Selecting/discarding specific blocks from resnet model

[Min-Je_Choi]

Hi everyone,
I’m currently trying to reproduce the results from “Residual networks behave like ensembles of relatively shallow networks” (https://arxiv.org/abs/1605.06431) with pytorch, and was wondering if there is any way to select specific layers (e.g. inputs / outputs of the 40th residual block) within a network. The closest to an answer that I can think of is either using loss.creator.previous_functions[0][0] in a chained manner, or manually listing the hundreds of residual layers when building a model and assigning a return function to each of them. Any advice out of pytorch would also be greatly appreciated. Thanks

[jhjungCode (JongHwa Jung)]

I think this is a answer of your question

How to extract features of an image from a trained model

Hi all, I try examples/imagenet of pytorch. It is awesome and easy to train, but I wonder how can I forward an image and get the feature extraction result? After I train with examples/imagenet/main.py, I get model as, model_best.pth.tar And I load this file with model = torch.load('model_best.pth.tar') which gives me a dict. How can I use forward method to get a feature (like fc7 layer’s output of Alexnet) of an image with this model?

[Min-Je_Choi]

Thanks for the link. This seems to be a good starting point. I also learned that one can easily reach the weight values using
model.state_dict().keys()
which lists the keys of all weights and biases belonging to our model, then access their values with
model.state_dict()['key_name_of_weight']

[apaszke (Adam Paszke)]

You might find .items() or .values() useful too.