Why are UNet up blocks channels "flipped"?

[AlejandroBaron]

Hello I'm checking the UNet implementation and when instantiating the up blocks, I see the input channels are "flipped" from my understanding of UNets, in the "up" side/second half

diffusers/src/diffusers/models/unets/unet_2d.py

Lines 209 to 214 in 08bf754

	reversed_block_out_channels = list(reversed(block_out_channels))
	output_channel = reversed_block_out_channels[0]
	for i, up_block_type in enumerate(up_block_types):
	prev_output_channel = output_channel
	output_channel = reversed_block_out_channels[i]
	input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]

So, if block_out_channels is let's say (32,64,128), then pairs of input_channel and output_channel would be

input_channel	output_channel	prev_channels
64	128	128
32	64	128
32	32	64

Shouldn't the up half of the UNet be bottlenecking the channels (i.e. input_channels>output_channels? I get that then when instantiating each layer the input channels are increased by the skip connection/prev_channels but still this looks weird to me

EDIT:

To support my theory, I think the way I comment this is how it's done in the StabilityAI repo
https://github.com/Stability-AI/stablediffusion/blob/main/ldm/modules/diffusionmodules/openaimodel.py

[tolgacangoz]

I think you understood correctly. What is your question exactly, could you elaborate?

[AlejandroBaron]

@StandardAI thanks for answering. My question is that the channels should be

input=>output
128 => 64
64 => 32
32 => 32

but as you can see in my table, it's the "flipped case"