BCEWithLogitsLoss training predication

vision
1

Hello

I am using BCEWithLogitsLoss to train a UNet model, the input images are in a RGB format and target images are binary (0 to 1)

When training the model I looked at the shape of inputs in the batch, are they are returning:

torch.Size([8, 3, 250, 250]) - input RGB image
torch.Size([8, 1, 250, 250]) - target binary image

over training the predicated outcome has a shape of:
torch.Size([8, 1, 250, 250])

pred = model(images)
print(pred.shape)
loss = torch.nn.BCEWithLogitsLoss()(pred, true_masks)

I just wondered that normal for the predicated outcome to include the batch size, being 8 in this case?

Thank you

2

Yes, the model is expected to return an output for each sample. Since you have passed 8 samples as a batch to the model also 8 outputs are expected in the output batch.

3

Thank you

Within a batch for-loop it will be calling each function line 8 times behind the scenes? For example:

loss.backward()
optimizer.step()

If we look at the output of: pred = model(images), will this show the logits numbers? Rather than the sigmoid outcomes?

Thank you

4

No, since PyTorch layers accept batched inputs and will launch single kernels for each operation.

It depends on your model and if it’s returning raw logits from e.g. a linear layer of if a torch.sigmoid is applied on the output before returning it.

5

Sorry, I meant over training if we have:

pred = model(images),
loss = torch.nn.BCEWithLogitsLoss()(pred, true_masks)

would the output of pred be the logits numbers, so much wider range of numbers than 0 to 1? I just noticed the output over training has minus values, so just trying to understand what it represents. I have check the true_masks and they are binary with a max of 1.

Thanks