Configurable metrics #230
Configurable metrics #230
Conversation
djdameln
added
Enhancement
| pixel_f1 = F1(num_classes=1, compute_on_step=False, threshold=self.hparams.model.threshold.pixel_default) | ||
| self.image_metrics = MetricCollection([image_auroc, image_f1], prefix="image_").cpu() | ||
| self.pixel_metrics = MetricCollection([pixel_auroc, pixel_f1], prefix="pixel_").cpu() | ||
| self.image_metrics, self.pixel_metrics = get_metrics(self.hparams) |
There was a problem hiding this comment.
Is there something planned for classification models which do not have pixel metrics? When I removed the pixel metric key from the config file it threw error for padim
There was a problem hiding this comment.
This should work:
metrics:
image:
- F1
- AUROC
pixel: []
There was a problem hiding this comment.
Not sure why my comments disappeared from here, but ideally
metrics:
image:
- F1
- AUROCshould also work fine.
|
Works like a charm! |
| from .min_max import MinMax | ||
| from .optimal_f1 import OptimalF1 | ||
|
|
||
| __all__ = ["AUROC", "OptimalF1", "AdaptiveThreshold", "AnomalyScoreDistribution", "MinMax"] | ||
|
|
||
|
|
||
| def get_metrics(config: Union[ListConfig, DictConfig]) -> Tuple[AnomalibMetricCollection, AnomalibMetricCollection]: |
There was a problem hiding this comment.
We might need to modify this for LightningCLI
| pixel_f1 = F1(num_classes=1, compute_on_step=False, threshold=self.hparams.model.threshold.pixel_default) | ||
| self.image_metrics = MetricCollection([image_auroc, image_f1], prefix="image_").cpu() | ||
| self.pixel_metrics = MetricCollection([pixel_auroc, pixel_f1], prefix="pixel_").cpu() | ||
| self.image_metrics, self.pixel_metrics = get_metrics(self.hparams) |
There was a problem hiding this comment.
Not sure why my comments disappeared from here, but ideally
metrics:
image:
- F1
- AUROCshould also work fine.
|
Might be out of the scope of this PR but we can have a look at this PR Lightning-AI/torchmetrics#709 which updates the base metrics only once. Maybe we can find a way to automatically group metrics. |
|
Can you fix torchmetrics version in this PR as well |
Description
Changes
Checklist