About the detail of the ResNet50 backbone and FPN

[HardboiledHu]

Thanks for your work! Your Inference Code is working and I have generated excellent fake face using your code and model. After reading the article and the code of megafs.py, I have some question.

1. The number of Residual blocks is [7,6,3,2] in the article, but it is [3, 4, 6, 3, 2] in the megafs.py and actually, it is exactly same with ResNet50 network.
2. About the output of FPN which are p4,p3,p2,p1 from top to bottom. There are 3 ouput accepting the feature from the upper layer in the article, but the number is 2 in the code.
3. what the meaning of omega_only?
4. Do you directly resize the 1024x1024 input to 256?
5. what the meaning of CelebAMask-HQ? Because I didn't see any information of CelebAMask-HQ in the article.

[zyainfal]

1. [3, 4, 6, 3, 2] is the ResNet50 setting as we used the pretrained model to initialize the backbone. And [7,6,3,2] indicates how many blocks before c1, c2, c3, and c4.
2. I see the difference. The code is the correct version.
3. omega_only means the difference between latent space W+(omega only) and W++(omega + constant input of stylegan2).
4. Yes, this saves GPU memory a lot.
5. CelebAMask-HQ provides the labeled masks of faces in CelebA-HQ during the inference. We don't use masks in training so it is neglected.

[HardboiledHu]

1. [3, 4, 6, 3, 2] is the ResNet50 setting as we used the pretrained model to initialize the backbone. And [7,6,3,2] indicates how many blocks before c1, c2, c3, and c4.
2. I see the difference. The code is the correct version.
3. omega_only means the difference between latent space W+(omega only) and W++(omega + constant input of stylegan2).
4. Yes, this saves GPU memory a lot.
5. CelebAMask-HQ provides the labeled masks of faces in CelebA-HQ during the inference. We don't use masks in training so it is neglected.

But I print the model of HieRFE in inference.py and found the blocks are [3,4,6,3,2]. Shouldn't it be [7.6.3.2]?

[zyainfal]

It should be [3,4,6,3,2] because it is the definition of ResNet50.

For inference, please see the following codes in megafs.py (L119-L213):

def _forward_impl(self, x):
    # See note [TorchScript super()]
    x = self.conv1(x)
    x = self.bn1(x)
    x = self.relu(x)
    x = self.maxpool(x)
    x = self.layer1(x)
    c1 = self.layer2(x)
    c2 = self.layer3(c1)
    c3 = self.layer4(c2)
    if not self.omega_only:
        c4 = self.layer5(c3)
        return c1, c2, c3, c4
    return c1, c2, c3

where [7,6,3,2] means [3+4, 6, 3, 2] denoted by [self.layer1+self.layer2, self.layer3, self.layer4]

[HardboiledHu]

It should be [3,4,6,3,2] because it is the definition of ResNet50.

For inference, please see the following codes in megafs.py (L119-L213):

def _forward_impl(self, x):
    # See note [TorchScript super()]
    x = self.conv1(x)
    x = self.bn1(x)
    x = self.relu(x)
    x = self.maxpool(x)
    x = self.layer1(x)
    c1 = self.layer2(x)
    c2 = self.layer3(c1)
    c3 = self.layer4(c2)
    if not self.omega_only:
        c4 = self.layer5(c3)
        return c1, c2, c3, c4
    return c1, c2, c3

where [7,6,3,2] means [3+4, 6, 3, 2] denoted by [self.layer1+self.layer2, self.layer3, self.layer4]

ok,I got it. The first Residual blocks is layer1+layer2 = 3+4 = 7 blocks.