Joint-Bilateral-Learning
Joint-Bilateral-Learning copied to clipboard
mousecpn
An unofficial implementation of Joint Bilateral Learning for Real-time Universal photorealistic Style Transfer
你好，非常感谢您的工作！在对比论文阅读您的代码时，我发现了在VGG.py中style loss好像有些问题，具体是在 `loss(self, output, content, style, alpha=1.0)`处计算了两次relu1_1特征的[style loss](https://github.com/mousecpn/Joint-Bilateral-Learning/blob/master/VGG.py#L143) ```python loss_s = self.calc_style_loss(out_feats[0], style_feats[0]) for i in range(0, 4): loss_s += self.calc_style_loss(out_feats[i], style_feats[i]) ``` 我反复阅读了论文，其中并没有提到要单独对relu1_1算两次loss，这里是怎么回事呢？ 谢谢啦！
你好，这篇利用HDRNET来做样式迁移的应用，我也在复现。目前有一些效果。文章细节缺失，代码也没有公开。复现挺难的。从文章看，要用6次AdaIN，然后如果正则损失过小，也会产生很多偏色。所以regW我设置为100000。最后的效果其实更像是整体颜色的迁移，因为hdrnet本身就是保边滤波，无法迁移样式。
RuntimeError: Error(s) in loading state_dict for Model: Missing key(s) in state_dict: "Biliteral_Grid.SB1.conv1.conv2d.weight", "Biliteral_Grid.SB1.conv1.conv2d.bias", "Biliteral_Grid.SB1.conv2.conv2d.weight", "Biliteral_Grid.SB1.conv2.conv2d.bias", "Biliteral_Grid.SB1.conv_short.weight", "Biliteral_Grid.SB1.conv_short.bias", "Biliteral_Grid.SB2.conv1.conv2d.weight", "Biliteral_Grid.SB2.conv1.conv2d.bias", "Biliteral_Grid.SB2.conv2.conv2d.weight", "Biliteral_Grid.SB2.conv2.conv2d.bias", "Biliteral_Grid.SB2.conv_short.weight", "Biliteral_Grid.SB2.conv_short.bias", "Biliteral_Grid.SB3.conv1.conv2d.weight", "Biliteral_Grid.SB3.conv1.conv2d.bias", "Biliteral_Grid.SB3.conv2.conv2d.weight", "Biliteral_Grid.SB3.conv2.conv2d.bias", "Biliteral_Grid.SB3.conv_short.weight", "Biliteral_Grid.SB3.conv_short.bias", "Biliteral_Grid.conv1.conv2d.weight",...
Hi, I'm interested with this model. But I can't run the test without your pretrained weights. Appreciated if you can offer the pretrained weights. Thx. My email: [email protected]
### add supplementary material: https://link.springer.com/chapter/10.1007%2F978-3-030-58598-3_20 the supplementary material 1 (zip 80044 KB) is still missing.
Hello,@mousecpn, I have a question ,should you help me ? the question is when the network in inference need vgg network to get features. when it deploy on mobile device,...
If my test picture is a picture of any resolution size, such as 1k, 2k, can I omit this step? transform = transforms.Compose([ transforms.Resize((512, 512), Image.BICUBIC), transforms.ToTensor() ]) if not...
Hello, What are used as the training content and style images? Are coco_2014_train dataset used as the content images? How's your performance?
