MASS objective + transform #16
Conversation
|
smoketested on unpc. requires a thorough review. |
onmt/tests/test_transform.py
Outdated
| @@ -22,22 +22,22 @@ def test_transform_register(self): | |||
| "sentencepiece", | |||
| "bpe", | |||
| "onmt_tokenize", | |||
| "bart", | |||
| "ae_noise", | |||
There was a problem hiding this comment.
The transform is named ae_noise but the file is named denoising. Would it be better to use only one of the names in both locations?
| @@ -461,7 +461,7 @@ def _gradient_accumulation_over_lang_pairs( | |||
| seen_comm_batches.add(comm_batch) | |||
| if self.norm_method == "tokens": | |||
| num_tokens = ( | |||
| batch.tgt[1:, :, 0].ne(self.train_loss_md[f'trainloss{metadata.tgt_lang}'].padding_idx).sum() | |||
| batch.labels[1:, :, 0].ne(self.train_loss_md[f'trainloss{metadata.tgt_lang}'].padding_idx).sum() | |||
| ) | |||
| normalization += num_tokens.item() | |||
| else: | |||
There was a problem hiding this comment.
Later on line 484 tgt is used: tgt_outer = batch.tgt.
If I read this correctly, the idea is that tgt_outer is used only as the input sequence to the decoder. The whole batch is then given to the loss function self.train_loss_md which will take care of using the label sequence instead.
This is not easy to figure out, because the loss function is wrapped in several layers of onmt scaffolding. A comment would be nice.
There was a problem hiding this comment.
Yes.
Full disclosure: I did minimal edits. I don't know why the code doesn't use batch.tgt directly line 490. I try not to fix what's not broken.
| @@ -328,6 +330,7 @@ def _make_shard_state(self, batch, output, range_, attns=None): | |||
| shard_state = { | |||
| "output": output, | |||
| "target": batch.tgt[range_start:range_end, :, 0], | |||
| "labels": batch.labels[range_start:range_end, :, 0], | |||
There was a problem hiding this comment.
It may not actually be necessary to have separate "target" and "labels" here. As this is the loss, it is only concerned with comparing the output to the desired labels. The decoder input ("target") shouldn't be needed, and as far as I can tell is not used.
This way is clearer, though. Also, slicing a tensor and then not touching the slice should not incur much of a cost, as it will be just a view, so removing the "target" might not make any difference.
There was a problem hiding this comment.
I'm strongly in favor of the currently implemented approach here (I'd rather anything that exits the batchers to have roughly the same attributes)
There was a problem hiding this comment.
Yes, consistent naming and structure is desirable, and is a good enough motivation for keeping target here.
On the topic of consistent naming: I noticed that there are some places (at least in the _stats method) where the name target is used, but it now actually refers to labels. _stats takes gtruth, which was changed in this PR.
| def __init__(self, opts): | ||
| super().__init__(opts) | ||
| self.denoising_objective = opts.denoising_objective |
There was a problem hiding this comment.
We should validate that mask-random is not used with the MASS objective.
In BART, it is possible to occasionally substitute a random token instead of the mask token. This slightly alleviates the tendency to copy all unmasked tokens verbatim.
In MASS, substituting a random token on the source side would also affect the target sequence: the random token would be complemented into a mask token, and would not contribute to the loss. This does not make sense.
There was a problem hiding this comment.
so random_ratio > 0
closes #15
Essentially, adds a
labelskey to batches to distinguish between decoder-side inputs (tgtandtargets') and decoder-side labels (currently only in use for MASS, but can be helpful for other decoder-side denoising schemes).labelsare initialized as a copy oftargetunless otherwise specified by a transform.With this change, the MASS noising transform devolves into a slightly more complex formatting in a BART noising transform.