Metric accuracy change from PL 1.1.8 to 1.2.3

[kielnino]

Hi, I have a classical binary segmentation model and log its accuracy with the pl.metrics.Accuracy class.

def __init__(self):
    [....]
    self.seg_val_metric = pl.metrics.Accuracy()

def validation_step(self, batch, batch_idx):
    x, y = batch
    preds = self(x)
    [....]
    true_labels = y[1].view(y[1].size(0), -1).int()
    pred_labels = (torch.sigmoid(preds[1]).view(preds[1].size(0), -1) > 0.5).int()
    self.log('val_binary_segmentation_accuracy', self.seg_val_metric(pred_labels, true_labels))

def validation_epoch_end(self, outputs):
    self.log('val_binary_segmentation_accuracy', self.seg_val_metric.compute())

This works fine, but when I upgrade from PL 1.1.8 to 1.2.3 without any code changes, the accurcy is much lower (and smoother). So I'm wondering, if there is an API-change to the metrics class or somewhere else that I missed:

The same holds for a custom metric (subclass of the base Metric). Any hint or help would be great!

[kielnino]

I think the mystery is solved with #6696 now. I did'nt reset the metrics properly, which is required from 1.2.x on.