can not reproduce the result

[ceo1207]

i run this code to train a model twice, each time use 1w+ iterations.
But i can not reproduce the result of the paper. when i visualize the result, the result is terriable.

is there something wrong?
i just rescale and clip the result, but when i see the result, it just like a Color Inverted image. I try to change the order of the result, it does not work.

def show_img(img_data):
    img_data = np.squeeze(img_data)
    # img_data = img_data[:,:,::-1]
    img_data = img_data*255
    img_data = img_data.astype("int")
    img_data = np.clip(img_data, 0, 255)
    plt.imshow(img_data)
    plt.show()
def show_result(model, device):
    files_path  = "/home/ceo1207/Datasets/sr_test/Set5/"
    files = os.listdir(files_path)
    for item in files:
        img_array = np.asarray(Image.open(os.path.join(files_path, item)))
        plt.imshow(img_array)
        plt.show()
        img_array = img_array.transpose((2, 0, 1))
        img_array = img_array[np.newaxis, :, :, :]
        img_array = img_array.astype("float32") / 255.0
        x = torch.from_numpy(img_array).float()
        x = x.to(device)
        y_fake = model.gen_g(x)
        # y_fake.detach().cpu().numpy().transpose(0, 2, 3, 1)
        y_fake = y_fake.detach().cpu().numpy().transpose(0, 2, 3, 1)
        show_img(y_fake)

[shadow2496]

First of all, sorry for the late reply.

The problem of result images you mentioned is the one that we are frequently facing as well.
Though it is not included in the paper, we consider the absence of the content consistency loss in the target domain Y (DSLR-quality) as the most plausible cause of the problem and are planning to update the repo in several weeks so that the model can reach an ideal optimum.
Currently, we are focusing on improving the readability of our code and error checking, since we can't completely exclude coding mistakes from causes.

If you have any other ideas or suggestions about the problem, please share with us.
Thank you.

[jxust01]

Are discriminators not frozen when training generators?