Running feature extraction seperate from classification

I want to know if there is a method of running feature extraction of resnet101 separately and then running classification separately

I assume you are referring to the conv blocks as feature extractors and the linear layers as the classifier.
If so, you could create a custom nn.Module and rewrite the forward method.
E.g. you could wrap the feature extraction part and the classifier into separate modules and could use them in a standalone manner. (The best approach depends on your actual use case and how you would like to use it)