Reproducing results of the paper with iMAML

[zaccharieramzi]

Reproducing results of the paper with iMAML

[zaccharieramzi]

Ok makes total sense!
Problem is that even with the original implementation I cannot retrieve the numbers stated in the paper.

Will let you know if I get to anything with my own implementation.

[prolearner]

Hi @zaccharieramzi,
thanks a lot for your interest in the code :).

I did not try to reproduce the exact numbers in the iMAML paper, but just to implement the main components of the experimental setup. The purpose of the iMAML example is just to showcase what it can be achieved with the hypergradient approximation methods implemented here and possibly be used as a starting point for further research or to apply the method to different data.

As you noticed there are some differences in the implementation which happened mostly by chance. You can try to modify the implementation to match the one described by the original paper and code to see if this improves accuracy. I'd try first with larger batch size and using the same network (with correct placement of batch norm), possibly tweaking the learning rate. I don't think initialising the final linear layer weights randomly instead of to zero would make much difference.

The approximation I'm referring to here is the hypergradient approximation that they use, which relies on the conjugate gradient method to solve the linear system.