How to get separate conv feature map from pretrained ResNet?

1

PyTorch supplied pretrained ResNet with different depths. The ResNet has the “BasicBlock” or “Bottleneck” structure. I want to ask that how to get separate conv feature maps from these pretrained ResNet (for example, restnet18 or resnet50)?

2

your question is not clear at all. What do you mean by “get separate conv feature maps”.
Also, there are a few posts explaining how to extract intermediate outputs of resnet, please read those.

3

What I mean is “Extracting image features” from different layers in ResNet, and I read this post.
I just wonder how this can be done in ResNet. One possible way is to construct a partial model which only use the front layers for forwarding. Are there other ways?
By the way, the Torch version of ResNet has a nice script
to extract features from ResNet. If PyTorch also has a corresponding one will be better.

4

In your case, attaching a hook to the layer you want seems easier.

model = models.resnet18()

outputs = []
def hook(module, input, output):
    outputs.append(output)

model.layer2.conv1.register_forward_hook (hook)

(Typing from the phone, the code is an approximation and might contain errors)

6 Likes
5

Why do we have output as one of the inputs in the hook?

6

Here is what i do:

class Resnet50Extractor(nn.Module):
    def __init__(self, submodule, extracted_layer):
        super(Resnet50Extractor, self).__init__()
        self.submodule = submodule
        self.extracted_layer = extracted_layer

    def forward(self, x):
        if self.extracted_layer == 'maxpool':                       
            modules = list(self.submodule.children())[:4]
        elif self.extracted_layer == 'inner-layer-3':                     
            modules = list(self.submodule.children())[:6]
            third_module = list(self.submodule.children())[6]
            third_module_modules = list(third_module.children())[:3]    # take the first three inner modules
            third_module = nn.Sequential(*third_module_modules)
            modules.append(third_module)
        elif self.extracted_layer == 'layer-3':                     
            modules = list(self.submodule.children())[:7]
        else:                                               # after avg-pool
            modules = list(self.submodule.children())[:9]

        self.submodule = nn.Sequential(*modules)
        x = self.submodule(x)
        return x

then call like below:

model_ft = models.resnet50(pretrained=True)
extractor = Resnet50Extractor(model_ft, extracted_layer)
features = extractor(input_tensor)
2 Likes
7

is there any pros or cons between using your method comparing to register_forward_hook