About SVR

[lyxhope]

Hi thanks for your work. Can I ask a question in Chinese？ 我想尝试一下关于SVR的工作，是不是像训练和测试gan那样，训练时把gan的训练换成一个ResNet的训练，测试时把gan生成latent code的部分换成用ResNet从图像来生成就可以？关于这部分您能提供相关的代码吗？

[rgga-16]

I would like to follow up on this. I'm also interested in this part.

[ChrisWu1997]

Yes, we use ResNet18 as the image encoder to map images to latent codes. The training objective is simply MSE, and only the image encoder got optimized. Sorry that I couldn't find the full training code anymore, but the image encoder should have structure as below. As I recall, PQ-Net's SVR results are not as good as those works specifically targeting SVR (e.g., DISN).

import torch.nn as nn
from torchvision.models import resnet18

class ImageEncoder(nn.Module):
    def __init__(self, z_dim=512):
        super(ImageEncoder, self).__init__()
        resnet = resnet18(pretrained=True)
        modules = list(resnet.children())[:-1]
        self.resnet = nn.Sequential(*modules)
        self.fc = nn.Sequential(nn.Linear(512, z_dim))

    def forward(self, x):
        feature = self.resnet(x)
        out = self.fc(feature.squeeze())
        return out

[rgga-16]

Got it. Thanks for the code. So based on how I understood the paper, do I input the image feature into the decoder part of the Seq2SeqAE? DL beginner here.

[ChrisWu1997]

Correct.