PGGAN blending issue

[GongXinyuu]

pytorch_GAN_zoo/models/networks/progressive_conv_net.py

Line 345 in 382a968

mergeLayer = self.alpha > 0 and len(self.scaleLayers) > 1

Should this line be mergeLayer = self.alpha > 0 and len(self.fromRGBLayers) > 1?

As the previous line calculate y on the condition of

pytorch_GAN_zoo/models/networks/progressive_conv_net.py

Line 336 in 382a968

if self.alpha > 0 and len(self.fromRGBLayers) > 1:

[Molugan]

You're right, the size of fromRGBLayers and scaleLayers are supposed to be the same at each step, but it's not super clean to refer to both in the code. I'll clean that up.

[LiUzHiAn]

Hi,

I don't think the size of fromRGBLayers and scaleLayers are the exactly same. Since the scale0 was not added to the scaleLayers.

pytorch_GAN_zoo/models/networks/progressive_conv_net.py

Lines 264 to 273 in b75dee4

	self.groupScaleZero.append(EqualizedConv2d(dimEntryScale0, depthScale0,
	3, padding=1,
	equalized=equalizedlR,
	initBiasToZero=initBiasToZero))
	self.groupScaleZero.append(EqualizedLinear(depthScale0 * 16,
	depthScale0,
	equalized=equalizedlR,
	initBiasToZero=initBiasToZero))

But the RGBLayer for the scale0 was added.

pytorch_GAN_zoo/models/networks/progressive_conv_net.py

Lines 254 to 257 in b75dee4

	self.fromRGBLayers.append(EqualizedConv2d(dimInput, depthScale0, 1,
	equalized=equalizedlR,
	initBiasToZero=initBiasToZero))

So, mergeLayer = self.alpha > 0 and len(self.fromRGBLayers) > 1 may be more reasonable. Did I misunderstand?

Thanks for your work.