Utilizing resnet_50.pth for 3D Feature Map Extraction

[aeinkoupaei]

Utilizing resnet_50.pth for 3D Feature Map Extraction

[Ram2314]

Hi! @aeinkoupaei Did you figure out if this is the correct approach to get the feature map from a image? Also needing to use this make a feature map. Is there any reason you made num_seg_classes=1?

[aeinkoupaei]

Hi @Ram2314,

To use a pre-trained ResNet model for extracting 3D feature maps, you'll need to focus on the ResNet class within the resnet.py file. Here's what to change:
1- Modifying the init function of ResNet class: Delete the entire self.conv_seg block. This removes the unnecessary layers for our feature extraction task.
2- Modifying the forward function of ResNet class: Delete this line: x = self.conv_seg(x). This ensures the model doesn't perform the final segmentation prediction, but instead outputs the feature map before that stage.

Here's an example of how to use a pre-trained ResNet-10 model for feature extraction:

resnet_10 = resnet10(shortcut_type='B', no_cuda=True)
pretrain = torch.load("resnet_10_23dataset.pth")
pretrained_dict = pretrain['state_dict']
new_state_dict = OrderedDict()
for k, v in pretrained_dict.items():
name = k[7:]
new_state_dict[name] = v

resnet_10.load_state_dict(new_state_dict)

[Ram2314]

@aeinkoupaei Awesome thanks! Also did you figure out the length, width, height issue? Seems I can set it to anything and it will work?