Fixes issue #543

[vfdev-5]

PR fixes #543 and probably #554

Description:
Previous CM implementation suffered from the problem if target contains non-contiguous indices.
New implementation is almost taken from torchvision's https://github.com/pytorch/vision/blob/master/references/segmentation/utils.py#L75-L117

In addition, this implementation is more robust when trained in fp16. Previous implementation used torch.matmul which is patched by nvidia/apex confusion matrix goes easily to inf/nan.

!!! This commit also removes the case of targets as (batchsize, num_categories, ...) where num_categories excludes background class.
Confusion matrix computation is possible almost similarly for (batchsize, ...), but when target is all zero (0, ..., 0) = no classes (background class), then confusion matrix does not count any true/false predictions.

• PR breaks BC

Check list:

• New tests are added (if a new feature is added)
• New doc strings: description and/or example code are in RST format
• Documentation is updated (if required)

---

Target indices ouside of [0, num_classes] are ignored as a mask (y >= 0) & (y < num_classes) is applied. This can be generalized with a custom method isin method :

import torch
a = torch.randint(0, 255, size=(10000, ))
b = list(range(5, 25))

def isin(x, y):
    max_value = x.max() + 1
    x_mask = torch.zeros(max_value, dtype=torch.uint8)
    x_mask[x] = 1
    y_mask = torch.zeros(max_value, dtype=torch.uint8)
    y_mask[y] = 1
    mask = x_mask & y_mask
    return mask[x]

mask = isin(a, b)
a[mask]

[vfdev-5]

cc @TheCodez

[TheCodez]

LGTM. I stopped working on my implementation because I couldn't get the tests to pass, so thanks for implementing 👍