fairseq/models/speech_to_text/modules/augmented_memory_attention.py (5 lines): - line 83: # TODO: fix positional embedding - line 100: # TODO: Consider mask here - line 142: # TODO reseach new sum_query method - line 260: # TODO: positional embedding on memory - line 263: # TODO: need to fix here fairseq/tasks/online_backtranslation.py (5 lines): - line 234: # TODO: could we do the BT using denoise sample ? - line 255: # TODO: should we shuffle ? we are already sorting batch by sizes so ? - line 494: # TODO: allow more complex mapping - line 528: # TODO: Could we translate to several language at once ? - line 641: # TODO: we should reuse the pretrained model dict which already has fairseq/data/multilingual/multilingual_data_manager.py (5 lines): - line 707: # TODO: Unifiy with alter_dataset_langtok - line 721: # TODO: Unifiy with alter_dataset_langtok - line 836: # TODO: handle modified lang toks for mined data and dae data - line 946: # TODO: to extend with extra datasets and keys and loop over different shard data paths - line 1087: # TODO: to investigate why TransformEosLangPairDataset doesn't work with ConcatDataset fairseq/tasks/multilingual_language_modeling.py (4 lines): - line 50: # TODO common var add to parent - line 125: # TODO: legacy parameter kept for compatibility - line 137: # TODO common vars below add to parent - line 427: # TODO: add an option for shrinking all size ratios to below 1 fairseq/model_parallel/models/pipeline_parallel_transformer/layers.py (3 lines): - line 163: # HACK for now, need to fix (TODO sidgoyal) - line 436: # TODO remove this once we update apex with the fix - line 499: # TODO: add back prev_self_attn_state, prev_attn_state, fairseq/models/transformer/transformer_config.py (2 lines): - line 108: # TODO should really be in the encoder config - line 114: # TODO should really be in the decoder config fairseq/iterative_refinement_generator.py (2 lines): - line 108: # TODO: iterative refinement generator does not support ensemble for now. - line 129: # TODO: better encoder inputs? fairseq/tasks/multilingual_translation.py (2 lines): - line 373: # TODO make summing of the sample sizes configurable - line 400: # TODO make summing of the sample sizes configurable fairseq/models/speech_to_text/modules/emformer.py (2 lines): - line 1814: scaled_init=True, # TODO: use constant for now. - line 1840: # TODO: make it configurable from the args fairseq/models/roberta/enc_dec.py (2 lines): - line 106: # TODO: hide setting "encoder_attn" layers behind a flag. - line 139: # TODO: this would become easier if encoder/decoder where using a similar fairseq/data/iterators.py (2 lines): - line 507: # TODO: Below is a lazy implementation which discard the final batch regardless - line 549: # TODO: [Hack] Here the grouped iterator modifies the base iterator size so that fairseq/data/round_robin_zip_datasets.py (1 line): - line 87: # TODO make it configurable whether to use max() or sum() here fairseq_cli/generate.py (1 line): - line 403: # TODO: replace this workaround with refactoring of `AudioPretraining` fairseq/optim/nag.py (1 line): - line 22: # TODO common vars in parent class fairseq/data/audio/data_cfg.py (1 line): - line 171: # TODO: move this into individual transforms fairseq/models/nat/nonautoregressive_transformer.py (1 line): - line 394: # TODO: implementing length-beam fairseq/data/audio/speech_to_speech_dataset.py (1 line): - line 204: "tgt_speaker": tgt_speakers, # TODO: unify "speaker" and "tgt_speaker" fairseq/optim/cpu_adam.py (1 line): - line 49: # TODO common vars below in parent fairseq/data/transform_eos_concat_langpair_dataset.py (1 line): - line 127: # TODO: support different padding direction on target side fairseq/models/transformer/transformer_base.py (1 line): - line 56: # -- TODO T96535332 fairseq/data/legacy/block_pair_dataset.py (1 line): - line 218: TODO: ids in skip_ids should be consecutive, we can extend it to more generic version later fairseq/data/legacy/masked_lm_dataset.py (1 line): - line 196: # TODO: Can we add deteminism without this constraint? fairseq/logging/progress_bar.py (1 line): - line 367: # TODO add hparams to Tensorboard fairseq/models/wav2vec/wav2vec2.py (1 line): - line 490: # FIXME: what happens if padding_count is specified? fairseq/tasks/semisupervised_translation.py (1 line): - line 419: # TODO make summing of the sample sizes configurable fairseq/models/speech_to_text/utils.py (1 line): - line 495: TODO: fairseq/criterions/legacy_masked_lm.py (1 line): - line 102: # TODO: Remove this after refactor of BERTModel fairseq/models/lstm.py (1 line): - line 450: # TODO make bias configurable fairseq/models/wav2vec/wav2vec2_asr.py (1 line): - line 586: # TODO: update this when transformer gets converted to dataclass configs fairseq/data/transform_eos_lang_pair_dataset.py (1 line): - line 82: # TODO: support different padding direction on target side fairseq/tasks/speech_to_speech.py (1 line): - line 57: # TODO: incorporate max_len_a and max_len_b fairseq/sequence_generator.py (1 line): - line 463: # TODO replace `nonzero(as_tuple=False)` after TorchScript supports it fairseq/utils.py (1 line): - line 223: # TODO: Very rare cases where the replacement is '' should be handled gracefully fairseq/criterions/wav2vec_criterion.py (1 line): - line 219: # FIXME: revert when gather based xla reduction is implemented fairseq/data/noising.py (1 line): - line 121: # TODO: speed up the following loop fairseq/optim/adam.py (1 line): - line 39: # TODO common vars below in parent fairseq/models/nat/insertion_transformer.py (1 line): - line 181: # TODO: decoding for InsertionTransformer fairseq/tasks/language_modeling.py (1 line): - line 96: # TODO common vars below add to parent fairseq/modules/positional_embedding.py (1 line): - line 21: # TODO: The right place for this offset would be inside