something about β about esrgan

[xinntao]

Thanks for reporting the problem :-)

The code is what we use, so use the code version. It seems that we put the beta in the wrong path. We will update the figure in the paper later.

[IPNUISTlegal]

i am noob in tensorflow.

ESRGAN/block.py

Line 208 in ce16ee8

return x5.mul(0.2) + x

The channel of x5 (data_format ，HWC) reduces to the original one-fifth of x5 to add skip connection with x in your torch framework. Is any similar operation or function in tensorflow?
appreciate that(#^.^#)

[xinntao]

I am not familar with tensorflow. You may try 0.2 * x5 directly. It should support this operation.

[IPNUISTlegal]

I am not sure about whether your code uses a 1×1 convolution as bottleneck layer before each 3×3 convolution in your denseblock. It seems to no 1×1 convolution in network architecture of your paper.

[xinntao]

@IPNUISTlegal As you can see from the code, we do not use 1x1 convolution in the denseblock.