When creating a CoarseTransformerTrainer with data_max_length_seconds (rather than data_max_length), you will get the following type error:
File "audiolm-pytorch/train.py", line 48, in make_coarse_trainer
trainer = CoarseTransformerTrainer(
File "<@beartype(audiolm_pytorch.trainer.CoarseTransformerTrainer.__init__) at 0x7f056897a4c0>", line 196, in __init__
File "audiolm-pytorch/audiolm_pytorch/trainer.py", line 1090, in __init__
self.ds = SoundDataset(
File "<@beartype(audiolm_pytorch.data.SoundDataset.__init__) at 0x7f05689709d0>", line 58, in __init__
beartype.roar.BeartypeCallHintParamViolation: Method audiolm_pytorch.data.SoundDataset.__init__() parameter max_length=(160000, 240000) violates type hint typing.Optional[int], as tuple (160000, 240000) not <class "builtins.NoneType"> or int.
This is because, indeed, SoundDataset expects max_length of type either None or int. But CoarseTransformerTrainer attempts to pass it a tuple. In line trainer.py:1089 we have:
if exists(data_max_length_seconds):
data_max_length = tuple(data_max_length_seconds * hz for hz in (wav2vec.target_sample_hz, codec.target_sample_hz))
self.ds = SoundDataset(
folder,
max_length = data_max_length,
Note that this issue is only for Coarse. For convenience, I list how data_max_length is derived from data_max_length_seconds in the other trainers:
max_length: Optional[int] = None, # max length would apply to the highest target_sample_hz, if there are multiple
Therefore my guess is that CoarseTransformerTrainer should pass the max() rather than the tuple itself. But perhaps it would be cleaner to change the type to Union[int, Tuple[int, ...]] (as in target_sample_hz) and take the max() in the SoundDataset construct, if a tuple is passed.
When creating a
CoarseTransformerTrainerwithdata_max_length_seconds(rather thandata_max_length), you will get the following type error:This is because, indeed,
SoundDatasetexpectsmax_lengthof type eitherNoneorint. ButCoarseTransformerTrainerattempts to pass it a tuple. In line trainer.py:1089 we have:Note that this issue is only for Coarse. For convenience, I list how
data_max_lengthis derived fromdata_max_length_secondsin the other trainers:In the
SoundDatasetconstructor we haveTherefore my guess is that
CoarseTransformerTrainershould pass themax()rather than the tuple itself. But perhaps it would be cleaner to change the type toUnion[int, Tuple[int, ...]](as intarget_sample_hz) and take themax()in theSoundDatasetconstruct, if a tuple is passed.Thoughts?