I want to parallelize the inner loop of MAML.
Each inner loop of the MAML will produce individual loss along with individual gradient graphs,
and after the iteration, I have to aggregate the losses followed by backpropagation.
My naive idea is replacing the loop to map.
To do this, I guess I need to aggregate the loss from multiple threads.
Is it possible to aggregate graphs from worker threads and then do the backprop at once?