Add conv layer instead of linear layer to meet the original paper

[seungwooYoo]

Hi,

Thanks for sharing the codes of Image-to-Image Translation with CGAN.

I found that the last layer of discriminator network consists of linear layer instead of conv layer, which is written in the appendix of original paper (Appendix 5.1.2) :

After the last layer, a convolution is applied to map to a 1 dimensional output, followed by a Sigmoid function.

I changed a little of your implementation to meet the original ones. If my suggestion seems not correct, then do not care about the pull request. Thanks,

[yenchenlin]

You are correct, I'll test this and then merge it if it works.

[yenchenlin]

Hello @seungwooYoo , I've tried the code in this PR and also try to enhance the discriminator to 6-layer. However, the results seems more blurry than before. How is the result on your side using code in this PR?

[seungwooYoo]

Actually, I haven't compared with the previous codes. I'll test and tell you the differences.

[yenchenlin]

@seungwooYoo thanks.

[seungwooYoo]

@yenchenlin It seems that the result is little bit much blurred than before. Seems like cGAN is not working well when compared with Figure 4. Do you have any idea to solve that issue ? I'll retest using L1 penalty as zero and see the results after optimization which meet the example of figure 4.

[yenchenlin]

No, I haven't come out a solution yet 😞 Please let me know if you do.

[seungwooYoo]

@yenchenlin Yes I will. I'll test some ideas in the following weekend :)

[eyaler]

original pix2pix implementation does not have bn on last convolution

[nightrome]

Neither does this... https://github.com/yenchenlin/pix2pix-tensorflow/blob/master/model.py#L280 https://github.com/yenchenlin/pix2pix-tensorflow/blob/master/model.py#L211

[eyaler]

also the you are missing the args: d_h=1, d_w=1