Can dice and bce loss work on a multi-class task? about pytorch-unet

[ckolluru]

Why are the metrics multiplied by the batch size, added cumulatively and then divided by the total number of samples during printing?

It seems like this will print a scaled version of the average metric value (from print_metrics), depending on the batch size. Please correct me if I'm wrong.

[usuyama]

You're right. The current code uses BCE, so each pixel can have multiple classes i.e. multi-labels.

To make it a single class for each pixel i.e. multi-class, you can use CE. I think you need to use reshape/view to 2d.

[sarmientoj24]

Hi, I am having a problem dealing with a multi-class task where dimensions are like these:

MASK TARGET: torch.Size([4, 1, 600, 900, 3])
OUTPUT: torch.Size([4, 5, 600, 900]

@ckolluru Have you created your loss function for multiclass already?

[shahzad-ali]

Why are the metrics multiplied by the batch size, added cumulatively and then divided by the total number of samples during printing?

The reason is some batches (i.e. the last batch) may have fewer training examples than all the other batches. Dividing the product metric * batch_size by total_samples is a better estimate of a used metric for a complete epoch. Skimming through the example Training a Classifier from PyTorch tutorials reveals that the same strategy was used in section "4. Train the network" is the important one where the statistics were printed.

It seems like this will print a scaled version of the average metric value (from print_metrics), depending on the batch size. Please correct me if I'm wrong.

That's true! You may get a better insight into this topic by reading How is the loss for an epoch reported?

[AlphonsG]

If you want to use cross entropy make sure you're not applying sigmoid function beforehand.