jik0730 / maml-in-pytorch Goto Github PK

Can you provide instructions on how to organize the datasets such that it can be used in your code?

hey I am trying to run the model on MiniImageNet, do I just have to point the code the location of downloaded dataset ? is any preprocessing required ?

Thanks for your implementation.
I notice that in function "train_single_task", you manually update parameters of the model with SGD formula, but then there is a loop which executes nearly the same operation to update the parameters, therefore parameters are actually updated (num_train_updates + 1) times. Does this additional operation (out of the loop) have any special meaning? Can we put all the update operations into one loop so that update number is equal to num_train_updates?
Besides, I don't understand your annotation "we manually compute adpated parameters since optimizer.step() operates in-place", could you please explain it?

Thanks for your excellent reproduction. I wonder if you could provide a miniImageNet experiment params json file. Thank you very much.