from torchmetrics.functional import true_negatives, false_positives

[nahondobruno]

iam new to torchmetrics and Iam inquiring which of these is the right implementation torchmetrics.
Thanks

from torchmetrics.functional import true_negatives, false_positives
TN = true_negatives(mincaps[~no_step_mask], actual_mincaps[~no_step_mask], num_classes=2).sum()
FP = false_positives(mincaps[~no_step_mask], actual_mincaps[~no_step_mask], num_classes=2).sum()
self.mincaps_FP += FP
self.mincaps_TN += TN

OR

import torchmetrics
TN = torchmetrics.functional.confusion_matrix(mincaps[~no_step_mask], actual_mincaps[~no_step_mask], num_classes=2)[0, 0]
FP = torchmetrics.functional.confusion_matrix(mincaps[~no_step_mask], actual_mincaps[~no_step_mask], num_classes=2)[0, 1]
self.mincaps_FP += FP
self.mincaps_TN += TN

[Borda]

Welcome to TorchMetrics. If you’re wondering which implementation of metrics like true_negatives or false_positives to use, the recommended way is to import them directly from the torchmetrics.functional module as you mentioned:

from torchmetrics.functional import true_negatives, false_positives

These functional APIs provide stateless metric calculations that you can use directly on tensors. Alternatively, TorchMetrics also offers metric classes (e.g., TrueNegatives, FalsePositives) in torchmetrics which maintain state across batches and are useful for aggregation during training loops.
If you’re just starting out and want quick metric computations, using the functional APIs is a good approach. As your pipeline grows, consider switching to the metric classes for more integration with PyTorch Lightning or custom training loops.