cfg: $vocoder_cfg_json 22 | dur_prediction: True 23 | -------------------------------------------------------------------------------- /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/speech_text_joint_to_text/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 | for file in os.listdir(os.path.dirname(__file__)): 11 | if file.endswith(".py") and not file.startswith("_"): 12 | criterion_name = file[: file.find(".py")] 13 | importlib.import_module( 14 | "examples.speech_text_joint_to_text.criterions." + criterion_name 15 | ) 16 | -------------------------------------------------------------------------------- /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/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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/speech_matrix/data_helper/config.yaml: -------------------------------------------------------------------------------- 1 | audio_root: audios 2 | input_channels: 1 3 | input_feat_per_channel: 80 4 | output_channels: 1 5 | output_feat_per_channel: 1 6 | output_feat_reduction_rate: 0 7 | output_sample_rate: 16000 8 | specaugment: 9 | freq_mask_F: 27 10 | freq_mask_N: 1 11 | time_mask_N: 1 12 | time_mask_T: 100 13 | time_mask_p: 1.0 14 | time_wrap_W: 0 15 | transforms: 16 | _eval: 17 | - utterance_cmvn 18 | _train: 19 | - utterance_cmvn 20 | - specaugment 21 | vocoder: 22 | dur_prediction: true 23 | model_path: vocoder.pt 24 | speaker: false 25 | type: code_hifigan 26 | -------------------------------------------------------------------------------- /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}/viterbi 9 | sweep: 10 | dir: ${common_eval.results_path} 11 | subdir: viterbi 12 | 13 | task: 14 | _name: hubert_pretraining 15 | single_target: true 16 | fine_tuning: true 17 | data: ??? 18 | normalize: ??? 19 | 20 | decoding: 21 | type: viterbi 22 | unique_wer_file: true 23 | common_eval: 24 | results_path: ??? 25 | path: ??? 26 | post_process: letter 27 | dataset: 28 | max_tokens: 1100000 29 | gen_subset: ??? 30 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/operators/utils.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 // @manual=//caffe2:torch_extension 12 | 13 | #define CHECK_CUDA(x) \ 14 | TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor") 15 | #define CHECK_CONTIGUOUS(x) \ 16 | TORCH_CHECK(x.is_contiguous(), #x " must be contiguous") 17 | #define CHECK_INPUT(x) \ 18 | CHECK_CUDA(x); \ 19 | CHECK_CONTIGUOUS(x) 20 | -------------------------------------------------------------------------------- /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: /checkpoint/${env:USER}/${env:PREFIX}/${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: 3000000 20 | gen_subset: test 21 | distributed_training: 22 | distributed_world_size: 1 23 | decoding: 24 | beam: 5 25 | type: viterbi 26 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_vtt.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test.yaml 2 | dataset: 3 | meta_processor: MSRVTTMetaProcessor 4 | test_path: data/msrvtt/MSRVTT_JSFUSION_test.csv 5 | video_processor: VideoProcessor 6 | vfeat_dir: data/feat/feat_vtt_s3d 7 | text_processor: MSRVTTTextProcessor 8 | num_iso_layer: 12 9 | model: 10 | model_cls: MMFusionJoint 11 | mm_encoder_cls: MMBertForJoint 12 | eval: 13 | save_path: runs/task/vtt/eval 14 | fairseq: 15 | # read code and find what is the checkpoint arg. 16 | common_eval: 17 | path: runs/task/vtt/checkpoint_last.pt 18 | metric: RetrievalMetric 19 | predictor: RetrievalPredictor 20 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /examples/speech_to_speech/benchmarking/configs/3StageS2ST.yaml: -------------------------------------------------------------------------------- 1 | general: 2 | dataset_path: $npy_dataset 3 | cpu: True 4 | model_type: 3StageS2ST 5 | max_len_a: 2 6 | max_len_b: 500 7 | dataset_size: 1 8 | 9 | stage1: 10 | data: $data_bin_stage1 11 | task: speech_to_text 12 | path: $checkpoint_stage1 13 | config_yaml: config.yaml 14 | max_len_a: 2 15 | max_len_b: 500 16 | 17 | stage2: 18 | data: $data_bin_stage2 19 | task: translation 20 | path: $checkpoint_stage2 21 | config_yaml: config.yaml 22 | 23 | 24 | stage2: 25 | data: $data_bin_stage3 26 | task: text_to_speech 27 | path: $checkpoint_stage3 28 | config_yaml: config.yaml 29 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_vttqa.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test.yaml 2 | dataset: 3 | meta_processor: MSRVTTQAMetaProcessor 4 | test_path: data/msrvtt-qa/MSR_MC_test.csv 5 | video_processor: VideoProcessor 6 | vfeat_dir: data/feat/feat_vtt_s3d 7 | text_processor: MSRVTTQATextProcessor 8 | aligner: MSRVTTQAAligner 9 | num_iso_layer: 12 10 | model: 11 | model_cls: MMFusionJoint 12 | mm_encoder_cls: MMBertForJoint 13 | eval: 14 | save_path: runs/task/vttqa/eval 15 | fairseq: 16 | # read code and find what is the checkpoint arg. 17 | common_eval: 18 | path: runs/task/vttqa/checkpoint_last.pt 19 | metric: QAMetric 20 | predictor: QAPredictor 21 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/vttqa.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/ft.yaml 2 | dataset: 3 | meta_processor: MSRVTTMetaProcessor 4 | train_path: data/msrvtt/MSRVTT_train.csv 5 | dup: 20 6 | val_path: data/msrvtt/MSRVTT_JSFUSION_test.csv 7 | vfeat_dir: data/feat/feat_vtt_s3d 8 | text_processor: MSRVTTTextProcessor 9 | json_path: data/msrvtt/MSRVTT_data.json 10 | aligner: DSAligner 11 | num_iso_layer: 12 12 | model: 13 | model_cls: MMFusionJoint 14 | mm_encoder_cls: MMBertForJoint 15 | loss: 16 | loss_cls: V2TContraLoss 17 | fairseq: 18 | dataset: 19 | batch_size: 128 20 | optimization: 21 | max_epoch: 5 22 | checkpoint: 23 | save_dir: runs/task/vttqa 24 | -------------------------------------------------------------------------------- /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/models/speech_to_speech/modules/ctc_decoder.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 torch import nn 7 | 8 | from fairseq.models import FairseqEncoder 9 | 10 | 11 | class CTCDecoder(FairseqEncoder): 12 | def __init__(self, dictionary, in_dim): 13 | super().__init__(dictionary) 14 | self.proj = nn.Linear(in_dim, len(dictionary)) 15 | 16 | def forward(self, src_tokens, src_lengths=None, **kwargs): 17 | encoder_out = self.proj(src_tokens) 18 | return {"encoder_out": encoder_out} 19 | -------------------------------------------------------------------------------- /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) -------------------------------------------------------------------------------- /examples/MMPT/projects/task/default.yaml: -------------------------------------------------------------------------------- 1 | # this yaml cannot be run alone. you must use `how2.yaml`, `vtt.yaml` etc for training. 2 | dataset: 3 | video_processor: VideoProcessor 4 | bert_name: bert-base-uncased 5 | fairseq: 6 | common: 7 | tensorboard_logdir: run 8 | log_interval: 1000 9 | dataset: 10 | num_workers: 4 11 | optimization: 12 | lr: [ 0.00005 ] 13 | clip_norm: 2.0 14 | optimizer: adam 15 | adam_betas: (0.9, 0.98) 16 | lr_scheduler: polynomial_decay 17 | total_num_update: 1000000 # backward compatible on fairseq 1.0.0a0+af0389f for reproducibility. 18 | warmup_updates: 1000 19 | weight_decay: 0.0 20 | ddp_backend: no_c10d 21 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /fairseq/data/huffman/__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 .huffman_coder import HuffmanCodeBuilder, HuffmanCoder 7 | from .huffman_mmap_indexed_dataset import ( 8 | HuffmanMMapIndex, 9 | HuffmanMMapIndexedDataset, 10 | HuffmanMMapIndexedDatasetBuilder, 11 | vocab_file_path, 12 | ) 13 | 14 | __all__ = [ 15 | "HuffmanCoder", 16 | "HuffmanCodeBuilder", 17 | "HuffmanMMapIndexedDatasetBuilder", 18 | "HuffmanMMapIndexedDataset", 19 | "HuffmanMMapIndex", 20 | "vocab_file_path", 21 | ] 22 | -------------------------------------------------------------------------------- /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/models/ema/__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 .ema import EMA 10 | 11 | 12 | def build_ema(model, cfg, device): 13 | return EMA(model, cfg, device) 14 | 15 | 16 | # automatically import any Python files in the models/ema/ directory 17 | for file in sorted(os.listdir(os.path.dirname(__file__))): 18 | if file.endswith(".py") and not file.startswith("_"): 19 | file_name = file[: file.find(".py")] 20 | importlib.import_module("fairseq.models.ema." + file_name) 21 | -------------------------------------------------------------------------------- /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/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/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/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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/youcookcap.yaml: -------------------------------------------------------------------------------- 1 | # finetuning for youcook captioning. 2 | includes: projects/task/ft.yaml 3 | dataset: 4 | meta_processor: YoucookNLGMetaProcessor 5 | train_path: data/youcook/train_list.txt 6 | val_path: data/youcook/val_list.txt 7 | trainval_annotation: data/youcook/youcookii_annotations_trainval.json 8 | video_processor: YoucookVideoProcessor 9 | vfeat_dir: data/feat/feat_youcook_s3d 10 | text_processor: NLGTextProcessor 11 | aligner: DSNLGAligner 12 | model: 13 | model_cls: MMFusionNLG 14 | mm_encoder_cls: MMBertForNLG 15 | loss: 16 | loss_cls: NLGLoss 17 | fairseq: 18 | dataset: 19 | batch_size: 128 20 | optimization: 21 | max_epoch: 10 22 | checkpoint: 23 | save_dir: runs/task/youcookcap 24 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /tests/speech/test_s2t_transformer.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 tests.speech import TestFairseqSpeech 8 | 9 | 10 | class TestS2TTransformer(TestFairseqSpeech): 11 | def setUp(self): 12 | self.set_up_librispeech() 13 | 14 | def test_librispeech_s2t_transformer_s_checkpoint(self): 15 | self.base_test( 16 | ckpt_name="librispeech_transformer_s.pt", 17 | reference_score=9, 18 | arg_overrides={"config_yaml": "cfg_librispeech.yaml"}, 19 | ) 20 | 21 | 22 | if __name__ == "__main__": 23 | unittest.main() 24 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /tests/speech/test_s2t_conformer.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 tests.speech import TestFairseqSpeech 8 | 9 | 10 | class TestS2TConformer(TestFairseqSpeech): 11 | def setUp(self): 12 | self.set_up_librispeech() 13 | 14 | def test_librispeech_s2t_conformer_s_checkpoint(self): 15 | self.base_test( 16 | ckpt_name="librispeech_conformer_rel_pos_s.pt", 17 | reference_score=12, 18 | arg_overrides={"config_yaml": "cfg_librispeech.yaml"}, 19 | ) 20 | 21 | 22 | if __name__ == "__main__": 23 | unittest.main() 24 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/how2.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/default.yaml 2 | task_type: sweep_big 3 | slurm_config: big 4 | dataset: 5 | meta_processor: ShardedHow2MetaProcessor 6 | train_path: data/how2/how2_s3d_train.lst 7 | val_path: data/how2/how2_s3d_val.lst 8 | video_processor: ShardedVideoProcessor 9 | vfeat_dir: data/feat/feat_how2_s3d_shard_small 10 | text_processor: ShardedTextProcessor 11 | tfeat_dir: data/feat/feat_how2_s3d_shard_small/raw_caption_dedup.bert-base-uncased. 12 | aligner: FixedLenAligner 13 | # disable direct running of this yaml 14 | eval: 15 | save_path: runs/task 16 | fairseq: 17 | checkpoint: 18 | save_dir: runs/task 19 | save_interval_updates: 1024 20 | keep_interval_updates: 2 21 | keep_last_epochs: 30 22 | 23 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_didemo_zs.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test.yaml 2 | dataset: 3 | meta_processor: DiDeMoMetaProcessor 4 | test_path: data/didemo/test_data.json 5 | video_processor: VideoProcessor 6 | vfeat_dir: data/feat/feat_didemo_s3d 7 | text_processor: DiDeMoTextProcessor 8 | aligner: DiDeMoAligner 9 | num_iso_layer: 12 10 | model: 11 | model_cls: MMFusionSeparate 12 | mm_encoder_cls: 13 | video_encoder_cls: MMBertForEncoder 14 | text_encoder_cls: BertModel 15 | num_hidden_video_layers: 6 16 | eval: 17 | save_path: runs/task/didemo_zs/eval 18 | fairseq: 19 | # read code and find what is the checkpoint arg. 20 | common_eval: 21 | path: runs/task/checkpoint_best.pt 22 | metric: DiDeMoMetric 23 | predictor: DiDeMoPredictor 24 | -------------------------------------------------------------------------------- /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/MMPT/mmpt/processors/__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 | from .processor import * 6 | 7 | from .how2processor import * 8 | from .how2retriprocessor import * 9 | 10 | from .dsprocessor import * 11 | 12 | try: 13 | from .rawvideoprocessor import * 14 | from .codecprocessor import * 15 | from .webvidprocessor import * 16 | from .expprocessor import * 17 | from .exphow2processor import * 18 | from .exphow2retriprocessor import * 19 | from .expcodecprocessor import * 20 | from .expfeatureencoder import * 21 | from .expdsprocessor import * 22 | except ImportError: 23 | pass 24 | -------------------------------------------------------------------------------- /examples/MMPT/setup.py: -------------------------------------------------------------------------------- 1 | import setuptools 2 | 3 | with open("README.md", "r") as fh: 4 | long_description = fh.read() 5 | 6 | setuptools.setup( 7 | name="mmpt", 8 | version="0.0.1", 9 | author="Hu Xu, Po-yao Huang", 10 | author_email="huxu@fb.com", 11 | description="A package for multimodal pretraining.", 12 | long_description=long_description, 13 | long_description_content_type="text/markdown", 14 | url="https://github.com/pytorch/fairseq/examples/MMPT", 15 | packages=setuptools.find_packages(), 16 | install_requires=[ 17 | ], 18 | classifiers=[ 19 | "Programming Language :: Python :: 3", 20 | "License :: CC-BY-NC", 21 | "Operating System :: OS Independent", 22 | ], 23 | python_requires='>=3.6', 24 | ) 25 | -------------------------------------------------------------------------------- /examples/speech_matrix/mined_train_sets/download_speech_to_unit.py: -------------------------------------------------------------------------------- 1 | import os 2 | import argparse 3 | from examples.speech_matrix.data_helper.data_cfg import ( 4 | DOWNLOAD_HUB, 5 | manifest_key, 6 | ) 7 | 8 | 9 | def download_s2u_manifests(save_root): 10 | # download 11 | manifest_dl = f"{DOWNLOAD_HUB}/{manifest_key}.tar.gz" 12 | os.system(f"wget {manifest_dl} -P {save_root}") 13 | # unzip 14 | os.system(f"tar -zxvf {save_root}/{manifest_key}.tar.gz -C {save_root}/") 15 | 16 | 17 | if __name__ == "__main__": 18 | 19 | parser = argparse.ArgumentParser() 20 | parser.add_argument("--save-root", type=str, required=True) 21 | args = parser.parse_args() 22 | 23 | # download speech_to_unit manifests 24 | download_s2u_manifests(args.save_root) 25 | -------------------------------------------------------------------------------- /examples/textless_nlp/gslm/metrics/README.md: -------------------------------------------------------------------------------- 1 | # GSLM Metrics 2 | 3 | ## ASR Metrics 4 | The suite of metrics here uses an ASR model to transcribe the synthesized speech into text, and then uses text-based metrics. We also use word error rate from ASR transcription itself as one of the metrics. [More details](asr_metrics) 5 | 6 | ## ABX Metrics 7 | We use [ABX](https://www.semanticscholar.org/paper/ABX-Discriminability-Measures-and-Applications-Schatz/13d3537228f728c1063cc83743cb118bba3367a0) to evaluate how well-separated phonetic categories are with quantized representations. [More details](abx_metrics) 8 | 9 | ## sWUGGY and sBLIMP 10 | We refer to [ZeroSpeech challenge](https://www.zerospeech.com/2021/track_s.html#scoring-based-metrics) for details on the sWUGGY and sBLIMP metrics. 11 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/coin.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/ft.yaml 2 | task_type: sweep_big 3 | dataset: 4 | meta_processor: COINActionSegmentationMetaProcessor 5 | train_path: data/coin/COIN.json 6 | val_path: data/coin/COIN.json 7 | vfeat_dir: data/feat/feat_coin_s3d 8 | video_processor: VideoProcessor 9 | text_processor: COINActionSegmentationTextProcessor 10 | aligner: COINActionSegmentationAligner 11 | num_iso_layer: 12 12 | sliding_window: 8 13 | sliding_window_size: 32 14 | model: 15 | model_cls: MMFusionActionSegmentation 16 | mm_encoder_cls: MMBertForTokenClassification 17 | loss: 18 | loss_cls: CrossEntropy 19 | fairseq: 20 | dataset: 21 | batch_size: 1 22 | optimization: 23 | max_epoch: 8 24 | checkpoint: 25 | save_dir: runs/task/coin 26 | -------------------------------------------------------------------------------- /fairseq/models/wav2vec/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 math 7 | import torch.nn.functional as F 8 | 9 | 10 | def pad_to_multiple(x, multiple, dim=-1, value=0): 11 | # Inspired from https://github.com/lucidrains/local-attention/blob/master/local_attention/local_attention.py#L41 12 | if x is None: 13 | return None, 0 14 | tsz = x.size(dim) 15 | m = tsz / multiple 16 | remainder = math.ceil(m) * multiple - tsz 17 | if m.is_integer(): 18 | return x, 0 19 | pad_offset = (0,) * (-1 - dim) * 2 20 | 21 | return F.pad(x, (*pad_offset, 0, remainder), value=value), remainder 22 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/vtt.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/ft.yaml 2 | dataset: 3 | meta_processor: MSRVTTMetaProcessor 4 | train_path: data/msrvtt/MSRVTT_train.csv 5 | jsfusion_path: data/msrvtt/MSRVTT_JSFUSION_test.csv 6 | full_test_path: data/msrvtt/MSRVTT_FULL_test.csv 7 | dup: 20 8 | val_path: data/msrvtt/MSRVTT_JSFUSION_test.csv 9 | vfeat_dir: data/feat/feat_vtt_s3d 10 | text_processor: MSRVTTTextProcessor 11 | json_path: data/msrvtt/MSRVTT_data.json 12 | aligner: DSAligner 13 | num_iso_layer: 12 14 | model: 15 | model_cls: MMFusionJoint 16 | mm_encoder_cls: MMBertForJoint 17 | loss: 18 | loss_cls: T2VContraLoss 19 | fairseq: 20 | dataset: 21 | batch_size: 256 22 | optimization: 23 | max_epoch: 10 24 | checkpoint: 25 | save_dir: runs/task/vtt 26 | -------------------------------------------------------------------------------- /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/MMPT/projects/task/test_youcookcap.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test.yaml 2 | dataset: 3 | meta_processor: YoucookNLGMetaProcessor 4 | test_path: data/youcook/val_list.txt 5 | trainval_annotation: data/youcook/youcookii_annotations_trainval.json 6 | video_processor: YoucookVideoProcessor 7 | vfeat_dir: data/feat/feat_youcook_s3d 8 | text_processor: NLGTextProcessor 9 | aligner: DSNLGAligner 10 | model: 11 | model_cls: MMFusionNLG 12 | mm_encoder_cls: MMBertForNLG 13 | max_decode_length: 24 14 | eval: 15 | save_path: runs/task/youcookcap/eval 16 | fairseq: 17 | # read code and find what is the checkpoint arg. 18 | common_eval: 19 | path: runs/task/youcookcap/checkpoint_best.pt 20 | metric: NLGMetric 21 | predictor: NLGPredictor 22 | gen_param: 23 | num_beams: 5 24 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_coin.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test.yaml 2 | dataset: 3 | split: test 4 | test_path: data/coin/COIN.json 5 | meta_processor: COINActionSegmentationMetaProcessor 6 | vfeat_dir: data/feat/feat_coin_s3d 7 | video_processor: VideoProcessor 8 | text_processor: COINActionSegmentationTextProcessor 9 | aligner: COINActionSegmentationAligner 10 | num_iso_layer: 12 11 | sliding_window: 16 12 | sliding_window_size: 32 13 | model: 14 | model_cls: MMFusionActionSegmentation 15 | mm_encoder_cls: MMBertForTokenClassification 16 | eval: 17 | save_path: runs/task/coin/eval 18 | fairseq: 19 | dataset: 20 | batch_size: 1 21 | common_eval: 22 | path: runs/task/coin/checkpoint_best.pt 23 | metric: COINActionSegmentationMetric 24 | predictor: COINPredictor 25 | -------------------------------------------------------------------------------- /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 | 7 | import os 8 | import importlib 9 | from fairseq import registry 10 | 11 | ( 12 | build_monotonic_attention, 13 | register_monotonic_attention, 14 | MONOTONIC_ATTENTION_REGISTRY, 15 | _, 16 | ) = registry.setup_registry("--simul-type") 17 | 18 | for file in sorted(os.listdir(os.path.dirname(__file__))): 19 | if file.endswith(".py") and not file.startswith("_"): 20 | model_name = file[: file.find(".py")] 21 | importlib.import_module( 22 | "examples.simultaneous_translation.modules." + model_name 23 | ) 24 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/operators/alignment_train_cuda.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 "alignment_train_cuda.h" 10 | #include "utils.h" 11 | 12 | namespace { 13 | 14 | void alignmentTrainCUDA( 15 | const torch::Tensor& p_choose, 16 | torch::Tensor& alpha, 17 | float eps) { 18 | CHECK_INPUT(p_choose); 19 | CHECK_INPUT(alpha); 20 | 21 | alignmentTrainCUDAWrapper(p_choose, alpha, eps); 22 | } 23 | 24 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 25 | m.def( 26 | "alignment_train_cuda", 27 | &alignmentTrainCUDA, 28 | "expected_alignment_from_p_choose (CUDA)"); 29 | } 30 | 31 | } // namespace 32 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /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/textless_nlp/gslm/speech2unit/clustering/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 | from typing import List, Tuple 7 | 8 | 9 | def get_audio_files(manifest_path: str) -> Tuple[str, List[str], List[int]]: 10 | fnames, sizes = [], [] 11 | with open(manifest_path, "r") as f: 12 | root_dir = f.readline().strip() 13 | for line in f: 14 | items = line.strip().split("\t") 15 | assert ( 16 | len(items) == 2 17 | ), f"File must have two columns separated by tab. Got {line}" 18 | fnames.append(items[0]) 19 | sizes.append(int(items[1])) 20 | return root_dir, fnames, sizes 21 | -------------------------------------------------------------------------------- /examples/textless_nlp/gslm/unit2speech/tacotron2/symbols.py: -------------------------------------------------------------------------------- 1 | """ from https://github.com/keithito/tacotron """ 2 | 3 | ''' 4 | Defines the set of symbols used in text input to the model. 5 | 6 | The default is a set of ASCII characters that works well for English or text that has been run through Unidecode. For other data, you can modify _characters. See TRAINING_DATA.md for details. ''' 7 | from . import cmudict 8 | 9 | _pad = '_' 10 | _punctuation = '!\'(),.:;? ' 11 | _special = '-' 12 | _letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz' 13 | 14 | # Prepend "@" to ARPAbet symbols to ensure uniqueness (some are the same as uppercase letters): 15 | _arpabet = ['@' + s for s in cmudict.valid_symbols] 16 | 17 | # Export all symbols: 18 | symbols = [_pad] + list(_special) + list(_punctuation) + list(_letters) + _arpabet 19 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/MMPT/projects/mtm/vlm/test_vtt.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: DSAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: MSRVTTMetaProcessor 9 | test_path: data/msrvtt/MSRVTT_JSFUSION_test.csv 10 | vfeat_dir: data/feat/feat_vtt_s3d 11 | text_processor: MSRVTTTextProcessor 12 | num_iso_layer: 12 13 | max_video_len: 32 14 | max_len: 96 15 | fairseq: 16 | dataset: 17 | batch_size: 256 18 | valid_subset: test 19 | num_workers: 2 20 | common_eval: 21 | path: runs/mtm/vlm/vtt/checkpoint_last.pt 22 | model: 23 | model_cls: MMFusionJoint 24 | mm_encoder_cls: MMBertForJoint 25 | use_seg_emb: true 26 | eval: 27 | save_path: runs/mtm/vlm/vtt/eval 28 | metric: RetrievalMetric 29 | predictor: RetrievalPredictor 30 | -------------------------------------------------------------------------------- /examples/MMPT/projects/mtm/vlm/test_vttqa.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: MSRVTTQAAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: MSRVTTQAMetaProcessor 9 | test_path: data/msrvtt-qa/MSR_MC_test.csv 10 | vfeat_dir: data/feat/feat_vtt_s3d 11 | text_processor: MSRVTTQATextProcessor 12 | num_iso_layer: 12 13 | max_video_len: 32 14 | max_len: 96 15 | fairseq: 16 | dataset: 17 | batch_size: 256 18 | valid_subset: test 19 | num_workers: 2 20 | common_eval: 21 | path: runs/mtm/vlm/vttqa/checkpoint_last.pt 22 | model: 23 | model_cls: MMFusionJoint 24 | mm_encoder_cls: MMBertForJoint 25 | use_seg_emb: true 26 | eval: 27 | save_path: runs/mtm/vlm/vttqa/eval 28 | metric: QAMetric 29 | predictor: QAPredictor 30 | -------------------------------------------------------------------------------- /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/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: 10 9 | epsilon: 0.025 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 | decoding.silweight: 28 | type: range 29 | bounds: [ -8.0, 0.0 ] 30 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/youcook.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/ft.yaml 2 | dataset: 3 | meta_processor: YoucookMetaProcessor 4 | train_path: data/youcook/youcook_train.pkl 5 | val_path: data/youcook/youcook_val.pkl 6 | trainval_annotation: data/youcook/youcookii_annotations_trainval.json 7 | use_annotation_text: True 8 | video_processor: YoucookVideoProcessor 9 | vfeat_dir: data/feat/feat_youcook_s3d # /checkpoint/huxu/feat/youcook_vmz # /checkpoint/prarora/berniehuang/feat_youcook_vmz 10 | text_processor: TextProcessor 11 | aligner: DSAligner 12 | num_iso_layer: 12 13 | model: 14 | model_cls: MMFusionJoint 15 | mm_encoder_cls: MMBertForJoint 16 | loss: 17 | loss_cls: T2VContraLoss 18 | fairseq: 19 | dataset: 20 | batch_size: 128 21 | optimization: 22 | max_epoch: 10 23 | checkpoint: 24 | save_dir: runs/task/youcook 25 | 26 | -------------------------------------------------------------------------------- /examples/MMPT/projects/task/test_youcook.yaml: -------------------------------------------------------------------------------- 1 | includes: projects/task/test.yaml 2 | dataset: 3 | meta_processor: YoucookMetaProcessor 4 | test_path: data/youcook/youcook_val.pkl 5 | trainval_annotation: data/youcook/youcookii_annotations_trainval.json 6 | use_annotation_text: True 7 | video_processor: YoucookVideoProcessor 8 | vfeat_dir: data/feat/feat_youcook_s3d # /checkpoint/huxu/feat/youcook_vmz # /checkpoint/prarora/berniehuang/feat_youcook_vmz 9 | text_processor: TextProcessor 10 | aligner: DSAligner 11 | num_iso_layer: 12 12 | model: 13 | model_cls: MMFusionJoint 14 | mm_encoder_cls: MMBertForJoint 15 | eval: 16 | save_path: runs/task/youcook/eval 17 | fairseq: 18 | # read code and find what is the checkpoint arg. 19 | common_eval: 20 | path: runs/task/youcook/checkpoint_last.pt 21 | metric: RetrievalMetric 22 | predictor: RetrievalPredictor 23 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 ( 8 | fsdp_enable_wrap, 9 | fsdp_wrap, 10 | FullyShardedDataParallel, 11 | ) 12 | from .legacy_distributed_data_parallel import LegacyDistributedDataParallel 13 | from .module_proxy_wrapper import ModuleProxyWrapper 14 | from .tpu_distributed_data_parallel import TPUDistributedDataParallel 15 | 16 | 17 | __all__ = [ 18 | "DistributedTimeoutWrapper", 19 | "fsdp_enable_wrap", 20 | "fsdp_wrap", 21 | "FullyShardedDataParallel", 22 | "LegacyDistributedDataParallel", 23 | "ModuleProxyWrapper", 24 | "TPUDistributedDataParallel", 25 | ] 26 | -------------------------------------------------------------------------------- /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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/textless_nlp/gslm/unit2speech/multiproc.py: -------------------------------------------------------------------------------- 1 | import os 2 | import time 3 | import torch 4 | import sys 5 | import subprocess 6 | 7 | argslist = list(sys.argv)[1:] 8 | log_dir = argslist[-1] 9 | num_gpus = torch.cuda.device_count() 10 | argslist.append('--n_gpus={}'.format(num_gpus)) 11 | workers = [] 12 | job_id = time.strftime("%Y_%m_%d-%H%M%S") 13 | argslist.append("--group_name=group_{}".format(job_id)) 14 | 15 | print("GPU log directory is {}".format(log_dir)) 16 | os.makedirs(log_dir, exist_ok=True) 17 | for i in range(num_gpus): 18 | argslist.append('--rank={}'.format(i)) 19 | stdout = None if i == 0 else open("{}/{}_GPU_{}.log".format(log_dir, job_id, i), 20 | "w") 21 | print(argslist) 22 | p = subprocess.Popen([str(sys.executable)]+argslist, stdout=stdout) 23 | workers.append(p) 24 | argslist = argslist[:-1] 25 | 26 | for p in workers: 27 | p.wait() 28 | -------------------------------------------------------------------------------- /tests/test_hf_hub.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 unittest 8 | 9 | import torch 10 | 11 | try: 12 | import huggingface_hub 13 | except ImportError: 14 | huggingface_hub = None 15 | 16 | from fairseq.checkpoint_utils import load_model_ensemble_and_task_from_hf_hub 17 | 18 | 19 | @unittest.skipIf(not huggingface_hub, "Requires huggingface_hub install") 20 | class TestHuggingFaceHub(unittest.TestCase): 21 | @torch.no_grad() 22 | def test_hf_fastspeech2(self): 23 | hf_model_id = "facebook/fastspeech2-en-ljspeech" 24 | models, cfg, task = load_model_ensemble_and_task_from_hf_hub(hf_model_id) 25 | self.assertTrue(len(models) > 0) 26 | 27 | 28 | if __name__ == "__main__": 29 | unittest.main() 30 | -------------------------------------------------------------------------------- /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/wmt21/README.md: -------------------------------------------------------------------------------- 1 | # WMT 21 2 | 3 | This page provides pointers to the models of Facebook AI's WMT'21 news translation task submission [(Tran et al., 2021)](https://arxiv.org/abs/2108.03265). 4 | 5 | ## Single best dense models 6 | 7 | Model | Description | Download 8 | ---|---|--- 9 | `wmt21.dense-24-wide.X-En` | X-En | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt21.dense-24-wide.X-En.tar.gz) 10 | `wmt21.dense-24-wide.En-X` | En-X | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt21.dense-24-wide.En-X.tar.gz) 11 | 12 | ## Example usage 13 | 14 | See eval.sh 15 | 16 | 17 | ## Citation 18 | ```bibtex 19 | @inproceedings{tran2021facebook 20 | title={Facebook AI’s WMT21 News Translation Task Submission}, 21 | author={Chau Tran and Shruti Bhosale and James Cross and Philipp Koehn and Sergey Edunov and Angela Fan}, 22 | booktitle={Proc. of WMT}, 23 | year={2021}, 24 | } 25 | ``` 26 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/retri/videoclip/test_vttqa_zs.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: MSRVTTQAAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: MSRVTTQAMetaProcessor 9 | test_path: data/msrvtt-qa/MSR_MC_test.csv 10 | vfeat_dir: data/feat/feat_vtt_s3d 11 | text_processor: MSRVTTQATextProcessor 12 | num_iso_layer: 12 13 | max_video_len: 32 14 | max_len: 96 15 | fairseq: 16 | dataset: 17 | batch_size: 256 18 | valid_subset: test 19 | num_workers: 2 20 | common_eval: 21 | path: runs/retri/videoclip/checkpoint_best.pt 22 | model: 23 | model_cls: MMFusionSeparate 24 | mm_encoder_cls: null 25 | video_encoder_cls: MMBertForEncoder 26 | text_encoder_cls: BertModel 27 | num_hidden_video_layers: 6 28 | eval: 29 | save_path: runs/retri/videoclip/vttqa_zs/eval 30 | metric: QAMetric 31 | predictor: QAPredictor 32 | -------------------------------------------------------------------------------- /examples/MMPT/projects/retri/videoclip/test_didemo_zs.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: DiDeMoAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: DiDeMoMetaProcessor 9 | test_path: data/didemo/test_data.json 10 | vfeat_dir: data/feat/feat_didemo_s3d 11 | text_processor: DiDeMoTextProcessor 12 | num_iso_layer: 12 13 | max_video_len: 32 14 | max_len: 96 15 | fairseq: 16 | dataset: 17 | batch_size: 256 18 | valid_subset: test 19 | num_workers: 2 20 | common_eval: 21 | path: runs/retri/videoclip/checkpoint_best.pt 22 | model: 23 | model_cls: MMFusionSeparate 24 | mm_encoder_cls: null 25 | video_encoder_cls: MMBertForEncoder 26 | text_encoder_cls: BertModel 27 | num_hidden_video_layers: 6 28 | eval: 29 | save_path: runs/retri/videoclip/didemo_zs/eval 30 | metric: DiDeMoMetric 31 | predictor: DiDeMoPredictor 32 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/retri/videoclip/test_vtt_zs.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: DSAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: MSRVTTMetaProcessor 9 | test_path: data/msrvtt/MSRVTT_JSFUSION_test.csv 10 | vfeat_dir: data/feat/feat_vtt_s3d 11 | text_processor: MSRVTTTextProcessor 12 | num_iso_layer: 12 13 | max_video_len: 32 14 | max_len: 96 15 | fairseq: 16 | dataset: 17 | batch_size: 256 18 | valid_subset: test 19 | num_workers: 2 20 | common_eval: 21 | path: runs/retri/videoclip/checkpoint_best.pt 22 | model: 23 | model_cls: MMFusionSeparate 24 | mm_encoder_cls: null 25 | video_encoder_cls: MMBertForEncoder 26 | text_encoder_cls: BertModel 27 | num_hidden_video_layers: 6 28 | eval: 29 | save_path: runs/retri/videoclip/vtt_zs/eval 30 | metric: RetrievalMetric 31 | predictor: RetrievalPredictor 32 | -------------------------------------------------------------------------------- /examples/MMPT/projects/retri/videoclip/test_vttqa_videoclip.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: MSRVTTQAAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: MSRVTTQAMetaProcessor 9 | test_path: data/msrvtt-qa/MSR_MC_test.csv 10 | vfeat_dir: data/feat/feat_vtt_s3d 11 | text_processor: MSRVTTQATextProcessor 12 | num_iso_layer: 12 13 | max_video_len: 32 14 | max_len: 96 15 | fairseq: 16 | dataset: 17 | batch_size: 256 18 | valid_subset: test 19 | num_workers: 2 20 | common_eval: 21 | path: runs/retri/videoclip/vttqa/checkpoint_last.pt 22 | model: 23 | model_cls: MMFusionSeparate 24 | mm_encoder_cls: null 25 | video_encoder_cls: MMBertForEncoder 26 | text_encoder_cls: BertModel 27 | num_hidden_video_layers: 6 28 | eval: 29 | save_path: runs/retri/videoclip/vttqa/eval 30 | metric: QAMetric 31 | predictor: QAPredictor 32 | -------------------------------------------------------------------------------- /tests/speech/test_xm_transformer.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 tests.speech import TestFairseqSpeech 8 | 9 | 10 | class TestXMTransformer(TestFairseqSpeech): 11 | def setUp(self): 12 | self.set_up_sotasty_es_en() 13 | 14 | def test_sotasty_es_en_600m_checkpoint(self): 15 | self.base_test( 16 | ckpt_name="xm_transformer_600m_es_en_md.pt", 17 | reference_score=30.42, 18 | score_delta=0.2, 19 | max_tokens=3_000_000, 20 | max_positions=(1_000_000, 1_024), 21 | dataset="sotasty_es_en_test_ted", 22 | arg_overrides={"config_yaml": "cfg_es_en.yaml"}, 23 | score_type="bleu", 24 | ) 25 | 26 | 27 | if __name__ == "__main__": 28 | unittest.main() 29 | -------------------------------------------------------------------------------- /examples/MMPT/projects/retri/videoclip/test_vtt_videoclip.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: DSAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: MSRVTTMetaProcessor 9 | test_path: data/msrvtt/MSRVTT_JSFUSION_test.csv 10 | vfeat_dir: data/feat/feat_vtt_s3d 11 | text_processor: MSRVTTTextProcessor 12 | num_iso_layer: 12 13 | max_video_len: 32 14 | max_len: 96 15 | fairseq: 16 | dataset: 17 | batch_size: 256 18 | valid_subset: test 19 | num_workers: 2 20 | common_eval: 21 | path: runs/retri/videoclip/vtt/checkpoint_last.pt 22 | model: 23 | model_cls: MMFusionSeparate 24 | mm_encoder_cls: null 25 | video_encoder_cls: MMBertForEncoder 26 | text_encoder_cls: BertModel 27 | num_hidden_video_layers: 6 28 | eval: 29 | save_path: runs/retri/videoclip/vtt/eval 30 | metric: RetrievalMetric 31 | predictor: RetrievalPredictor 32 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/mtm/vlm/test_youcook.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: YoucookVideoProcessor 6 | aligner: DSAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: YoucookMetaProcessor 9 | test_path: data/youcook/youcook_val.pkl 10 | trainval_annotation: data/youcook/youcookii_annotations_trainval.json 11 | use_annotation_text: true 12 | vfeat_dir: data/feat/feat_youcook_s3d 13 | text_processor: TextProcessor 14 | num_iso_layer: 12 15 | max_video_len: 32 16 | max_len: 96 17 | fairseq: 18 | dataset: 19 | batch_size: 256 20 | valid_subset: test 21 | num_workers: 2 22 | common_eval: 23 | path: runs/mtm/vlm/youcook/checkpoint_last.pt 24 | model: 25 | model_cls: MMFusionJoint 26 | mm_encoder_cls: MMBertForJoint 27 | use_seg_emb: true 28 | eval: 29 | save_path: runs/mtm/vlm/youcook/eval 30 | metric: RetrievalMetric 31 | predictor: RetrievalPredictor 32 | -------------------------------------------------------------------------------- /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/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/modules/dynamicconv_layer/dynamiconv_cpu.cpp: -------------------------------------------------------------------------------- 1 | #include 2 | #include 3 | 4 | std::vector 5 | dynamicconv_cpu_forward(float* input, float* filters, int padding_l); 6 | 7 | std::vector dynamicconv_cpu_backward( 8 | float* gradOutput, 9 | int padding_l, 10 | float* input, 11 | float* filters); 12 | 13 | std::vector 14 | dynamicconv_forward(float* input, float* filters, int padding_l) { 15 | return dynamicconv_cpu_forward(input, filters, padding_l); 16 | } 17 | 18 | std::vector dynamicconv_backward( 19 | float* gradOutput, 20 | int padding_l, 21 | float* input, 22 | float* filters) { 23 | return dynamicconv_cpu_backward(gradOutput, padding_l, input, filters); 24 | } 25 | 26 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 27 | m.def("forward", &dynamicconv_forward, "dynamicconv forward (CPU)"); 28 | m.def("backward", &dynamicconv_backward, "dynamicconv backward (CPU)"); 29 | } 30 | -------------------------------------------------------------------------------- /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.beam}_th${decoding.beamthreshold}_lmw${decoding.lmweight}_wrd${decoding.wordscore}_sil${decoding.silweight} 9 | sweep: 10 | dir: ${common_eval.results_path} 11 | subdir: beam${decoding.beam}_th${decoding.beamthreshold}_lmw${decoding.lmweight}_wrd${decoding.wordscore}_sil${decoding.silweight} 12 | 13 | task: 14 | _name: hubert_pretraining 15 | single_target: true 16 | fine_tuning: true 17 | data: ??? 18 | normalize: ??? 19 | 20 | decoding: 21 | type: kenlm 22 | lexicon: ??? 23 | lmpath: ??? 24 | beamthreshold: 100 25 | beam: 500 26 | lmweight: 2 27 | wordscore: -1 28 | silweight: 0 29 | unique_wer_file: true 30 | common_eval: 31 | results_path: ??? 32 | path: ??? 33 | post_process: letter 34 | dataset: 35 | max_tokens: 1100000 36 | gen_subset: ??? 37 | -------------------------------------------------------------------------------- /examples/MMPT/projects/mtm/vlm/test_coin.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: COINActionSegmentationAligner 7 | bert_name: bert-base-uncased 8 | test_path: data/coin/COIN.json 9 | meta_processor: COINActionSegmentationMetaProcessor 10 | vfeat_dir: data/feat/feat_coin_s3d 11 | text_processor: COINActionSegmentationTextProcessor 12 | num_iso_layer: 12 13 | sliding_window: 16 14 | sliding_window_size: 32 15 | max_video_len: 32 16 | max_len: 96 17 | fairseq: 18 | dataset: 19 | batch_size: 1 20 | valid_subset: test 21 | num_workers: 2 22 | common_eval: 23 | path: runs/mtm/vlm/coin/checkpoint_best.pt 24 | model: 25 | model_cls: MMFusionActionSegmentation 26 | mm_encoder_cls: MMBertForTokenClassification 27 | use_seg_emb: true 28 | eval: 29 | save_path: runs/mtm/vlm/coin/eval 30 | metric: COINActionSegmentationMetric 31 | predictor: COINPredictor 32 | -------------------------------------------------------------------------------- /examples/MMPT/projects/mtm/vlm/test_youcookcap.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: YoucookVideoProcessor 6 | aligner: DSNLGAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: YoucookNLGMetaProcessor 9 | test_path: data/youcook/val_list.txt 10 | trainval_annotation: data/youcook/youcookii_annotations_trainval.json 11 | vfeat_dir: data/feat/feat_youcook_s3d 12 | text_processor: NLGTextProcessor 13 | max_video_len: 32 14 | max_len: 96 15 | fairseq: 16 | dataset: 17 | batch_size: 256 18 | valid_subset: test 19 | num_workers: 2 20 | common_eval: 21 | path: runs/mtm/vlm/youcookcap/checkpoint_best.pt 22 | model: 23 | model_cls: MMFusionNLG 24 | mm_encoder_cls: MMBertForNLG 25 | max_decode_length: 24 26 | use_seg_emb: true 27 | eval: 28 | save_path: runs/mtm/vlm/youcookcap/eval 29 | metric: NLGMetric 30 | predictor: NLGPredictor 31 | gen_param: 32 | num_beams: 5 33 | -------------------------------------------------------------------------------- /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.beam}_th${decoding.beamthreshold}_lmw${decoding.lmweight}_wrd${decoding.wordscore}_sil${decoding.silweight} 9 | sweep: 10 | dir: ${common_eval.results_path} 11 | subdir: beam${decoding.beam}_th${decoding.beamthreshold}_lmw${decoding.lmweight}_wrd${decoding.wordscore}_sil${decoding.silweight} 12 | 13 | task: 14 | _name: hubert_pretraining 15 | single_target: true 16 | fine_tuning: true 17 | data: ??? 18 | normalize: ??? 19 | 20 | decoding: 21 | type: fairseqlm 22 | lexicon: ??? 23 | lmpath: ??? 24 | beamthreshold: 25 25 | beam: 500 26 | lmweight: 2 27 | wordscore: -1 28 | silweight: 0 29 | unique_wer_file: true 30 | common_eval: 31 | results_path: ??? 32 | path: ??? 33 | post_process: letter 34 | dataset: 35 | max_tokens: 1100000 36 | gen_subset: ??? 37 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | static PyMethodDef method_def[] = {{NULL, NULL, 0, NULL}}; // NOLINT 12 | 13 | static struct PyModuleDef module_def = { 14 | PyModuleDef_HEAD_INIT, 15 | "libbleu", /* name of module */ 16 | // NOLINTNEXTLINE 17 | NULL, /* module documentation, may be NULL */ 18 | -1, /* size of per-interpreter state of the module, 19 | or -1 if the module keeps state in global variables. */ 20 | method_def}; // NOLINT 21 | 22 | #if PY_MAJOR_VERSION == 2 23 | PyMODINIT_FUNC init_libbleu() 24 | #else 25 | PyMODINIT_FUNC PyInit_libbleu() 26 | #endif 27 | { 28 | PyObject* m = PyModule_Create(&module_def); 29 | if (!m) { 30 | return NULL; 31 | } 32 | return m; 33 | } 34 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/mmpt/tasks/vlmtask.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 | import torch 6 | 7 | from .task import Task 8 | 9 | 10 | class VLMTask(Task): 11 | """A VLM task for reproducibility. 12 | the collator split subsamples into two sub-batches. 13 | This has should have no logic changes. 14 | but changed the randomness in frame masking. 15 | """ 16 | 17 | def flat_subsample(self, tensor): 18 | size = tensor.size() 19 | if len(size) >= 2: 20 | batch_size = size[0] * (size[1] // 2) 21 | expanded_size = ( 22 | (batch_size, 2) + size[2:] if len(size) > 2 23 | else (batch_size, 2) 24 | ) 25 | tensor = tensor.view(expanded_size) 26 | tensor = torch.cat([tensor[:, 0], tensor[:, 1]], dim=0) 27 | return tensor 28 | -------------------------------------------------------------------------------- /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/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/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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/MMPT/scripts/video_feature_extractor/random_sequence_shuffler.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. All Rights Reserved 2 | 3 | import numpy as np 4 | 5 | from torch.utils.data.sampler import Sampler 6 | 7 | 8 | class RandomSequenceSampler(Sampler): 9 | 10 | def __init__(self, n_sample, seq_len): 11 | self.n_sample = n_sample 12 | self.seq_len = seq_len 13 | 14 | def _pad_ind(self, ind): 15 | zeros = np.zeros(self.seq_len - self.n_sample % self.seq_len) 16 | ind = np.concatenate((ind, zeros)) 17 | return ind 18 | 19 | def __iter__(self): 20 | idx = np.arange(self.n_sample) 21 | if self.n_sample % self.seq_len != 0: 22 | idx = self._pad_ind(idx) 23 | idx = np.reshape(idx, (-1, self.seq_len)) 24 | np.random.shuffle(idx) 25 | idx = np.reshape(idx, (-1)) 26 | return iter(idx.astype(int)) 27 | 28 | def __len__(self): 29 | return self.n_sample + (self.seq_len - self.n_sample % self.seq_len) 30 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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/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/attention_head_selection/src/loss/attention_head_selection.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 math 7 | 8 | import torch 9 | from torch.nn.modules.loss import _Loss 10 | 11 | 12 | class HeadSelectionLoss(_Loss): 13 | 14 | def __init__(self, args): 15 | super().__init__() 16 | self.args = args 17 | self.kl_weight = getattr(args, "kl_weight", 0.0) 18 | 19 | def forward(self, head_samples, sample_sizes, prior=0.5, eps=1e-7): 20 | """ 21 | head_scores: (num_tasks, num_layers, num_heads) 22 | sample_sizes: (num_tasks, ) 23 | """ 24 | kl_loss = (head_samples * (torch.log(head_samples + eps) - math.log(prior))).sum(-1).sum(-1) 25 | kl_loss /= (torch.numel(head_samples) / head_samples.size(0)) 26 | kl_loss = self.kl_weight * torch.matmul(kl_loss, sample_sizes) 27 | return kl_loss 28 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/speech_text_joint_to_text/configs/mustc_noise.list: -------------------------------------------------------------------------------- 1 | "(Applause) NOISE 2 | "(Laughter) VOICE 3 | "(Laughter)" VOICE 4 | (Applause) NOISE 5 | (Applause). NOISE 6 | (Audience) VOICE 7 | (Audio) NOISE 8 | (Beat) NOISE 9 | (Beatboxing) VOICE 10 | (Beep) NOISE 11 | (Beeps) NOISE 12 | (Cheering) VOICE 13 | (Cheers) VOICE 14 | (Claps) NOISE 15 | (Clicking) NOISE 16 | (Clunk) NOISE 17 | (Coughs) NOISE 18 | (Drums) NOISE 19 | (Explosion) NOISE 20 | (Gasps) VOICE 21 | (Guitar) NOISE 22 | (Honk) NOISE 23 | (Laugher) VOICE 24 | (Laughing) VOICE 25 | (Laughs) VOICE 26 | (Laughter) VOICE 27 | (Laughter). VOICE 28 | (Laughter)... VOICE 29 | (Mumbling) VOICE 30 | (Music) NOISE 31 | (Noise) NOISE 32 | (Recording) VOICE 33 | (Ringing) NOISE 34 | (Shouts) VOICE 35 | (Sigh) VOICE 36 | (Sighs) VOICE 37 | (Silence) NOISE 38 | (Singing) VOICE 39 | (Sings) VOICE 40 | (Spanish) VOICE 41 | (Static) NOISE 42 | (Tones) NOISE 43 | (Trumpet) NOISE 44 | (Video) NOISE 45 | (Video): NOISE 46 | (Voice-over) NOISE 47 | (Whistle) NOISE 48 | (Whistling) NOISE 49 | (video): NOISE 50 | -------------------------------------------------------------------------------- /examples/MMPT/projects/retri/videoclip/test_coin_zs.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: COINActionSegmentationAligner 7 | bert_name: bert-base-uncased 8 | test_path: data/coin/COIN.json 9 | meta_processor: COINActionSegmentationMetaProcessor 10 | vfeat_dir: data/feat/feat_coin_s3d 11 | text_processor: COINActionSegmentationTextProcessor 12 | num_iso_layer: 12 13 | sliding_window: 16 14 | sliding_window_size: 32 15 | max_video_len: 32 16 | max_len: 96 17 | fairseq: 18 | dataset: 19 | batch_size: 1 20 | valid_subset: test 21 | num_workers: 2 22 | common_eval: 23 | path: runs/retri/videoclip/checkpoint_best.pt 24 | model: 25 | model_cls: MMFusionSeparate 26 | mm_encoder_cls: null 27 | video_encoder_cls: MMBertForEncoder 28 | text_encoder_cls: BertModel 29 | num_hidden_video_layers: 6 30 | eval: 31 | save_path: runs/retri/videoclip/coin_zs/eval 32 | metric: COINActionSegmentationMetric 33 | predictor: COINZSPredictor 34 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/retri/videoclip/test_youcook_zs.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: YoucookVideoProcessor 6 | aligner: DSAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: YoucookMetaProcessor 9 | test_path: data/youcook/youcook_val.pkl 10 | trainval_annotation: data/youcook/youcookii_annotations_trainval.json 11 | use_annotation_text: true 12 | vfeat_dir: data/feat/feat_youcook_s3d 13 | text_processor: TextProcessor 14 | num_iso_layer: 12 15 | max_video_len: 32 16 | max_len: 96 17 | fairseq: 18 | dataset: 19 | batch_size: 256 20 | valid_subset: test 21 | num_workers: 2 22 | common_eval: 23 | path: runs/retri/videoclip/checkpoint_best.pt 24 | model: 25 | model_cls: MMFusionSeparate 26 | mm_encoder_cls: null 27 | video_encoder_cls: MMBertForEncoder 28 | text_encoder_cls: BertModel 29 | num_hidden_video_layers: 6 30 | eval: 31 | save_path: runs/retri/videoclip/youcook_zs/eval 32 | metric: RetrievalMetric 33 | predictor: RetrievalPredictor 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/MMPT/projects/retri/videoclip/test_youcook_videoclip.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: YoucookVideoProcessor 6 | aligner: DSAligner 7 | bert_name: bert-base-uncased 8 | meta_processor: YoucookMetaProcessor 9 | test_path: data/youcook/youcook_val.pkl 10 | trainval_annotation: data/youcook/youcookii_annotations_trainval.json 11 | use_annotation_text: true 12 | vfeat_dir: data/feat/feat_youcook_s3d 13 | text_processor: TextProcessor 14 | num_iso_layer: 12 15 | max_video_len: 32 16 | max_len: 96 17 | fairseq: 18 | dataset: 19 | batch_size: 256 20 | valid_subset: test 21 | num_workers: 2 22 | common_eval: 23 | path: runs/retri/videoclip/youcook/checkpoint_last.pt 24 | model: 25 | model_cls: MMFusionSeparate 26 | mm_encoder_cls: null 27 | video_encoder_cls: MMBertForEncoder 28 | text_encoder_cls: BertModel 29 | num_hidden_video_layers: 6 30 | eval: 31 | save_path: runs/retri/videoclip/youcook/eval 32 | metric: RetrievalMetric 33 | predictor: RetrievalPredictor 34 | -------------------------------------------------------------------------------- /examples/MMPT/mmpt/tasks/milncetask.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 .task import Task 9 | 10 | 11 | class MILNCETask(Task): 12 | def reshape_subsample(self, sample): 13 | if ( 14 | hasattr(self.config.dataset, "subsampling") 15 | and self.config.dataset.subsampling is not None 16 | and self.config.dataset.subsampling > 1 17 | ): 18 | for key in sample: 19 | if torch.is_tensor(sample[key]): 20 | tensor = self.flat_subsample(sample[key]) 21 | if key in ["caps", "cmasks"]: 22 | size = tensor.size() 23 | batch_size = size[0] * size[1] 24 | expanded_size = (batch_size,) + size[2:] 25 | tensor = tensor.view(expanded_size) 26 | sample[key] = tensor 27 | return sample 28 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /examples/MMPT/projects/retri/videoclip/test_coin_videoclip.yaml: -------------------------------------------------------------------------------- 1 | slurm_config: big 2 | task_type: local_predict 3 | dataset: 4 | split: test 5 | video_processor: VideoProcessor 6 | aligner: COINActionSegmentationAligner 7 | bert_name: bert-base-uncased 8 | test_path: data/coin/COIN.json 9 | meta_processor: COINActionSegmentationMetaProcessor 10 | vfeat_dir: data/feat/feat_coin_s3d 11 | text_processor: COINActionSegmentationTextProcessor 12 | num_iso_layer: 12 13 | sliding_window: 16 14 | sliding_window_size: 32 15 | max_video_len: 32 16 | max_len: 96 17 | fairseq: 18 | dataset: 19 | batch_size: 1 20 | valid_subset: test 21 | num_workers: 2 22 | common_eval: 23 | path: runs/retri/videoclip/coin/checkpoint_best.pt 24 | model: 25 | model_cls: MMFusionSeparateActionSegmentation 26 | mm_encoder_cls: null 27 | video_encoder_cls: MMBertForTokenClassification 28 | text_encoder_cls: BertModel 29 | num_hidden_video_layers: 6 30 | eval: 31 | save_path: runs/retri/videoclip/coin/eval 32 | metric: COINActionSegmentationMetric 33 | predictor: COINPredictor 34 | -------------------------------------------------------------------------------- /examples/textless_nlp/gslm/speech2unit/pretrained/logmel_feature_reader.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 soundfile as sf 7 | import torch 8 | import torchaudio.compliance.kaldi as kaldi 9 | 10 | 11 | class LogMelFeatureReader: 12 | """ 13 | Wrapper class to run inference on HuBERT model. 14 | Helps extract features for a given audio file. 15 | """ 16 | 17 | def __init__(self, *args, **kwargs): 18 | self.num_mel_bins = kwargs.get("num_mel_bins", 80) 19 | self.frame_length = kwargs.get("frame_length", 25.0) 20 | 21 | def get_feats(self, file_path): 22 | wav, sr = sf.read(file_path) 23 | feats = torch.from_numpy(wav).float() 24 | feats = kaldi.fbank( 25 | feats.unsqueeze(0), 26 | num_mel_bins=self.num_mel_bins, 27 | frame_length=self.frame_length, 28 | sample_frequency=sr, 29 | ) 30 | return feats 31 | -------------------------------------------------------------------------------- /examples/wmt21/scripts/replace-unicode-punctuation.perl: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env perl 2 | # 3 | # This file is part of moses. Its use is licensed under the GNU Lesser General 4 | # Public License version 2.1 or, at your option, any later version. 5 | 6 | use warnings; 7 | use strict; 8 | 9 | while (@ARGV) { 10 | $_ = shift; 11 | /^-b$/ && ($| = 1, next); # not buffered (flush each line) 12 | } 13 | 14 | #binmode(STDIN, ":utf8"); 15 | #binmode(STDOUT, ":utf8"); 16 | 17 | while() { 18 | s/，/,/g; 19 | s/。 */. /g; 20 | s/、/,/g; 21 | s/”/"/g; 22 | s/“/"/g; 23 | s/∶/:/g; 24 | s/：/:/g; 25 | s/？/\?/g; 26 | s/《/"/g; 27 | s/》/"/g; 28 | s/）/\)/g; 29 | s/！/\!/g; 30 | s/（/\(/g; 31 | s/；/;/g; 32 | s/１/1/g; 33 | s/」/"/g; 34 | s/「/"/g; 35 | s/０/0/g; 36 | s/３/3/g; 37 | s/２/2/g; 38 | s/５/5/g; 39 | s/６/6/g; 40 | s/９/9/g; 41 | s/７/7/g; 42 | s/８/8/g; 43 | s/４/4/g; 44 | s/． */. /g; 45 | s/～/\~/g; 46 | s/’/\'/g; 47 | s/…/\.\.\./g; 48 | s/━/\-/g; 49 | s/〈/\/g; 51 | s/【/\[/g; 52 | s/】/\]/g; 53 | s/％/\%/g; 54 | print $_; 55 | } 56 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | def test_no_iopath(self): 12 | from .test_reproducibility import TestReproducibility 13 | 14 | with mock.patch.dict("sys.modules", {"iopath": None}): 15 | # reuse reproducibility tests, which are e2e tests that should cover 16 | # most checkpoint related functionality 17 | TestReproducibility._test_reproducibility(self, "test_reproducibility") 18 | 19 | def test_no_supports_rename(self): 20 | from .test_reproducibility import TestReproducibility 21 | 22 | with mock.patch("fairseq.file_io.PathManager.supports_rename") as mock_fn: 23 | mock_fn.return_value = False 24 | TestReproducibility._test_reproducibility(self, "test_reproducibility") 25 | 26 | 27 | if __name__ == "__main__": 28 | unittest.main() 29 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | 7 | from dataclasses import dataclass 8 | 9 | from fairseq.dataclass import FairseqDataclass 10 | from fairseq.scoring import BaseScorer, register_scorer 11 | 12 | 13 | @dataclass 14 | class ChrFScorerConfig(FairseqDataclass): 15 | pass 16 | 17 | 18 | @register_scorer("chrf", dataclass=ChrFScorerConfig) 19 | class ChrFScorer(BaseScorer): 20 | def __init__(self, args): 21 | super(ChrFScorer, self).__init__(args) 22 | import sacrebleu 23 | 24 | self.sacrebleu = sacrebleu 25 | 26 | def add_string(self, ref, pred): 27 | self.ref.append(ref) 28 | self.pred.append(pred) 29 | 30 | def score(self, order=4): 31 | return self.result_string(order).score 32 | 33 | def result_string(self, order=4): 34 | if order != 4: 35 | raise NotImplementedError 36 | return self.sacrebleu.corpus_chrf(self.pred, [self.ref]).format() 37 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) 2023 Rongjie Huang 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 | -------------------------------------------------------------------------------- /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/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/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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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_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 | -------------------------------------------------------------------------------- /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 | --------------------------------------------------------------------------------