Accuracy metric for preds at half precision is zero with pl=1.0.8

[luzuku]

🐛 Bug

The accuracy metric is wrong if preds are given with half precision. See example.

To Reproduce

import torch
from pytorch_lightning.metrics import Accuracy

acc = Accuracy(threshold=0.5)
target = torch.Tensor([1, 1, 0, 0])
preds = torch.Tensor([0.7, 0.4, 0.8, 0.4])

print(acc(preds, target))  -> 0.5
print(acc(preds.half(), target))  -> 0.0

Expected behavior

The accuracy metric should not fail silently. Either an Error needs to be raised when preds are half precision or it should work correctly.

Environment

• PyTorch Version (e.g., 1.0): 1.7.0
• OS (e.g., Linux): Linux
• How you installed PyTorch (conda, pip, source): conda
• Build command you used (if compiling from source):
• Python version: 3.8
• CUDA/cuDNN version: 10.2
• GPU models and configuration: ...
• Any other relevant information:

Additional context

This might already be fixed in master. I filed the issue regardless because I don't have time to check.

[github-actions]

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

[luzuku]

It is indeed fixed in master with the change from preds.dtype == torch.float to preds.is_floating_point(). It is also a duplicate of #4840. Sorry, my bad.

[tadejsv]

It is fixed in master but would be broken again in #4838, so thanks for catching this :)

[tadejsv]

@luzuku And one small note: if you create tensors with torch.Tensor, the created tensor will be a float tensor. Targets as floats will not be supported as inputs in classification functions anymore. If you want to preserve dtypes, create the tensor with torch.tensor (note that tensor is not capitalized)