NaN loss for no positive pair from SupCon

[srsawant34]

The SupCon losses throw NaN values for input where we don't have any positive pair for an anchor.

For example:

from losses import SupConLoss
import torch
import torch.nn.functional as F

# define loss with a temperature `temp`
temp = 0.8
criterion = SupConLoss(temperature=temp)

# features: [bsz, n_views, f_dim]
# `n_views` is the number of crops from each image
# better be L2 normalized in f_dim dimension
features = torch.randn(4,1,3)
features = F.normalize(features, p=2, dim=1)
# labels: [bsz]
labels = torch.tensor([0,1,1,3], dtype=float)

# SupContrast
loss = criterion(features, labels)

Outputs:

tensor([    nan, 36.9528, 36.9528,     nan])