Doubt about the filters of transposed convolution?
Closed this issue · 4 comments
I read your paper and in this paper, the first two transposed convolution is drawn as:
It seems that the output of first two transposed convolutions are a 1x32x1024 tensor and a 1x64x1024 tensor. However in your code the transposed convolutions are:
filters = 64
for _ in range(int(np.log(upscaling_factor) / np.log(2))):
x0 = up_block(x0, filters, strides=(2,1)
Which make the output of transposed convolutions are a 64x32x1024 tensor and a 64x64x1024 tensor.
Which is right?
(I am not familiar with tensorflow, and maybe i misunderstood the code).
@EpsAvlc
The code seems correct to me. A stride of (2,1) will upscale the row number of the image.
@TixiaoShan May be I didn't explain my doubt well.
Whether the filter bank of the first two transposed convolution layers is 1 or 64?
In the picture, it seems 1. In the code, it seems 64.
@EpsAvlc
Ah, I see what you mean. You are right, in the paper, it's 1. But it's 64 in the code. 64 should give better accuracy.
Got it. Thank you.