Bug In "Art Generation with Neural Style Transfer"

[niunuinui]

Hi, I learn a lot from your code! Thanks a lot!

I think there is a bug in the save_image function of Art Generation with Neural Style Transfer code. Since .detach() share the same memory, we should clone the input image first or the change will apply to the original image and mess the training.

Bellow is the fixed code for save_image

def save_image(path, imageO):
    # Un-normalize the image so that it looks good
    image=imageO.clone().detach()
    image = image.cpu().numpy().transpose(0, 2, 3, 1)
    image += CONFIG.MEANS
    image = np.clip(image[0], 0, 255).astype('uint8')

[r1ght0us]

Hi, I learn a lot from your code! Thanks a lot!

I think there is a bug in the save_image function of Art Generation with Neural Style Transfer code. Since .detach() share the same memory, we should clone the input image first or the change will apply to the original image and mess the training.

Bellow is the fixed code for save_image

def save_image(path, imageO):
    # Un-normalize the image so that it looks good
    image=imageO.clone().detach()
    image = image.cpu().numpy().transpose(0, 2, 3, 1)
    image += CONFIG.MEANS
    image = np.clip(image[0], 0, 255).astype('uint8')

牛逼好兄弟，终于找到原因了！！！！