rework metrics logic to support states #2391
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #2391 +/- ##
==========================================
- Coverage 93.75% 93.74% -0.01%
==========================================
Files 138 138
Lines 14343 14338 -5
==========================================
- Hits 13447 13441 -6
- Misses 896 897 +1 ☔ View full report in Codecov by Sentry. |
|
|
||
| def update(self, preds, target) -> None: | ||
| if preds.shape != target.shape: | ||
| raise ValueError("preds and target must have the same shape") |
There was a problem hiding this comment.
| raise ValueError("preds and target must have the same shape") | |
| raise ValueError(f"preds and target must have the same shape, but got {preds.shape} for preds and {target.shape} for target.") |
There was a problem hiding this comment.
Thanks, updated. Note that this is just a test class, this error isn't raised in the training loop itself
There was a problem hiding this comment.
Thanks for this great PR @tRosenflanz 🚀
It looks really good, just had some minor suggestions.
After those have been addressed, we can merge :)
| return loss | ||
|
|
||
| def _compute_metrics(self, metrics): | ||
| res = metrics.compute() |
There was a problem hiding this comment.
Can you move this method to where method _update_metrics() is defined?
Also, we can skip as done in _update_metrics()
| res = metrics.compute() | |
| if not len(metrics): | |
| return | |
| res = metrics.compute() |
There was a problem hiding this comment.
Moved and "skipped" when needed. For some reason linter doesn't add empty line by default btw
| if preds.shape != target.shape: | ||
| raise ValueError( | ||
| "preds and target must have the same shape " | ||
| f"but got {preds.shape} for preds and {target.shape} for target." | ||
| ) |
There was a problem hiding this comment.
Not required for this test
| if preds.shape != target.shape: | |
| raise ValueError( | |
| "preds and target must have the same shape " | |
| f"but got {preds.shape} for preds and {target.shape} for target." | |
| ) |
| def __init__(self, **kwargs): | ||
| super().__init__(**kwargs) |
There was a problem hiding this comment.
Not required for this test
| def __init__(self, **kwargs): | |
| super().__init__(**kwargs) | |
| def __init__(self): | |
| super().__init__() |
There was a problem hiding this comment.
removed. Note that this might fail in future if you start adding some args for devices or such
| **tfm_kwargs, | ||
| ) | ||
| model.fit(self.series) | ||
| assert model.model.trainer.logged_metrics["train_NumsCalled"] != 1 |
There was a problem hiding this comment.
Sorry, just saw this now. We should probably check that it is > 1 since 0 would be incorrect.
| assert model.model.trainer.logged_metrics["train_NumsCalled"] != 1 | |
| assert model.model.trainer.logged_metrics["train_NumsCalled"] > 1 |
There was a problem hiding this comment.
Good call, fixed. This tests that update_metrics works correctly as well
Checklist before merging this PR:
Fixes #2390, fixes #2389
Summary
Instead of doing forward/log on metrics on each step it only updates the state at every step and does the computation/log/reset at the end of the epoch.