Summary: 33 instances, 32 unique

Text	Count
# TODO: add file channel	1
# TODO: clean this up once fully migrated to new data handler design	1
# TODO: remove this try block after the new release by fairseq that	1
# ARBABU TODO : convert this to a enum	1
# TODO: This is a _HORRIBLE_ patch related to #208	2
TODO: This is too slow, improve the performance	1
# TODO: segment_embeddings?	1
# TODO these two configs are only kept only to be backward comptible with	1
# TODO: this is only to satisfy the TorchScript compiler.	1
# TODO: add back after migration	1
TODO: Even though very similar, 'FloatListTensorizer' currently does not support this vanilla case for tensorization of List[float].	1
# TODO merge this step into add_batch_stats once all data	1
# TODO (@mttk): Populate classs with default values of special symbols	1
# TODO : Positional embeddings needs to be tested in AR mode	1
# TODO: temporary fix fairseq dependency, remove after fairseq new release.	1
# TODO (geoffreygoh): Using config for such a purpose is really a hack,	1
# TODO: deprecate this	1
# TODO : Verify incremental generation for AR mode	1
# TODO change print_metrics function to __str__ T33522209	1
# TODO: @stevenliu use PathManagerFactory after it's released to PyPI	1
# TODO this function can be merged with batch_context once data migration is done	1
# TODO this method can be removed by moving Channel construction to Task	1
BiLSTMSlotAttention.Config,  # TODO: make default when sorting solved	1
# TODO: refactor metric reporting and remove this hack	1
# TODO: T53664090 @stevenliu save & load state_dict() of optimizer and scheduler	1
# TODO: remove GPU0 report	1
# TODO: Remove assumption that only 1 +ve passage is sample per question.	1
# TODO: T57433776 remove once FairSeq support PathManager	1
# TODO this part should be handled more elegantly	1
# TODO: support remaining features (see OutputRecord)	1
# TODO save prev_pcontext for causal attention	1
TODO: Current implementation has a single layer conv-maxpool operation.	1