I am trying to retrain AlexNet, however, in order to check if I could improve accuracy, I would like to feed it a scalar information, namely before the last fully connected layer. Nonetheless, I cannot find any way to flag “input layers” using pyTorch, so I have no idea how to achieve this. Does anybody tried anything similar?
You can simply modify the
forward method of your model and pass an arbitrary number of parameters.
PyTorch does not use specific “input” tensors etc.