Anyone encounter getting nan cxloss?

[gitlabspy]

contextualLoss/CX/CSFlow.py

Line 12 in 863af53

def __calculate_CS(self, scaled_distances, axis_for_normalization = TensorAxis.C):

This op is very unstable, the value goes inf after exp, any solution?

[FeiiYin]

Same on me.

[FeiiYin]

Try this [issue].

[FeiiYin]

Add clamp() in compute_relative_distance() to prevent the exp index too large, which works for me.

def compute_relative_distance(dist_raw):
    dist_min, _ = torch.min(dist_raw, dim=2, keepdim=True)
    dist_tilde = dist_raw / (dist_min + 1e-5)
    dist_tilde = torch.clamp(dist_tilde, max=10., min=-10)
    return dist_tilde