PyTorch workaround for masking cross entropy loss

masked_cross_entropy.py

def _sequence_mask(sequence_length, max_len=None):
    if max_len is None:
        max_len = sequence_length.data.max()
    batch_size = sequence_length.size(0)
    seq_range = torch.range(0, max_len - 1).long()
    seq_range_expand = seq_range.unsqueeze(0).expand(batch_size, max_len)
    seq_range_expand = Variable(seq_range_expand)
    if sequence_length.is_cuda:
        seq_range_expand = seq_range_expand.cuda()
    seq_length_expand = (sequence_length.unsqueeze(1)
                         .expand_as(seq_range_expand))
    return seq_range_expand < seq_length_expand


def compute_loss(logits, target, length):
    """
    Args:
        logits: A Variable containing a FloatTensor of size
            (batch, max_len, num_classes) which contains the
            unnormalized probability for each class.
        target: A Variable containing a LongTensor of size
            (batch, max_len) which contains the index of the true
            class for each corresponding step.
        length: A Variable containing a LongTensor of size (batch,)
            which contains the length of each data in a batch.

    Returns:
        loss: An average loss value masked by the length.
    """

    # logits_flat: (batch * max_len, num_classes)
    logits_flat = logits.view(-1, logits.size(-1))
    # log_probs_flat: (batch * max_len, num_classes)
    log_probs_flat = functional.log_softmax(logits_flat)
    # target_flat: (batch * max_len, 1)
    target_flat = target.view(-1, 1)
    # losses_flat: (batch * max_len, 1)
    losses_flat = -torch.gather(log_probs_flat, dim=1, index=target_flat)
    # losses: (batch, max_len)
    losses = losses_flat.view(*target.size())
    # mask: (batch, max_len)
    mask = _sequence_mask(sequence_length=length, max_len=target.size(1))
    losses = losses * mask.float()
    loss = losses.sum() / length.float().sum()
    return loss

@viig99
Hi, it seems that functional.cross_entropy still doesn't support >2D input.
I think log_softmax + gather calls can be merged into one cross_entropy call with reduce=False, and I expect there might be
some performance gain.
I will update this gist soon.

I (genuinely) wonder how this is different from using:

weight = torch.ones(vocab_size)
weight[pad_idx] = 0.0
crit = nn.CrossEntropy(weight=weight)
crit(output, targets)

I seem to get the same numbers (assuming that you have padded every sequence with pad_idx up to the maximum sentence length in the batch.

Interesting solution, @emanjavacas!

@emanjavacas, @lzfelix
yeah its a good one, and shouldnt this be even more "correct"?

criterion = nn.CrossEntropyLoss(ignore_index=pad_idx)

i hope this was created for masking and no other purpose?

I (genuinely) wonder how this is different from using:

weight = torch.ones(vocab_size)
weight[pad_idx] = 0.0
crit = nn.CrossEntropy(weight=weight)
crit(output, targets)

I seem to get the same numbers (assuming that you have padded every sequence with pad_idx up to the maximum sentence length in the batch.

masking in the proposed gist is per-sample, i.e. telling you if each sample should have a loss or not, while your weight here is per-class.