R2s from .predict() and .test() do not match

[giusmatera]

Hi, I found what it seems to be a bug:

I have

from torchmetrics import MeanSquaredError, R2Score

	def test_step(self, batch, batch_idx):
	
	    batch = batch.to(torch.float32)
	
	    pooled_output = batch[
	        :, :-1
	    ].float()  # All columns except the last
	    targets = batch[:, -1].float()
	
	    predictions = self(pooled_output).squeeze(-1)
	
	    score_r2 = self.test_r2(predictions, targets)  # Compute R2
	    self.log("test_r2", score_r2, on_epoch=True)
	
	    return score_r2

but the test r2 score this step gives differs from what I retrieve directly from predictions and actual values.

To clarify, the r2 score in results = trainer.test(regressor_model, test_dataloader) is different from what I get from

predictions = trainer.predict(regressor_model, test_dataloader)
torchmetrics.R2Score()(predictions, error)

Also, torchmetrics and sklearn r2 functions match in the latter task. How come? Thanks a lot.

Info:

• torchmetrics' version is 1.4.2

[github-actions]

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