Trainer multi label #7191
Merged
Trainer multi label #7191
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
sgugger
commented
Sep 16, 2020
| Tuple[Optional[float], Optional[torch.Tensor], Optional[torch.Tensor]]: | ||
| A tuple with the loss, logits and labels (each being optional). | ||
| """ | ||
| has_labels = any(inputs.get(k) is not None for k in ["labels", "lm_labels", "masked_lm_labels"]) |
Collaborator
Author
There was a problem hiding this comment.
No need to test for old deprecated argument names since they have all been changed in the lib and the user can now set their own name if they have an old model they are still using.
Codecov Report
@@ Coverage Diff @@
## master #7191 +/- ##
==========================================
- Coverage 80.86% 79.41% -1.46%
==========================================
Files 169 169
Lines 32293 32322 +29
==========================================
- Hits 26115 25668 -447
- Misses 6178 6654 +476
Continue to review full report at Codecov.
|
LysandreJik
approved these changes
Sep 17, 2020
Member
LysandreJik
left a comment
LysandreJik
left a comment
There was a problem hiding this comment.
LGTM, very nice addition.
Comment on lines
+162
to
+171
| def nested_xla_mesh_reduce(tensors, name): | ||
| if is_torch_tpu_available(): | ||
| import torch_xla.core.xla_model as xm | ||
|
|
||
| if isinstance(tensors, (list, tuple)): | ||
| return type(tensors)(nested_xla_mesh_reduce(t, f"{name}_{i}") for i, t in enumerate(tensors)) | ||
| return xm.mesh_reduce(name, tensors, torch.cat) | ||
| else: | ||
| raise ImportError("Torch xla must be installed to use `nested_xla_mesh_reduce`") | ||
|
|
Member
There was a problem hiding this comment.
Did you get a chance to test this on TPU?
Collaborator
Author
There was a problem hiding this comment.
No, was planning to ask you about it this morning.
Zigur
pushed a commit
to Zigur/transformers
that referenced
this pull request
Oct 26, 2020
* Trainer accep multiple labels * Missing import * Fix dosctrings
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is a follow-up from #7126. The same kinds of models that can output multiple predictions expect multiple labels (not named "labels") so the evaluation code needs to be changed for this. To support models built by users, I added a
label_namesfield in theTrainingArgumentswhich contain the label names. It then defaults to["labels"]for most models,["start_positions", "end_positions"]for question answering models if the user does not set it to work seamlessly for all Transformers models.I ended up writing a few util functions that concat/numpify for tensors or nested lists/tuples of tensors to avoid testing everywhere in
Trainer, I think the design is cleaner this way and it also supports model with crazy outputs (if we setoutput_attentions=Truefor instance). I also added a test for the multiple labels predictions.