TorchMetrics wrapper around another library's metrics

[ibro45]

Hi!

I would like to have a TorchMetrics wrapper around MONAI metrics. I initially implemented this, but I realized that it does not work properly, and I assume it is because of lack of .add_state():

import torchmetrics


class MonaiToTorchMetricsWrapper(torchmetrics.Metric):
    def __init__(self, monai_metric_class, **kwargs):
        super().__init__()
        self.monai_metric = monai_metric_class(**kwargs)

    def update(self, preds, target):
        return self.monai_metric(preds, target)

    def compute(self):
        return self.monai_metric.aggregate()

    def reset(self):
        self.monai_metric.reset()

How would you go about doing that?

[ibro45]

Here's exactly what's happening:

• First row - torchmetrics Dice
• Second row - MONAI Dice
• On the right - metric at a particular step. The two metrics are almost the same, there are some differences due to different implementations.
• On the left - metric over an epoch. You can see that the torchmetrics and MONAI Dice differ here. Specifically, MONAI Dice is not corrected correctly for the epoch. Note, that the MONAI Dice is wrapped into a torchmetrics Metric, as posted above.

Validation

Training

[nkaenzig]

One issue in your wrapper class is that in the .reset() method, you should also call torchmetric's reset:

    def reset(self):
        super().reset()
        self. monai_metric.reset()

Also make sure to configure aggregations / reductions correctly. I think monai's implementation using reduction="mean", ignore_empty=False is equivalent to using average="macro" in torchmetrics.Dice ("micro" is the default there).