Shard eval dataset and aggregate eval metrics #10
Conversation
This file contains 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
Also, instead of calling `eval_loss.item()` every time do summation with tensors on device.
taylanbil
reviewed
Apr 1, 2020
examples/run_glue_tpu.py
Outdated
| @@ -256,14 +257,29 @@ def evaluate(args, model, tokenizer, prefix="", disable_logging=False): | |||
| preds = np.squeeze(preds) | |||
| result = compute_metrics(eval_task, preds, out_label_ids) | |||
| results.update(result) | |||
| results['eval_loss'] = eval_loss.item() | |||
|
|
|||
| # Average all metrics from each shard | |||
There was a problem hiding this comment.
what are some of the metrics? does it make sense to avg all of them? some metrics may be additive.
There was a problem hiding this comment.
f1, accuracy, eval_loss, acc_and_f1 (avg of the two)
I checked they all make sense averaged.
There was a problem hiding this comment.
Yep ignore that :D Not additive as discussed. Will update PR.
There was a problem hiding this comment.
I'd be surprised if global f1 == np.mean(local f1s). That's probably not true, let's verify on paper what the true formula to get global f1 is.
There was a problem hiding this comment.
Updated to sync the pred/label tensors directly instead (these shouldn't be that big for finetuning tasks; single integer 0/1 per example). This way we don't have to have some custom aggregation per metric and don't touch upstream core code.
| logger.info(" %s = %s", key, str(result[key])) | ||
| writer.write("%s = %s\n" % (key, str(result[key]))) | ||
| tb_writer.add_scalar(key, result[key]) | ||
| if xm.is_master_ordinal(): |
There was a problem hiding this comment.
is everything being logged here already on cpu?
There was a problem hiding this comment.
maybe let's add a comment? It's a subtle point that can be missed by code readers.
jysohn23
commented
Apr 1, 2020
examples/run_glue_tpu.py
Outdated
| @@ -256,14 +257,29 @@ def evaluate(args, model, tokenizer, prefix="", disable_logging=False): | |||
| preds = np.squeeze(preds) | |||
| result = compute_metrics(eval_task, preds, out_label_ids) | |||
| results.update(result) | |||
| results['eval_loss'] = eval_loss.item() | |||
|
|
|||
| # Average all metrics from each shard | |||
There was a problem hiding this comment.
f1, accuracy, eval_loss, acc_and_f1 (avg of the two)
I checked they all make sense averaged.
| logger.info(" %s = %s", key, str(result[key])) | ||
| writer.write("%s = %s\n" % (key, str(result[key]))) | ||
| tb_writer.add_scalar(key, result[key]) | ||
| if xm.is_master_ordinal(): |
As brought up during review some metrics like f1 cannot be aggregated via averaging. GLUE task metrics depends largely on the dataset, so instead we sync the prediction and label tensors so that the metrics can be computed accurately on those instead.
taylanbil
approved these changes
Apr 2, 2020
jysohn23
added a commit
to jysohn23/transformers
that referenced
this pull request
Apr 10, 2020
* Initial commit to get BERT + run_glue.py on TPU * Add README section for TPU and address comments. * Cleanup TPU bits from run_glue.py (pytorch-tpu#3) TPU runner is currently implemented in: https://github.com/pytorch-tpu/transformers/blob/tpu/examples/run_glue_tpu.py. We plan to upstream this directly into `huggingface/transformers` (either `master` or `tpu`) branch once it's been more thoroughly tested. * Cleanup TPU bits from run_glue.py TPU runner is currently implemented in: https://github.com/pytorch-tpu/transformers/blob/tpu/examples/run_glue_tpu.py. We plan to upstream this directly into `huggingface/transformers` (either `master` or `tpu`) branch once it's been more thoroughly tested. * No need to call `xm.mark_step()` explicitly (pytorch-tpu#4) Since for gradient accumulation we're accumulating on batches from `ParallelLoader` instance which on next() marks the step itself. * Resolve R/W conflicts from multiprocessing (pytorch-tpu#5) * Add XLNet in list of models for `run_glue_tpu.py` (pytorch-tpu#6) * Add RoBERTa to list of models in TPU GLUE (pytorch-tpu#7) * Add RoBERTa and DistilBert to list of models in TPU GLUE (pytorch-tpu#8) * Use barriers to reduce duplicate work/resources (pytorch-tpu#9) * Shard eval dataset and aggregate eval metrics (pytorch-tpu#10) * Shard eval dataset and aggregate eval metrics Also, instead of calling `eval_loss.item()` every time do summation with tensors on device. * Change defaultdict to float * Reduce the pred, label tensors instead of metrics As brought up during review some metrics like f1 cannot be aggregated via averaging. GLUE task metrics depends largely on the dataset, so instead we sync the prediction and label tensors so that the metrics can be computed accurately on those instead. * Only use tb_writer from master (pytorch-tpu#11) * Apply huggingface black code formatting * Style * Remove `--do_lower_case` as example uses cased * Add option to specify tensorboard logdir This is needed for our testing framework which checks regressions against key metrics writtern by the summary writer. * Using configuration for `xla_device` * Prefix TPU specific comments. * num_cores clarification and namespace eval metrics * Cache features file under `args.cache_dir` Instead of under `args.data_dir`. This is needed as our test infra uses data_dir with a read-only filesystem. * Rename `run_glue_tpu` to `run_tpu_glue` Co-authored-by: LysandreJik <lysandre.debut@reseau.eseo.fr>
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.
Also, instead of calling
eval_loss.item()every time do summation withtensors on device.