fyi - custom resnet50 - reference code fc / maxpool

[johndpope]

https://github.com/search?q=repo%3AAlessandroRuzzi%2FGazeNeRF%20resnet50&type=code

https://github.com/AlessandroRuzzi/GazeNeRF/blob/eb0a5f1f625bcb40809139628e4bcf31a9b30e55/gaze_estimation/gaze_estimator_resnet.py#L11

this nn got a bit butchered.. layer4 ripped out - probably completely wrong.

the code above could be better approach.

# we need custom resnet blocks - so use the ResNet50  es.shape: torch.Size([1, 512, 1, 1])
# n.b. emoportraits reduced this from 512 -> 128 dim - these are feature maps / identity fingerprint of image 
class CustomResNet50(nn.Module):
    def __init__(self, *args, **kwargs):
        super().__init__()
        resnet = models.resnet50(*args, **kwargs)
        self.conv1 = resnet.conv1
        self.bn1 = resnet.bn1
      #  self.relu = resnet.relu
        self.maxpool = resnet.maxpool
        self.layer1 = resnet.layer1
        self.layer2 = resnet.layer2
        self.layer3 = resnet.layer3
        # Remove the last residual block (layer4)
        # self.layer4 = resnet.layer4

        # Add an adaptive average pooling layer
        self.adaptive_avg_pool = nn.AdaptiveAvgPool2d(FEATURE_SIZE_AVG_POOL)

        # Add a 1x1 convolutional layer to reduce the number of channels to 512
        self.conv_reduce = nn.Conv2d(1024, 512, kernel_size=1)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = F.relu(x)
        x = self.maxpool(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        # Remove the forward pass through layer4
        # x = self.layer4(x)

        # Apply adaptive average pooling
        x = self.adaptive_avg_pool(x)

        # Apply the 1x1 convolutional layer to reduce the number of channels
        x = self.conv_reduce(x)

        return x

[johndpope]

leave this for now.