NameError: name 'T' is not defined

[Modius22]

🐛 Bug

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
Input In [6], in <cell line: 1>()
----> 1 datamodule = AudioClassificationData.from_folders(
      2     train_folder="/workspaces/audio-ml/data/i/urban8k_images/train",
      3     val_folder="/workspaces/audio-ml/data/i/urban8k_images/val",
      4     transform_kwargs=dict(spectrogram_size=(64, 64)),
      5     batch_size=4,
      6 )

File /usr/local/lib/python3.10/site-packages/flash/audio/classification/data.py:286, in AudioClassificationData.from_folders(cls, train_folder, val_folder, test_folder, predict_folder, train_transform, val_transform, test_transform, predict_transform, target_formatter, input_cls, transform_kwargs, **data_module_kwargs)
    181 """Load the :class:`~flash.audio.classification.data.AudioClassificationData` from folders containing
    182 spectrogram images.
    183 
   (...)
    277     >>> shutil.rmtree("predict_folder")
    278 """
    280 ds_kw = dict(
    281     target_formatter=target_formatter,
    282     transform_kwargs=transform_kwargs,
    283     input_transforms_registry=cls.input_transforms_registry,
    284 )
--> 286 train_input = input_cls(RunningStage.TRAINING, train_folder, transform=train_transform, **ds_kw)
    287 ds_kw["target_formatter"] = getattr(train_input, "target_formatter", None)
    289 return cls(
    290     train_input,
    291     input_cls(RunningStage.VALIDATING, val_folder, transform=val_transform, **ds_kw),
   (...)
    294     **data_module_kwargs,
    295 )

File /usr/local/lib/python3.10/site-packages/flash/core/data/io/input.py:134, in _wrap_init.<locals>.wrapper(self, *args, **kwargs)
    132 @functools.wraps(fn)
    133 def wrapper(self, *args, **kwargs):
--> 134     fn(self, *args, **kwargs)
    135     _validate_input(self)

File /usr/local/lib/python3.10/site-packages/flash/core/data/io/input.py:182, in InputBase.__init__(self, running_stage, transform, transform_kwargs, input_transforms_registry, *args, **kwargs)
    171 def __init__(
    172     self,
    173     running_stage: RunningStage,
   (...)
    178     **kwargs: Any,
    179 ) -> None:
    180     from flash.core.data.io.input_transform import create_transform
--> 182     self.transform = create_transform(
    183         transform,
    184         running_stage,
    185         input_transforms_registry or self.input_transforms_registry,
    186         transform_kwargs,
    187     )
    188     super().__init__(running_stage=running_stage)
    190     self.data = None

File /usr/local/lib/python3.10/site-packages/flash/core/data/io/input_transform.py:1052, in create_transform(transform, running_stage, input_transforms_registry, transform_kwargs)
   1049     return transform
   1051 if inspect.isclass(transform) and issubclass(transform, InputTransform):
-> 1052     return transform(running_stage=running_stage, **transform_kwargs)
   1054 if isinstance(transform, partial):
   1055     return transform(running_stage=running_stage, **transform_kwargs)

File <string>:7, in __init__(self, running_stage, spectrogram_size, time_mask_param, freq_mask_param)

File /usr/local/lib/python3.10/site-packages/flash/core/data/io/input_transform.py:93, in InputTransform.__post_init__(self)
     91 self._collate_in_worker_from_transform: Optional[bool] = None
     92 self._transform = None
---> 93 self._transform = self._check_transforms(self._resolve_transforms(self.running_stage), self.running_stage)
     95 # Hack
     96 Properties.__init__(self, running_stage=self.running_stage)

File /usr/local/lib/python3.10/site-packages/flash/core/data/io/input_transform.py:935, in InputTransform._resolve_transforms(self, running_stage)
    933 # get associated transform
    934 try:
--> 935     fn = getattr(self, method_name)()
    936 except AttributeError as e:
    937     raise AttributeError(str(e) + ". Hint: Call super().__init__(...) after setting all attributes.")

File /usr/local/lib/python3.10/site-packages/flash/audio/classification/input_transform.py:44, in AudioClassificationInputTransform.train_input_per_sample_transform(self)
     41 if self.freq_mask_param is not None:
     42     transforms.append(TAudio.FrequencyMasking(freq_mask_param=self.freq_mask_param))
---> 44 transforms += [T.ToTensor(), T.Resize(self.spectrogram_size)]
     45 return T.Compose(transforms)

NameError: name 'T' is not defined

To Reproduce

https://lightning-flash.readthedocs.io/en/latest/reference/audio_classification.html Step 2 with the datamodul dosn't work

Code sample

python notebook: https://github.com/Modius22/audio-ml/blob/main/src/flash-prototyping.ipynb

For reproducing: clone github repo; open the project with vscode and devcontainer extension; in the devcontainer use the ./start_jupyter.sh to start the jupyter server

Environment

python:3.10 docker images as devcontainer in vs code

general project: https://github.com/Modius22/audio-ml python notebook: https://github.com/Modius22/audio-ml/blob/main/src/flash-prototyping.ipynb

[krshrimali]

Hi, @Modius22 - Thanks for raising the issue. Can you check if you have torchvision installed? Try installing using: pip install torchvision, and try running the script again?

From the user experience perspective, we should also raise an error if the library is not installed. I'll take a note of it. Thanks!

[Modius22]

Hi @krshrimali,

Thx for the hint. That resolved my problem =)

[krshrimali]

Awesome, thanks for letting me know. 😄

[jayqi]

Hi @krshrimali, ran into this problem as well and was confused until finding this issue. Am I correct in understanding that #1318 fixes this, in terms of providing a more clear error?

[krshrimali]

Hi, @jayqi - Thank you for the question. In terms of providing a more clear error, this hasn't been fixed. I'll re-open this issue, and I think a refactor will be good in Flash to add better errors if the module is not found. In case you are interested to pick this up, please feel free (also let me know so that I can assign you the issue). Else, I'll try to scope this for the next week, and create an issue and a PR around this. 🎉

Sorry that you had to face this issue!

cc: @ethanwharris - What do you think about the refactor to raise errors instead of running into: `NameError "T" not defined``.

[ethanwharris]

Hey @krshrimali Yeah it looks like we're just missing some coverage with the requires decorator here: https://github.com/Lightning-AI/lightning-flash/blob/master/flash/audio/classification/input_transform.py

Probably we also need to add torchvision to the audio requirements since the AudioClassificationData won't work without it