I can't understand the generator network of the Train Generative Adversarial Network (GAN) example

Question

煜浩汪 2021-8-11

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/896417-i-can-t-understand-the-generator-network-of-the-train-generative-adversarial-network-gan-example

回答： Ben 2024-4-9

In the first layer of transposed convolution, why does it not use 'Stride',2,'Cropping','same' like the following layers? I think the width and height of the input is 4x4 so it needs to be enlarged by a factor of 2 for 4 times to get a size of 64x64. But in this code, I think it only enlarges the input by a factor of 2 for 3 times. Could you please help me understand this?

By the way, why it says 'Converts the random vectors of size 100 to 7-by-7-by-128 arrays using a project and reshape layer' but the projection size in the code is [4 4 512]?

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Ben 2024-4-9

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/896417-i-can-t-understand-the-generator-network-of-the-train-generative-adversarial-network-gan-example#answer_1438581

The documentation for transposedConv2dLayer states in the Algorithms section that the input is padded with zeros up to "filter edge size minus 1". I think this explains why the 4x4 input becomes 8x8. To see this, call analyzeNetwork(netG) on the generator in the example.

The statement "Converts the random vectors of size 100 to 7-by-7-by-128 arrays using a project and reshape layer" isn't in the current version of the example, perhaps it was a typo in an old version.