Generalized DICE Metric

[vikashg]

Describe the bug
When I try to compute GeneralizedDiceMetric. I get the following error. I must point out that if I replace the 'GeneralizedDiceMetricwithDiceMetric` it works fine.

  File "train_4.py", line 120, in main
    dice_metric(y_pred = val_outputs.to(dtype=torch.int32), y = val_labels.to(dtype=torch.int32))
  File "/opt/conda/lib/python3.8/site-packages/monai/metrics/metric.py", line 344, in __call__
    ret = super().__call__(y_pred=y_pred, y=y, **kwargs)
  File "/opt/conda/lib/python3.8/site-packages/monai/metrics/metric.py", line 77, in __call__
    return self._compute_tensor(y_pred.detach(), y_, **kwargs)
  File "/opt/conda/lib/python3.8/site-packages/monai/metrics/generalized_dice.py", line 77, in _compute_tensor
    return compute_generalized_dice(
  File "/opt/conda/lib/python3.8/site-packages/monai/metrics/generalized_dice.py", line 173, in compute_generalized_dice
    (y_pred_o == 0)[denom_zeros],
  File "/opt/conda/lib/python3.8/site-packages/monai/data/meta_tensor.py", line 297, in __torch_function__
    ret = MetaTensor.update_meta(ret, func, args, kwargs)
  File "/opt/conda/lib/python3.8/site-packages/monai/data/meta_tensor.py", line 237, in update_meta
    ret_meta = decollate_batch(args[0], detach=False)[batch_idx]
  File "/opt/conda/lib/python3.8/site-packages/monai/data/meta_tensor.py", line 276, in __torch_function__
    ret = super().__torch_function__(func, types, args, kwargs)
  File "/opt/conda/lib/python3.8/site-packages/torch/_tensor.py", line 1295, in __torch_function__
    ret = func(*args, **kwargs)
TypeError: only integer tensors of a single element can be converted to an index

To Reproduce
I am sharing a relevant code snippet

post_trans = Compose([Activations(sigmoid=True), AsDiscrete(threshold=0.5)])
dice_metric = GeneralizedDiceScore(include_background=False, reduction="mean_batch")
for val_data in val_loader:
	val_image = val_data["image"].to(device)
	val_labels = val_data["mask"].to(device)
	_val_outputs = model(val_image)
	val_outputs = post_trans(_val_outputs)
	dice_metric(y_pred = val_outputs, y = val_labels)

The shape of val_outputs is [4, 1, 256, 256, 24] and is a binary tensor.

Expected behavior
I expect the behavior to be the same DiceMetric

[KumoLiu]

Hi @vikashg, I can't reproduce the error, could you please share more information?

from monai.metrics import GeneralizedDiceScore
import torch

val_outputs = torch.zeros(4, 1, 9, 9, 9)
val_outputs[:, :, 3:4, ...] = 1
val_labels = torch.zeros(4, 1, 9, 9, 9)
val_labels[:, :, 3:5, ...] = 1
dice_metric = GeneralizedDiceScore(include_background=False, reduction="mean_batch")
dice_metric(y_pred = val_outputs, y = val_labels)

Thanks!

[vikashg]

Hi @KumoLiu,
Thanks for looking into it Here is a gist for the training
https://gist.github.com/vikashg/0f2131aec1b18eef20073a8ec59df894

Here is the sample image
https://drive.google.com/drive/folders/1l0kicu68z6oYw-fCbgYbVD5Z4Pz5Rkoj?usp=sharing

[KumoLiu]

Hi @vikashg, I tried with your demo but no error report. May I know your MONAI version? It may caused by your environment.

[vikashg]

hmm weird

I still get the error. My version is 1.2.0

[KumoLiu]

Hi @vikashg, looks like a bug fixed by #6781.
Could you please try the dev branch or weekly version? Thanks!

[vikashg]

Yes it works on the dev branch. Thanks for the tip.