Help..I have a try of pytorch and meet pixel problem

[Oldpan]

It's a nice work and I want to use pytorch to re-accomplish it. But during the first step 'IndependentMapping' I meet a pixel mess problem...

I do clamp and use mask image to do backward but the pixels seem to be out of range. The loss function I use are content loss,gram loss, and TV loss. I haven't use histogram loss. The model I use is Vgg19 from pytorch model zoo which the data range is 0-1.I'm sure the image I put into is right format(RGB) and when I tune the style weight or content the result changes a bit. I have no clear idea where the problem is. Can you help me?Thanks! 貌似我该说中文？

[luanfujun]

For the purpose that more people can read it, I will still reply using English but you can contact me by email if you still have troubles in reproducing it in PyTorch...😄

Firstly, the VGG-19 requires input image to be [0, 255] and then...

Secondly, did you subtract the mean pixel?

One easier way to debug might be starting from a PyTorch version of style transfer implementation such as this one from original author Leon Gatys: https://github.com/leongatys/PytorchNeuralStyleTransfer

[Oldpan]

Wow, so fast.. The image I input is [0-1]. I use the code in pytorch:

def toTensor(img):
    assert type(img) == np.ndarray,'the img type is {}, but ndarry expected'.format(type(img))
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = torch.from_numpy(img.transpose((2, 0, 1))).unsqueeze(0).clone()
    return img.float().div(255.0)

And I also do the normalization to the image.

cnn_normalization_mean = torch.tensor([0.485, 0.456, 0.406]).to(device)
cnn_normalization_std = torch.tensor([0.229, 0.224, 0.225]).to(device)

# create a module to normalize input image so we can easily put it in a
# nn.Sequential
class Normalization(nn.Module):
    def __init__(self, mean, std):
        super(Normalization, self).__init__()
        # .view the mean and std to make them [C x 1 x 1] so that they can
        # directly work with image Tensor of shape [B x C x H x W].
        # B is batch size. C is number of channels. H is height and W is width.
        self.mean = torch.tensor(mean).view(-1, 1, 1)
        self.std = torch.tensor(std).view(-1, 1, 1)

    def forward(self, img):
        # normalize img
        return (img - self.mean) / self.std

But the problem still exists. So I'm trying to figure out ... Yeah,if I can't fix thie problem maybe I should start from Pytorch version of other's. Uh.uhuh..I think I need to contact you with email.

[luanfujun]

I think VGG-19 requires input to be [0, 255] subtracted by mean pixel?

[Oldpan]

thanks!

I had sent you a gmail the day before yesterday and I don't know if you have received it.

The model I use is from Pytorch and the VGG-19 model requires input to be [0-1]...I think the model I use is right because it works in common style transfer.

I compared two ways to load the image: PIL vs OpenCV-python, and I found something weird but I'm still not sure where the problem comes from.

I'm confused about the code below from 'neural_gram.lua':

      if name == style_layers[next_style_idx] then
        print("Setting up style layer  ", i, ":", layer.name)
        local gram   = GramMatrix():float():cuda()
        local input  = net:forward(content_image_caffe):clone()
        local target = net:forward(style_image_caffe):clone()
        local mask   = mask_image:clone():repeatTensor(1,1,1):expandAs(target):cuda()

       # if I don't use histogram match and just input target to gram-loss function
       # like this : target_gram = gram:forward(target):clone()

        local match, correspondence = 
            cuda_utils.patchmatch_r(input, target, params.patchmatch_size, 1)
        match:cmul(mask)
        local target_gram = gram:forward(match):clone()

        target_gram:div(mask:sum())
        local norm = params.normalize_gradients
        local loss_module = nn.StyleLoss(params.style_weight, target_gram, norm, mask_image):float():cuda()
        net:add(loss_module)
        table.insert(style_losses, loss_module)
        next_style_idx = next_style_idx + 1
      end

I haven't use histogram and just input activation layer from style-image to gram-loss function. If I do this, the image I produced is totally wrong? Or just not looks so well but still right? Thank you for helping me!

[luanfujun]

If only style loss but no histogram loss the quality will be less good but not like the one you posted. There might still have some bug so I would first ensure entire image style transfer works and then debug masked region using unit tests.

[Oldpan]

Yeah...The entire image style transfer works. So I'm still debugging in mask mode and when I change the style layers and content layers position the result changes drastically.

[Oldpan]

I check my code over again and expel all the bugs I could find There are some of my results... When I change the style weight and content weight the result also changes.Some times the image seemed to get better. But I still can't find the proper weight to produce a nice image... [sad][sad][sad]

[Oldpan]

After a long parameters tuning the first step is almost working. But when I run patchmatch like patchmatch_r() function I often run out of memory. Does this function needs memory more than 10GB? Thanks~