Using conv_only in Convolutional layers

[Can-Zhao]

Hi,

I found that conv_only=True is almost always used in MONAI Generative Models, like

GenerativeModels/generative/networks/nets/autoencoderkl.py

Line 137 in b1c1a45

conv_only=True,

While in official Stable Diffusion code, bias=True is used as this is the default in torch.nn.Conv2d like in stable diffusion v1 and stable diffusion v2.

I was wondering if this modification was expected?

Thank you!!

[Can-Zhao]

@Warvito @ericspod What do you think? Thank you!

[ericspod]

The conv_only=True argument means that the block will not have activation/regularisation/dropout layers, ie. just the convolutional layer, but this will still have the bias term if the bias argument is not changed from its default of True.

[Can-Zhao]

Got it! Thank you for the explanation!