4326582bb1e68480ef79a02abbf4bfacc3aadede,layers/losses.py,MSELossMasked,forward,#MSELossMasked#Any#Any#Any#,39

Before Change

            loss: An average loss value masked by the length.
        
        input = input.contiguous()
        target = target.contiguous()

        // logits_flat: (batch * max_len, dim)
        input = input.view(-1, input.shape[-1])
        // target_flat: (batch * max_len, dim)
        target_flat = target.view(-1, target.shape[-1])
        // losses_flat: (batch * max_len, dim)
        losses_flat = functional.mse_loss(
            input, target_flat, size_average=False, reduce=False)
        // losses: (batch, max_len, dim)
        losses = losses_flat.view(*target.size())

        // mask: (batch, max_len, 1)

After Change

            loss: An average loss value masked by the length.
        
        // mask: (batch, max_len, 1)
        mask = sequence_mask(
            sequence_length=length, max_len=target.size(1)).unsqueeze(2).float()
        mask = mask.expand_as(input)
        loss = functional.mse_loss(
            input * mask, target * mask, reduction="sum")
        loss = loss / mask.sum()

Instances