Use torchmetrics in baal.metrics

[Dref360]

We should swap most of our metrics for torchmetrics this would make our code more robust.

Only ECE is special I think and we should keep our own implementation.

[GeorgePearse]

I'd like to take this if no one's got started on it?

[Dref360]

Sounds good, what I was thinking was the following:

• Add torchmetrics as a dependency (poetry add torchmetrics)
• Make sure that baal.modelwrapper.ModelWrapper._update_metrics and _reset_metrics works with PL metrics and ours for backward compatibility.
• Update experiment scripts to match the new PL.

[Dref360]

I just wrote the following Baal metrics that leverage torchmetrics.

We have very similar APIs so I think its a good idea to scrap Baal metrics in favor of torchmetrics. Wyt?

 from typing import List
 
 import numpy as np
 import torch
 from torchmetrics import Accuracy
 from torchmetrics.functional import auc
 
 from baal.utils.metrics import Metrics
 
 
 class AccuracyAUC(Metrics):
     def __init__(self, thresholds: List[float], **kwargs):
         self.thresholds = thresholds
         self._accs = [Accuracy(threshold=th) for th in thresholds]
         super().__init__(**kwargs)
 
     def reset(self):
         for acc in self._accs:
             acc.reset()
 
     def update(self, output=None, target=None):
         for acc in self._accs:
             acc.update(output, target)
 
     @property
     def value(self):
         return auc(torch.FloatTensor(self.thresholds), torch.FloatTensor([acc.compute() for acc in self._accs]),
                    reorder=False).item()
 
 
 random_probs = torch.softmax(torch.randn(100, 10), dim=1)
 random_target = torch.randint(0, 10, (100,))
 
 met = AccuracyAUC(np.linspace(0, 1, 10))
 met.update(random_probs, random_target)
 print(met.value)

[Dref360]

Hello, I would have time to work on it this week if you are not already started.