CompositionalMetric only returns value for last batch

[Callidior]

🐛 Bug

When using a compositional metric, i.e., metric_a + metric_b or 2 * metric_a with forward() instead of update(), the value returned by .compute() will be based only on the last batch.

This is because the default implementation of Metric.forward(), which is inherited by CompositionalMetric, calls reset(), which in turn resets all child metrics. Their state, however, is not restored by forward() afterwards.

My proposed fix would be to override forward() in CompositionalMetric and I will file a PR implementing this.

To Reproduce

Code sample

import torch
import torchmetrics
from torchmetrics.metric import CompositionalMetric

# Example predictions and targets
pred = torch.tensor([0.4, 0.2, 0.6, 0.7, 0.5])
target = torch.tensor([0.5, 0.4, 0.4, 0.3, 0.6])

# Compute base metric in batches
metric = torchmetrics.MeanSquaredError()
metric(preds=pred[:3], target=target[:3])
metric(preds=pred[3:], target=target[3:])
val = metric.compute()
print(val)

# Compute compositional metric in batches
cm = 1 * torchmetrics.MeanSquaredError()
cm(preds=pred[:3], target=target[:3])
cm(preds=pred[3:], target=target[3:])
val = cm.compute()
print(val)

Expected behavior

Two identical values are printed:

tensor(0.0520)
tensor(0.0520)

Actual (faulty) behavior

The compositional metric only returns the value for the last batch:

tensor(0.0520)
tensor(0.0850)

Environment

• PyTorch Version (e.g., 1.0): 1.10.0
• OS (e.g., Linux): Linux
• How you installed PyTorch (conda, pip, source): conda
• Build command you used (if compiling from source):
• Python version: 3.8
• CUDA/cuDNN version: 11.1
• GPU models and configuration:
• Any other relevant information:

Additional context

[github-actions]

Hi! thanks for your contribution!, great first issue!