Cyclegan Notebook Results Pixelated

Part 2 & Alumni (2018)
1

Thank you for making this course! I have learned a lot from both Jeremy and the community and hope to contribute myself as well.

I am trying to run the cyclegan and cyclegan_ws notebooks from course-v3 in dl2. The output images are pixelated, even if I do the same amount of training that is the default in the notebook (100 iterations). I tried about 300 iterations too. Is this just a problem of not training enough or are there other changes I should make to the code in the notebook? Attached Here is a reference of what I am getting as an output.

00%20PM
Screen Shot 2019-11-01 at 3.43.00 PM.jpg1908×472 227 KB

Thank You!

2

Have you change the notebook at all? I have not went trough the notebook, but have implmented a cyclegan based network. Your predictions look too regular to be a GAN output, and I would think they are an issue with incorrectly decoding your tensors to images, or something like that.

3

No, I have not made any changes. I ran the notebook as is. I wanted to make changes once I got it working, but the output is not what I expected.

4

I’m having the exact same problem with the horse/zebra dataset, which I trained for 100 epochs, as per the notebook.
Screenshot%20from%202019-11-02%2011-02-31

5

I’m having the same issue. I trained it for just 7 epochs and the top and left parts of the image are pixelated. The only change I made to the notebook is using bs=2 b/c bs=4 wouldn’t fit in memory.