Apply mask softmax

13

I had to implement something similar. My approach was the following (where mask is a tensor of 1s and 0s indicating the entries to be removed):

def masked_softmax(vec, mask, dim=1):
    masked_vec = vec * mask.float()
    max_vec = torch.max(masked_vec, dim=dim, keepdim=True)[0]
    exps = torch.exp(masked_vec-max_vec)
    masked_exps = exps * mask.float()
    masked_sums = masked_exps.sum(dim, keepdim=True)
    zeros=(masked_sums == 0)
    masked_sums += zeros.float()
    return masked_exps/masked_sums
4 Likes