Some metrics fail on GPU

[Johanmkr]

TL;DR Can everyone test their metric on some GPU, most are failing on mine.

Issue

Hi all, I discovered this in my own code but it could affect more people:

In my metric class, Precision, I declare new tensors like this:

true_oh = torch.zeros(y_true.size(0), self.num_classes).scatter_(
1, y_true.unsqueeze(1), 1
)
pred_oh = torch.zeros(y_pred.size(0), self.num_classes).scatter_(
1, y_pred.unsqueeze(1), 1
)

However, this will cause the code to fail on the GPU since these are allocated to CPU memory.

I get the same issue with recall @salomaestro.

For Entropy I get a different GPU error @Seilmast:
TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

And for f1 @sot176 I receive the following error:
/CollaborativeCoding/metrics/F1.py", line 161, in returnmetric
self.y_pred = torch.cat(self.y_pred)
^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: zero-dimensional tensor (at position 0) cannot be concatenated

Accuracy seems to work well on both cpu and gpu, well done @hzavadil98.

Suggestion

Either:

1. We may pass the device variable to the metric classes
2. Avoid allocating memory to new tensors within the metric classes.

Any thoughts?

[Johanmkr]

This is resolved by PR #103, sorry for calling the alarm.