Reproducing Results

[acerdur]

Hi,

I am running photo_transfer.py with the given content and style images. All the settings are in default. However, the results are not satisfying as shown. Here is what I get for the first pair:

And for the second:

Are the provided model weights up-to-date? Or is the default d_control variable in photo_transfer.py in the optimal configuration? Are the results produced with the given settings? Many thanks.

[ayushmankumar7]

Hey which version of Pytorch did you use?

[oferidan1]

Hi, I am running the pre-trained model and getting same results as acerdur above. did you succeed to fix it?

I am using pytorch version 0.4.1. using the original code I got an error "RuntimeError: Given groups=1, weight of size [3, 3, 1, 1], expected input[1, 512, 768, 3] to have 3 channels, but got 512 channels instead" due to dimension mismatch between opencv and pytorch images. I added below code to fix it and it runs with success, but now get similar images as above with artifacts. content = np.transpose(content, (2, 0, 1)) style = np.transpose(style, (2, 0, 1))

can you please assist? thanks in advanced. Ofer

[oferidan1]

found a fix - after installing pytorch 0.4.1 + torchvision 0.2.2 results are good.