torchmetrics.image.fid FrechetInceptionDistance normalize argument not working

[Jelle-Plomp]

🐛 Bug

torchmetrics.image.fid FrechetInceptionDistance normalize argument is not working. The description states that if normalize=True, then the input images are expected to be of float type. Yet if I input float I still get the ValueError: Expecting image as torch.Tensor with dtype=torch.uint8

To Reproduce

The following is adapted from the example provided with the docs, now using random images of dtype float. [(https://torchmetrics.readthedocs.io/en/latest/image/frechet_inception_distance.html)]

Steps to reproduce the behavior...

Code sample

import torch
_ = torch.manual_seed(123)
from torchmetrics.image.fid import FrechetInceptionDistance
fid = FrechetInceptionDistance(feature=64, normalize=True)
imgs_dist1 = torch.rand((100, 3, 299, 299), dtype=torch.float)
imgs_dist2 = torch.rand((100, 3, 299, 299), dtype=torch.float)
fid.update(imgs_dist1, real=True)
fid.update(imgs_dist2, real=False)
fid.compute()

### Traceback
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Input In [9], in <cell line: 10>()
      8 print(imgs_dist1.size())
      9 imgs_dist2 = torch.rand((100, 3, 299, 299), dtype=torch.float)
---> 10 fid.update(imgs_dist1, real=True)
     11 fid.update(imgs_dist2, real=False)
     12 fid.compute()

File ~\anaconda3\lib\site-packages\torchmetrics\metric.py:391, in Metric._wrap_update.<locals>.wrapped_func(*args, **kwargs)
    389 with torch.set_grad_enabled(self._enable_grad):
    390     try:
--> 391         update(*args, **kwargs)
    392     except RuntimeError as err:
    393         if "Expected all tensors to be on" in str(err):

File ~\anaconda3\lib\site-packages\torchmetrics\image\fid.py:253, in FrechetInceptionDistance.update(self, imgs, real)
    246 def update(self, imgs: Tensor, real: bool) -> None:  # type: ignore
    247     """Update the state with extracted features.
    248 
    249     Args:
    250         imgs: tensor with images feed to the feature extractor
    251         real: bool indicating if ``imgs`` belong to the real or the fake distribution
    252     """
--> 253     features = self.inception(imgs)
    255     if real:
    256         self.real_features.append(features)

File ~\anaconda3\lib\site-packages\torch\nn\modules\module.py:1102, in Module._call_impl(self, *input, **kwargs)
   1098 # If we don't have any hooks, we want to skip the rest of the logic in
   1099 # this function, and just call forward.
   1100 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1101         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102     return forward_call(*input, **kwargs)
   1103 # Do not call functions when jit is used
   1104 full_backward_hooks, non_full_backward_hooks = [], []

File ~\anaconda3\lib\site-packages\torchmetrics\image\fid.py:57, in NoTrainInceptionV3.forward(self, x)
     56 def forward(self, x: Tensor) -> Tensor:
---> 57     out = super().forward(x)
     58     return out[0].reshape(x.shape[0], -1)

File ~\anaconda3\lib\site-packages\torch_fidelity\feature_extractor_inceptionv3.py:92, in FeatureExtractorInceptionV3.forward(self, x)
     91 def forward(self, x):
---> 92     vassert(torch.is_tensor(x) and x.dtype == torch.uint8, 'Expecting image as torch.Tensor with dtype=torch.uint8')
     93     features = {}
     94     remaining_features = self.features_list.copy()

File ~\anaconda3\lib\site-packages\torch_fidelity\helpers.py:9, in vassert(truecond, message)
      7 def vassert(truecond, message):
      8     if not truecond:
----> 9         raise ValueError(message)

ValueError: Expecting image as torch.Tensor with dtype=torch.uint8

Environment

• TorchMetrics version 0.9.3 (conda)
• Python 3.9.12 & PyTorch 1.10.2
• torch fidelity installed using pip install torch-fidelity

[github-actions]

Hi! thanks for your contribution!, great first issue!

[stancld]

Hi @Jelle-Plomp, thanks for raising the issue. I've run your code on the master and also on the latest 0.10.2 release. Found out I got the same error as you for the latest release. However, everything seems to work fine on master (likely thanks to #1246). You can either install torchmetrics from the source, or we are also planning to come with a new release soon.

cc: @Borda @SkafteNicki

[SkafteNicki]

I can confirm @stancld that on master it is fixed/feature has been added. @Jelle-Plomp I think the confusion is that you had navigated to the latest page for our documentation which includes changes to master that have still not being release in a official version
https://torchmetrics.readthedocs.io/en/latest/image/frechet_inception_distance.html
However, if we look at the stable page
https://torchmetrics.readthedocs.io/en/stable/image/frechet_inception_distance.html
(which is the default landing page) it correctly shows that this feature is not included in v0.9.3 of torchmetrics.

Closing issue.