Request to avoid loss to NaN

[u-shiori]

Please change the minimum value of the square root to the epsilon calculator, as shown below.

↓ Before change (current)

arcface-pytorch/models/metrics.py

Line 38 in 47ace80

sine = torch.sqrt((1.0 - torch.pow(cosine, 2)).clamp(0, 1))

↓ After change

sine = torch.sqrt((1.0 - torch.pow(cosine, 2)).clamp(1e-6, 1))

The differential function of √x, 1/(2√x), becomes inf near x=0, i.e., an exploding gradient.

This results in a loss of nan as reported in #81.

So please fix it so that sqrt does not take 0.