the problem of the generated images

[qianxiao111]

I used this method to train the translation between undistorted images and distorted images, but I found some strange ripples in the generated images. I would like to ask you what caused this? The number of trained images is 300.

[ssnl]

300 is very likely too small a dataset.

[qianxiao111]

300 is very likely too small a dataset.

However, the test results is not all the images are like this, only a few are like this.

[junyanz]

A few inferior results are expected. Reasons could be that your test input images are different from the training set images. (out of distribution)

[John1231983]

I meet the problem before. It due to learning conv in upsampling path. You have two choices (1) training longer. (2) replace the conv2d transpose with upsampling followed by pad and conv3x3. Detail in #382