Frist parameter in downsamplers should be 'out_channels', instead of 'in_channels'

[chenguolin]

diffusers/src/diffusers/models/unet_blocks.py

Line 619 in 9bca402

in_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"

All the frist parameters of Downsample2D-like classes in self.downsamplers are in_channels.

However, the inputs of these self.downsamplers are the outputs of Resnets, whose number of channels should be out_channels.

Though these wrong parameters do not affect the results, since in_channels is assigned as out_channels in Resnet iteration:

diffusers/src/diffusers/models/unet_blocks.py

Line 597 in 9bca402

in_channels = in_channels if i == 0 else out_channels

, they may cause confusion for code reading.

By the way, the first parameters of Upsample2D-like classes are out_channels, which is correct I think:

diffusers/src/diffusers/models/unet_blocks.py

Line 1200 in 9bca402

self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])

[patrickvonplaten]

Hey @chenguolin,

I agree that the naming is not ideal here, I've never managed to find time to clean up the naming. Do you have good ideas for how we should change the naming? Would you like to open a PR maybe? :-)

[chenguolin]

OK. I'll pull a request to clean up the naming in these classes later.