├── tests ├── __init__.py ├── speech_recognition │ ├── __init__.py │ ├── test_cross_entropy.py │ └── test_collaters.py ├── test_iterators.py ├── test_file_io.py ├── test_character_token_embedder.py ├── test_convtbc.py ├── test_multihead_attention.py ├── test_concat_dataset.py ├── test_memory_efficient_fp16.py ├── test_sparse_multihead_attention.py └── test_metrics.py ├── fairseq_cli └── __init__.py ├── scripts ├── __init__.py ├── spm_train.py ├── compound_split_bleu.sh ├── sacrebleu_pregen.sh ├── convert_dictionary.lua ├── compare_namespaces.py ├── read_binarized.py ├── spm_decode.py ├── shard_docs.py ├── count_docs.py └── split_train_valid_docs.py ├── fairseq ├── logging │ └── __init__.py ├── data │ ├── audio │ │ └── __init__.py │ ├── num_samples_dataset.py │ ├── id_dataset.py │ ├── legacy │ │ ├── __init__.py │ │ └── masked_lm_dictionary.py │ ├── offset_tokens_dataset.py │ ├── roll_dataset.py │ ├── strip_token_dataset.py │ ├── encoders │ │ ├── space_tokenizer.py │ │ ├── characters.py │ │ ├── nltk_tokenizer.py │ │ ├── __init__.py │ │ ├── bytes.py │ │ ├── utils.py │ │ ├── fastbpe.py │ │ ├── byte_bpe.py │ │ ├── hf_byte_bpe.py │ │ ├── sentencepiece_bpe.py │ │ ├── gpt2_bpe.py │ │ ├── subword_nmt_bpe.py │ │ ├── byte_utils.py │ │ ├── hf_bert_bpe.py │ │ └── moses_tokenizer.py │ ├── raw_label_dataset.py │ ├── lru_cache_dataset.py │ ├── sort_dataset.py │ ├── list_dataset.py │ ├── numel_dataset.py │ ├── colorize_dataset.py │ ├── pad_dataset.py │ ├── truncate_dataset.py │ ├── prepend_dataset.py │ ├── append_token_dataset.py │ ├── prepend_token_dataset.py │ ├── replace_dataset.py │ ├── base_wrapper_dataset.py │ ├── concat_sentences_dataset.py │ ├── data_utils_fast.pyx │ ├── subsample_dataset.py │ ├── fairseq_dataset.py │ ├── plasma_utils.py │ └── lm_context_window_dataset.py ├── models │ ├── huggingface │ │ └── __init__.py │ ├── bart │ │ └── __init__.py │ ├── nat │ │ └── __init__.py │ ├── roberta │ │ ├── __init__.py │ │ ├── model_camembert.py │ │ └── model_xlmr.py │ ├── fairseq_encoder.py │ ├── composite_encoder.py │ ├── model_utils.py │ └── distributed_fairseq_model.py ├── model_parallel │ ├── __init__.py │ ├── models │ │ └── __init__.py │ ├── modules │ │ ├── __init__.py │ │ └── transformer_layer.py │ ├── criterions │ │ └── __init__.py │ └── megatron_trainer.py ├── modules │ ├── lightconv_layer │ │ ├── __init__.py │ │ ├── setup.py │ │ ├── lightconv_cuda.cpp │ │ └── lightconv_cuda.cuh │ ├── dynamicconv_layer │ │ ├── __init__.py │ │ ├── setup.py │ │ ├── dynamiconv_cpu.cpp │ │ ├── dynamicconv_cuda.cuh │ │ └── dynamicconv_cuda.cpp │ ├── grad_multiply.py │ ├── unfold.py │ ├── gelu.py │ ├── fp32_group_norm.py │ ├── scalar_bias.py │ ├── positional_embedding.py │ ├── conv_tbc.py │ ├── layer_norm.py │ ├── sparse_transformer_sentence_encoder_layer.py │ ├── cross_entropy.py │ ├── beamable_mm.py │ ├── learned_positional_embedding.py │ ├── adaptive_input.py │ ├── __init__.py │ └── transformer_sentence_encoder_layer.py ├── benchmark │ └── __init__.py ├── tokenizer.py ├── clib │ ├── libnat_cuda │ │ ├── edit_dist.h │ │ └── binding.cpp │ └── libbleu │ │ └── module.cpp ├── criterions │ ├── __init__.py │ ├── cross_entropy.py │ └── masked_lm.py ├── optim │ ├── lr_scheduler │ │ ├── __init__.py │ │ ├── fairseq_lr_scheduler.py │ │ ├── fixed_schedule.py │ │ ├── triangular_lr_scheduler.py │ │ └── inverse_square_root_schedule.py │ ├── __init__.py │ ├── adagrad.py │ ├── sgd.py │ ├── adadelta.py │ └── fused_lamb.py ├── __init__.py ├── tasks │ ├── translation_from_pretrained_xlm.py │ ├── audio_pretraining.py │ └── __init__.py ├── pdb.py ├── incremental_decoding_utils.py └── registry.py ├── docs ├── docutils.conf ├── requirements.txt ├── _static │ └── theme_overrides.css ├── modules.rst ├── Makefile ├── criterions.rst ├── make.bat ├── optim.rst ├── lr_scheduler.rst ├── index.rst ├── data.rst ├── tasks.rst ├── command_line_tools.rst └── overview.rst ├── pyproject.toml ├── score.py ├── train.py ├── eval_lm.py ├── generate.py ├── validate.py ├── interactive.py ├── preprocess.py ├── LICENSE └── hubconf.py /tests/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /fairseq_cli/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /scripts/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /fairseq/logging/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /fairseq/data/audio/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /tests/speech_recognition/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /docs/docutils.conf: -------------------------------------------------------------------------------- 1 | [writers] 2 | option-limit=0 3 | -------------------------------------------------------------------------------- /docs/requirements.txt: -------------------------------------------------------------------------------- 1 | sphinx<2.0 2 | sphinx-argparse 3 | -------------------------------------------------------------------------------- /pyproject.toml: -------------------------------------------------------------------------------- 1 | [build-system] 2 | requires = ["setuptools", "wheel", "cython"] 3 | build-backend = "setuptools.build_meta" 4 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | from .hf_gpt2 import * # 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, modules, models # 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 | -------------------------------------------------------------------------------- /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/modules/dynamicconv_layer/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from .dynamicconv_layer import DynamicconvLayer # noqa 7 | -------------------------------------------------------------------------------- /fairseq/models/nat/__init__.py: -------------------------------------------------------------------------------- 1 | from .fairseq_nat_model import * 2 | from .nonautoregressive_transformer import * 3 | from .nat_crf_transformer import * 4 | from .iterative_nonautoregressive_transformer import * 5 | from .cmlm_transformer import * 6 | from .levenshtein_transformer import * 7 | from .insertion_transformer import * 8 | -------------------------------------------------------------------------------- /score.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 | from fairseq_cli.score import cli_main 8 | 9 | 10 | if __name__ == '__main__': 11 | cli_main() 12 | -------------------------------------------------------------------------------- /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 | from fairseq_cli.train import cli_main 8 | 9 | 10 | if __name__ == '__main__': 11 | cli_main() 12 | -------------------------------------------------------------------------------- /eval_lm.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 | from fairseq_cli.eval_lm import cli_main 8 | 9 | 10 | if __name__ == '__main__': 11 | cli_main() 12 | -------------------------------------------------------------------------------- /generate.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 | from fairseq_cli.generate import cli_main 8 | 9 | 10 | if __name__ == '__main__': 11 | cli_main() 12 | -------------------------------------------------------------------------------- /validate.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 | from fairseq_cli.validate import cli_main 8 | 9 | 10 | if __name__ == '__main__': 11 | cli_main() 12 | -------------------------------------------------------------------------------- /interactive.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 | from fairseq_cli.interactive import cli_main 8 | 9 | 10 | if __name__ == '__main__': 11 | cli_main() 12 | -------------------------------------------------------------------------------- /preprocess.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 | from fairseq_cli.preprocess import cli_main 8 | 9 | 10 | if __name__ == '__main__': 11 | cli_main() 12 | -------------------------------------------------------------------------------- /fairseq/benchmark/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | # import models/tasks to register them 7 | from . import ( # noqa 8 | dummy_lm, 9 | dummy_masked_lm, 10 | dummy_model, 11 | ) 12 | -------------------------------------------------------------------------------- /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 .model_camembert import * # noqa 9 | from .model_xlmr import * # 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 | SPACE_NORMALIZER = re.compile(r"\s+") 9 | 10 | 11 | def tokenize_line(line): 12 | line = SPACE_NORMALIZER.sub(" ", line) 13 | line = line.strip() 14 | return line.split() 15 | -------------------------------------------------------------------------------- /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 | 11 | def __getitem__(self, index): 12 | return 1 13 | 14 | def __len__(self): 15 | return 0 16 | 17 | def collater(self, samples): 18 | return sum(samples) 19 | -------------------------------------------------------------------------------- /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/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 | for file in 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('fairseq.model_parallel.models.' + model_name) 14 | -------------------------------------------------------------------------------- /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 | 13 | def __getitem__(self, index): 14 | return index 15 | 16 | def __len__(self): 17 | return 0 18 | 19 | def collater(self, samples): 20 | return torch.tensor(samples) 21 | -------------------------------------------------------------------------------- /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 .masked_lm_dictionary import BertDictionary, MaskedLMDictionary 7 | from .block_pair_dataset import BlockPairDataset 8 | from .masked_lm_dataset import MaskedLMDataset 9 | 10 | __all__ = [ 11 | 'BertDictionary', 12 | 'BlockPairDataset', 13 | 'MaskedLMDataset', 14 | 'MaskedLMDictionary', 15 | ] 16 | -------------------------------------------------------------------------------- /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 | 11 | def __init__(self, dataset, offset): 12 | super().__init__(dataset) 13 | self.offset = offset 14 | 15 | def __getitem__(self, idx): 16 | return self.dataset[idx] + self.offset 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 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | 6 | from .multihead_attention import ModelParallelMultiheadAttention 7 | from .transformer_layer import ModelParallelTransformerEncoderLayer, ModelParallelTransformerDecoderLayer 8 | 9 | __all__ = [ 10 | 'ModelParallelMultiheadAttention', 11 | 'ModelParallelTransformerEncoderLayer', 12 | 'ModelParallelTransformerDecoderLayer', 13 | ] 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 | 13 | def __init__(self, dataset, shifts): 14 | super().__init__(dataset) 15 | self.shifts = shifts 16 | 17 | def __getitem__(self, index): 18 | item = self.dataset[index] 19 | return torch.roll(item, self.shifts) 20 | -------------------------------------------------------------------------------- /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 | 11 | def __init__(self, dataset, id_to_strip): 12 | super().__init__(dataset) 13 | self.id_to_strip = id_to_strip 14 | 15 | def __getitem__(self, index): 16 | item = self.dataset[index] 17 | return item[item.ne(self.id_to_strip)] 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 os.listdir(os.path.dirname(__file__)): 12 | if file.endswith('.py') and not file.startswith('_'): 13 | module = file[:file.find('.py')] 14 | importlib.import_module('fairseq.model_parallel.criterions.' + module) 15 | -------------------------------------------------------------------------------- /fairseq/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(x, (0, 0, 0, 0, padding_l, kernel_size - 1 - padding_l), value=pad_value) 14 | x = x.as_strided((T, B, C, kernel_size), (B*C, C, 1, B*C)) 15 | else: 16 | x = x.unsqueeze(3) 17 | return x 18 | -------------------------------------------------------------------------------- /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 CUDAExtension, BuildExtension 9 | 10 | setup( 11 | name='lightconv_layer', 12 | ext_modules=[ 13 | CUDAExtension('lightconv_cuda', [ 14 | 'lightconv_cuda.cpp', 15 | 'lightconv_cuda_kernel.cu', 16 | ]), 17 | ], 18 | cmdclass={ 19 | 'build_ext': BuildExtension 20 | }) 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 | 10 | 11 | @register_tokenizer('space') 12 | class SpaceTokenizer(object): 13 | 14 | def __init__(self, source_lang=None, target_lang=None): 15 | self.space_tok = re.compile(r"\s+") 16 | 17 | def encode(self, x: str) -> str: 18 | return self.space_tok.sub(' ', x) 19 | 20 | def decode(self, x: str) -> str: 21 | return x 22 | -------------------------------------------------------------------------------- /fairseq/data/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 | 13 | def __init__(self, labels): 14 | super().__init__() 15 | self.labels = labels 16 | 17 | def __getitem__(self, index): 18 | return self.labels[index] 19 | 20 | def __len__(self): 21 | return len(self.labels) 22 | 23 | def collater(self, samples): 24 | return torch.tensor(samples) 25 | -------------------------------------------------------------------------------- /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) -------------------------------------------------------------------------------- /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 | 13 | def __init__(self, dataset, token=None): 14 | super().__init__(dataset) 15 | 16 | @lru_cache(maxsize=8) 17 | def __getitem__(self, index): 18 | return self.dataset[index] 19 | 20 | @lru_cache(maxsize=8) 21 | def collater(self, samples): 22 | return self.dataset.collater(samples) 23 | -------------------------------------------------------------------------------- /scripts/sacrebleu_pregen.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=$4 13 | 14 | echo 'Cloning Moses github repository (for tokenization scripts)...' 15 | git clone https://github.com/moses-smt/mosesdecoder.git 16 | 17 | SCRIPTS=mosesdecoder/scripts 18 | DETOKENIZER=$SCRIPTS/tokenizer/detokenizer.perl 19 | 20 | grep ^H $GEN \ 21 | | sed 's/^H\-//' \ 22 | | sort -n -k 1 \ 23 | | cut -f 3 \ 24 | | perl $DETOKENIZER -l $TGTLANG \ 25 | | sed "s/ - /-/g" \ 26 | > $GEN.sorted.detok 27 | 28 | sacrebleu --test-set $TESTSET --language-pair "${SRCLANG}-${TGTLANG}" < $GEN.sorted.detok 29 | -------------------------------------------------------------------------------- /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 | 13 | def __init__(self, dataset, sort_order): 14 | super().__init__(dataset) 15 | if not isinstance(sort_order, (list, tuple)): 16 | sort_order = [sort_order] 17 | self.sort_order = sort_order 18 | 19 | assert all(len(so) == len(dataset) for so in sort_order) 20 | 21 | def ordered_indices(self): 22 | return np.lexsort(self.sort_order) 23 | -------------------------------------------------------------------------------- /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 CUDAExtension, BuildExtension 9 | 10 | setup( 11 | name='dynamicconv_layer', 12 | ext_modules=[ 13 | CUDAExtension( 14 | name='dynamicconv_cuda', 15 | sources=[ 16 | 'dynamicconv_cuda.cpp', 17 | 'dynamicconv_cuda_kernel.cu', 18 | ], 19 | ), 20 | ], 21 | cmdclass={ 22 | 'build_ext': BuildExtension 23 | }) 24 | -------------------------------------------------------------------------------- /fairseq/clib/libnat_cuda/edit_dist.h: -------------------------------------------------------------------------------- 1 | /** 2 | * Copyright 2017-present, Facebook, Inc. 3 | * All rights reserved. 4 | * 5 | * This source code is licensed under the license found in the 6 | * LICENSE file in the root directory of this source tree. 7 | */ 8 | 9 | #pragma once 10 | 11 | #include 12 | 13 | torch::Tensor LevenshteinDistanceCuda( 14 | torch::Tensor source, 15 | torch::Tensor target, 16 | torch::Tensor source_length, 17 | torch::Tensor target_length); 18 | 19 | torch::Tensor GenerateDeletionLabelCuda( 20 | torch::Tensor source, 21 | torch::Tensor operations); 22 | 23 | std::pair GenerateInsertionLabelCuda( 24 | torch::Tensor source, 25 | torch::Tensor operations); 26 | -------------------------------------------------------------------------------- /fairseq/modules/gelu.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | """ 6 | See "Gaussian Error Linear Units (GELUs)" by Dan Hendrycks and Kevin Gimpel with 7 | the corresponding GitHub repo: https://github.com/hendrycks/GELUs 8 | """ 9 | 10 | import math 11 | 12 | import torch 13 | import torch.nn as nn 14 | 15 | 16 | def gelu_accurate(x): 17 | if not hasattr(gelu_accurate, "_a"): 18 | gelu_accurate._a = math.sqrt(2 / math.pi) 19 | return ( 20 | 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3)))) 21 | ) 22 | 23 | 24 | def gelu(x: torch.Tensor) -> torch.Tensor: 25 | return torch.nn.functional.gelu(x.float()).type_as(x) 26 | -------------------------------------------------------------------------------- /fairseq/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 | SPACE = chr(32) 10 | SPACE_ESCAPE = chr(9601) 11 | 12 | 13 | @register_bpe('characters') 14 | class Characters(object): 15 | def __init__(self, args): 16 | pass 17 | 18 | @staticmethod 19 | def add_args(parser): 20 | pass 21 | 22 | @staticmethod 23 | def encode(x: str) -> str: 24 | escaped = x.replace(SPACE, SPACE_ESCAPE) 25 | return SPACE.join(list(escaped)) 26 | 27 | @staticmethod 28 | def decode(x: str) -> str: 29 | return x.replace(SPACE, '').replace(SPACE_ESCAPE, SPACE) 30 | -------------------------------------------------------------------------------- /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 | 8 | 9 | @register_tokenizer('nltk') 10 | class NLTKTokenizer(object): 11 | 12 | def __init__(self, source_lang=None, target_lang=None): 13 | try: 14 | from nltk.tokenize import word_tokenize 15 | self.word_tokenize = word_tokenize 16 | except ImportError: 17 | raise ImportError('Please install nltk with: pip install nltk') 18 | 19 | def encode(self, x: str) -> str: 20 | return ' '.join(self.word_tokenize(x)) 21 | 22 | def decode(self, x: str) -> str: 23 | return x 24 | -------------------------------------------------------------------------------- /tests/test_iterators.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 | 8 | from fairseq.data import iterators 9 | 10 | 11 | class TestIterators(unittest.TestCase): 12 | 13 | def test_counting_iterator(self): 14 | x = list(range(10)) 15 | itr = iterators.CountingIterator(x) 16 | self.assertTrue(itr.has_next()) 17 | self.assertEqual(next(itr), 0) 18 | self.assertEqual(next(itr), 1) 19 | itr.skip(3) 20 | self.assertEqual(next(itr), 5) 21 | itr.skip(3) 22 | self.assertEqual(next(itr), 9) 23 | self.assertFalse(itr.has_next()) 24 | 25 | 26 | if __name__ == '__main__': 27 | unittest.main() 28 | -------------------------------------------------------------------------------- /fairseq/modules/fp32_group_norm.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | """ 6 | Layer norm done in fp32 (for fp16 training) 7 | """ 8 | 9 | import torch.nn as nn 10 | import torch.nn.functional as F 11 | 12 | 13 | class Fp32GroupNorm(nn.GroupNorm): 14 | def __init__(self, *args, **kwargs): 15 | super().__init__(*args, **kwargs) 16 | 17 | def forward(self, input): 18 | output = F.group_norm( 19 | input.float(), 20 | self.num_groups, 21 | self.weight.float() if self.weight is not None else None, 22 | self.bias.float() if self.bias is not None else None, 23 | self.eps, 24 | ) 25 | return output.type_as(input) 26 | -------------------------------------------------------------------------------- /fairseq/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 os.listdir(os.path.dirname(__file__)): 27 | if file.endswith('.py') and not file.startswith('_'): 28 | module = file[:file.find('.py')] 29 | importlib.import_module('fairseq.data.encoders.' + module) 30 | -------------------------------------------------------------------------------- /fairseq/data/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 | 11 | def __init__(self, dataset, sizes=None): 12 | super().__init__(dataset) 13 | self._sizes = sizes 14 | 15 | def __iter__(self): 16 | for x in self.dataset: 17 | yield x 18 | 19 | def collater(self, samples): 20 | return samples 21 | 22 | @property 23 | def sizes(self): 24 | return self._sizes 25 | 26 | def num_tokens(self, index): 27 | return self.sizes[index] 28 | 29 | def size(self, index): 30 | return self.sizes[index] 31 | 32 | def set_epoch(self, epoch): 33 | pass 34 | -------------------------------------------------------------------------------- /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/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 | from fairseq import registry 10 | from fairseq.criterions.fairseq_criterion import FairseqCriterion, LegacyFairseqCriterion 11 | 12 | 13 | build_criterion, register_criterion, CRITERION_REGISTRY = registry.setup_registry( 14 | '--criterion', 15 | base_class=FairseqCriterion, 16 | default='cross_entropy', 17 | ) 18 | 19 | 20 | # automatically import any Python files in the criterions/ directory 21 | for file in os.listdir(os.path.dirname(__file__)): 22 | if file.endswith('.py') and not file.startswith('_'): 23 | module = file[:file.find('.py')] 24 | importlib.import_module('fairseq.criterions.' + module) 25 | -------------------------------------------------------------------------------- /fairseq/optim/lr_scheduler/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import importlib 7 | import os 8 | 9 | from fairseq import registry 10 | from fairseq.optim.lr_scheduler.fairseq_lr_scheduler import FairseqLRScheduler 11 | 12 | 13 | build_lr_scheduler, register_lr_scheduler, LR_SCHEDULER_REGISTRY = registry.setup_registry( 14 | '--lr-scheduler', 15 | base_class=FairseqLRScheduler, 16 | default='fixed', 17 | ) 18 | 19 | # automatically import any Python files in the optim/lr_scheduler/ directory 20 | for file in os.listdir(os.path.dirname(__file__)): 21 | if file.endswith('.py') and not file.startswith('_'): 22 | module = file[:file.find('.py')] 23 | importlib.import_module('fairseq.optim.lr_scheduler.' + module) 24 | -------------------------------------------------------------------------------- /fairseq/__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 | __all__ = ['pdb'] 7 | __version__ = '0.9.0' 8 | 9 | import sys 10 | 11 | # backwards compatibility to support `from fairseq.meters import AverageMeter` 12 | from fairseq.logging import meters, metrics, progress_bar # noqa 13 | sys.modules['fairseq.meters'] = meters 14 | sys.modules['fairseq.metrics'] = metrics 15 | sys.modules['fairseq.progress_bar'] = progress_bar 16 | 17 | import fairseq.criterions # noqa 18 | import fairseq.models # noqa 19 | import fairseq.modules # noqa 20 | import fairseq.optim # noqa 21 | import fairseq.optim.lr_scheduler # noqa 22 | import fairseq.pdb # noqa 23 | import fairseq.tasks # noqa 24 | 25 | import fairseq.benchmark # noqa 26 | import fairseq.model_parallel # noqa 27 | -------------------------------------------------------------------------------- /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 | 14 | def __init__(self, dataset, reduce=False): 15 | super().__init__(dataset) 16 | self.reduce = reduce 17 | 18 | def __getitem__(self, index): 19 | item = self.dataset[index] 20 | if torch.is_tensor(item): 21 | return torch.numel(item) 22 | else: 23 | return np.size(item) 24 | 25 | def __len__(self): 26 | return len(self.dataset) 27 | 28 | def collater(self, samples): 29 | if self.reduce: 30 | return sum(samples) 31 | else: 32 | return torch.tensor(samples) 33 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /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 | def __init__(self, dataset, color_getter): 14 | super().__init__(dataset) 15 | self.color_getter = color_getter 16 | 17 | def collater(self, samples): 18 | base_collate = super().collater(samples) 19 | if len(base_collate) > 0: 20 | base_collate["net_input"]["colors"] = torch.tensor( 21 | list(self.color_getter(self.dataset, s["id"]) for s in samples), 22 | dtype=torch.long, 23 | ) 24 | return base_collate 25 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /fairseq/clib/libbleu/module.cpp: -------------------------------------------------------------------------------- 1 | /** 2 | * Copyright 2017-present, Facebook, Inc. 3 | * All rights reserved. 4 | * 5 | * This source code is licensed under the license found in the 6 | * LICENSE file in the root directory of this source tree. 7 | */ 8 | 9 | #include 10 | 11 | 12 | static PyMethodDef method_def[] = { 13 | {NULL, NULL, 0, NULL} 14 | }; 15 | 16 | static struct PyModuleDef module_def = { 17 | PyModuleDef_HEAD_INIT, 18 | "libbleu", /* name of module */ 19 | NULL, /* module documentation, may be NULL */ 20 | -1, /* size of per-interpreter state of the module, 21 | or -1 if the module keeps state in global variables. */ 22 | method_def 23 | }; 24 | 25 | 26 | #if PY_MAJOR_VERSION == 2 27 | PyMODINIT_FUNC init_libbleu() 28 | #else 29 | PyMODINIT_FUNC PyInit_libbleu() 30 | #endif 31 | { 32 | PyObject *m = PyModule_Create(&module_def); 33 | if (!m) { 34 | return NULL; 35 | } 36 | return m; 37 | } 38 | -------------------------------------------------------------------------------- /fairseq/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 | 13 | def __init__(self, dataset, pad_idx, left_pad): 14 | super().__init__(dataset) 15 | self.pad_idx = pad_idx 16 | self.left_pad = left_pad 17 | 18 | def collater(self, samples): 19 | return data_utils.collate_tokens(samples, self.pad_idx, left_pad=self.left_pad) 20 | 21 | 22 | class LeftPadDataset(PadDataset): 23 | 24 | def __init__(self, dataset, pad_idx): 25 | super().__init__(dataset, pad_idx, left_pad=True) 26 | 27 | 28 | class RightPadDataset(PadDataset): 29 | 30 | def __init__(self, dataset, pad_idx): 31 | super().__init__(dataset, pad_idx, left_pad=False) 32 | -------------------------------------------------------------------------------- /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 (byte_encode, smart_byte_decode, 9 | SPACE, SPACE_ESCAPE) 10 | 11 | 12 | @register_bpe('bytes') 13 | class Bytes(object): 14 | def __init__(self, args): 15 | pass 16 | 17 | @staticmethod 18 | def add_args(parser): 19 | pass 20 | 21 | @staticmethod 22 | def encode(x: str) -> str: 23 | encoded = byte_encode(x) 24 | escaped = encoded.replace(SPACE, SPACE_ESCAPE) 25 | return SPACE.join(list(escaped)) 26 | 27 | @staticmethod 28 | def decode(x: str) -> str: 29 | unescaped = x.replace(SPACE, '').replace(SPACE_ESCAPE, SPACE) 30 | return smart_byte_decode(unescaped) 31 | -------------------------------------------------------------------------------- /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 | -------------------------------------------------------------------------------- /fairseq/data/truncate_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 TruncateDataset(BaseWrapperDataset): 12 | 13 | def __init__(self, dataset, truncation_length): 14 | super().__init__(dataset) 15 | assert truncation_length is not None 16 | self.truncation_length = truncation_length 17 | self.dataset = dataset 18 | 19 | def __getitem__(self, index): 20 | item = self.dataset[index] 21 | item_len = item.size(0) 22 | if item_len > self.truncation_length: 23 | item = item[:self.truncation_length] 24 | return item 25 | 26 | @property 27 | def sizes(self): 28 | return np.minimum(self.dataset.sizes, self.truncation_length) 29 | 30 | def __len__(self): 31 | return len(self.dataset) 32 | -------------------------------------------------------------------------------- /fairseq/modules/dynamicconv_layer/dynamiconv_cpu.cpp: -------------------------------------------------------------------------------- 1 | #include 2 | #include 3 | 4 | std::vector dynamicconv_cpu_forward( 5 | float* input, 6 | float* filters, 7 | int padding_l); 8 | 9 | std::vector dynamicconv_cpu_backward( 10 | float* gradOutput, 11 | int padding_l, 12 | float* input, 13 | float* filters); 14 | 15 | std::vector dynamicconv_forward( 16 | float* input, 17 | float* filters, 18 | int padding_l) { 19 | 20 | return dynamicconv_cpu_forward(input, filters, padding_l); 21 | } 22 | 23 | std::vector dynamicconv_backward( 24 | float* gradOutput, 25 | int padding_l, 26 | float* input, 27 | float* filters) { 28 | 29 | return dynamicconv_cpu_backward(gradOutput, padding_l, input, filters); 30 | } 31 | 32 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 33 | m.def("forward", &dynamicconv_forward, "dynamicconv forward (CPU)"); 34 | m.def("backward", &dynamicconv_backward, "dynamicconv backward (CPU)"); 35 | } 36 | -------------------------------------------------------------------------------- /fairseq/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 | def is_beginning_of_word(i): 14 | if i < dictionary.nspecial: 15 | # special elements are always considered beginnings 16 | return True 17 | tok = dictionary[i] 18 | if tok.startswith('madeupword'): 19 | return True 20 | try: 21 | return bpe.is_beginning_of_word(tok) 22 | except ValueError: 23 | return True 24 | mask_whole_words = torch.ByteTensor(list( 25 | map(is_beginning_of_word, range(len(dictionary))) 26 | )) 27 | return mask_whole_words 28 | return None 29 | -------------------------------------------------------------------------------- /fairseq/modules/scalar_bias.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | # 6 | 7 | import torch 8 | 9 | 10 | class ScalarBias(torch.autograd.Function): 11 | """ 12 | Adds a vector of scalars, used in self-attention mechanism to allow 13 | the model to optionally attend to this vector instead of the past 14 | """ 15 | 16 | @staticmethod 17 | def forward(ctx, input, dim, bias_init): 18 | size = list(input.size()) 19 | size[dim] += 1 20 | output = input.new(*size).fill_(bias_init) 21 | output.narrow(dim, 1, size[dim] - 1).copy_(input) 22 | ctx.dim = dim 23 | return output 24 | 25 | @staticmethod 26 | def backward(ctx, grad): 27 | return grad.narrow(ctx.dim, 1, grad.size(ctx.dim) - 1), None, None 28 | 29 | 30 | def scalar_bias(input, dim, bias_init=0): 31 | return ScalarBias.apply(input, dim, bias_init) 32 | -------------------------------------------------------------------------------- /fairseq/data/prepend_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import numpy as np 7 | import torch 8 | 9 | from . import BaseWrapperDataset 10 | 11 | 12 | class PrependDataset(BaseWrapperDataset): 13 | def __init__(self, dataset, prepend_getter, ensure_first_token_is=None): 14 | super().__init__(dataset) 15 | self.prepend_getter = prepend_getter 16 | self.ensure_first_token = ensure_first_token_is 17 | 18 | def __getitem__(self, idx): 19 | item = self.dataset[idx] 20 | is_tuple = isinstance(item, tuple) 21 | src = item[0] if is_tuple else item 22 | 23 | assert self.ensure_first_token is None or src[0] == self.ensure_first_token 24 | prepend_idx = self.prepend_getter(self.dataset, idx) 25 | assert isinstance(prepend_idx, int) 26 | src[0] = prepend_idx 27 | item = tuple((src,) + item[1:]) if is_tuple else src 28 | return item 29 | -------------------------------------------------------------------------------- /fairseq/optim/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import importlib 7 | import os 8 | 9 | from fairseq import registry 10 | from fairseq.optim.fairseq_optimizer import FairseqOptimizer 11 | from fairseq.optim.fp16_optimizer import FP16Optimizer, MemoryEfficientFP16Optimizer 12 | from fairseq.optim.bmuf import FairseqBMUF # noqa 13 | 14 | 15 | __all__ = [ 16 | 'FairseqOptimizer', 17 | 'FP16Optimizer', 18 | 'MemoryEfficientFP16Optimizer', 19 | ] 20 | 21 | 22 | build_optimizer, register_optimizer, OPTIMIZER_REGISTRY = registry.setup_registry( 23 | '--optimizer', 24 | base_class=FairseqOptimizer, 25 | default='nag', 26 | ) 27 | 28 | 29 | # automatically import any Python files in the optim/ directory 30 | for file in os.listdir(os.path.dirname(__file__)): 31 | if file.endswith('.py') and not file.startswith('_'): 32 | module = file[:file.find('.py')] 33 | importlib.import_module('fairseq.optim.' + module) 34 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | MIT License 2 | 3 | Copyright (c) Facebook, Inc. and its affiliates. 4 | 5 | Permission is hereby granted, free of charge, to any person obtaining a copy 6 | of this software and associated documentation files (the "Software"), to deal 7 | in the Software without restriction, including without limitation the rights 8 | to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 9 | copies of the Software, and to permit persons to whom the Software is 10 | furnished to do so, subject to the following conditions: 11 | 12 | The above copyright notice and this permission notice shall be included in all 13 | copies or substantial portions of the Software. 14 | 15 | THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 | IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 | FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 18 | AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 | LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 | OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 | SOFTWARE. 22 | -------------------------------------------------------------------------------- /docs/lr_scheduler.rst: -------------------------------------------------------------------------------- 1 | .. role:: hidden 2 | :class: hidden-section 3 | 4 | .. _Learning Rate Schedulers: 5 | 6 | Learning Rate Schedulers 7 | ======================== 8 | 9 | Learning Rate Schedulers update the learning rate over the course of training. 10 | Learning rates can be updated after each update via :func:`step_update` or at 11 | epoch boundaries via :func:`step`. 12 | 13 | .. automodule:: fairseq.optim.lr_scheduler 14 | :members: 15 | 16 | .. autoclass:: fairseq.optim.lr_scheduler.FairseqLRScheduler 17 | :members: 18 | :undoc-members: 19 | 20 | .. autoclass:: fairseq.optim.lr_scheduler.cosine_lr_scheduler.CosineSchedule 21 | :members: 22 | :undoc-members: 23 | .. autoclass:: fairseq.optim.lr_scheduler.fixed_schedule.FixedSchedule 24 | :members: 25 | :undoc-members: 26 | .. autoclass:: fairseq.optim.lr_scheduler.inverse_square_root_schedule.InverseSquareRootSchedule 27 | :members: 28 | :undoc-members: 29 | .. autoclass:: fairseq.optim.lr_scheduler.reduce_lr_on_plateau.ReduceLROnPlateau 30 | :members: 31 | :undoc-members: 32 | .. autoclass:: fairseq.optim.lr_scheduler.triangular_lr_scheduler.TriangularSchedule 33 | :members: 34 | :undoc-members: 35 | -------------------------------------------------------------------------------- /docs/index.rst: -------------------------------------------------------------------------------- 1 | .. fairseq documentation master file, created by 2 | sphinx-quickstart on Fri Aug 17 21:45:30 2018. 3 | You can adapt this file completely to your liking, but it should at least 4 | contain the root `toctree` directive. 5 | 6 | :github_url: https://github.com/pytorch/fairseq 7 | 8 | 9 | fairseq documentation 10 | ===================== 11 | 12 | Fairseq is a sequence modeling toolkit written in `PyTorch 13 | `_ that allows researchers and developers to 14 | train custom models for translation, summarization, language modeling and other 15 | text generation tasks. 16 | 17 | .. toctree:: 18 | :maxdepth: 1 19 | :caption: Getting Started 20 | 21 | getting_started 22 | command_line_tools 23 | 24 | .. toctree:: 25 | :maxdepth: 1 26 | :caption: Extending Fairseq 27 | 28 | overview 29 | tutorial_simple_lstm 30 | tutorial_classifying_names 31 | 32 | .. toctree:: 33 | :maxdepth: 2 34 | :caption: Library Reference 35 | 36 | tasks 37 | models 38 | criterions 39 | optim 40 | lr_scheduler 41 | data 42 | modules 43 | 44 | 45 | Indices and tables 46 | ================== 47 | 48 | * :ref:`genindex` 49 | * :ref:`search` 50 | -------------------------------------------------------------------------------- /scripts/compare_namespaces.py: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env python 2 | """Helper script to compare two argparse.Namespace objects.""" 3 | 4 | from argparse import Namespace # noqa 5 | 6 | 7 | def main(): 8 | 9 | ns1 = eval(input('Namespace 1: ')) 10 | ns2 = eval(input('Namespace 2: ')) 11 | 12 | def keys(ns): 13 | ks = set() 14 | for k in dir(ns): 15 | if not k.startswith('_'): 16 | ks.add(k) 17 | return ks 18 | 19 | k1 = keys(ns1) 20 | k2 = keys(ns2) 21 | 22 | def print_keys(ks, ns1, ns2=None): 23 | for k in ks: 24 | if ns2 is None: 25 | print('{}\t{}'.format(k, getattr(ns1, k, None))) 26 | else: 27 | print('{}\t{}\t{}'.format(k, getattr(ns1, k, None), getattr(ns2, k, None))) 28 | 29 | print('Keys unique to namespace 1:') 30 | print_keys(k1 - k2, ns1) 31 | print() 32 | 33 | print('Keys unique to namespace 2:') 34 | print_keys(k2 - k1, ns2) 35 | print() 36 | 37 | print('Overlapping keys with different values:') 38 | ks = [k for k in k1 & k2 if getattr(ns1, k, 'None') != getattr(ns2, k, 'None')] 39 | print_keys(ks, ns1, ns2) 40 | print() 41 | 42 | 43 | if __name__ == '__main__': 44 | main() 45 | -------------------------------------------------------------------------------- /fairseq/data/append_token_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import numpy as np 7 | import torch 8 | 9 | from . import BaseWrapperDataset 10 | 11 | 12 | class AppendTokenDataset(BaseWrapperDataset): 13 | 14 | def __init__(self, dataset, token=None): 15 | super().__init__(dataset) 16 | self.token = token 17 | if token is not None: 18 | self._sizes = np.array(dataset.sizes) + 1 19 | else: 20 | self._sizes = dataset.sizes 21 | 22 | def __getitem__(self, idx): 23 | item = self.dataset[idx] 24 | if self.token is not None: 25 | item = torch.cat([item, item.new([self.token])]) 26 | return item 27 | 28 | @property 29 | def sizes(self): 30 | return self._sizes 31 | 32 | def num_tokens(self, index): 33 | n = self.dataset.num_tokens(index) 34 | if self.token is not None: 35 | n += 1 36 | return n 37 | 38 | def size(self, index): 39 | n = self.dataset.size(index) 40 | if self.token is not None: 41 | n += 1 42 | return n 43 | -------------------------------------------------------------------------------- /fairseq/data/encoders/fastbpe.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | from fairseq import file_utils 7 | from fairseq.data.encoders import register_bpe 8 | 9 | 10 | @register_bpe('fastbpe') 11 | class fastBPE(object): 12 | 13 | @staticmethod 14 | def add_args(parser): 15 | # fmt: off 16 | parser.add_argument('--bpe-codes', type=str, 17 | help='path to fastBPE BPE') 18 | # fmt: on 19 | 20 | def __init__(self, args): 21 | if args.bpe_codes is None: 22 | raise ValueError('--bpe-codes is required for --bpe=subword_nmt') 23 | codes = file_utils.cached_path(args.bpe_codes) 24 | try: 25 | import fastBPE 26 | self.bpe = fastBPE.fastBPE(codes) 27 | self.bpe_symbol = "@@ " 28 | except ImportError: 29 | raise ImportError('Please install fastBPE with: pip install fastBPE') 30 | 31 | def encode(self, x: str) -> str: 32 | return self.bpe.apply([x])[0] 33 | 34 | def decode(self, x: str) -> str: 35 | return (x + ' ').replace(self.bpe_symbol, '').rstrip() 36 | -------------------------------------------------------------------------------- /fairseq/tasks/translation_from_pretrained_xlm.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.legacy.masked_lm_dictionary import MaskedLMDictionary 7 | from fairseq.tasks.translation import TranslationTask 8 | 9 | from . import register_task 10 | 11 | 12 | @register_task("translation_from_pretrained_xlm") 13 | class TranslationFromPretrainedXLMTask(TranslationTask): 14 | """ 15 | Same as TranslationTask except use the MaskedLMDictionary class so that 16 | we can load data that was binarized with the MaskedLMDictionary class. 17 | 18 | This task should be used for the entire training pipeline when we want to 19 | train an NMT model from a pretrained XLM checkpoint: binarizing NMT data, 20 | training NMT with the pretrained XLM checkpoint, and subsequent evaluation 21 | of that trained model. 22 | """ 23 | 24 | @classmethod 25 | def load_dictionary(cls, filename): 26 | """Load the masked LM dictionary from the filename 27 | 28 | Args: 29 | filename (str): the filename 30 | """ 31 | return MaskedLMDictionary.load(filename) 32 | -------------------------------------------------------------------------------- /fairseq/data/prepend_token_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import numpy as np 7 | import torch 8 | 9 | from . import BaseWrapperDataset 10 | 11 | 12 | class PrependTokenDataset(BaseWrapperDataset): 13 | 14 | def __init__(self, dataset, token=None): 15 | super().__init__(dataset) 16 | self.token = token 17 | if token is not None: 18 | self._sizes = np.array(dataset.sizes) + 1 19 | else: 20 | self._sizes = dataset.sizes 21 | 22 | def __getitem__(self, idx): 23 | item = self.dataset[idx] 24 | if self.token is not None: 25 | item = torch.cat([item.new([self.token]), item]) 26 | return item 27 | 28 | @property 29 | def sizes(self): 30 | return self._sizes 31 | 32 | def num_tokens(self, index): 33 | n = self.dataset.num_tokens(index) 34 | if self.token is not None: 35 | n += 1 36 | return n 37 | 38 | def size(self, index): 39 | n = self.dataset.size(index) 40 | if self.token is not None: 41 | n += 1 42 | return n 43 | -------------------------------------------------------------------------------- /fairseq/models/roberta/model_camembert.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 | CamemBERT: a Tasty French Language Model 7 | """ 8 | 9 | from fairseq.models import register_model 10 | 11 | from .hub_interface import RobertaHubInterface 12 | from .model import RobertaModel 13 | 14 | 15 | @register_model('camembert') 16 | class CamembertModel(RobertaModel): 17 | 18 | @classmethod 19 | def hub_models(cls): 20 | return { 21 | 'camembert.v0': 'http://dl.fbaipublicfiles.com/fairseq/models/camembert.v0.tar.gz', 22 | } 23 | 24 | @classmethod 25 | def from_pretrained(cls, model_name_or_path, checkpoint_file='model.pt', data_name_or_path='.', bpe='sentencepiece', **kwargs): 26 | from fairseq import hub_utils 27 | x = hub_utils.from_pretrained( 28 | model_name_or_path, 29 | checkpoint_file, 30 | data_name_or_path, 31 | archive_map=cls.hub_models(), 32 | bpe=bpe, 33 | load_checkpoint_heads=True, 34 | **kwargs, 35 | ) 36 | return RobertaHubInterface(x['args'], x['task'], x['models'][0]) 37 | -------------------------------------------------------------------------------- /fairseq/pdb.py: -------------------------------------------------------------------------------- 1 | # Copyright (c) Facebook, Inc. and its affiliates. 2 | # 3 | # This source code is licensed under the MIT license found in the 4 | # LICENSE file in the root directory of this source tree. 5 | 6 | import multiprocessing 7 | import os 8 | import pdb 9 | import sys 10 | 11 | 12 | __all__ = ['set_trace'] 13 | 14 | 15 | _stdin = [None] 16 | _stdin_lock = multiprocessing.Lock() 17 | try: 18 | _stdin_fd = sys.stdin.fileno() 19 | except Exception: 20 | _stdin_fd = None 21 | 22 | 23 | class MultiprocessingPdb(pdb.Pdb): 24 | """A Pdb wrapper that works in a multiprocessing environment. 25 | 26 | Usage: `from fairseq import pdb; pdb.set_trace()` 27 | """ 28 | 29 | def __init__(self): 30 | pdb.Pdb.__init__(self, nosigint=True) 31 | 32 | def _cmdloop(self): 33 | stdin_bak = sys.stdin 34 | with _stdin_lock: 35 | try: 36 | if _stdin_fd is not None: 37 | if not _stdin[0]: 38 | _stdin[0] = os.fdopen(_stdin_fd) 39 | sys.stdin = _stdin[0] 40 | self.cmdloop() 41 | finally: 42 | sys.stdin = stdin_bak 43 | 44 | 45 | def set_trace(): 46 | pdb = MultiprocessingPdb() 47 | pdb.set_trace(sys._getframe().f_back) 48 | -------------------------------------------------------------------------------- /fairseq/models/roberta/model_xlmr.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 | Unsupervised Cross-lingual Representation Learning at Scale 7 | """ 8 | 9 | from fairseq.models import register_model 10 | 11 | from .hub_interface import RobertaHubInterface 12 | from .model import RobertaModel 13 | 14 | 15 | @register_model('xlmr') 16 | class XLMRModel(RobertaModel): 17 | 18 | @classmethod 19 | def hub_models(cls): 20 | return { 21 | 'xlmr.base': 'http://dl.fbaipublicfiles.com/fairseq/models/xlmr.base.tar.gz', 22 | 'xlmr.large': 'http://dl.fbaipublicfiles.com/fairseq/models/xlmr.large.tar.gz', 23 | } 24 | 25 | @classmethod 26 | def from_pretrained(cls, model_name_or_path, checkpoint_file='model.pt', data_name_or_path='.', bpe='sentencepiece', **kwargs): 27 | from fairseq import hub_utils 28 | x = hub_utils.from_pretrained( 29 | model_name_or_path, 30 | checkpoint_file, 31 | data_name_or_path, 32 | archive_map=cls.hub_models(), 33 | bpe=bpe, 34 | load_checkpoint_heads=True, 35 | **kwargs, 36 | ) 37 | return RobertaHubInterface(x['args'], x['task'], x['models'][0]) 38 | -------------------------------------------------------------------------------- /fairseq/modules/positional_embedding.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 as nn 7 | from .learned_positional_embedding import LearnedPositionalEmbedding 8 | from .sinusoidal_positional_embedding import SinusoidalPositionalEmbedding 9 | 10 | 11 | def PositionalEmbedding( 12 | num_embeddings: int, 13 | embedding_dim: int, 14 | padding_idx: int, 15 | learned: bool = False, 16 | ): 17 | if learned: 18 | # if padding_idx is specified then offset the embedding ids by 19 | # this index and adjust num_embeddings appropriately 20 | # TODO: The right place for this offset would be inside 21 | # LearnedPositionalEmbedding. Move this there for a cleaner implementation. 22 | if padding_idx is not None: 23 | num_embeddings = num_embeddings + padding_idx + 1 24 | m = LearnedPositionalEmbedding(num_embeddings, embedding_dim, padding_idx) 25 | nn.init.normal_(m.weight, mean=0, std=embedding_dim ** -0.5) 26 | if padding_idx is not None: 27 | nn.init.constant_(m.weight[padding_idx], 0) 28 | else: 29 | m = SinusoidalPositionalEmbedding( 30 | embedding_dim, padding_idx, init_size=num_embeddings + padding_idx + 1, 31 | ) 32 | return m 33 | -------------------------------------------------------------------------------- /fairseq/optim/adagrad.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.optim 7 | 8 | from . import FairseqOptimizer, register_optimizer 9 | 10 | 11 | @register_optimizer('adagrad') 12 | class Adagrad(FairseqOptimizer): 13 | def __init__(self, args, params): 14 | super().__init__(args) 15 | self._optimizer = torch.optim.Adagrad(params, **self.optimizer_config) 16 | 17 | @staticmethod 18 | def add_args(parser): 19 | """Add optimizer-specific arguments to the parser.""" 20 | # fmt: off 21 | parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', 22 | help='weight decay') 23 | # fmt: on 24 | 25 | @property 26 | def optimizer_config(self): 27 | """ 28 | Return a kwarg dictionary that will be used to override optimizer 29 | args stored in checkpoints. This allows us to load a checkpoint and 30 | resume training using a different set of optimizer args, e.g., with a 31 | different learning rate. 32 | """ 33 | return { 34 | 'lr': self.args.lr[0], 35 | 'weight_decay': self.args.weight_decay, 36 | } 37 | 38 | @property 39 | def supports_flat_params(self): 40 | return True 41 | -------------------------------------------------------------------------------- /docs/data.rst: -------------------------------------------------------------------------------- 1 | .. role:: hidden 2 | :class: hidden-section 3 | 4 | .. module:: fairseq.data 5 | 6 | Data Loading and Utilities 7 | ========================== 8 | 9 | .. _datasets: 10 | 11 | Datasets 12 | -------- 13 | 14 | **Datasets** define the data format and provide helpers for creating 15 | mini-batches. 16 | 17 | .. autoclass:: fairseq.data.FairseqDataset 18 | :members: 19 | .. autoclass:: fairseq.data.LanguagePairDataset 20 | :members: 21 | .. autoclass:: fairseq.data.MonolingualDataset 22 | :members: 23 | 24 | **Helper Datasets** 25 | 26 | These datasets wrap other :class:`fairseq.data.FairseqDataset` instances and 27 | provide additional functionality: 28 | 29 | .. autoclass:: fairseq.data.BacktranslationDataset 30 | :members: 31 | .. autoclass:: fairseq.data.ConcatDataset 32 | :members: 33 | .. autoclass:: fairseq.data.ResamplingDataset 34 | :members: 35 | .. autoclass:: fairseq.data.RoundRobinZipDatasets 36 | :members: 37 | .. autoclass:: fairseq.data.TransformEosDataset 38 | :members: 39 | 40 | 41 | Dictionary 42 | ---------- 43 | 44 | .. autoclass:: fairseq.data.Dictionary 45 | :members: 46 | 47 | 48 | Iterators 49 | --------- 50 | 51 | .. autoclass:: fairseq.data.CountingIterator 52 | :members: 53 | .. autoclass:: fairseq.data.EpochBatchIterator 54 | :members: 55 | .. autoclass:: fairseq.data.GroupedIterator 56 | :members: 57 | .. autoclass:: fairseq.data.ShardedIterator 58 | :members: 59 | -------------------------------------------------------------------------------- /fairseq/data/encoders/byte_bpe.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 import file_utils 8 | from fairseq.data.encoders import register_bpe 9 | from fairseq.data.encoders.byte_utils import (byte_encode, smart_byte_decode, 10 | SPACE, SPACE_ESCAPE) 11 | 12 | 13 | @register_bpe('byte_bpe') 14 | class ByteBPE(object): 15 | @staticmethod 16 | def add_args(parser): 17 | # fmt: off 18 | parser.add_argument('--sentencepiece-model-path', type=str, 19 | help='path to sentencepiece model') 20 | # fmt: on 21 | 22 | def __init__(self, args): 23 | vocab = file_utils.cached_path(args.sentencepiece_model_path) 24 | try: 25 | import sentencepiece as spm 26 | self.sp = spm.SentencePieceProcessor() 27 | self.sp.Load(vocab) 28 | except ImportError: 29 | raise ImportError('Please install sentencepiece with: pip install sentencepiece') 30 | 31 | def encode(self, x: str) -> str: 32 | byte_encoded = byte_encode(x) 33 | return SPACE.join(self.sp.EncodeAsPieces(byte_encoded)) 34 | 35 | @staticmethod 36 | def decode(x: str) -> str: 37 | unescaped = x.replace(SPACE, '').replace(SPACE_ESCAPE, SPACE) 38 | return smart_byte_decode(unescaped) 39 | -------------------------------------------------------------------------------- /fairseq/modules/conv_tbc.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 torch.nn.modules.utils import _single 8 | 9 | 10 | class ConvTBC(torch.nn.Module): 11 | """1D convolution over an input of shape (time x batch x channel) 12 | 13 | The implementation uses gemm to perform the convolution. This implementation 14 | is faster than cuDNN for small kernel sizes. 15 | """ 16 | def __init__(self, in_channels, out_channels, kernel_size, padding=0): 17 | super(ConvTBC, self).__init__() 18 | self.in_channels = in_channels 19 | self.out_channels = out_channels 20 | self.kernel_size = _single(kernel_size) 21 | self.padding = _single(padding) 22 | 23 | self.weight = torch.nn.Parameter(torch.Tensor( 24 | self.kernel_size[0], in_channels, out_channels)) 25 | self.bias = torch.nn.Parameter(torch.Tensor(out_channels)) 26 | 27 | def forward(self, input): 28 | return torch.conv_tbc(input.contiguous(), self.weight, self.bias, self.padding[0]) 29 | 30 | def __repr__(self): 31 | s = ('{name}({in_channels}, {out_channels}, kernel_size={kernel_size}' 32 | ', padding={padding}') 33 | if self.bias is None: 34 | s += ', bias=False' 35 | s += ')' 36 | return s.format(name=self.__class__.__name__, **self.__dict__) 37 | -------------------------------------------------------------------------------- /fairseq/optim/lr_scheduler/fairseq_lr_scheduler.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 FairseqOptimizer 7 | 8 | 9 | class FairseqLRScheduler(object): 10 | 11 | def __init__(self, args, optimizer): 12 | super().__init__() 13 | if not isinstance(optimizer, FairseqOptimizer): 14 | raise ValueError('optimizer must be an instance of FairseqOptimizer') 15 | self.args = args 16 | self.optimizer = optimizer 17 | self.best = None 18 | 19 | @staticmethod 20 | def add_args(parser): 21 | """Add arguments to the parser for this LR scheduler.""" 22 | pass 23 | 24 | def state_dict(self): 25 | """Return the LR scheduler state dict.""" 26 | return {'best': self.best} 27 | 28 | def load_state_dict(self, state_dict): 29 | """Load an LR scheduler state dict.""" 30 | self.best = state_dict['best'] 31 | 32 | def step(self, epoch, val_loss=None): 33 | """Update the learning rate at the end of the given epoch.""" 34 | if val_loss is not None: 35 | if self.best is None: 36 | self.best = val_loss 37 | else: 38 | self.best = min(self.best, val_loss) 39 | 40 | def step_update(self, num_updates): 41 | """Update the learning rate after each update.""" 42 | return self.optimizer.get_lr() 43 | -------------------------------------------------------------------------------- /fairseq/data/replace_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 ReplaceDataset(BaseWrapperDataset): 10 | """Replaces tokens found in the dataset by a specified replacement token 11 | 12 | Args: 13 | dataset (~torch.utils.data.Dataset): dataset to replace tokens in 14 | replace_map(Dictionary[int,int]): map of token to replace -> replacement token 15 | offsets (List[int]): do not replace tokens before (from left if pos, right if neg) this offset. should be 16 | as many as the number of objects returned by the underlying dataset __getitem__ method. 17 | """ 18 | 19 | def __init__(self, dataset, replace_map, offsets): 20 | super().__init__(dataset) 21 | assert len(replace_map) > 0 22 | self.replace_map = replace_map 23 | self.offsets = offsets 24 | 25 | def __getitem__(self, index): 26 | item = self.dataset[index] 27 | is_tuple = isinstance(item, tuple) 28 | srcs = item if is_tuple else [item] 29 | 30 | for offset, src in zip(self.offsets, srcs): 31 | for k, v in self.replace_map.items(): 32 | src_off = src[offset:] if offset >= 0 else src[:offset] 33 | src_off.masked_fill_(src_off == k, v) 34 | 35 | item = srcs if is_tuple else srcs[0] 36 | return item 37 | -------------------------------------------------------------------------------- /scripts/read_binarized.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 argparse 8 | 9 | from fairseq.data import data_utils, Dictionary, indexed_dataset 10 | 11 | 12 | def get_parser(): 13 | parser = argparse.ArgumentParser( 14 | description='writes text from binarized file to stdout') 15 | # fmt: off 16 | parser.add_argument('--dataset-impl', help='dataset implementation', 17 | choices=indexed_dataset.get_available_dataset_impl()) 18 | parser.add_argument('--dict', metavar='FP', help='dictionary containing known words', default=None) 19 | parser.add_argument('--input', metavar='FP', required=True, help='binarized file to read') 20 | # fmt: on 21 | 22 | return parser 23 | 24 | 25 | def main(): 26 | parser = get_parser() 27 | args = parser.parse_args() 28 | 29 | dictionary = Dictionary.load(args.dict) if args.dict is not None else None 30 | dataset = data_utils.load_indexed_dataset( 31 | args.input, 32 | dictionary, 33 | dataset_impl=args.dataset_impl, 34 | default='lazy', 35 | ) 36 | 37 | for tensor_line in dataset: 38 | if dictionary is None: 39 | line = ' '.join([str(int(x)) for x in tensor_line]) 40 | else: 41 | line = dictionary.string(tensor_line) 42 | 43 | print(line) 44 | 45 | 46 | if __name__ == '__main__': 47 | main() 48 | -------------------------------------------------------------------------------- /tests/speech_recognition/test_cross_entropy.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 examples.speech_recognition.criterions.cross_entropy_acc import CrossEntropyWithAccCriterion 8 | from .asr_test_base import CrossEntropyCriterionTestBase 9 | 10 | 11 | class CrossEntropyWithAccCriterionTest(CrossEntropyCriterionTestBase): 12 | def setUp(self): 13 | self.criterion_cls = CrossEntropyWithAccCriterion 14 | super().setUp() 15 | 16 | def test_cross_entropy_all_correct(self): 17 | sample = self.get_test_sample(correct=True, soft_target=False, aggregate=False) 18 | loss, sample_size, logging_output = self.criterion( 19 | self.model, sample, "sum", log_probs=True 20 | ) 21 | assert logging_output["correct"] == 20 22 | assert logging_output["total"] == 20 23 | assert logging_output["sample_size"] == 20 24 | assert logging_output["ntokens"] == 20 25 | 26 | def test_cross_entropy_all_wrong(self): 27 | sample = self.get_test_sample(correct=False, soft_target=False, aggregate=False) 28 | loss, sample_size, logging_output = self.criterion( 29 | self.model, sample, "sum", log_probs=True 30 | ) 31 | assert logging_output["correct"] == 0 32 | assert logging_output["total"] == 20 33 | assert logging_output["sample_size"] == 20 34 | assert logging_output["ntokens"] == 20 35 | -------------------------------------------------------------------------------- /fairseq/data/base_wrapper_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 torch.utils.data.dataloader import default_collate 7 | 8 | from . import FairseqDataset 9 | 10 | 11 | class BaseWrapperDataset(FairseqDataset): 12 | 13 | def __init__(self, dataset): 14 | super().__init__() 15 | self.dataset = dataset 16 | 17 | def __getitem__(self, index): 18 | return self.dataset[index] 19 | 20 | def __len__(self): 21 | return len(self.dataset) 22 | 23 | def collater(self, samples): 24 | if hasattr(self.dataset, 'collater'): 25 | return self.dataset.collater(samples) 26 | else: 27 | return default_collate(samples) 28 | 29 | @property 30 | def sizes(self): 31 | return self.dataset.sizes 32 | 33 | def num_tokens(self, index): 34 | return self.dataset.num_tokens(index) 35 | 36 | def size(self, index): 37 | return self.dataset.size(index) 38 | 39 | def ordered_indices(self): 40 | return self.dataset.ordered_indices() 41 | 42 | @property 43 | def supports_prefetch(self): 44 | return getattr(self.dataset, 'supports_prefetch', False) 45 | 46 | def prefetch(self, indices): 47 | self.dataset.prefetch(indices) 48 | 49 | def set_epoch(self, epoch): 50 | super().set_epoch(epoch) 51 | if hasattr(self.dataset, 'set_epoch'): 52 | self.dataset.set_epoch(epoch) 53 | -------------------------------------------------------------------------------- /fairseq/optim/sgd.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.optim 7 | 8 | from . import FairseqOptimizer, register_optimizer 9 | 10 | 11 | @register_optimizer('sgd') 12 | class SGD(FairseqOptimizer): 13 | def __init__(self, args, params): 14 | super().__init__(args) 15 | self._optimizer = torch.optim.SGD(params, **self.optimizer_config) 16 | 17 | @staticmethod 18 | def add_args(parser): 19 | """Add optimizer-specific arguments to the parser.""" 20 | # fmt: off 21 | parser.add_argument('--momentum', default=0.0, type=float, metavar='M', 22 | help='momentum factor') 23 | parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', 24 | help='weight decay') 25 | # fmt: on 26 | 27 | @property 28 | def optimizer_config(self): 29 | """ 30 | Return a kwarg dictionary that will be used to override optimizer 31 | args stored in checkpoints. This allows us to load a checkpoint and 32 | resume training using a different set of optimizer args, e.g., with a 33 | different learning rate. 34 | """ 35 | return { 36 | 'lr': self.args.lr[0], 37 | 'momentum': self.args.momentum, 38 | 'weight_decay': self.args.weight_decay, 39 | } 40 | 41 | @property 42 | def supports_flat_params(self): 43 | return True 44 | -------------------------------------------------------------------------------- /tests/test_file_io.py: -------------------------------------------------------------------------------- 1 | # This source code is licensed under the MIT license found in the 2 | # LICENSE file in the root directory of this source tree. 3 | 4 | import sys 5 | import tempfile 6 | import os 7 | import shutil 8 | 9 | from typing import Optional 10 | 11 | import unittest 12 | from unittest.mock import MagicMock 13 | 14 | 15 | class TestFileIO(unittest.TestCase): 16 | 17 | _tmpdir: Optional[str] = None 18 | _tmpfile: Optional[str] = None 19 | _tmpfile_contents = "Hello, World" 20 | 21 | @classmethod 22 | def setUpClass(cls) -> None: 23 | cls._tmpdir = tempfile.mkdtemp() 24 | with open(os.path.join(cls._tmpdir, "test.txt"), "w") as f: 25 | cls._tmpfile = f.name 26 | f.write(cls._tmpfile_contents) 27 | f.flush() 28 | 29 | @classmethod 30 | def tearDownClass(cls) -> None: 31 | # Cleanup temp working dir. 32 | if cls._tmpdir is not None: 33 | shutil.rmtree(cls._tmpdir) # type: ignore 34 | 35 | def test_file_io(self): 36 | from fairseq.file_io import PathManager 37 | with PathManager.open(os.path.join(self._tmpdir, "test.txt"), "r") as f: 38 | s = f.read() 39 | self.assertEqual(s, self._tmpfile_contents) 40 | 41 | def test_file_io_oss(self): 42 | # Mock fvcore to simulate oss environment. 43 | sys.modules['fvcore'] = MagicMock() 44 | from fairseq.file_io import PathManager 45 | with PathManager.open(os.path.join(self._tmpdir, "test.txt"), "r") as f: 46 | s = f.read() 47 | self.assertEqual(s, self._tmpfile_contents) 48 | -------------------------------------------------------------------------------- /hubconf.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 functools 7 | 8 | from fairseq.hub_utils import BPEHubInterface as bpe # noqa 9 | from fairseq.hub_utils import TokenizerHubInterface as tokenizer # noqa 10 | from fairseq.models import MODEL_REGISTRY 11 | 12 | 13 | dependencies = [ 14 | 'numpy', 15 | 'regex', 16 | 'requests', 17 | 'torch', 18 | ] 19 | 20 | 21 | # torch.hub doesn't build Cython components, so if they are not found then try 22 | # to build them here 23 | try: 24 | import fairseq.data.token_block_utils_fast 25 | except (ImportError, ModuleNotFoundError): 26 | try: 27 | import cython 28 | import os 29 | from setuptools import sandbox 30 | sandbox.run_setup( 31 | os.path.join(os.path.dirname(__file__), 'setup.py'), 32 | ['build_ext', '--inplace'], 33 | ) 34 | except (ImportError, ModuleNotFoundError): 35 | print( 36 | 'Unable to build Cython components. Please make sure Cython is ' 37 | 'installed if the torch.hub model you are loading depends on it.' 38 | ) 39 | 40 | 41 | for _model_type, _cls in MODEL_REGISTRY.items(): 42 | for model_name in _cls.hub_models().keys(): 43 | globals()[model_name] = functools.partial( 44 | _cls.from_pretrained, 45 | model_name, 46 | ) 47 | # to simplify the interface we only expose named models 48 | # globals()[_model_type] = _cls.from_pretrained 49 | -------------------------------------------------------------------------------- /docs/tasks.rst: -------------------------------------------------------------------------------- 1 | .. role:: hidden 2 | :class: hidden-section 3 | 4 | .. module:: fairseq.tasks 5 | 6 | .. _Tasks: 7 | 8 | Tasks 9 | ===== 10 | 11 | Tasks store dictionaries and provide helpers for loading/iterating over 12 | Datasets, initializing the Model/Criterion and calculating the loss. 13 | 14 | Tasks can be selected via the ``--task`` command-line argument. Once selected, a 15 | task may expose additional command-line arguments for further configuration. 16 | 17 | Example usage:: 18 | 19 | # setup the task (e.g., load dictionaries) 20 | task = fairseq.tasks.setup_task(args) 21 | 22 | # build model and criterion 23 | model = task.build_model(args) 24 | criterion = task.build_criterion(args) 25 | 26 | # load datasets 27 | task.load_dataset('train') 28 | task.load_dataset('valid') 29 | 30 | # iterate over mini-batches of data 31 | batch_itr = task.get_batch_iterator( 32 | task.dataset('train'), max_tokens=4096, 33 | ) 34 | for batch in batch_itr: 35 | # compute the loss 36 | loss, sample_size, logging_output = task.get_loss( 37 | model, criterion, batch, 38 | ) 39 | loss.backward() 40 | 41 | 42 | Translation 43 | ----------- 44 | 45 | .. autoclass:: fairseq.tasks.translation.TranslationTask 46 | 47 | .. _language modeling: 48 | 49 | Language Modeling 50 | ----------------- 51 | 52 | .. autoclass:: fairseq.tasks.language_modeling.LanguageModelingTask 53 | 54 | 55 | Adding new tasks 56 | ---------------- 57 | 58 | .. autofunction:: fairseq.tasks.register_task 59 | .. autoclass:: fairseq.tasks.FairseqTask 60 | :members: 61 | :undoc-members: 62 | -------------------------------------------------------------------------------- /fairseq/modules/layer_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 | import torch 7 | import torch.nn as nn 8 | import torch.nn.functional as F 9 | 10 | 11 | try: 12 | from apex.normalization import FusedLayerNorm as _FusedLayerNorm 13 | 14 | has_fused_layernorm = True 15 | 16 | class FusedLayerNorm(_FusedLayerNorm): 17 | @torch.jit.unused 18 | def forward(self, x): 19 | if not x.is_cuda: 20 | return super().forward(x) 21 | else: 22 | with torch.cuda.device(x.device): 23 | return super().forward(x) 24 | 25 | except ImportError: 26 | has_fused_layernorm = False 27 | 28 | 29 | def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True, export=False): 30 | if not export and torch.cuda.is_available() and has_fused_layernorm: 31 | return FusedLayerNorm(normalized_shape, eps, elementwise_affine) 32 | return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine) 33 | 34 | 35 | class Fp32LayerNorm(nn.LayerNorm): 36 | def __init__(self, *args, **kwargs): 37 | super().__init__(*args, **kwargs) 38 | 39 | def forward(self, input): 40 | output = F.layer_norm( 41 | input.float(), 42 | self.normalized_shape, 43 | self.weight.float() if self.weight is not None else None, 44 | self.bias.float() if self.bias is not None else None, 45 | self.eps, 46 | ) 47 | return output.type_as(input) 48 | -------------------------------------------------------------------------------- /fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cuh: -------------------------------------------------------------------------------- 1 | /** 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 | #include 9 | #include 10 | 11 | #include 12 | #include 13 | #include 14 | 15 | #include 16 | #include 17 | #include 18 | #include 19 | #include 20 | #include 21 | 22 | #include 23 | #include 24 | #include 25 | 26 | #define SHFL_MASK 0xffffffff 27 | 28 | template 29 | __global__ 30 | void dynamicconv_forward_kernel(const scalar_t* input, 31 | const scalar_t* weight, 32 | int minibatch, 33 | int sequenceLength, 34 | int numFeatures, 35 | int numFiltersInBlock, 36 | int numHeads, 37 | scalar_t* output); 38 | 39 | template 40 | __global__ 41 | void dynamicconv_backward_kernel( 42 | const scalar_t* gradOutput, // B * C * T 43 | const scalar_t* input, // B * C * T 44 | const scalar_t* weight, 45 | int minibatch, 46 | int sequenceLength, 47 | int numFeatures, 48 | int numFiltersInBlock, 49 | int numHeads, 50 | scalar_t* gradWeight, 51 | scalar_t* gradInput); // B * H * k * T 52 | -------------------------------------------------------------------------------- /fairseq/modules/lightconv_layer/lightconv_cuda.cpp: -------------------------------------------------------------------------------- 1 | /** 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 | #include 9 | #include 10 | 11 | std::vector lightconv_cuda_forward( 12 | at::Tensor input, 13 | at::Tensor filters, 14 | int padding_l); 15 | 16 | std::vector lightconv_cuda_backward( 17 | at::Tensor gradOutput, 18 | int padding_l, 19 | at::Tensor input, 20 | at::Tensor filters); 21 | 22 | 23 | #define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") 24 | #define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") 25 | #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) 26 | 27 | std::vector lightconv_forward( 28 | at::Tensor input, 29 | at::Tensor filters, 30 | int padding_l) { 31 | 32 | CHECK_INPUT(input); 33 | CHECK_INPUT(filters); 34 | 35 | return lightconv_cuda_forward(input, filters, padding_l); 36 | } 37 | 38 | std::vector lightconv_backward( 39 | at::Tensor gradOutput, 40 | int padding_l, 41 | at::Tensor input, 42 | at::Tensor filters) { 43 | 44 | CHECK_INPUT(gradOutput); 45 | CHECK_INPUT(input); 46 | CHECK_INPUT(filters); 47 | 48 | return lightconv_cuda_backward(gradOutput, padding_l, input, filters); 49 | } 50 | 51 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 52 | m.def("forward", &lightconv_forward, "lighconv forward (CUDA)"); 53 | m.def("backward", &lightconv_backward, "lighconv backward (CUDA)"); 54 | } 55 | -------------------------------------------------------------------------------- /fairseq/data/encoders/hf_byte_bpe.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_bpe 7 | 8 | 9 | @register_bpe('hf_byte_bpe') 10 | class HuggingFaceByteLevelBPE(object): 11 | 12 | @staticmethod 13 | def add_args(parser): 14 | # fmt: off 15 | parser.add_argument('--bpe-merges', help='path to merges.txt') 16 | parser.add_argument('--bpe-vocab', help='path to vocab.json') 17 | parser.add_argument('--bpe-add-prefix-space', action='store_true', 18 | help='add prefix space before encoding') 19 | # fmt: on 20 | 21 | def __init__(self, args): 22 | try: 23 | from tokenizers import ByteLevelBPETokenizer 24 | except ImportError: 25 | raise ImportError( 26 | 'Please install huggingface/tokenizers with: ' 27 | 'pip install tokenizers' 28 | ) 29 | 30 | self.bpe = ByteLevelBPETokenizer( 31 | args.bpe_vocab, 32 | args.bpe_merges, 33 | add_prefix_space=getattr(args, 'bpe_add_prefix_space', False), 34 | ) 35 | 36 | def encode(self, x: str) -> str: 37 | return ' '.join(map(str, self.bpe.encode(x).ids)) 38 | 39 | def decode(self, x: str) -> str: 40 | return self.bpe.decode([ 41 | int(tok) if tok not in {'', ''} else tok 42 | for tok in x.split() 43 | ]) 44 | 45 | def is_beginning_of_word(self, x: str) -> bool: 46 | return self.decode(x).startswith(' ') 47 | -------------------------------------------------------------------------------- /scripts/spm_decode.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 argparse 11 | 12 | import sentencepiece as spm 13 | 14 | 15 | def main(): 16 | parser = argparse.ArgumentParser() 17 | parser.add_argument("--model", required=True, 18 | help="sentencepiece model to use for decoding") 19 | parser.add_argument("--input", required=True, help="input file to decode") 20 | parser.add_argument("--input_format", choices=["piece", "id"], default="piece") 21 | args = parser.parse_args() 22 | 23 | sp = spm.SentencePieceProcessor() 24 | sp.Load(args.model) 25 | 26 | if args.input_format == "piece": 27 | def decode(l): 28 | return "".join(sp.DecodePieces(l)) 29 | elif args.input_format == "id": 30 | def decode(l): 31 | return "".join(sp.DecodeIds(l)) 32 | else: 33 | raise NotImplementedError 34 | 35 | def tok2int(tok): 36 | # remap reference-side (represented as <>) to 0 37 | return int(tok) if tok != "<>" else 0 38 | 39 | with open(args.input, "r", encoding="utf-8") as h: 40 | for line in h: 41 | if args.input_format == "id": 42 | print(decode(list(map(tok2int, line.rstrip().split())))) 43 | elif args.input_format == "piece": 44 | print(decode(line.rstrip().split())) 45 | 46 | if __name__ == "__main__": 47 | main() 48 | -------------------------------------------------------------------------------- /fairseq/modules/sparse_transformer_sentence_encoder_layer.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.modules import TransformerSentenceEncoderLayer 7 | from fairseq.modules.sparse_multihead_attention import SparseMultiheadAttention 8 | 9 | 10 | class SparseTransformerSentenceEncoderLayer(TransformerSentenceEncoderLayer): 11 | """ 12 | Implements a Sprase Transformer Encoder Layer (see SparseMultiheadAttention) 13 | """ 14 | 15 | def __init__( 16 | self, 17 | embedding_dim: int = 768, 18 | ffn_embedding_dim: int = 3072, 19 | num_attention_heads: int = 8, 20 | dropout: float = 0.1, 21 | attention_dropout: float = 0.1, 22 | activation_dropout: float = 0.1, 23 | activation_fn: str = 'relu', 24 | export: bool = False, 25 | is_bidirectional: bool = True, 26 | stride: int = 32, 27 | expressivity: int = 8, 28 | ) -> None: 29 | 30 | super().__init__( 31 | embedding_dim, ffn_embedding_dim, num_attention_heads, dropout, 32 | attention_dropout, activation_dropout, activation_fn, export 33 | ) 34 | 35 | self.self_attn = SparseMultiheadAttention( 36 | self.embedding_dim, 37 | num_attention_heads, 38 | dropout=attention_dropout, 39 | add_bias_kv=False, 40 | add_zero_attn=False, 41 | self_attention=True, 42 | is_bidirectional=is_bidirectional, 43 | stride=stride, 44 | expressivity=expressivity, 45 | ) 46 | -------------------------------------------------------------------------------- /fairseq/modules/dynamicconv_layer/dynamicconv_cuda.cpp: -------------------------------------------------------------------------------- 1 | /** 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 | #include 9 | #include 10 | 11 | std::vector dynamicconv_cuda_forward( 12 | at::Tensor input, 13 | at::Tensor filters, 14 | int padding_l); 15 | 16 | std::vector dynamicconv_cuda_backward( 17 | at::Tensor gradOutput, 18 | int padding_l, 19 | at::Tensor input, 20 | at::Tensor filters); 21 | 22 | 23 | #define CHECK_CUDA(x) AT_ASSERTM(x.type().is_cuda(), #x " must be a CUDA tensor") 24 | #define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") 25 | #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) 26 | 27 | std::vector dynamicconv_forward( 28 | at::Tensor input, 29 | at::Tensor filters, 30 | int padding_l) { 31 | 32 | CHECK_INPUT(input); 33 | CHECK_INPUT(filters); 34 | 35 | return dynamicconv_cuda_forward(input, filters, 36 | padding_l); 37 | } 38 | 39 | std::vector dynamicconv_backward( 40 | at::Tensor gradOutput, 41 | int padding_l, 42 | at::Tensor input, 43 | at::Tensor filters) { 44 | 45 | CHECK_INPUT(gradOutput); 46 | CHECK_INPUT(input); 47 | CHECK_INPUT(filters); 48 | 49 | return dynamicconv_cuda_backward(gradOutput, padding_l, 50 | input, filters); 51 | } 52 | 53 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 54 | m.def("forward", &dynamicconv_forward, "dynamicconv forward (CUDA)"); 55 | m.def("backward", &dynamicconv_backward, "dynamicconv backward (CUDA)"); 56 | } 57 | -------------------------------------------------------------------------------- /scripts/shard_docs.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 | Split a large file into shards while respecting document boundaries. Documents 8 | should be separated by a single empty line. 9 | """ 10 | 11 | import argparse 12 | import contextlib 13 | 14 | 15 | def main(): 16 | parser = argparse.ArgumentParser() 17 | parser.add_argument('input') 18 | parser.add_argument('--num-shards', type=int) 19 | args = parser.parse_args() 20 | 21 | assert args.num_shards is not None and args.num_shards > 1 22 | 23 | with open(args.input, 'r', encoding='utf-8') as h: 24 | with contextlib.ExitStack() as stack: 25 | outputs = [ 26 | stack.enter_context(open(args.input + ".shard" + str(i), "w", encoding="utf-8")) 27 | for i in range(args.num_shards) 28 | ] 29 | 30 | doc = [] 31 | first_doc = [True]*args.num_shards 32 | 33 | def output_doc(i): 34 | if not first_doc[i]: 35 | outputs[i].write("\n") 36 | first_doc[i] = False 37 | for line in doc: 38 | outputs[i].write(line) 39 | doc.clear() 40 | 41 | num_docs = 0 42 | for line in h: 43 | if line.strip() == "": # empty line indicates new document 44 | output_doc(num_docs % args.num_shards) 45 | num_docs += 1 46 | else: 47 | doc.append(line) 48 | output_doc(num_docs % args.num_shards) 49 | 50 | 51 | if __name__ == '__main__': 52 | main() 53 | -------------------------------------------------------------------------------- /fairseq/data/encoders/sentencepiece_bpe.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 import file_utils 7 | from fairseq.data.encoders import register_bpe 8 | 9 | 10 | @register_bpe('sentencepiece') 11 | class SentencepieceBPE(object): 12 | 13 | @staticmethod 14 | def add_args(parser): 15 | # fmt: off 16 | parser.add_argument('--sentencepiece-vocab', type=str, 17 | help='path to sentencepiece vocab') 18 | # fmt: on 19 | 20 | def __init__(self, args): 21 | vocab = file_utils.cached_path(args.sentencepiece_vocab) 22 | try: 23 | import sentencepiece as spm 24 | self.sp = spm.SentencePieceProcessor() 25 | self.sp.Load(vocab) 26 | except ImportError: 27 | raise ImportError('Please install sentencepiece with: pip install sentencepiece') 28 | 29 | def encode(self, x: str) -> str: 30 | return ' '.join(self.sp.EncodeAsPieces(x)) 31 | 32 | def decode(self, x: str) -> str: 33 | return x.replace(' ', '').replace('\u2581', ' ').strip() 34 | 35 | def is_beginning_of_word(self, x: str) -> bool: 36 | if x in ['', '', '', '']: 37 | # special elements are always considered beginnings 38 | # HACK: this logic is already present in fairseq/tasks/masked_lm.py 39 | # but these special tokens are also contained in the sentencepiece 40 | # vocabulary which causes duplicate special tokens. This hack makes 41 | # sure that they are all taken into account. 42 | return True 43 | return x.startswith('\u2581') 44 | -------------------------------------------------------------------------------- /fairseq/data/legacy/masked_lm_dictionary.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 Dictionary 7 | 8 | 9 | class MaskedLMDictionary(Dictionary): 10 | """ 11 | Dictionary for Masked Language Modelling tasks. This extends Dictionary by 12 | adding the mask symbol. 13 | """ 14 | def __init__( 15 | self, 16 | pad='', 17 | eos='', 18 | unk='', 19 | mask='', 20 | ): 21 | super().__init__(pad, eos, unk) 22 | self.mask_word = mask 23 | self.mask_index = self.add_symbol(mask) 24 | self.nspecial = len(self.symbols) 25 | 26 | def mask(self): 27 | """Helper to get index of mask symbol""" 28 | return self.mask_index 29 | 30 | 31 | class BertDictionary(MaskedLMDictionary): 32 | """ 33 | Dictionary for BERT task. This extends MaskedLMDictionary by adding support 34 | for cls and sep symbols. 35 | """ 36 | def __init__( 37 | self, 38 | pad='', 39 | eos='', 40 | unk='', 41 | mask='', 42 | cls='', 43 | sep='' 44 | ): 45 | super().__init__(pad, eos, unk, mask) 46 | self.cls_word = cls 47 | self.sep_word = sep 48 | self.cls_index = self.add_symbol(cls) 49 | self.sep_index = self.add_symbol(sep) 50 | self.nspecial = len(self.symbols) 51 | 52 | def cls(self): 53 | """Helper to get index of cls symbol""" 54 | return self.cls_index 55 | 56 | def sep(self): 57 | """Helper to get index of sep symbol""" 58 | return self.sep_index 59 | -------------------------------------------------------------------------------- /fairseq/data/concat_sentences_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 ConcatSentencesDataset(FairseqDataset): 12 | 13 | def __init__(self, *datasets): 14 | super().__init__() 15 | self.datasets = datasets 16 | assert all(len(ds) == len(datasets[0]) for ds in datasets), \ 17 | 'datasets must have the same length' 18 | 19 | def __getitem__(self, index): 20 | return torch.cat([ds[index] for ds in self.datasets]) 21 | 22 | def __len__(self): 23 | return len(self.datasets[0]) 24 | 25 | def collater(self, samples): 26 | return self.datasets[0].collater(samples) 27 | 28 | @property 29 | def sizes(self): 30 | return sum(ds.sizes for ds in self.datasets) 31 | 32 | def num_tokens(self, index): 33 | return sum(ds.num_tokens(index) for ds in self.datasets) 34 | 35 | def size(self, index): 36 | return sum(ds.size(index) for ds in self.datasets) 37 | 38 | def ordered_indices(self): 39 | return self.datasets[0].ordered_indices() 40 | 41 | @property 42 | def supports_prefetch(self): 43 | return any( 44 | getattr(ds, 'supports_prefetch', False) for ds in self.datasets 45 | ) 46 | 47 | def prefetch(self, indices): 48 | for ds in self.datasets: 49 | if getattr(ds, 'supports_prefetch', False): 50 | ds.prefetch(indices) 51 | 52 | def set_epoch(self, epoch): 53 | super().set_epoch(epoch) 54 | for ds in self.datasets: 55 | if hasattr(ds, 'set_epoch'): 56 | ds.set_epoch(epoch) 57 | -------------------------------------------------------------------------------- /fairseq/modules/cross_entropy.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 logging 7 | 8 | import torch 9 | import torch.nn.functional as F 10 | 11 | 12 | logger = logging.getLogger(__name__) 13 | 14 | 15 | def _cross_entropy_pytorch(logits, target, ignore_index=None, reduction='mean'): 16 | lprobs = F.log_softmax(logits, dim=-1, dtype=torch.float32) 17 | return F.nll_loss( 18 | lprobs, target, ignore_index=ignore_index, reduction=reduction, 19 | ) 20 | 21 | 22 | try: 23 | from apex.contrib import xentropy 24 | 25 | logger.info('using fused cross entropy') 26 | 27 | def cross_entropy(logits, target, ignore_index=-100, reduction='mean'): 28 | if logits.device == torch.device('cpu'): 29 | return _cross_entropy_pytorch(logits, target, ignore_index, reduction) 30 | else: 31 | half_to_float = (logits.dtype == torch.half) 32 | losses = xentropy.SoftmaxCrossEntropyLoss.apply( 33 | logits, target, 0.0, ignore_index, half_to_float, 34 | ) 35 | if reduction == 'sum': 36 | return losses.sum() 37 | elif reduction == 'mean': 38 | if ignore_index >= 0: 39 | return losses.sum() / target.ne(ignore_index).sum() 40 | else: 41 | return losses.mean() 42 | elif reduction == 'none': 43 | return losses 44 | else: 45 | raise NotImplementedError 46 | 47 | except ImportError: 48 | 49 | def cross_entropy(logits, target, ignore_index=-100, reduction='mean'): 50 | return _cross_entropy_pytorch(logits, target, ignore_index, reduction) 51 | -------------------------------------------------------------------------------- /fairseq/data/encoders/gpt2_bpe.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 import file_utils 7 | from fairseq.data.encoders import register_bpe 8 | 9 | from .gpt2_bpe_utils import get_encoder 10 | 11 | 12 | DEFAULT_ENCODER_JSON = 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json' 13 | DEFAULT_VOCAB_BPE = 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe' 14 | 15 | 16 | @register_bpe('gpt2') 17 | class GPT2BPE(object): 18 | 19 | @staticmethod 20 | def add_args(parser): 21 | # fmt: off 22 | parser.add_argument('--gpt2-encoder-json', type=str, 23 | default=DEFAULT_ENCODER_JSON, 24 | help='path to encoder.json') 25 | parser.add_argument('--gpt2-vocab-bpe', type=str, 26 | default=DEFAULT_VOCAB_BPE, 27 | help='path to vocab.bpe') 28 | # fmt: on 29 | 30 | def __init__(self, args): 31 | encoder_json = file_utils.cached_path( 32 | getattr(args, 'gpt2_encoder_json', DEFAULT_ENCODER_JSON) 33 | ) 34 | vocab_bpe = file_utils.cached_path( 35 | getattr(args, 'gpt2_vocab_bpe', DEFAULT_VOCAB_BPE) 36 | ) 37 | self.bpe = get_encoder(encoder_json, vocab_bpe) 38 | 39 | def encode(self, x: str) -> str: 40 | return ' '.join(map(str, self.bpe.encode(x))) 41 | 42 | def decode(self, x: str) -> str: 43 | return self.bpe.decode([ 44 | int(tok) if tok not in {'', ''} else tok 45 | for tok in x.split() 46 | ]) 47 | 48 | def is_beginning_of_word(self, x: str) -> bool: 49 | return self.decode(x).startswith(' ') 50 | -------------------------------------------------------------------------------- /fairseq/data/encoders/subword_nmt_bpe.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 import file_utils 7 | from fairseq.data.encoders import register_bpe 8 | 9 | 10 | @register_bpe('subword_nmt') 11 | class SubwordNMTBPE(object): 12 | 13 | @staticmethod 14 | def add_args(parser): 15 | # fmt: off 16 | parser.add_argument('--bpe-codes', type=str, 17 | help='path to subword NMT BPE') 18 | parser.add_argument('--bpe-separator', default='@@', 19 | help='BPE separator') 20 | # fmt: on 21 | 22 | def __init__(self, args): 23 | if args.bpe_codes is None: 24 | raise ValueError('--bpe-codes is required for --bpe=subword_nmt') 25 | codes = file_utils.cached_path(args.bpe_codes) 26 | try: 27 | from subword_nmt import apply_bpe 28 | bpe_parser = apply_bpe.create_parser() 29 | bpe_args = bpe_parser.parse_args([ 30 | '--codes', codes, 31 | '--separator', args.bpe_separator, 32 | ]) 33 | self.bpe = apply_bpe.BPE( 34 | bpe_args.codes, 35 | bpe_args.merges, 36 | bpe_args.separator, 37 | None, 38 | bpe_args.glossaries, 39 | ) 40 | self.bpe_symbol = bpe_args.separator + ' ' 41 | except ImportError: 42 | raise ImportError('Please install subword_nmt with: pip install subword-nmt') 43 | 44 | def encode(self, x: str) -> str: 45 | return self.bpe.process_line(x) 46 | 47 | def decode(self, x: str) -> str: 48 | return (x + ' ').replace(self.bpe_symbol, '').rstrip() 49 | -------------------------------------------------------------------------------- /tests/test_character_token_embedder.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 | import unittest 8 | 9 | from fairseq.data import Dictionary 10 | from fairseq.modules import CharacterTokenEmbedder 11 | 12 | 13 | class TestCharacterTokenEmbedder(unittest.TestCase): 14 | def test_character_token_embedder(self): 15 | vocab = Dictionary() 16 | vocab.add_symbol('hello') 17 | vocab.add_symbol('there') 18 | 19 | embedder = CharacterTokenEmbedder(vocab, [(2, 16), (4, 32), (8, 64), (16, 2)], 64, 5, 2) 20 | 21 | test_sents = [['hello', 'unk', 'there'], ['there'], ['hello', 'there']] 22 | max_len = max(len(s) for s in test_sents) 23 | input = torch.LongTensor(len(test_sents), max_len + 2).fill_(vocab.pad()) 24 | for i in range(len(test_sents)): 25 | input[i][0] = vocab.eos() 26 | for j in range(len(test_sents[i])): 27 | input[i][j + 1] = vocab.index(test_sents[i][j]) 28 | input[i][j + 2] = vocab.eos() 29 | embs = embedder(input) 30 | 31 | assert embs.size() == (len(test_sents), max_len + 2, 5) 32 | self.assertAlmostEqual(embs[0][0], embs[1][0]) 33 | self.assertAlmostEqual(embs[0][0], embs[0][-1]) 34 | self.assertAlmostEqual(embs[0][1], embs[2][1]) 35 | self.assertAlmostEqual(embs[0][3], embs[1][1]) 36 | 37 | embs.sum().backward() 38 | assert embedder.char_embeddings.weight.grad is not None 39 | 40 | def assertAlmostEqual(self, t1, t2): 41 | self.assertEqual(t1.size(), t2.size(), "size mismatch") 42 | self.assertLess((t1 - t2).abs().max(), 1e-6) 43 | 44 | 45 | if __name__ == '__main__': 46 | unittest.main() 47 | -------------------------------------------------------------------------------- /fairseq/data/encoders/byte_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 re 7 | 8 | WHITESPACE_NORMALIZER = re.compile(r'\s+') 9 | SPACE = chr(32) 10 | SPACE_ESCAPE = chr(9601) 11 | # excluding non-breaking space (160) here 12 | PRINTABLE_LATIN = set( 13 | list(range(32, 126 + 1)) + list(range(161, 172 + 1)) + 14 | list(range(174, 255 + 1)) 15 | ) 16 | BYTE_TO_BCHAR = { 17 | b: chr(b) if b in PRINTABLE_LATIN else chr(256 + b) for b in range(256) 18 | } 19 | BCHAR_TO_BYTE = {bc: b for b, bc in BYTE_TO_BCHAR.items()} 20 | 21 | 22 | def byte_encode(x: str) -> str: 23 | normalized = WHITESPACE_NORMALIZER.sub(SPACE, x) 24 | return ''.join([BYTE_TO_BCHAR[b] for b in normalized.encode('utf-8')]) 25 | 26 | 27 | def byte_decode(x: str) -> str: 28 | try: 29 | return bytes([BCHAR_TO_BYTE[bc] for bc in x]).decode('utf-8') 30 | except ValueError: 31 | return '' 32 | 33 | 34 | def smart_byte_decode(x: str) -> str: 35 | output = byte_decode(x) 36 | if output == '': 37 | # DP the best recovery (max valid chars) if it's broken 38 | n_bytes = len(x) 39 | f = [0 for _ in range(n_bytes + 1)] 40 | pt = [0 for _ in range(n_bytes + 1)] 41 | for i in range(1, n_bytes + 1): 42 | f[i], pt[i] = f[i - 1], i - 1 43 | for j in range(1, min(4, i) + 1): 44 | if f[i - j] + 1 > f[i] and len(byte_decode(x[i - j: i])) > 0: 45 | f[i], pt[i] = f[i - j] + 1, i - j 46 | cur_pt = n_bytes 47 | while cur_pt > 0: 48 | if f[cur_pt] == f[pt[cur_pt]] + 1: 49 | output = byte_decode(x[pt[cur_pt]: cur_pt]) + output 50 | cur_pt = pt[cur_pt] 51 | return output 52 | -------------------------------------------------------------------------------- /tests/test_convtbc.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 | import unittest 8 | from fairseq.modules import ConvTBC 9 | import torch.nn as nn 10 | 11 | 12 | class TestConvTBC(unittest.TestCase): 13 | 14 | def test_convtbc(self): 15 | # ksz, in_channels, out_channels 16 | conv_tbc = ConvTBC(4, 5, kernel_size=3, padding=1) 17 | # out_channels, in_channels, ksz 18 | conv1d = nn.Conv1d(4, 5, kernel_size=3, padding=1) 19 | 20 | conv_tbc.weight.data.copy_(conv1d.weight.data.transpose(0, 2)) 21 | conv_tbc.bias.data.copy_(conv1d.bias.data) 22 | 23 | input_tbc = torch.randn(7, 2, 4, requires_grad=True) 24 | input1d = input_tbc.data.transpose(0, 1).transpose(1, 2) 25 | input1d.requires_grad = True 26 | 27 | output_tbc = conv_tbc(input_tbc) 28 | output1d = conv1d(input1d) 29 | 30 | self.assertAlmostEqual(output_tbc.data.transpose(0, 1).transpose(1, 2), output1d.data) 31 | 32 | grad_tbc = torch.randn(output_tbc.size()) 33 | grad1d = grad_tbc.transpose(0, 1).transpose(1, 2).contiguous() 34 | 35 | output_tbc.backward(grad_tbc) 36 | output1d.backward(grad1d) 37 | 38 | self.assertAlmostEqual(conv_tbc.weight.grad.data.transpose(0, 2), conv1d.weight.grad.data) 39 | self.assertAlmostEqual(conv_tbc.bias.grad.data, conv1d.bias.grad.data) 40 | self.assertAlmostEqual(input_tbc.grad.data.transpose(0, 1).transpose(1, 2), input1d.grad.data) 41 | 42 | def assertAlmostEqual(self, t1, t2): 43 | self.assertEqual(t1.size(), t2.size(), "size mismatch") 44 | self.assertLess((t1 - t2).abs().max(), 1e-4) 45 | 46 | 47 | if __name__ == '__main__': 48 | unittest.main() 49 | -------------------------------------------------------------------------------- /fairseq/incremental_decoding_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 Dict, Optional 7 | import uuid 8 | 9 | from torch import Tensor 10 | 11 | 12 | class FairseqIncrementalState(object): 13 | 14 | def __init__(self, *args, **kwargs): 15 | super().__init__(*args, **kwargs) 16 | self.init_incremental_state() 17 | 18 | def init_incremental_state(self): 19 | self._incremental_state_id = str(uuid.uuid4()) 20 | 21 | def _get_full_incremental_state_key(self, key: str) -> str: 22 | return "{}.{}".format(self._incremental_state_id, key) 23 | 24 | def get_incremental_state( 25 | self, 26 | incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], 27 | key: str, 28 | ) -> Optional[Dict[str, Optional[Tensor]]]: 29 | """Helper for getting incremental state for an nn.Module.""" 30 | full_key = self._get_full_incremental_state_key(key) 31 | if incremental_state is None or full_key not in incremental_state: 32 | return None 33 | return incremental_state[full_key] 34 | 35 | def set_incremental_state( 36 | self, 37 | incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], 38 | key: str, 39 | value: Dict[str, Optional[Tensor]], 40 | ) -> Optional[Dict[str, Dict[str, Optional[Tensor]]]]: 41 | """Helper for setting incremental state for an nn.Module.""" 42 | if incremental_state is not None: 43 | full_key = self._get_full_incremental_state_key(key) 44 | incremental_state[full_key] = value 45 | return incremental_state 46 | 47 | 48 | def with_incremental_state(cls): 49 | cls.__bases__ = (FairseqIncrementalState,) + tuple(b for b in cls.__bases__ if b != FairseqIncrementalState) 50 | return cls 51 | -------------------------------------------------------------------------------- /fairseq/modules/beamable_mm.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 | import torch.nn as nn 8 | 9 | 10 | class BeamableMM(nn.Module): 11 | """This module provides an optimized MM for beam decoding with attention. 12 | 13 | It leverage the fact that the source-side of the input is replicated beam 14 | times and the target-side of the input is of width one. This layer speeds up 15 | inference by replacing the inputs {(bsz x 1 x nhu), (bsz x sz2 x nhu)} 16 | with smaller inputs {(bsz/beam x beam x nhu), (bsz/beam x sz2 x nhu)}. 17 | """ 18 | def __init__(self, beam_size=None): 19 | super(BeamableMM, self).__init__() 20 | self.beam_size = beam_size 21 | 22 | def forward(self, input1, input2): 23 | if ( 24 | not self.training and # test mode 25 | self.beam_size is not None and # beam size is set 26 | input1.dim() == 3 and # only support batched input 27 | input1.size(1) == 1 # single time step update 28 | ): 29 | bsz, beam = input1.size(0), self.beam_size 30 | 31 | # bsz x 1 x nhu --> bsz/beam x beam x nhu 32 | input1 = input1[:, 0, :].unfold(0, beam, beam).transpose(2, 1) 33 | 34 | # bsz x sz2 x nhu --> bsz/beam x sz2 x nhu 35 | input2 = input2.unfold(0, beam, beam)[:, :, :, 0] 36 | 37 | # use non batched operation if bsz = beam 38 | if input1.size(0) == 1: 39 | output = torch.mm(input1[0, :, :], input2[0, :, :]) 40 | else: 41 | output = input1.bmm(input2) 42 | return output.view(bsz, 1, -1) 43 | else: 44 | return input1.bmm(input2) 45 | 46 | def set_beam_size(self, beam_size): 47 | self.beam_size = beam_size 48 | -------------------------------------------------------------------------------- /fairseq/optim/adadelta.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.optim 7 | 8 | from . import FairseqOptimizer, register_optimizer 9 | 10 | 11 | @register_optimizer('adadelta') 12 | class Adadelta(FairseqOptimizer): 13 | def __init__(self, args, params): 14 | super().__init__(args) 15 | self._optimizer = torch.optim.Adadelta(params, **self.optimizer_config) 16 | 17 | @staticmethod 18 | def add_args(parser): 19 | """Add optimizer-specific arguments to the parser.""" 20 | # fmt: off 21 | parser.add_argument('--adadelta-rho', type=float, default=0.9, metavar='RHO', 22 | help='coefficient used for computing a running average of squared gradients') 23 | parser.add_argument('--adadelta-eps', type=float, default=1e-6, metavar='EPS', 24 | help='term added to the denominator to improve numerical stability') 25 | parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', 26 | help='weight decay') 27 | parser.add_argument('--anneal-eps', action='store_true', help='flag to anneal eps') 28 | # fmt: on 29 | 30 | @property 31 | def optimizer_config(self): 32 | """ 33 | Return a kwarg dictionary that will be used to override optimizer 34 | args stored in checkpoints. This allows us to load a checkpoint and 35 | resume training using a different set of optimizer args, e.g., with a 36 | different learning rate. 37 | """ 38 | return { 39 | 'lr': self.args.lr[0], 40 | 'rho': self.args.adadelta_rho, 41 | 'eps': self.args.adadelta_eps, 42 | 'weight_decay': self.args.weight_decay, 43 | } 44 | 45 | @property 46 | def supports_flat_params(self): 47 | return True 48 | -------------------------------------------------------------------------------- /fairseq/data/encoders/hf_bert_bpe.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_bpe 7 | 8 | 9 | @register_bpe('bert') 10 | class BertBPE(object): 11 | 12 | @staticmethod 13 | def add_args(parser): 14 | # fmt: off 15 | parser.add_argument('--bpe-cased', action='store_true', 16 | help='set for cased BPE', 17 | default=False) 18 | parser.add_argument('--bpe-vocab-file', type=str, 19 | help='bpe vocab file.') 20 | # fmt: on 21 | 22 | def __init__(self, args): 23 | try: 24 | from pytorch_transformers import BertTokenizer 25 | from pytorch_transformers.tokenization_utils import clean_up_tokenization 26 | except ImportError: 27 | raise ImportError( 28 | 'Please install 1.0.0 version of pytorch_transformers' 29 | 'with: pip install pytorch-transformers' 30 | ) 31 | 32 | if 'bpe_vocab_file' in args: 33 | self.bert_tokenizer = BertTokenizer( 34 | args.bpe_vocab_file, 35 | do_lower_case=not args.bpe_cased 36 | ) 37 | else: 38 | vocab_file_name = 'bert-base-cased' if args.bpe_cased else 'bert-base-uncased' 39 | self.bert_tokenizer = BertTokenizer.from_pretrained(vocab_file_name) 40 | self.clean_up_tokenization = clean_up_tokenization 41 | 42 | def encode(self, x: str) -> str: 43 | return ' '.join(self.bert_tokenizer.tokenize(x)) 44 | 45 | def decode(self, x: str) -> str: 46 | return self.clean_up_tokenization( 47 | self.bert_tokenizer.convert_tokens_to_string(x.split(' ')) 48 | ) 49 | 50 | def is_beginning_of_word(self, x: str) -> bool: 51 | return not x.startswith('##') 52 | -------------------------------------------------------------------------------- /scripts/count_docs.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 | Count the number of documents and average number of lines and tokens per 8 | document in a large file. Documents should be separated by a single empty line. 9 | """ 10 | 11 | import argparse 12 | import gzip 13 | import sys 14 | 15 | import numpy as np 16 | 17 | 18 | def main(): 19 | parser = argparse.ArgumentParser() 20 | parser.add_argument('input') 21 | parser.add_argument('--gzip', action='store_true') 22 | args = parser.parse_args() 23 | 24 | def gopen(): 25 | if args.gzip: 26 | return gzip.open(args.input, 'r') 27 | else: 28 | return open(args.input, 'r', encoding='utf-8') 29 | 30 | num_lines = [] 31 | num_toks = [] 32 | with gopen() as h: 33 | num_docs = 1 34 | num_lines_in_doc = 0 35 | num_toks_in_doc = 0 36 | for i, line in enumerate(h): 37 | if len(line.strip()) == 0: # empty line indicates new document 38 | num_docs += 1 39 | num_lines.append(num_lines_in_doc) 40 | num_toks.append(num_toks_in_doc) 41 | num_lines_in_doc = 0 42 | num_toks_in_doc = 0 43 | else: 44 | num_lines_in_doc += 1 45 | num_toks_in_doc += len(line.rstrip().split()) 46 | if i % 1000000 == 0: 47 | print(i, file=sys.stderr, end="", flush=True) 48 | elif i % 100000 == 0: 49 | print(".", file=sys.stderr, end="", flush=True) 50 | print(file=sys.stderr, flush=True) 51 | 52 | print("found {} docs".format(num_docs)) 53 | print("average num lines per doc: {}".format(np.mean(num_lines))) 54 | print("average num toks per doc: {}".format(np.mean(num_toks))) 55 | 56 | 57 | if __name__ == '__main__': 58 | main() 59 | -------------------------------------------------------------------------------- /fairseq/optim/fused_lamb.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.optim import FairseqOptimizer, register_optimizer 7 | 8 | 9 | @register_optimizer('lamb') 10 | class FairseqLAMB(FairseqOptimizer): 11 | """LAMB optimizer.""" 12 | 13 | def __init__(self, args, params): 14 | super().__init__(args) 15 | try: 16 | from apex.optimizers import FusedLAMB 17 | self._optimizer = FusedLAMB(params, **self.optimizer_config) 18 | except ImportError: 19 | raise ImportError('Please install apex to use LAMB optimizer') 20 | 21 | @staticmethod 22 | def add_args(parser): 23 | """Add optimizer-specific arguments to the parser.""" 24 | # fmt: off 25 | parser.add_argument('--lamb-betas', default='(0.9, 0.999)', metavar='B', 26 | help='betas for LAMB optimizer') 27 | parser.add_argument('--lamb-eps', type=float, default=1e-8, metavar='D', 28 | help='epsilon for LAMB optimizer') 29 | parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD', 30 | help='weight decay') 31 | # fmt: on 32 | 33 | @property 34 | def optimizer_config(self): 35 | """ 36 | Return a kwarg dictionary that will be used to override optimizer 37 | args stored in checkpoints. This allows us to load a checkpoint and 38 | resume training using a different set of optimizer args, e.g., with a 39 | different learning rate. 40 | """ 41 | return { 42 | 'lr': self.args.lr[0], 43 | 'betas': eval(self.args.lamb_betas), 44 | 'eps': self.args.lamb_eps, 45 | 'weight_decay': self.args.weight_decay, 46 | } 47 | 48 | @property 49 | def supports_flat_params(self): 50 | return False 51 | -------------------------------------------------------------------------------- /fairseq/clib/libnat_cuda/binding.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 | /* 10 | This code is partially adpoted from https://github.com/1ytic/pytorch-edit-distance 11 | */ 12 | 13 | #include "edit_dist.h" 14 | #include 15 | 16 | #ifndef TORCH_CHECK 17 | #define TORCH_CHECK AT_CHECK 18 | #endif 19 | 20 | #define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), #x " must be a CUDA tensor") 21 | #define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous") 22 | #define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) 23 | 24 | 25 | torch::Tensor LevenshteinDistance( 26 | torch::Tensor source, 27 | torch::Tensor target, 28 | torch::Tensor source_length, 29 | torch::Tensor target_length) { 30 | 31 | CHECK_INPUT(source); 32 | CHECK_INPUT(target); 33 | CHECK_INPUT(source_length); 34 | CHECK_INPUT(target_length); 35 | return LevenshteinDistanceCuda(source, target, source_length, target_length); 36 | } 37 | 38 | torch::Tensor GenerateDeletionLabel( 39 | torch::Tensor source, 40 | torch::Tensor operations) { 41 | 42 | CHECK_INPUT(source); 43 | CHECK_INPUT(operations); 44 | return GenerateDeletionLabelCuda(source, operations); 45 | } 46 | 47 | std::pair GenerateInsertionLabel( 48 | torch::Tensor target, 49 | torch::Tensor operations) { 50 | 51 | CHECK_INPUT(target); 52 | CHECK_INPUT(operations); 53 | return GenerateInsertionLabelCuda(target, operations); 54 | } 55 | 56 | PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { 57 | m.def("levenshtein_distance", &LevenshteinDistance, "Levenshtein distance"); 58 | m.def("generate_deletion_labels", &GenerateDeletionLabel, "Generate Deletion Label"); 59 | m.def("generate_insertion_labels", &GenerateInsertionLabel, "Generate Insertion Label"); 60 | } 61 | -------------------------------------------------------------------------------- /fairseq/models/fairseq_encoder.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 as nn 7 | from typing import List, NamedTuple, Optional 8 | from torch import Tensor 9 | 10 | EncoderOut = NamedTuple( 11 | "EncoderOut", 12 | [ 13 | ("encoder_out", Tensor), # T x B x C 14 | ("encoder_padding_mask", Tensor), # B x T 15 | ("encoder_embedding", Tensor), # B x T x C 16 | ("encoder_states", Optional[List[Tensor]]), # List[T x B x C] 17 | ("src_tokens", Optional[Tensor]), # B x T 18 | ("src_lengths", Optional[Tensor]), # B x 1 19 | ], 20 | ) 21 | 22 | 23 | class FairseqEncoder(nn.Module): 24 | """Base class for encoders.""" 25 | 26 | def __init__(self, dictionary): 27 | super().__init__() 28 | self.dictionary = dictionary 29 | 30 | def forward(self, src_tokens, src_lengths=None, **kwargs): 31 | """ 32 | Args: 33 | src_tokens (LongTensor): tokens in the source language of shape 34 | `(batch, src_len)` 35 | src_lengths (LongTensor): lengths of each source sentence of shape 36 | `(batch)` 37 | """ 38 | raise NotImplementedError 39 | 40 | def reorder_encoder_out(self, encoder_out, new_order): 41 | """ 42 | Reorder encoder output according to `new_order`. 43 | 44 | Args: 45 | encoder_out: output from the ``forward()`` method 46 | new_order (LongTensor): desired order 47 | 48 | Returns: 49 | `encoder_out` rearranged according to `new_order` 50 | """ 51 | raise NotImplementedError 52 | 53 | def max_positions(self): 54 | """Maximum input length supported by the encoder.""" 55 | return 1e6 # an arbitrary large number 56 | 57 | def upgrade_state_dict(self, state_dict): 58 | """Upgrade a (possibly old) state dict for new versions of fairseq.""" 59 | return state_dict 60 | -------------------------------------------------------------------------------- /fairseq/models/composite_encoder.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.models import FairseqEncoder 7 | 8 | 9 | class CompositeEncoder(FairseqEncoder): 10 | """ 11 | A wrapper around a dictionary of :class:`FairseqEncoder` objects. 12 | 13 | We run forward on each encoder and return a dictionary of outputs. The first 14 | encoder's dictionary is used for initialization. 15 | 16 | Args: 17 | encoders (dict): a dictionary of :class:`FairseqEncoder` objects. 18 | """ 19 | 20 | def __init__(self, encoders): 21 | super().__init__(next(iter(encoders.values())).dictionary) 22 | self.encoders = encoders 23 | for key in self.encoders: 24 | self.add_module(key, self.encoders[key]) 25 | 26 | def forward(self, src_tokens, src_lengths): 27 | """ 28 | Args: 29 | src_tokens (LongTensor): tokens in the source language of shape 30 | `(batch, src_len)` 31 | src_lengths (LongTensor): lengths of each source sentence of shape 32 | `(batch)` 33 | 34 | Returns: 35 | dict: 36 | the outputs from each Encoder 37 | """ 38 | encoder_out = {} 39 | for key in self.encoders: 40 | encoder_out[key] = self.encoders[key](src_tokens, src_lengths) 41 | return encoder_out 42 | 43 | def reorder_encoder_out(self, encoder_out, new_order): 44 | """Reorder encoder output according to new_order.""" 45 | for key in self.encoders: 46 | encoder_out[key] = self.encoders[key].reorder_encoder_out(encoder_out[key], new_order) 47 | return encoder_out 48 | 49 | def max_positions(self): 50 | return min(self.encoders[key].max_positions() for key in self.encoders) 51 | 52 | def upgrade_state_dict(self, state_dict): 53 | for key in self.encoders: 54 | self.encoders[key].upgrade_state_dict(state_dict) 55 | return state_dict 56 | -------------------------------------------------------------------------------- /fairseq/data/encoders/moses_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 | 8 | 9 | @register_tokenizer('moses') 10 | class MosesTokenizer(object): 11 | 12 | @staticmethod 13 | def add_args(parser): 14 | # fmt: off 15 | parser.add_argument('--moses-source-lang', metavar='SRC', 16 | help='source language') 17 | parser.add_argument('--moses-target-lang', metavar='TARGET', 18 | help='target language') 19 | parser.add_argument('--moses-no-dash-splits', action='store_true', default=False, 20 | help='don\'t apply dash split rules') 21 | parser.add_argument('--moses-no-escape', action='store_true', default=False, 22 | help='don\'t perform HTML escaping on apostrophy, quotes, etc.') 23 | # fmt: on 24 | 25 | def __init__(self, args): 26 | self.args = args 27 | 28 | if getattr(args, 'moses_source_lang', None) is None: 29 | args.moses_source_lang = getattr(args, 'source_lang', 'en') 30 | if getattr(args, 'moses_target_lang', None) is None: 31 | args.moses_target_lang = getattr(args, 'target_lang', 'en') 32 | 33 | try: 34 | from sacremoses import MosesTokenizer, MosesDetokenizer 35 | self.tok = MosesTokenizer(args.moses_source_lang) 36 | self.detok = MosesDetokenizer(args.moses_target_lang) 37 | except ImportError: 38 | raise ImportError('Please install Moses tokenizer with: pip install sacremoses') 39 | 40 | def encode(self, x: str) -> str: 41 | return self.tok.tokenize( 42 | x, 43 | aggressive_dash_splits=(not self.args.moses_no_dash_splits), 44 | return_str=True, 45 | escape=(not self.args.moses_no_escape), 46 | ) 47 | 48 | def decode(self, x: str) -> str: 49 | return self.detok.detokenize(x.split()) 50 | -------------------------------------------------------------------------------- /tests/test_multihead_attention.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 | import unittest 8 | from fairseq.modules.multihead_attention import MultiheadAttention 9 | 10 | 11 | class TestMultiheadAttention(unittest.TestCase): 12 | def test_append_prev_key_padding_mask(self): 13 | bsz = 1 14 | src_len = 4 15 | 16 | cases = [ 17 | # no padding mask 18 | (None, None, None), 19 | # current padding mask only 20 | ( 21 | torch.tensor([[1]]).bool(), 22 | None, 23 | torch.tensor([[0, 0, 0, 1]]).bool(), 24 | ), 25 | # previous padding mask only 26 | ( 27 | None, 28 | torch.tensor([[0, 1, 0]]).bool(), 29 | torch.tensor([[0, 1, 0, 0]]).bool(), 30 | ), 31 | # both padding masks 32 | ( 33 | torch.tensor([[1]]).bool(), 34 | torch.tensor([[0, 1, 0]]).bool(), 35 | torch.tensor([[0, 1, 0, 1]]).bool(), 36 | ), 37 | ] 38 | for c in cases: 39 | key_padding_mask = MultiheadAttention._append_prev_key_padding_mask( 40 | c[0], 41 | c[1], 42 | batch_size=bsz, 43 | src_len=src_len, 44 | static_kv=False, 45 | ) 46 | 47 | if key_padding_mask is not None: 48 | self.assertTrue( 49 | torch.all(torch.eq(key_padding_mask, c[2])), 50 | f'Unexpected resultant key padding mask: {key_padding_mask}' 51 | f' given current: {c[0]} and previous: {c[1]}', 52 | ) 53 | self.assertEqual(key_padding_mask.size(0), bsz) 54 | self.assertEqual(key_padding_mask.size(1), src_len) 55 | else: 56 | self.assertIsNone(c[2]) 57 | 58 | 59 | if __name__ == '__main__': 60 | unittest.main() 61 | -------------------------------------------------------------------------------- /docs/command_line_tools.rst: -------------------------------------------------------------------------------- 1 | .. _Command-line Tools: 2 | 3 | Command-line Tools 4 | ================== 5 | 6 | Fairseq provides several command-line tools for training and evaluating models: 7 | 8 | - :ref:`fairseq-preprocess`: Data pre-processing: build vocabularies and binarize training data 9 | - :ref:`fairseq-train`: Train a new model on one or multiple GPUs 10 | - :ref:`fairseq-generate`: Translate pre-processed data with a trained model 11 | - :ref:`fairseq-interactive`: Translate raw text with a trained model 12 | - :ref:`fairseq-score`: BLEU scoring of generated translations against reference translations 13 | - :ref:`fairseq-eval-lm`: Language model evaluation 14 | 15 | 16 | .. _fairseq-preprocess: 17 | 18 | fairseq-preprocess 19 | ~~~~~~~~~~~~~~~~~~ 20 | .. automodule:: preprocess 21 | 22 | .. argparse:: 23 | :module: fairseq.options 24 | :func: get_preprocessing_parser 25 | :prog: fairseq-preprocess 26 | 27 | 28 | .. _fairseq-train: 29 | 30 | fairseq-train 31 | ~~~~~~~~~~~~~ 32 | .. automodule:: train 33 | 34 | .. argparse:: 35 | :module: fairseq.options 36 | :func: get_training_parser 37 | :prog: fairseq-train 38 | 39 | 40 | .. _fairseq-generate: 41 | 42 | fairseq-generate 43 | ~~~~~~~~~~~~~~~~ 44 | .. automodule:: generate 45 | 46 | .. argparse:: 47 | :module: fairseq.options 48 | :func: get_generation_parser 49 | :prog: fairseq-generate 50 | 51 | 52 | .. _fairseq-interactive: 53 | 54 | fairseq-interactive 55 | ~~~~~~~~~~~~~~~~~~~ 56 | .. automodule:: interactive 57 | 58 | .. argparse:: 59 | :module: fairseq.options 60 | :func: get_interactive_generation_parser 61 | :prog: fairseq-interactive 62 | 63 | 64 | .. _fairseq-score: 65 | 66 | fairseq-score 67 | ~~~~~~~~~~~~~ 68 | .. automodule:: score 69 | 70 | .. argparse:: 71 | :module: fairseq_cli.score 72 | :func: get_parser 73 | :prog: fairseq-score 74 | 75 | 76 | .. _fairseq-eval-lm: 77 | 78 | fairseq-eval-lm 79 | ~~~~~~~~~~~~~~~ 80 | .. automodule:: eval_lm 81 | 82 | .. argparse:: 83 | :module: fairseq.options 84 | :func: get_eval_lm_parser 85 | :prog: fairseq-eval-lm 86 | -------------------------------------------------------------------------------- /tests/test_concat_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 unittest 7 | 8 | import torch 9 | from fairseq.data import LanguagePairDataset, TokenBlockDataset 10 | from fairseq.data.concat_dataset import ConcatDataset 11 | from tests.test_train import mock_dict 12 | 13 | 14 | class TestConcatDataset(unittest.TestCase): 15 | def setUp(self): 16 | d = mock_dict() 17 | tokens_1 = torch.LongTensor([1]).view(1, -1) 18 | tokens_ds1 = TokenBlockDataset( 19 | tokens_1, 20 | sizes=[tokens_1.size(-1)], 21 | block_size=1, 22 | pad=0, 23 | eos=1, 24 | include_targets=False, 25 | ) 26 | self.dataset_1 = LanguagePairDataset( 27 | tokens_ds1, tokens_ds1.sizes, d, shuffle=False 28 | ) 29 | tokens_2 = torch.LongTensor([2]).view(1, -1) 30 | tokens_ds2 = TokenBlockDataset( 31 | tokens_2, 32 | sizes=[tokens_2.size(-1)], 33 | block_size=1, 34 | pad=0, 35 | eos=1, 36 | include_targets=False, 37 | ) 38 | self.dataset_2 = LanguagePairDataset( 39 | tokens_ds2, tokens_ds2.sizes, d, shuffle=False 40 | ) 41 | 42 | def test_concat_dataset_basics(self): 43 | d = ConcatDataset( 44 | [self.dataset_1, self.dataset_2] 45 | ) 46 | assert(len(d) == 2) 47 | assert(d[0]['source'][0] == 1) 48 | assert(d[1]['source'][0] == 2) 49 | 50 | d = ConcatDataset( 51 | [self.dataset_1, self.dataset_2], sample_ratios=[1, 2] 52 | ) 53 | assert(len(d) == 3) 54 | assert(d[0]['source'][0] == 1) 55 | assert(d[1]['source'][0] == 2) 56 | assert(d[2]['source'][0] == 2) 57 | 58 | d = ConcatDataset( 59 | [self.dataset_1, self.dataset_2], sample_ratios=[2, 1] 60 | ) 61 | assert(len(d) == 3) 62 | assert(d[0]['source'][0] == 1) 63 | assert(d[1]['source'][0] == 1) 64 | assert(d[2]['source'][0] == 2) 65 | -------------------------------------------------------------------------------- /fairseq/model_parallel/megatron_trainer.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 | Train a network across multiple GPUs. 8 | """ 9 | 10 | from fairseq import distributed_utils 11 | from fairseq.trainer import Trainer 12 | 13 | try: 14 | from fairseq.model_parallel.megatron.mpu import ( 15 | get_data_parallel_group, 16 | get_data_parallel_rank, 17 | get_data_parallel_world_size, 18 | get_model_parallel_group, 19 | get_model_parallel_src_rank, 20 | ) 21 | has_megatron_submodule = True 22 | except (ImportError, ModuleNotFoundError): 23 | has_megatron_submodule = False 24 | 25 | 26 | class MegatronTrainer(Trainer): 27 | """Main class for model parallel with data parallel training. 28 | """ 29 | def __init__(self, args, task, model, criterion): 30 | if not has_megatron_submodule: 31 | raise ImportError( 32 | '\n\nPlease install the megatron submodule:' 33 | '\n\n git submodule update --init ' 34 | 'fairseq/model_parallel/megatron' 35 | ) 36 | super().__init__(args, task, model, criterion) 37 | 38 | @property 39 | def data_parallel_world_size(self): 40 | return get_data_parallel_world_size() 41 | 42 | @property 43 | def data_parallel_process_group(self): 44 | return get_data_parallel_group() 45 | 46 | @property 47 | def data_parallel_rank(self): 48 | return get_data_parallel_rank() 49 | 50 | @property 51 | def is_data_parallel_master(self): 52 | return get_model_parallel_src_rank() == 0 53 | 54 | def clip_grad_norm(self, clip_norm): 55 | def _aggregate_model_parallel_grad_norm(total_norm): 56 | total_norm = total_norm ** 2 57 | distributed_utils.all_reduce(total_norm, group=get_model_parallel_group()) 58 | total_norm = total_norm ** 0.5 59 | return total_norm 60 | return self.optimizer.clip_grad_norm( 61 | clip_norm, 62 | aggregate_norm_fn=_aggregate_model_parallel_grad_norm, 63 | ) 64 | -------------------------------------------------------------------------------- /fairseq/data/data_utils_fast.pyx: -------------------------------------------------------------------------------- 1 | # cython: language_level=3 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 numpy as np 8 | 9 | cimport cython 10 | cimport numpy as np 11 | 12 | DTYPE = np.int64 13 | ctypedef np.int64_t DTYPE_t 14 | 15 | 16 | cdef _is_batch_full(list batch, long num_tokens, long max_tokens, long max_sentences): 17 | if len(batch) == 0: 18 | return 0 19 | if max_sentences > 0 and len(batch) == max_sentences: 20 | return 1 21 | if max_tokens > 0 and num_tokens > max_tokens: 22 | return 1 23 | return 0 24 | 25 | 26 | @cython.cdivision(True) 27 | cpdef list batch_by_size_fast( 28 | np.ndarray[DTYPE_t, ndim=1] indices, 29 | num_tokens_fn, 30 | long max_tokens, 31 | long max_sentences, 32 | int bsz_mult, 33 | ): 34 | cdef long sample_len = 0 35 | cdef list sample_lens = [] 36 | cdef list batch = [] 37 | cdef list batches = [] 38 | cdef long mod_len 39 | cdef long i 40 | cdef long idx 41 | cdef long num_tokens 42 | cdef DTYPE_t[:] indices_view = indices 43 | 44 | for i in range(len(indices_view)): 45 | idx = indices_view[i] 46 | num_tokens = num_tokens_fn(idx) 47 | sample_lens.append(num_tokens) 48 | sample_len = max(sample_len, num_tokens) 49 | 50 | assert max_tokens <= 0 or sample_len <= max_tokens, ( 51 | "sentence at index {} of size {} exceeds max_tokens " 52 | "limit of {}!".format(idx, sample_len, max_tokens) 53 | ) 54 | num_tokens = (len(batch) + 1) * sample_len 55 | 56 | if _is_batch_full(batch, num_tokens, max_tokens, max_sentences): 57 | mod_len = max( 58 | bsz_mult * (len(batch) // bsz_mult), 59 | len(batch) % bsz_mult, 60 | ) 61 | batches.append(batch[:mod_len]) 62 | batch = batch[mod_len:] 63 | sample_lens = sample_lens[mod_len:] 64 | sample_len = max(sample_lens) if len(sample_lens) > 0 else 0 65 | batch.append(idx) 66 | if len(batch) > 0: 67 | batches.append(batch) 68 | return batches 69 | -------------------------------------------------------------------------------- /tests/test_memory_efficient_fp16.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 logging 8 | import unittest 9 | 10 | import torch 11 | 12 | from fairseq.optim.adam import FairseqAdam 13 | from fairseq.optim.fp16_optimizer import MemoryEfficientFP16Optimizer 14 | 15 | 16 | @unittest.skipIf(not torch.cuda.is_available(), 'test requires a GPU') 17 | class TestMemoryEfficientFP16(unittest.TestCase): 18 | 19 | def setUp(self): 20 | logging.disable(logging.CRITICAL) 21 | 22 | def tearDown(self): 23 | logging.disable(logging.NOTSET) 24 | 25 | def test_load_state_dict(self): 26 | # define simple FP16 model 27 | model = torch.nn.Linear(5, 5).cuda().half() 28 | params = list(model.parameters()) 29 | 30 | # initialize memory efficient FP16 optimizer 31 | optimizer = FairseqAdam( 32 | argparse.Namespace( 33 | lr=[0.00001], 34 | adam_betas='(0.9, 0.999)', 35 | adam_eps=1e-8, 36 | weight_decay=0.0, 37 | ), 38 | params, 39 | ) 40 | me_optimizer = MemoryEfficientFP16Optimizer( 41 | argparse.Namespace( 42 | fp16_init_scale=1, 43 | fp16_scale_window=1, 44 | fp16_scale_tolerance=1, 45 | threshold_loss_scale=1, 46 | min_loss_scale=1e-4, 47 | ), 48 | params, 49 | optimizer, 50 | ) 51 | 52 | # optimizer state is created in the first step 53 | loss = model(torch.rand(5).cuda().half()).sum() 54 | me_optimizer.backward(loss) 55 | me_optimizer.step() 56 | 57 | # reload state 58 | state = me_optimizer.state_dict() 59 | me_optimizer.load_state_dict(state) 60 | for k, v in me_optimizer.optimizer.state.items(): 61 | self.assertTrue(k.dtype == torch.float16) 62 | for v_i in v.values(): 63 | if torch.is_tensor(v_i): 64 | self.assertTrue(v_i.dtype == torch.float32) 65 | 66 | 67 | if __name__ == '__main__': 68 | unittest.main() 69 | -------------------------------------------------------------------------------- /tests/speech_recognition/test_collaters.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 numpy as np 10 | import torch 11 | from examples.speech_recognition.data.collaters import Seq2SeqCollater 12 | 13 | 14 | class TestSeq2SeqCollator(unittest.TestCase): 15 | def test_collate(self): 16 | 17 | eos_idx = 1 18 | pad_idx = 0 19 | collater = Seq2SeqCollater( 20 | feature_index=0, label_index=1, pad_index=pad_idx, eos_index=eos_idx 21 | ) 22 | 23 | # 2 frames in the first sample and 3 frames in the second one 24 | frames1 = np.array([[7, 8], [9, 10]]) 25 | frames2 = np.array([[1, 2], [3, 4], [5, 6]]) 26 | target1 = np.array([4, 2, 3, eos_idx]) 27 | target2 = np.array([3, 2, eos_idx]) 28 | sample1 = {"id": 0, "data": [frames1, target1]} 29 | sample2 = {"id": 1, "data": [frames2, target2]} 30 | batch = collater.collate([sample1, sample2]) 31 | 32 | # collate sort inputs by frame's length before creating the batch 33 | self.assertTensorEqual(batch["id"], torch.tensor([1, 0])) 34 | self.assertEqual(batch["ntokens"], 7) 35 | self.assertTensorEqual( 36 | batch["net_input"]["src_tokens"], 37 | torch.tensor( 38 | [[[1, 2], [3, 4], [5, 6]], [[7, 8], [9, 10], [pad_idx, pad_idx]]] 39 | ), 40 | ) 41 | self.assertTensorEqual( 42 | batch["net_input"]["prev_output_tokens"], 43 | torch.tensor([[eos_idx, 3, 2, pad_idx], [eos_idx, 4, 2, 3]]), 44 | ) 45 | self.assertTensorEqual(batch["net_input"]["src_lengths"], torch.tensor([3, 2])) 46 | self.assertTensorEqual( 47 | batch["target"], 48 | torch.tensor([[3, 2, eos_idx, pad_idx], [4, 2, 3, eos_idx]]), 49 | ) 50 | self.assertEqual(batch["nsentences"], 2) 51 | 52 | def assertTensorEqual(self, t1, t2): 53 | self.assertEqual(t1.size(), t2.size(), "size mismatch") 54 | self.assertEqual(t1.ne(t2).long().sum(), 0) 55 | 56 | 57 | if __name__ == "__main__": 58 | unittest.main() 59 | -------------------------------------------------------------------------------- /fairseq/tasks/audio_pretraining.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 os 7 | 8 | from fairseq.data import FileAudioDataset 9 | from . import FairseqTask, register_task 10 | 11 | 12 | @register_task('audio_pretraining') 13 | class AudioPretrainingTask(FairseqTask): 14 | """ 15 | 16 | """ 17 | 18 | @staticmethod 19 | def add_args(parser): 20 | """Add task-specific arguments to the parser.""" 21 | parser.add_argument('data', help='path to data directory') 22 | parser.add_argument('--sample-rate', default=16000, type=int, 23 | help='target sample rate. audio files will be up/down sampled to this rate') 24 | parser.add_argument('--max-sample-size', default=None, type=int, 25 | help='max sample size to crop to for batching. default = min sample length') 26 | parser.add_argument('--min-sample-size', default=None, type=int, 27 | help='min sample size to crop to for batching. default = same as --max-sample-size') 28 | 29 | def __init__(self, args): 30 | super().__init__(args) 31 | 32 | @classmethod 33 | def setup_task(cls, args, **kwargs): 34 | """Setup the task (e.g., load dictionaries). 35 | 36 | Args: 37 | args (argparse.Namespace): parsed command-line arguments 38 | """ 39 | return cls(args) 40 | 41 | def load_dataset(self, split, **kwargs): 42 | """Load a given dataset split. 43 | 44 | Args: 45 | split (str): name of the split (e.g., train, valid, test) 46 | """ 47 | 48 | manifest = os.path.join(self.args.data, '{}.tsv'.format(split)) 49 | self.datasets[split] = FileAudioDataset(manifest, 50 | sample_rate=self.args.sample_rate, 51 | max_sample_size=self.args.max_sample_size, 52 | min_sample_size=self.args.min_sample_size) 53 | 54 | @property 55 | def target_dictionary(self): 56 | """Return the :class:`~fairseq.data.Dictionary` for the language 57 | model.""" 58 | return None 59 | -------------------------------------------------------------------------------- /fairseq/data/subsample_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 logging 7 | 8 | import numpy as np 9 | 10 | from . import BaseWrapperDataset 11 | 12 | 13 | logger = logging.getLogger(__name__) 14 | 15 | 16 | class SubsampleDataset(BaseWrapperDataset): 17 | """Subsamples a given dataset by a specified ratio. Subsampling is done on the number of examples 18 | 19 | Args: 20 | dataset (~torch.utils.data.Dataset): dataset to subsample 21 | size_ratio(float): the ratio to subsample to. must be between 0 and 1 (exclusive) 22 | """ 23 | 24 | def __init__(self, dataset, size_ratio): 25 | super().__init__(dataset) 26 | assert size_ratio < 1 27 | self.actual_size = np.ceil(len(dataset) * size_ratio).astype(int) 28 | self.indices = np.random.choice( 29 | list(range(len(self.dataset))), self.actual_size, replace=False 30 | ) 31 | logger.info( 32 | "subsampled dataset from {} to {} (ratio={})".format( 33 | len(self.dataset), self.actual_size, size_ratio 34 | ) 35 | ) 36 | 37 | def __getitem__(self, index): 38 | return self.dataset[self.indices[index]] 39 | 40 | def __len__(self): 41 | return self.actual_size 42 | 43 | def collater(self, samples): 44 | return self.dataset.collater(samples) 45 | 46 | @property 47 | def sizes(self): 48 | return self.dataset.sizes[self.indices] 49 | 50 | @property 51 | def name(self): 52 | return self.dataset.name 53 | 54 | def num_tokens(self, index): 55 | return self.dataset.num_tokens(self.indices[index]) 56 | 57 | def size(self, index): 58 | return self.dataset.size(self.indices[index]) 59 | 60 | def ordered_indices(self): 61 | """Return an ordered list of indices. Batches will be constructed based 62 | on this order.""" 63 | if self.shuffle: 64 | order = [np.random.permutation(len(self))] 65 | else: 66 | order = [np.arange(len(self))] 67 | order.append(self.sizes) 68 | return np.lexsort(order) 69 | 70 | def prefetch(self, indices): 71 | self.dataset.prefetch(self.indices[indices]) 72 | -------------------------------------------------------------------------------- /fairseq/data/fairseq_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.utils.data 8 | 9 | 10 | class EpochListening: 11 | """Mixin for receiving updates whenever the epoch increments.""" 12 | def set_epoch(self, epoch): 13 | """Will receive the updated epoch number at the beginning of the epoch. 14 | """ 15 | pass 16 | 17 | 18 | class FairseqDataset(torch.utils.data.Dataset, EpochListening): 19 | """A dataset that provides helpers for batching.""" 20 | 21 | def __getitem__(self, index): 22 | raise NotImplementedError 23 | 24 | def __len__(self): 25 | raise NotImplementedError 26 | 27 | def collater(self, samples): 28 | """Merge a list of samples to form a mini-batch. 29 | 30 | Args: 31 | samples (List[dict]): samples to collate 32 | 33 | Returns: 34 | dict: a mini-batch suitable for forwarding with a Model 35 | """ 36 | raise NotImplementedError 37 | 38 | def num_tokens(self, index): 39 | """Return the number of tokens in a sample. This value is used to 40 | enforce ``--max-tokens`` during batching.""" 41 | raise NotImplementedError 42 | 43 | def size(self, index): 44 | """Return an example's size as a float or tuple. This value is used when 45 | filtering a dataset with ``--max-positions``.""" 46 | raise NotImplementedError 47 | 48 | def ordered_indices(self): 49 | """Return an ordered list of indices. Batches will be constructed based 50 | on this order.""" 51 | return np.arange(len(self)) 52 | 53 | @property 54 | def supports_prefetch(self): 55 | """Whether this dataset supports prefetching.""" 56 | return False 57 | 58 | def attr(self, attr: str, index: int): 59 | return getattr(self, attr, None) 60 | 61 | def prefetch(self, indices): 62 | """Prefetch the data required for this epoch.""" 63 | raise NotImplementedError 64 | 65 | 66 | class FairseqIterableDataset(torch.utils.data.IterableDataset, EpochListening): 67 | """For datasets that need to be read sequentially, usually because the data 68 | is being streamed or otherwise can't be manipulated on a single machine. 69 | """ 70 | 71 | def __iter__(self): 72 | raise NotImplementedError 73 | -------------------------------------------------------------------------------- /fairseq/modules/learned_positional_embedding.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 Dict, Optional 7 | 8 | import torch 9 | import torch.nn as nn 10 | import torch.nn.functional as F 11 | from fairseq import utils 12 | from torch import Tensor 13 | 14 | 15 | class LearnedPositionalEmbedding(nn.Embedding): 16 | """ 17 | This module learns positional embeddings up to a fixed maximum size. 18 | Padding ids are ignored by either offsetting based on padding_idx 19 | or by setting padding_idx to None and ensuring that the appropriate 20 | position ids are passed to the forward function. 21 | """ 22 | 23 | def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int): 24 | super().__init__(num_embeddings, embedding_dim, padding_idx) 25 | self.onnx_trace = False 26 | if self.padding_idx is not None: 27 | self.max_positions = self.num_embeddings - self.padding_idx - 1 28 | else: 29 | self.max_positions = self.num_embeddings 30 | 31 | def forward( 32 | self, 33 | input: Tensor, 34 | incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, 35 | positions: Optional[Tensor] = None, 36 | ): 37 | """Input is expected to be of size [bsz x seqlen].""" 38 | assert (positions is None) or ( 39 | self.padding_idx is None 40 | ), "If positions is pre-computed then padding_idx should not be set." 41 | 42 | if positions is None: 43 | if incremental_state is not None: 44 | # positions is the same for every token when decoding a single step 45 | # Without the int() cast, it doesn't work in some cases when exporting to ONNX 46 | positions = torch.zeros( 47 | (1, 1), device=input.device, dtype=input.dtype 48 | ).fill_(int(self.padding_idx + input.size(1))) 49 | else: 50 | positions = utils.make_positions( 51 | input, self.padding_idx, onnx_trace=self.onnx_trace 52 | ) 53 | return F.embedding( 54 | positions, 55 | self.weight, 56 | self.padding_idx, 57 | self.max_norm, 58 | self.norm_type, 59 | self.scale_grad_by_freq, 60 | self.sparse, 61 | ) 62 | -------------------------------------------------------------------------------- /fairseq/modules/lightconv_layer/lightconv_cuda.cuh: -------------------------------------------------------------------------------- 1 | /** 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 | #include 9 | #include 10 | 11 | #include 12 | #include 13 | 14 | #include 15 | #include 16 | #include 17 | #include 18 | #include 19 | #include 20 | 21 | #include 22 | #include 23 | 24 | #define SHFL_MASK 0xffffffff 25 | 26 | template 27 | __global__ 28 | void lightconv_forward_kernel(const scalar_t* input, 29 | const scalar_t* filters, 30 | int minibatch, int sequenceLength, 31 | int numFeatures, int numFiltersInBlock, 32 | scalar_t* output); 33 | 34 | template 35 | __global__ 36 | void lightconv_grad_wrt_input_kernel( 37 | const scalar_t* input, 38 | const scalar_t* filters, 39 | int minibatch, 40 | int sequenceLength, 41 | int numFeatures, 42 | int numFiltersInBlock, 43 | scalar_t* output); 44 | 45 | template 46 | __global__ 47 | void lightconv_grad_wrt_weights_firstpass_short_kernel( 48 | const scalar_t* input, 49 | const scalar_t* gradInput, 50 | int minibatch, 51 | int sequenceLength, 52 | int numFeatures, 53 | int numFiltersInBlock, 54 | int numHeads, 55 | float* output); 56 | 57 | template 58 | __global__ 59 | void lightconv_grad_wrt_weights_secondpass_short_kernel( 60 | const float* input, 61 | const int minibatch, 62 | const int numFiltersInBlock, 63 | scalar_t* output); 64 | 65 | template 66 | __global__ 67 | void lightconv_grad_wrt_weights_firstpass_kernel( 68 | const scalar_t* input, 69 | const scalar_t* gradInput, 70 | int minibatch, 71 | int sequenceLength, 72 | int numFeatures, 73 | int numFiltersInBlock, 74 | float* output); 75 | 76 | template 77 | __global__ 78 | void lightconv_grad_wrt_weights_secondpass_kernel( 79 | const float* input, 80 | const int minibatch, 81 | const int numFiltersInBlock, 82 | scalar_t* output); 83 | 84 | -------------------------------------------------------------------------------- /fairseq/optim/lr_scheduler/fixed_schedule.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 FairseqLRScheduler, register_lr_scheduler 7 | 8 | 9 | @register_lr_scheduler('fixed') 10 | class FixedSchedule(FairseqLRScheduler): 11 | """Decay the LR on a fixed schedule.""" 12 | 13 | def __init__(self, args, optimizer): 14 | super().__init__(args, optimizer) 15 | 16 | # set defaults 17 | args.warmup_updates = getattr(args, 'warmup_updates', 0) or 0 18 | 19 | self.lr = args.lr[0] 20 | if args.warmup_updates > 0: 21 | self.warmup_factor = 1. / args.warmup_updates 22 | else: 23 | self.warmup_factor = 1 24 | 25 | @staticmethod 26 | def add_args(parser): 27 | """Add arguments to the parser for this LR scheduler.""" 28 | # fmt: off 29 | parser.add_argument('--force-anneal', '--fa', type=int, metavar='N', 30 | help='force annealing at specified epoch') 31 | parser.add_argument('--lr-shrink', default=0.1, type=float, metavar='LS', 32 | help='shrink factor for annealing, lr_new = (lr * lr_shrink)') 33 | parser.add_argument('--warmup-updates', default=0, type=int, metavar='N', 34 | help='warmup the learning rate linearly for the first N updates') 35 | # fmt: on 36 | 37 | def get_next_lr(self, epoch): 38 | lrs = self.args.lr 39 | if self.args.force_anneal is None or epoch < self.args.force_anneal: 40 | # use fixed LR schedule 41 | next_lr = lrs[min(epoch, len(lrs) - 1)] 42 | else: 43 | # annneal based on lr_shrink 44 | next_lr = lrs[-1] * self.args.lr_shrink ** (epoch + 1 - self.args.force_anneal) 45 | return next_lr 46 | 47 | def step(self, epoch, val_loss=None): 48 | """Update the learning rate at the end of the given epoch.""" 49 | super().step(epoch, val_loss) 50 | self.lr = self.get_next_lr(epoch) 51 | self.optimizer.set_lr(self.warmup_factor * self.lr) 52 | return self.optimizer.get_lr() 53 | 54 | def step_update(self, num_updates): 55 | """Update the learning rate after each update.""" 56 | if self.args.warmup_updates > 0 and num_updates < self.args.warmup_updates: 57 | self.warmup_factor = (num_updates + 1) / float(self.args.warmup_updates) 58 | self.optimizer.set_lr(self.warmup_factor * self.lr) 59 | return self.optimizer.get_lr() 60 | -------------------------------------------------------------------------------- /fairseq/modules/adaptive_input.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 | from torch import nn 9 | 10 | from typing import List 11 | 12 | 13 | class AdaptiveInput(nn.Module): 14 | 15 | def __init__( 16 | self, 17 | vocab_size: int, 18 | padding_idx: int, 19 | initial_dim: int, 20 | factor: float, 21 | output_dim: int, 22 | cutoff: List[int], 23 | ): 24 | super().__init__() 25 | 26 | if vocab_size > cutoff[-1]: 27 | cutoff = cutoff + [vocab_size] 28 | else: 29 | assert vocab_size == cutoff[ 30 | -1], 'cannot specify cutoff larger than vocab size' 31 | 32 | self.cutoff = cutoff 33 | self.embedding_dim = output_dim 34 | self.padding_idx = padding_idx 35 | 36 | self.embeddings = nn.ModuleList() 37 | for i in range(len(self.cutoff)): 38 | prev = self.cutoff[i - 1] if i > 0 else 0 39 | size = self.cutoff[i] - prev 40 | dim = int(initial_dim // (factor ** i)) 41 | seq = nn.Sequential( 42 | nn.Embedding(size, dim, self.padding_idx), 43 | nn.Linear(dim, output_dim, bias=False) 44 | ) 45 | self.embeddings.append(seq) 46 | self.padding_idx = None 47 | self.padding_idx = padding_idx 48 | 49 | def init_weights(m): 50 | if isinstance(m, nn.Embedding): 51 | nn.init.normal_(m.weight, mean=0, std=m.weight.shape[1] ** -0.5) 52 | nn.init.constant_(m.weight[padding_idx], 0) 53 | elif hasattr(m, 'weight'): 54 | nn.init.xavier_uniform_(m.weight) 55 | 56 | self.apply(init_weights) 57 | 58 | self.register_buffer('_float_tensor', torch.FloatTensor(1)) 59 | 60 | def weights_for_band(self, band: int): 61 | return self.embeddings[band][0].weight, self.embeddings[band][1].weight 62 | 63 | def forward(self, input: torch.Tensor): 64 | result = self._float_tensor.new(input.shape + (self.embedding_dim,)) 65 | for i in range(len(self.cutoff)): 66 | mask = input.lt(self.cutoff[i]) 67 | if i > 0: 68 | mask.mul_(input.ge(self.cutoff[i - 1])) 69 | chunk_input = input[mask] - self.cutoff[i - 1] 70 | else: 71 | chunk_input = input[mask] 72 | if mask.any(): 73 | result[mask] = self.embeddings[i](chunk_input) 74 | return result 75 | -------------------------------------------------------------------------------- /fairseq/models/model_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, Optional 7 | 8 | import torch 9 | from torch import Tensor 10 | 11 | 12 | @torch.jit.script 13 | def script_skip_tensor_list(x: List[Tensor], mask): 14 | res = [xi[mask] if xi.size(0) == mask.size(0) else xi[:, mask] for xi in x] 15 | outputs = [] 16 | for i, t in enumerate(res): 17 | if t.numel() != 0: 18 | outputs.append(t) 19 | else: 20 | outputs.append(x[i]) 21 | return outputs 22 | 23 | 24 | @torch.jit.script 25 | def script_skip_tensor(x: Tensor, mask): 26 | # None case 27 | if x.size(0) == 0: 28 | return x 29 | res = x[mask] if x.size(0) == mask.size(0) else x[:, mask] 30 | if res.numel() == 0: 31 | return x 32 | else: 33 | return res 34 | 35 | 36 | @torch.jit.script 37 | def expand_2d_or_3d_tensor(x, trg_dim: int, padding_idx: int): 38 | """ 39 | Expand 2D/3D tensor on dim=1 40 | """ 41 | if x is None: 42 | return None 43 | 44 | assert x.dim() == 2 or x.dim() == 3 45 | assert trg_dim >= x.size(1), (trg_dim, x.size()) 46 | if trg_dim == x.size(1): 47 | return x 48 | 49 | dims = [x.size(0), trg_dim - x.size(1)] 50 | if x.dim() == 3: 51 | dims.append(x.size(2)) 52 | x = torch.cat([x, torch.zeros(dims).to(x).fill_(padding_idx)], 1) 53 | 54 | return x 55 | 56 | 57 | @torch.jit.script 58 | def coalesce(x: Optional[Tensor], y: Tensor) -> Tensor: 59 | return x if x is not None else y 60 | 61 | 62 | @torch.jit.script 63 | def fill_tensors(x: Optional[Tensor], mask, y: Optional[Tensor], padding_idx: int) -> Optional[Tensor]: 64 | """ 65 | Filling tensor x with y at masked positions (dim=0). 66 | """ 67 | if x is None or x.size()[0] == 0 or y is None: 68 | return x 69 | assert x.dim() == y.dim() and mask.size(0) == x.size(0) 70 | assert x.dim() == 2 or (x.dim() == 3 and x.size(2) == y.size(2)) 71 | 72 | n_selected = mask.sum() 73 | if n_selected == 0: 74 | return x 75 | assert n_selected == y.size(0) 76 | if n_selected == x.size(0): 77 | return y 78 | 79 | if x.size(1) < y.size(1): 80 | x = expand_2d_or_3d_tensor(x, y.size(1), padding_idx) 81 | x[mask] = y 82 | elif x.size(1) > y.size(1): 83 | x[mask] = torch.tensor(padding_idx).type_as(x) 84 | if x.dim() == 2: 85 | x[mask, :y.size(1)] = y 86 | else: 87 | x[mask, :y.size(1), :] = y 88 | else: 89 | x[mask] = y 90 | return x 91 | -------------------------------------------------------------------------------- /tests/test_sparse_multihead_attention.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 | import unittest 8 | from fairseq.modules.sparse_multihead_attention import SparseMultiheadAttention 9 | 10 | 11 | class TestSparseMultiheadAttention(unittest.TestCase): 12 | def test_sparse_multihead_attention(self): 13 | attn_weights = torch.randn(1, 8, 8) 14 | bidirectional_sparse_mask = torch.tensor([ 15 | [0, 0, 0, 0, 0, float('-inf'), float('-inf'), 0], 16 | [0, 0, 0, 0, 0, float('-inf'), float('-inf'), 0], 17 | [0, 0, 0, 0, 0, float('-inf'), float('-inf'), 0], 18 | [0, 0, 0, 0, 0, float('-inf'), float('-inf'), 0], 19 | [float('-inf'), float('-inf'), float('-inf'), 0, 0, 0, 0, 0], 20 | [float('-inf'), float('-inf'), float('-inf'), 0, 0, 0, 0, 0], 21 | [float('-inf'), float('-inf'), float('-inf'), 0, 0, 0, 0, 0], 22 | [float('-inf'), float('-inf'), float('-inf'), 0, 0, 0, 0, 0] 23 | ]) 24 | 25 | bidirectional_attention = SparseMultiheadAttention(16, 1, stride=4, expressivity=1, is_bidirectional=True) 26 | bidirectional_attention_sparse_mask = bidirectional_attention.buffered_sparse_mask(attn_weights, 8, 8) 27 | torch.all(torch.eq(bidirectional_attention_sparse_mask, bidirectional_sparse_mask)) 28 | 29 | sparse_mask = torch.tensor([ 30 | [0, float('-inf'), float('-inf'), float('-inf'), float('-inf'), float('-inf'), 31 | float('-inf'), float('-inf')], 32 | [0, 0, float('-inf'), float('-inf'), float('-inf'), float('-inf'), float('-inf'), float('-inf')], 33 | [0, 0, 0, float('-inf'), float('-inf'), float('-inf'), float('-inf'), float('-inf')], 34 | [0, 0, 0, 0, float('-inf'), float('-inf'), float('-inf'), float('-inf')], 35 | [0, 0, 0, 0, 0, float('-inf'), float('-inf'), float('-inf')], 36 | [float('-inf'), float('-inf'), float('-inf'), 0, 0, 0, float('-inf'), float('-inf')], 37 | [float('-inf'), float('-inf'), float('-inf'), 0, 0, 0, 0, float('-inf')], 38 | [float('-inf'), float('-inf'), float('-inf'), 0, 0, 0, 0, 0], 39 | ]) 40 | 41 | attention = SparseMultiheadAttention(16, 1, stride=4, expressivity=1, is_bidirectional=False) 42 | attention_sparse_mask = attention.buffered_sparse_mask(attn_weights, 8, 8) 43 | 44 | torch.all(torch.eq(attention_sparse_mask, sparse_mask)) 45 | 46 | 47 | if __name__ == '__main__': 48 | unittest.main() 49 | -------------------------------------------------------------------------------- /fairseq/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 .adaptive_input import AdaptiveInput 7 | from .adaptive_softmax import AdaptiveSoftmax 8 | from .beamable_mm import BeamableMM 9 | from .character_token_embedder import CharacterTokenEmbedder 10 | from .conv_tbc import ConvTBC 11 | from .cross_entropy import cross_entropy 12 | from .downsampled_multihead_attention import DownsampledMultiHeadAttention 13 | from .dynamic_convolution import DynamicConv, DynamicConv1dTBC 14 | from .dynamic_crf_layer import DynamicCRF 15 | from .fp32_group_norm import Fp32GroupNorm 16 | from .gelu import gelu, gelu_accurate 17 | from .grad_multiply import GradMultiply 18 | from .gumbel_vector_quantizer import GumbelVectorQuantizer 19 | from .kmeans_vector_quantizer import KmeansVectorQuantizer 20 | from .layer_norm import Fp32LayerNorm, LayerNorm 21 | from .learned_positional_embedding import LearnedPositionalEmbedding 22 | from .lightweight_convolution import LightweightConv, LightweightConv1dTBC 23 | from .linearized_convolution import LinearizedConvolution 24 | from .multihead_attention import MultiheadAttention 25 | from .positional_embedding import PositionalEmbedding 26 | from .scalar_bias import ScalarBias 27 | from .sinusoidal_positional_embedding import SinusoidalPositionalEmbedding 28 | from .transformer_sentence_encoder_layer import TransformerSentenceEncoderLayer 29 | from .transformer_sentence_encoder import TransformerSentenceEncoder 30 | from .unfold import unfold1d 31 | from .transformer_layer import TransformerDecoderLayer, TransformerEncoderLayer 32 | from .vggblock import VGGBlock 33 | from .one_embed import OneEmbed 34 | 35 | __all__ = [ 36 | 'AdaptiveInput', 37 | 'AdaptiveSoftmax', 38 | 'BeamableMM', 39 | 'CharacterTokenEmbedder', 40 | 'ConvTBC', 41 | 'cross_entropy', 42 | 'DownsampledMultiHeadAttention', 43 | 'DynamicConv1dTBC', 44 | 'DynamicConv', 45 | 'DynamicCRF', 46 | 'Fp32GroupNorm', 47 | 'Fp32LayerNorm', 48 | 'gelu', 49 | 'gelu_accurate', 50 | 'GradMultiply', 51 | 'GumbelVectorQuantizer', 52 | 'KmeansVectorQuantizer', 53 | 'LayerNorm', 54 | 'LearnedPositionalEmbedding', 55 | 'LightweightConv1dTBC', 56 | 'LightweightConv', 57 | 'LinearizedConvolution', 58 | 'MultiheadAttention', 59 | 'PositionalEmbedding', 60 | 'ScalarBias', 61 | 'SinusoidalPositionalEmbedding', 62 | 'TransformerSentenceEncoderLayer', 63 | 'TransformerSentenceEncoder', 64 | 'TransformerDecoderLayer', 65 | 'TransformerEncoderLayer', 66 | 'VGGBlock', 67 | 'unfold1d', 68 | 'OneEmbed', 69 | ] 70 | -------------------------------------------------------------------------------- /fairseq/models/distributed_fairseq_model.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 inspect 7 | 8 | import torch.nn as nn 9 | 10 | from fairseq.legacy_distributed_data_parallel import LegacyDistributedDataParallel 11 | from fairseq.models import BaseFairseqModel 12 | 13 | 14 | def DistributedFairseqModel(args, model, process_group=None): 15 | """ 16 | Wrap a *model* to support distributed data parallel training. 17 | 18 | This is similar to the built-in DistributedDataParallel, but allows 19 | additional configuration of the DistributedDataParallel class to 20 | use, and also provides easier access to the wrapped model by 21 | forwarding requests for missing attributes to the wrapped model. 22 | 23 | Args: 24 | args (argparse.Namespace): fairseq args 25 | model (BaseFairseqModel): model to wrap 26 | """ 27 | # determine which DDP class to extend 28 | assert isinstance(model, nn.Module) 29 | if args.ddp_backend == 'c10d': 30 | ddp_class = nn.parallel.DistributedDataParallel 31 | init_kwargs = dict( 32 | module=model, 33 | device_ids=[args.device_id], 34 | output_device=args.device_id, 35 | broadcast_buffers=args.broadcast_buffers, 36 | bucket_cap_mb=args.bucket_cap_mb, 37 | process_group=process_group, 38 | ) 39 | # Maintain backward compatibility 40 | if 'check_reduction' in inspect.getargspec(ddp_class)[0]: 41 | init_kwargs['check_reduction'] = True 42 | if 'find_unused_parameters' in inspect.getargspec(ddp_class)[0]: 43 | init_kwargs['find_unused_parameters'] = args.find_unused_parameters 44 | elif args.ddp_backend == 'no_c10d': 45 | ddp_class = LegacyDistributedDataParallel 46 | init_kwargs = dict( 47 | module=model, 48 | world_size=args.distributed_world_size, 49 | buffer_size=2**28, 50 | process_group=process_group, 51 | ) 52 | else: 53 | raise ValueError('Unknown --ddp-backend: ' + args.ddp_backend) 54 | 55 | class _DistributedFairseqModel(ddp_class): 56 | """Extend DistributedDataParallel to check for missing 57 | attributes in the wrapped module.""" 58 | 59 | def __init__(self, *args, **kwargs): 60 | super().__init__(*args, **kwargs) 61 | 62 | def __getattr__(self, name): 63 | wrapped_module = super().__getattr__('module') 64 | if hasattr(wrapped_module, name): 65 | return getattr(wrapped_module, name) 66 | return super().__getattr__(name) 67 | 68 | return _DistributedFairseqModel(**init_kwargs) 69 | -------------------------------------------------------------------------------- /fairseq/model_parallel/modules/transformer_layer.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.modules import ( 7 | TransformerEncoderLayer, 8 | TransformerDecoderLayer, 9 | ) 10 | 11 | from fairseq.model_parallel.modules import ModelParallelMultiheadAttention 12 | 13 | try: 14 | from fairseq.model_parallel.megatron.mpu import ( 15 | ColumnParallelLinear, 16 | RowParallelLinear, 17 | ) 18 | has_megatron_submodule = True 19 | except (ImportError, ModuleNotFoundError): 20 | has_megatron_submodule = False 21 | 22 | 23 | class ModelParallelTransformerEncoderLayer(TransformerEncoderLayer): 24 | """Encoder layer block over multiple gpus. 25 | 26 | See "Megatron-LM: https://arxiv.org/pdf/1909.08053.pdf" for more details. 27 | """ 28 | 29 | def build_fc1(self, input_dim, output_dim): 30 | return ColumnParallelLinear(input_dim, output_dim, gather_output=False) 31 | 32 | def build_fc2(self, input_dim, output_dim): 33 | return RowParallelLinear(input_dim, output_dim, input_is_parallel=True) 34 | 35 | def build_self_attention(self, embed_dim, args, **unused_kwargs): 36 | return ModelParallelMultiheadAttention( 37 | embed_dim, 38 | args.encoder_attention_heads, 39 | dropout=args.attention_dropout, 40 | self_attention=True, 41 | ) 42 | 43 | 44 | class ModelParallelTransformerDecoderLayer(TransformerDecoderLayer): 45 | """Decoder layer block. 46 | 47 | See "Megatron-LM: https://arxiv.org/pdf/1909.08053.pdf" for more details. 48 | """ 49 | def build_fc1(self, input_dim, output_dim): 50 | return ColumnParallelLinear(input_dim, output_dim, gather_output=False) 51 | 52 | def build_fc2(self, input_dim, output_dim): 53 | return RowParallelLinear(input_dim, output_dim, input_is_parallel=True) 54 | 55 | def build_self_attention(self, embed_dim, args, **unused_kwargs): 56 | return ModelParallelMultiheadAttention( 57 | embed_dim=embed_dim, 58 | num_heads=args.decoder_attention_heads, 59 | dropout=args.attention_dropout, 60 | self_attention=not getattr(args, "cross_self_attention", False), 61 | ) 62 | 63 | def build_encoder_attention(self, embed_dim, args, **unused_kwargs): 64 | return ModelParallelMultiheadAttention( 65 | embed_dim=embed_dim, 66 | num_heads=args.decoder_attention_heads, 67 | kdim=getattr(args, "encoder_embed_dim", None), 68 | vdim=getattr(args, "encoder_embed_dim", None), 69 | dropout=args.attention_dropout, 70 | encoder_decoder_attention=True, 71 | ) 72 | -------------------------------------------------------------------------------- /fairseq/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 argparse 7 | import importlib 8 | import os 9 | 10 | from .fairseq_task import FairseqTask 11 | 12 | TASK_REGISTRY = {} 13 | TASK_CLASS_NAMES = set() 14 | 15 | 16 | def setup_task(args, **kwargs): 17 | return TASK_REGISTRY[args.task].setup_task(args, **kwargs) 18 | 19 | 20 | def register_task(name): 21 | """ 22 | New tasks can be added to fairseq with the 23 | :func:`~fairseq.tasks.register_task` function decorator. 24 | 25 | For example:: 26 | 27 | @register_task('classification') 28 | class ClassificationTask(FairseqTask): 29 | (...) 30 | 31 | .. note:: 32 | 33 | All Tasks must implement the :class:`~fairseq.tasks.FairseqTask` 34 | interface. 35 | 36 | Please see the 37 | 38 | Args: 39 | name (str): the name of the task 40 | """ 41 | 42 | def register_task_cls(cls): 43 | if name in TASK_REGISTRY: 44 | raise ValueError('Cannot register duplicate task ({})'.format(name)) 45 | if not issubclass(cls, FairseqTask): 46 | raise ValueError('Task ({}: {}) must extend FairseqTask'.format(name, cls.__name__)) 47 | if cls.__name__ in TASK_CLASS_NAMES: 48 | raise ValueError('Cannot register task with duplicate class name ({})'.format(cls.__name__)) 49 | TASK_REGISTRY[name] = cls 50 | TASK_CLASS_NAMES.add(cls.__name__) 51 | return cls 52 | 53 | return register_task_cls 54 | 55 | 56 | def get_task(name): 57 | return TASK_REGISTRY[name] 58 | 59 | 60 | # automatically import any Python files in the tasks/ directory 61 | tasks_dir = os.path.dirname(__file__) 62 | for file in os.listdir(tasks_dir): 63 | path = os.path.join(tasks_dir, file) 64 | if ( 65 | not file.startswith('_') 66 | and not file.startswith('.') 67 | and (file.endswith('.py') or os.path.isdir(path)) 68 | ): 69 | task_name = file[:file.find('.py')] if file.endswith('.py') else file 70 | importlib.import_module('fairseq.tasks.' + task_name) 71 | 72 | # expose `task_parser` for sphinx 73 | if task_name in TASK_REGISTRY: 74 | parser = argparse.ArgumentParser(add_help=False) 75 | group_task = parser.add_argument_group('Task name') 76 | # fmt: off 77 | group_task.add_argument('--task', metavar=task_name, 78 | help='Enable this task with: ``--task=' + task_name + '``') 79 | # fmt: on 80 | group_args = parser.add_argument_group('Additional command-line arguments') 81 | TASK_REGISTRY[task_name].add_args(group_args) 82 | globals()[task_name + '_parser'] = parser 83 | -------------------------------------------------------------------------------- /scripts/split_train_valid_docs.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 | Split a large file into a train and valid set while respecting document 8 | boundaries. Documents should be separated by a single empty line. 9 | """ 10 | 11 | import argparse 12 | import random 13 | import sys 14 | 15 | 16 | def main(): 17 | parser = argparse.ArgumentParser() 18 | parser.add_argument('input') 19 | parser.add_argument('sample_output', help='train output file') 20 | parser.add_argument('remainder_output', help='valid output file') 21 | parser.add_argument('-k', type=int, help="remainder size") 22 | parser.add_argument('--lines', action='store_true', 23 | help='split lines instead of docs') 24 | args = parser.parse_args() 25 | 26 | assert args.k is not None 27 | 28 | sample = [] 29 | remainder = [] 30 | num_docs = [0] 31 | 32 | def update_sample(doc): 33 | if len(sample) < args.k: 34 | sample.append(doc.copy()) 35 | else: 36 | i = num_docs[0] 37 | j = random.randrange(i + 1) 38 | if j < args.k: 39 | remainder.append(sample[j]) 40 | sample[j] = doc.copy() 41 | else: 42 | remainder.append(doc.copy()) 43 | num_docs[0] += 1 44 | doc.clear() 45 | 46 | with open(args.input, 'r', encoding='utf-8') as h: 47 | doc = [] 48 | for i, line in enumerate(h): 49 | if line.strip() == "": # empty line indicates new document 50 | update_sample(doc) 51 | else: 52 | doc.append(line) 53 | if args.lines: 54 | update_sample(doc) 55 | if i % 1000000 == 0: 56 | print(i, file=sys.stderr, end="", flush=True) 57 | elif i % 100000 == 0: 58 | print(".", file=sys.stderr, end="", flush=True) 59 | if len(doc) > 0: 60 | update_sample(doc) 61 | print(file=sys.stderr, flush=True) 62 | 63 | assert len(sample) == args.k 64 | 65 | with open(args.sample_output, 'w', encoding='utf-8') as out: 66 | first = True 67 | for doc in sample: 68 | if not first and not args.lines: 69 | out.write("\n") 70 | first = False 71 | for line in doc: 72 | out.write(line) 73 | 74 | with open(args.remainder_output, 'w', encoding='utf-8') as out: 75 | first = True 76 | for doc in remainder: 77 | if not first and not args.lines: 78 | out.write("\n") 79 | first = False 80 | for line in doc: 81 | out.write(line) 82 | 83 | 84 | if __name__ == '__main__': 85 | main() 86 | -------------------------------------------------------------------------------- /fairseq/registry.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 | 8 | 9 | REGISTRIES = {} 10 | 11 | 12 | def setup_registry( 13 | registry_name: str, 14 | base_class=None, 15 | default=None, 16 | ): 17 | assert registry_name.startswith('--') 18 | registry_name = registry_name[2:].replace('-', '_') 19 | 20 | REGISTRY = {} 21 | REGISTRY_CLASS_NAMES = set() 22 | 23 | # maintain a registry of all registries 24 | if registry_name in REGISTRIES: 25 | return # registry already exists 26 | REGISTRIES[registry_name] = { 27 | 'registry': REGISTRY, 28 | 'default': default, 29 | } 30 | 31 | def build_x(args, *extra_args, **extra_kwargs): 32 | choice = getattr(args, registry_name, None) 33 | if choice is None: 34 | return None 35 | cls = REGISTRY[choice] 36 | if hasattr(cls, 'build_' + registry_name): 37 | builder = getattr(cls, 'build_' + registry_name) 38 | else: 39 | builder = cls 40 | set_defaults(args, cls) 41 | return builder(args, *extra_args, **extra_kwargs) 42 | 43 | def register_x(name): 44 | 45 | def register_x_cls(cls): 46 | if name in REGISTRY: 47 | raise ValueError('Cannot register duplicate {} ({})'.format(registry_name, name)) 48 | if cls.__name__ in REGISTRY_CLASS_NAMES: 49 | raise ValueError( 50 | 'Cannot register {} with duplicate class name ({})'.format( 51 | registry_name, cls.__name__, 52 | ) 53 | ) 54 | if base_class is not None and not issubclass(cls, base_class): 55 | raise ValueError('{} must extend {}'.format(cls.__name__, base_class.__name__)) 56 | REGISTRY[name] = cls 57 | REGISTRY_CLASS_NAMES.add(cls.__name__) 58 | return cls 59 | 60 | return register_x_cls 61 | 62 | return build_x, register_x, REGISTRY 63 | 64 | 65 | def set_defaults(args, cls): 66 | """Helper to set default arguments based on *add_args*.""" 67 | if not hasattr(cls, 'add_args'): 68 | return 69 | parser = argparse.ArgumentParser(argument_default=argparse.SUPPRESS, allow_abbrev=False) 70 | cls.add_args(parser) 71 | # copied from argparse.py: 72 | defaults = argparse.Namespace() 73 | for action in parser._actions: 74 | if action.dest is not argparse.SUPPRESS: 75 | if not hasattr(defaults, action.dest): 76 | if action.default is not argparse.SUPPRESS: 77 | setattr(defaults, action.dest, action.default) 78 | for key, default_value in vars(defaults).items(): 79 | if not hasattr(args, key): 80 | setattr(args, key, default_value) 81 | -------------------------------------------------------------------------------- /tests/test_metrics.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 | import uuid 8 | 9 | from fairseq import metrics 10 | 11 | 12 | class TestMetrics(unittest.TestCase): 13 | 14 | def test_nesting(self): 15 | with metrics.aggregate() as a: 16 | metrics.log_scalar('loss', 1) 17 | with metrics.aggregate() as b: 18 | metrics.log_scalar('loss', 2) 19 | 20 | self.assertEqual(a.get_smoothed_values()['loss'], 1.5) 21 | self.assertEqual(b.get_smoothed_values()['loss'], 2) 22 | 23 | def test_new_root(self): 24 | with metrics.aggregate() as a: 25 | metrics.log_scalar('loss', 1) 26 | with metrics.aggregate(new_root=True) as b: 27 | metrics.log_scalar('loss', 2) 28 | 29 | self.assertEqual(a.get_smoothed_values()['loss'], 1) 30 | self.assertEqual(b.get_smoothed_values()['loss'], 2) 31 | 32 | def test_nested_new_root(self): 33 | with metrics.aggregate() as layer1: 34 | metrics.log_scalar('loss', 1) 35 | with metrics.aggregate(new_root=True) as layer2: 36 | metrics.log_scalar('loss', 2) 37 | with metrics.aggregate() as layer3: 38 | metrics.log_scalar('loss', 3) 39 | with metrics.aggregate(new_root=True) as layer4: 40 | metrics.log_scalar('loss', 4) 41 | metrics.log_scalar('loss', 1.5) 42 | 43 | self.assertEqual(layer4.get_smoothed_values()['loss'], 4) 44 | self.assertEqual(layer3.get_smoothed_values()['loss'], 3) 45 | self.assertEqual(layer2.get_smoothed_values()['loss'], 2.5) 46 | self.assertEqual(layer1.get_smoothed_values()['loss'], 1.25) 47 | 48 | def test_named(self): 49 | name = str(uuid.uuid4()) 50 | metrics.reset_meters(name) 51 | 52 | with metrics.aggregate(name): 53 | metrics.log_scalar('loss', 1) 54 | 55 | metrics.log_scalar('loss', 3) 56 | 57 | with metrics.aggregate(name): 58 | metrics.log_scalar('loss', 2) 59 | 60 | self.assertEqual(metrics.get_smoothed_values(name)['loss'], 1.5) 61 | 62 | def test_nested_duplicate_names(self): 63 | name = str(uuid.uuid4()) 64 | metrics.reset_meters(name) 65 | 66 | with metrics.aggregate(name): 67 | metrics.log_scalar('loss', 1) 68 | with metrics.aggregate() as other: 69 | with metrics.aggregate(name): 70 | metrics.log_scalar('loss', 2) 71 | metrics.log_scalar('loss', 6) 72 | 73 | self.assertEqual(metrics.get_smoothed_values(name)['loss'], 3) 74 | self.assertEqual(other.get_smoothed_values()['loss'], 2) 75 | 76 | 77 | if __name__ == '__main__': 78 | unittest.main() 79 | -------------------------------------------------------------------------------- /docs/overview.rst: -------------------------------------------------------------------------------- 1 | Overview 2 | ======== 3 | 4 | Fairseq can be extended through user-supplied `plug-ins 5 | `_. We support five kinds of 6 | plug-ins: 7 | 8 | - :ref:`Models` define the neural network architecture and encapsulate all of the 9 | learnable parameters. 10 | - :ref:`Criterions` compute the loss function given the model outputs and targets. 11 | - :ref:`Tasks` store dictionaries and provide helpers for loading/iterating over 12 | Datasets, initializing the Model/Criterion and calculating the loss. 13 | - :ref:`Optimizers` update the Model parameters based on the gradients. 14 | - :ref:`Learning Rate Schedulers` update the learning rate over the course of 15 | training. 16 | 17 | **Training Flow** 18 | 19 | Given a ``model``, ``criterion``, ``task``, ``optimizer`` and ``lr_scheduler``, 20 | fairseq implements the following high-level training flow:: 21 | 22 | for epoch in range(num_epochs): 23 | itr = task.get_batch_iterator(task.dataset('train')) 24 | for num_updates, batch in enumerate(itr): 25 | task.train_step(batch, model, criterion, optimizer) 26 | average_and_clip_gradients() 27 | optimizer.step() 28 | lr_scheduler.step_update(num_updates) 29 | lr_scheduler.step(epoch) 30 | 31 | where the default implementation for ``task.train_step`` is roughly:: 32 | 33 | def train_step(self, batch, model, criterion, optimizer, **unused): 34 | loss = criterion(model, batch) 35 | optimizer.backward(loss) 36 | return loss 37 | 38 | **Registering new plug-ins** 39 | 40 | New plug-ins are *registered* through a set of ``@register`` function 41 | decorators, for example:: 42 | 43 | @register_model('my_lstm') 44 | class MyLSTM(FairseqEncoderDecoderModel): 45 | (...) 46 | 47 | Once registered, new plug-ins can be used with the existing :ref:`Command-line 48 | Tools`. See the Tutorial sections for more detailed walkthroughs of how to add 49 | new plug-ins. 50 | 51 | **Loading plug-ins from another directory** 52 | 53 | New plug-ins can be defined in a custom module stored in the user system. In 54 | order to import the module, and make the plugin available to *fairseq*, the 55 | command line supports the ``--user-dir`` flag that can be used to specify a 56 | custom location for additional modules to load into *fairseq*. 57 | 58 | For example, assuming this directory tree:: 59 | 60 | /home/user/my-module/ 61 | └── __init__.py 62 | 63 | with ``__init__.py``:: 64 | 65 | from fairseq.models import register_model_architecture 66 | from fairseq.models.transformer import transformer_vaswani_wmt_en_de_big 67 | 68 | @register_model_architecture('transformer', 'my_transformer') 69 | def transformer_mmt_big(args): 70 | transformer_vaswani_wmt_en_de_big(args) 71 | 72 | it is possible to invoke the :ref:`fairseq-train` script with the new architecture with:: 73 | 74 | fairseq-train ... --user-dir /home/user/my-module -a my_transformer --task translation 75 | -------------------------------------------------------------------------------- /fairseq/data/plasma_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 subprocess 7 | import tempfile 8 | 9 | 10 | class PlasmaArray(object): 11 | """ 12 | Wrapper around numpy arrays that automatically moves the data to shared 13 | memory upon serialization. This is particularly helpful when passing numpy 14 | arrays through multiprocessing, so that data is not unnecessarily 15 | duplicated or pickled. 16 | """ 17 | 18 | def __init__(self, array): 19 | super().__init__() 20 | self.array = array 21 | self.disable = array.nbytes < 134217728 # disable for arrays <128MB 22 | self.object_id = None 23 | self.path = None 24 | 25 | # variables with underscores shouldn't be pickled 26 | self._client = None 27 | self._server = None 28 | self._server_tmp = None 29 | self._plasma = None 30 | 31 | @property 32 | def plasma(self): 33 | if self._plasma is None and not self.disable: 34 | try: 35 | import pyarrow.plasma as plasma 36 | self._plasma = plasma 37 | except ImportError: 38 | self._plasma = None 39 | return self._plasma 40 | 41 | def start_server(self): 42 | if self.plasma is None or self._server is not None: 43 | return 44 | assert self.object_id is None 45 | assert self.path is None 46 | self._server_tmp = tempfile.NamedTemporaryFile() 47 | self.path = self._server_tmp.name 48 | self._server = subprocess.Popen([ 49 | 'plasma_store', 50 | '-m', str(int(1.05 * self.array.nbytes)), 51 | '-s', self.path, 52 | ]) 53 | 54 | @property 55 | def client(self): 56 | if self._client is None: 57 | assert self.path is not None 58 | self._client = self.plasma.connect(self.path) 59 | return self._client 60 | 61 | def __getstate__(self): 62 | if self.plasma is None: 63 | return self.__dict__ 64 | if self.object_id is None: 65 | self.start_server() 66 | self.object_id = self.client.put(self.array) 67 | state = self.__dict__.copy() 68 | del state['array'] 69 | state['_client'] = None 70 | state['_server'] = None 71 | state['_server_tmp'] = None 72 | state['_plasma'] = None 73 | return state 74 | 75 | def __setstate__(self, state): 76 | self.__dict__.update(state) 77 | if self.plasma is None: 78 | return 79 | self.array = self.client.get(self.object_id) 80 | 81 | def __del__(self): 82 | if self._server is not None: 83 | self._server.kill() 84 | self._server = None 85 | self._server_tmp.close() 86 | self._server_tmp = None 87 | -------------------------------------------------------------------------------- /fairseq/optim/lr_scheduler/triangular_lr_scheduler.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 | from . import FairseqLRScheduler, register_lr_scheduler 9 | 10 | 11 | @register_lr_scheduler('triangular') 12 | class TriangularSchedule(FairseqLRScheduler): 13 | """Assign LR based on a triangular cyclical schedule. 14 | 15 | See https://arxiv.org/pdf/1506.01186.pdf for details. 16 | """ 17 | 18 | def __init__(self, args, optimizer): 19 | super().__init__(args, optimizer) 20 | if len(args.lr) > 1: 21 | raise ValueError( 22 | 'Cannot use a fixed learning rate schedule with triangular.' 23 | ' Consider --lr-scheduler=fixed instead.' 24 | ) 25 | 26 | lr = args.lr[0] 27 | 28 | assert args.max_lr > lr, 'max_lr must be more than lr' 29 | self.min_lr = lr 30 | self.max_lr = args.max_lr 31 | self.stepsize = args.lr_period_updates // 2 32 | self.lr_shrink = args.lr_shrink 33 | self.shrink_min = args.shrink_min 34 | 35 | # initial learning rate 36 | self.lr = self.min_lr 37 | self.optimizer.set_lr(self.lr) 38 | 39 | @staticmethod 40 | def add_args(parser): 41 | """Add arguments to the parser for this LR scheduler.""" 42 | # fmt: off 43 | parser.add_argument('--max-lr', required=True, type=float, metavar='LR', 44 | help='max learning rate, must be more than args.lr') 45 | parser.add_argument('--lr-period-updates', default=5000, type=float, metavar='LR', 46 | help='initial number of updates per period (cycle length)') 47 | parser.add_argument('--lr-shrink', default=0.1, type=float, metavar='LS', 48 | help='shrink factor for annealing') 49 | parser.add_argument('--shrink-min', action='store_true', 50 | help='if set, also shrinks min lr') 51 | # fmt: on 52 | 53 | def step(self, epoch, val_loss=None): 54 | """Update the learning rate at the end of the given epoch.""" 55 | super().step(epoch, val_loss) 56 | # we don't change the learning rate at epoch boundaries 57 | return self.optimizer.get_lr() 58 | 59 | def step_update(self, num_updates): 60 | """Update the learning rate after each update.""" 61 | cycle = math.floor(num_updates / (2 * self.stepsize)) 62 | 63 | lr_shrink = self.lr_shrink ** cycle 64 | max_lr = self.max_lr * lr_shrink 65 | if self.shrink_min: 66 | min_lr = self.min_lr * lr_shrink 67 | else: 68 | min_lr = self.min_lr 69 | 70 | x = abs(num_updates / self.stepsize - 2 * (cycle + 1) + 1) 71 | self.lr = min_lr + (max_lr - min_lr) * max(0, (1 - x)) 72 | 73 | self.optimizer.set_lr(self.lr) 74 | return self.lr 75 | -------------------------------------------------------------------------------- /fairseq/criterions/cross_entropy.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.nn.functional as F 9 | 10 | from fairseq import metrics, utils 11 | from fairseq.criterions import FairseqCriterion, register_criterion 12 | 13 | 14 | @register_criterion('cross_entropy') 15 | class CrossEntropyCriterion(FairseqCriterion): 16 | 17 | def __init__(self, task, sentence_avg): 18 | super().__init__(task) 19 | self.sentence_avg = sentence_avg 20 | 21 | def forward(self, model, sample, reduce=True): 22 | """Compute the loss for the given sample. 23 | 24 | Returns a tuple with three elements: 25 | 1) the loss 26 | 2) the sample size, which is used as the denominator for the gradient 27 | 3) logging outputs to display while training 28 | """ 29 | net_output = model(**sample['net_input']) 30 | loss, _ = self.compute_loss(model, net_output, sample, reduce=reduce) 31 | sample_size = sample['target'].size(0) if self.sentence_avg else sample['ntokens'] 32 | logging_output = { 33 | 'loss': loss.data, 34 | 'ntokens': sample['ntokens'], 35 | 'nsentences': sample['target'].size(0), 36 | 'sample_size': sample_size, 37 | } 38 | return loss, sample_size, logging_output 39 | 40 | def compute_loss(self, model, net_output, sample, reduce=True): 41 | lprobs = model.get_normalized_probs(net_output, log_probs=True) 42 | lprobs = lprobs.view(-1, lprobs.size(-1)) 43 | target = model.get_targets(sample, net_output).view(-1) 44 | loss = F.nll_loss( 45 | lprobs, 46 | target, 47 | ignore_index=self.padding_idx, 48 | reduction='sum' if reduce else 'none', 49 | ) 50 | return loss, loss 51 | 52 | @staticmethod 53 | def reduce_metrics(logging_outputs) -> None: 54 | """Aggregate logging outputs from data parallel training.""" 55 | loss_sum = sum(log.get('loss', 0) for log in logging_outputs) 56 | ntokens = sum(log.get('ntokens', 0) for log in logging_outputs) 57 | sample_size = sum(log.get('sample_size', 0) for log in logging_outputs) 58 | 59 | metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) 60 | if sample_size != ntokens: 61 | metrics.log_scalar('nll_loss', loss_sum / ntokens / math.log(2), ntokens, round=3) 62 | metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['nll_loss'].avg)) 63 | else: 64 | metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['loss'].avg)) 65 | 66 | @staticmethod 67 | def logging_outputs_can_be_summed() -> bool: 68 | """ 69 | Whether the logging outputs returned by `forward` can be summed 70 | across workers prior to calling `reduce_metrics`. Setting this 71 | to True will improves distributed training speed. 72 | """ 73 | return True 74 | -------------------------------------------------------------------------------- /fairseq/data/lm_context_window_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 fairseq.data.monolingual_dataset import MonolingualDataset 10 | 11 | from . import FairseqDataset 12 | 13 | 14 | class LMContextWindowDataset(FairseqDataset): 15 | """Wraps a MonolingualDataset and provides more context for evaluation.""" 16 | 17 | def __init__(self, dataset, tokens_per_sample, context_window, pad_idx): 18 | assert isinstance(dataset, MonolingualDataset) 19 | assert context_window > 0 20 | self.dataset = dataset 21 | self.tokens_per_sample = tokens_per_sample 22 | self.context_window = context_window 23 | self.pad_idx = pad_idx 24 | self.prev_tokens = np.empty([0]) 25 | 26 | def __getitem__(self, index): 27 | return self.dataset[index] 28 | 29 | def __len__(self): 30 | return len(self.dataset) 31 | 32 | def collater(self, samples): 33 | sample = self.dataset.collater(samples) 34 | 35 | pad = self.pad_idx 36 | max_sample_len = self.tokens_per_sample + self.context_window 37 | 38 | bsz, tsz = sample['net_input']['src_tokens'].shape 39 | start_idxs = [0] * bsz 40 | toks = sample['net_input']['src_tokens'] 41 | lengths = sample['net_input']['src_lengths'] 42 | tgt = sample['target'] 43 | new_toks = np.empty([bsz, tsz + self.context_window], dtype=np.int64) 44 | new_tgt = np.full([bsz, tsz + self.context_window], pad, dtype=np.int64) 45 | sample_lens = toks.ne(pad).long().sum(dim=1).cpu() 46 | for i in range(bsz): 47 | sample_len = sample_lens[i] 48 | extra = len(self.prev_tokens) + sample_len - max_sample_len 49 | if extra > 0: 50 | self.prev_tokens = self.prev_tokens[extra:] 51 | pads = np.full(self.context_window - len(self.prev_tokens), pad) 52 | new_toks[i] = np.concatenate([self.prev_tokens, toks[i].numpy(), pads]) 53 | new_tgt[i, len(self.prev_tokens):len(self.prev_tokens) + len(tgt[i])] = tgt[i] 54 | start_idxs[i] = len(self.prev_tokens) 55 | lengths[i] += len(self.prev_tokens) 56 | self.prev_tokens = new_toks[i][new_toks[i] != pad][-self.context_window:] 57 | sample['net_input']['src_tokens'] = torch.from_numpy(new_toks) 58 | sample['target'] = torch.from_numpy(new_tgt) 59 | sample['start_indices'] = start_idxs 60 | 61 | return sample 62 | 63 | def num_tokens(self, index): 64 | return self.dataset.num_tokens(index) 65 | 66 | def size(self, index): 67 | return self.dataset.size(index) 68 | 69 | def ordered_indices(self): 70 | # NOTE we don't shuffle the data to retain access to the previous dataset elements 71 | return np.arange(len(self.dataset)) 72 | 73 | @property 74 | def supports_prefetch(self): 75 | return getattr(self.dataset, 'supports_prefetch', False) 76 | 77 | def prefetch(self, indices): 78 | return self.dataset.prefetch(indices) 79 | -------------------------------------------------------------------------------- /fairseq/optim/lr_scheduler/inverse_square_root_schedule.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 FairseqLRScheduler, register_lr_scheduler 7 | 8 | 9 | @register_lr_scheduler('inverse_sqrt') 10 | class InverseSquareRootSchedule(FairseqLRScheduler): 11 | """Decay the LR based on the inverse square root of the update number. 12 | 13 | We also support a warmup phase where we linearly increase the learning rate 14 | from some initial learning rate (``--warmup-init-lr``) until the configured 15 | learning rate (``--lr``). Thereafter we decay proportional to the number of 16 | updates, with a decay factor set to align with the configured learning rate. 17 | 18 | During warmup:: 19 | 20 | lrs = torch.linspace(args.warmup_init_lr, args.lr, args.warmup_updates) 21 | lr = lrs[update_num] 22 | 23 | After warmup:: 24 | 25 | decay_factor = args.lr * sqrt(args.warmup_updates) 26 | lr = decay_factor / sqrt(update_num) 27 | """ 28 | 29 | def __init__(self, args, optimizer): 30 | super().__init__(args, optimizer) 31 | if len(args.lr) > 1: 32 | raise ValueError( 33 | 'Cannot use a fixed learning rate schedule with inverse_sqrt.' 34 | ' Consider --lr-scheduler=fixed instead.' 35 | ) 36 | warmup_end_lr = args.lr[0] 37 | if args.warmup_init_lr < 0: 38 | args.warmup_init_lr = 0 if args.warmup_updates > 0 else warmup_end_lr 39 | 40 | # linearly warmup for the first args.warmup_updates 41 | self.lr_step = (warmup_end_lr - args.warmup_init_lr) / args.warmup_updates 42 | 43 | # then, decay prop. to the inverse square root of the update number 44 | self.decay_factor = warmup_end_lr * args.warmup_updates**0.5 45 | 46 | # initial learning rate 47 | self.lr = args.warmup_init_lr 48 | self.optimizer.set_lr(self.lr) 49 | 50 | @staticmethod 51 | def add_args(parser): 52 | """Add arguments to the parser for this LR scheduler.""" 53 | # fmt: off 54 | parser.add_argument('--warmup-updates', default=4000, type=int, metavar='N', 55 | help='warmup the learning rate linearly for the first N updates') 56 | parser.add_argument('--warmup-init-lr', default=-1, type=float, metavar='LR', 57 | help='initial learning rate during warmup phase; default is args.lr') 58 | # fmt: on 59 | 60 | def step(self, epoch, val_loss=None): 61 | """Update the learning rate at the end of the given epoch.""" 62 | super().step(epoch, val_loss) 63 | # we don't change the learning rate at epoch boundaries 64 | return self.optimizer.get_lr() 65 | 66 | def step_update(self, num_updates): 67 | """Update the learning rate after each update.""" 68 | if num_updates < self.args.warmup_updates: 69 | self.lr = self.args.warmup_init_lr + num_updates*self.lr_step 70 | else: 71 | self.lr = self.decay_factor * num_updates**-0.5 72 | self.optimizer.set_lr(self.lr) 73 | return self.lr 74 | -------------------------------------------------------------------------------- /fairseq/criterions/masked_lm.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 | import torch.nn.functional as F 10 | 11 | from fairseq import metrics, modules, utils 12 | from fairseq.criterions import FairseqCriterion, register_criterion 13 | 14 | 15 | @register_criterion('masked_lm') 16 | class MaskedLmLoss(FairseqCriterion): 17 | """ 18 | Implementation for the loss used in masked language model (MLM) training. 19 | """ 20 | 21 | def forward(self, model, sample, reduce=True): 22 | """Compute the loss for the given sample. 23 | 24 | Returns a tuple with three elements: 25 | 1) the loss 26 | 2) the sample size, which is used as the denominator for the gradient 27 | 3) logging outputs to display while training 28 | """ 29 | # compute MLM loss 30 | masked_tokens = sample['target'].ne(self.padding_idx) 31 | 32 | # Rare: when all tokens are masked, project all tokens. 33 | # We use torch.where to avoid device-to-host transfers, 34 | # except on CPU where torch.where is not well supported 35 | # (see github.com/pytorch/pytorch/issues/26247). 36 | if masked_tokens.device == torch.device('cpu'): 37 | if not masked_tokens.any(): 38 | masked_tokens.fill_(True) 39 | else: 40 | masked_tokens = torch.where( 41 | masked_tokens.any(), 42 | masked_tokens, 43 | masked_tokens.new([True]), 44 | ) 45 | 46 | logits = model(**sample['net_input'], masked_tokens=masked_tokens)[0] 47 | targets = model.get_targets(sample, [logits]) 48 | targets = targets[masked_tokens] 49 | 50 | loss = modules.cross_entropy( 51 | logits.view(-1, logits.size(-1)), 52 | targets.view(-1), 53 | reduction='sum', 54 | ignore_index=self.padding_idx, 55 | ) 56 | 57 | sample_size = masked_tokens.int().sum() 58 | logging_output = { 59 | 'loss': loss.data, 60 | 'ntokens': sample['ntokens'], 61 | 'nsentences': sample['nsentences'], 62 | 'sample_size': sample_size, 63 | } 64 | return loss, sample_size, logging_output 65 | 66 | @staticmethod 67 | def reduce_metrics(logging_outputs) -> None: 68 | """Aggregate logging outputs from data parallel training.""" 69 | loss_sum = sum(log.get('loss', 0) for log in logging_outputs) 70 | sample_size = sum(log.get('sample_size', 0) for log in logging_outputs) 71 | 72 | metrics.log_scalar('loss', loss_sum / sample_size / math.log(2), sample_size, round=3) 73 | metrics.log_derived('ppl', lambda meters: utils.get_perplexity(meters['loss'].avg)) 74 | 75 | @staticmethod 76 | def logging_outputs_can_be_summed() -> bool: 77 | """ 78 | Whether the logging outputs returned by `forward` can be summed 79 | across workers prior to calling `reduce_metrics`. Setting this 80 | to True will improves distributed training speed. 81 | """ 82 | return True 83 | -------------------------------------------------------------------------------- /fairseq/modules/transformer_sentence_encoder_layer.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 typing import Optional 6 | 7 | import torch 8 | import torch.nn as nn 9 | import torch.nn.functional as F 10 | 11 | from fairseq import utils 12 | from fairseq.modules import ( 13 | LayerNorm, 14 | MultiheadAttention, 15 | ) 16 | 17 | 18 | class TransformerSentenceEncoderLayer(nn.Module): 19 | """ 20 | Implements a Transformer Encoder Layer used in BERT/XLM style pre-trained 21 | models. 22 | """ 23 | 24 | def __init__( 25 | self, 26 | embedding_dim: int = 768, 27 | ffn_embedding_dim: int = 3072, 28 | num_attention_heads: int = 8, 29 | dropout: float = 0.1, 30 | attention_dropout: float = 0.1, 31 | activation_dropout: float = 0.1, 32 | activation_fn: str = 'relu', 33 | export: bool = False, 34 | ) -> None: 35 | 36 | super().__init__() 37 | # Initialize parameters 38 | self.embedding_dim = embedding_dim 39 | self.dropout = dropout 40 | self.activation_dropout = activation_dropout 41 | 42 | # Initialize blocks 43 | self.activation_fn = utils.get_activation_fn(activation_fn) 44 | self.self_attn = MultiheadAttention( 45 | self.embedding_dim, 46 | num_attention_heads, 47 | dropout=attention_dropout, 48 | add_bias_kv=False, 49 | add_zero_attn=False, 50 | self_attention=True 51 | ) 52 | 53 | # layer norm associated with the self attention layer 54 | self.self_attn_layer_norm = LayerNorm(self.embedding_dim, export=export) 55 | self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim) 56 | self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim) 57 | 58 | # layer norm associated with the position wise feed-forward NN 59 | self.final_layer_norm = LayerNorm(self.embedding_dim, export=export) 60 | 61 | def forward( 62 | self, 63 | x: torch.Tensor, 64 | self_attn_mask: Optional[torch.Tensor] = None, 65 | self_attn_padding_mask: Optional[torch.Tensor] = None, 66 | ): 67 | """ 68 | LayerNorm is applied either before or after the self-attention/ffn 69 | modules similar to the original Transformer implementation. 70 | """ 71 | residual = x 72 | x, attn = self.self_attn( 73 | query=x, 74 | key=x, 75 | value=x, 76 | key_padding_mask=self_attn_padding_mask, 77 | need_weights=False, 78 | attn_mask=self_attn_mask, 79 | ) 80 | x = F.dropout(x, p=self.dropout, training=self.training) 81 | x = residual + x 82 | x = self.self_attn_layer_norm(x) 83 | 84 | residual = x 85 | x = self.activation_fn(self.fc1(x)) 86 | x = F.dropout(x, p=self.activation_dropout, training=self.training) 87 | x = self.fc2(x) 88 | x = F.dropout(x, p=self.dropout, training=self.training) 89 | x = residual + x 90 | x = self.final_layer_norm(x) 91 | return x, attn 92 | --------------------------------------------------------------------------------