How can I get the vector of features for each pixel using a Resnet?

how can I get vector of features (features extraction) for pixel by pixel using a Resnet 50 or a Resnet 101 (Pytorch)? My image resolution is 474x496, so I want to extract 474x496 vectors at the end.

I don’t think that’s possible without changing the model architecture, since the “standard” ResNets will decrease the spatial size of the activation maps, so that you won’t be able to directly map the features to the input pixels.
You could of course try to keep the spatial size equal by changing the conv layer setups and removing the pooling layers, but I would expect this model might not be that easy to train as the activations would be huge.