Can't reproduce the results, maybe sensitive to data?

[dejianchen1989]

I ran the training scripts directly, but can't reproduce the results.
It seems there is no significant difference between the original and generated images:

Is it sensitive to training data? I also used face-cropped celebA as src_data:

and face-cropped danbooru2018 as tgt_data:

Each dataset contains about 1600 images (for fast training)。
So, where is the problem? THX~

[dejianchen1989]

Also, it seems that the Initialization phrase suffers from checkerboard effect, as illustrated in
https://distill.pub/2016/deconv-checkerboard/.

I changed ConvTranspose2d to Upsample+Conv2d (as suggested in the above post), but the quality of image generated drops a lot.

[zhra46]

@dejianchen1989 i think the checkerboard effect happened at original papers too.
Besides, there is a error while computing:D_fake_loss = BCE_loss(D_fake, fake)

ValueError: Target and input must have the same number of elements. target nelement (8192) != input nelement (38400)

i found this is because the different shape between D_fake and Fake variables
should i resize all the natural image to 256,256 before training process?

[Ricelll]

@dejianchen1989 Hi, I can't reproduce the results with CelebA and Cartoon's imgs,either.Gen Loss and Con Loss dosn't decrease.

[syaoran13]

e = y[:, :, :, args.input_size:]
y = y[:, :, :, :args.input_size] here i'm getting error

[zwq11]

May I ask if your problem has been solved? My training data is not very different from the raw data either.