Manifold Alignment for Semantically Aligned Style Transfer

Hi, thanks for sharing the code! I can't find your pre-trained models. can you please share it? also - do you have plans to share your train code? Thanks. Ofer

Hi @NJUHuoJing, thanks for sharing your great code and congratulations on your amazing work. I have a question about the encoder that you applied in your framework, I can get similar results as yours when I use your code with your provided encoder (vgg_r51.pth). However, when I use the pre-trained VGG-19 model provided by pytorch-torchvision, I cannot get any reasonable result. I have tried the torchvision VGG-19 with the exact same layers as you. For instance, in the test_artistic.py code, I used the r41,r31,r21, and I defined the following code for the encoder.

from torchvision import models 
class VGG19_encoder(torch.nn.Module):
    def __init__(self,requires_grad=False,n_layers=[1,6,11,20,29]):       
        super(VGG19_encoder,self).__init__()
        vgg_pretrained_features=models.vgg19(pretrained=True).features   
        self.slice0=torch.nn.Sequential()
        self.slice1=torch.nn.Sequential()
        self.slice2=torch.nn.Sequential()
        self.slice3=torch.nn.Sequential()
        self.slice4=torch.nn.Sequential()
        for x in range(n_layers[0]):#relu1_1
            self.slice0.add_module(str(x),vgg_pretrained_features[x])        
        for x in range(n_layers[0],n_layers[1]): #relu2_1
            self.slice1.add_module(str(x),vgg_pretrained_features[x])        
        for x in range(n_layers[1],n_layers[2]): #relu3_1
            self.slice2.add_module(str(x),vgg_pretrained_features[x])
        for x in range(n_layers[2],n_layers[3]):#relu4_1
            self.slice3.add_module(str(x),vgg_pretrained_features[x])
        for x in range(n_layers[3],n_layers[4]):#relu5_1
            self.slice4.add_module(str(x),vgg_pretrained_features[x])        
        if not requires_grad:
            for param in self.parameters():
                param.requires_grad=False            
    def forward(self,x):
        h0=self.slice0(x)    #[1,64,512,512]
        h1=self.slice1(h0)   #[1,128,256,256]
        h2=self.slice2(h1)   #[1,256,128,128]
        h3=self.slice3(h2)   #[1,512,64,64]
        h4=self.slice4(h3)   #[1,512,32,32]
        return {'r11':h0,'r21':h1,'r31':h2,'r41':h3,'r51':h4}

Then, I defined the encoder as the following line and added it in multi_level_test by using just the r41,r31,r21 layers:

encoder = VGG19_encoder(n_layers=[1,6,11,20,29]).to(device)

I am wondering how did you train your encoder and what is the difference between your VGG-19 model and the VGG-19 model from torchvision since both of them are trained on ImageNet.

I would appreciate a lot any help that you could provide regarding this issue. Thank you very much.