Hi SimpleChains team!
I’m currently using SimpleChains and I want to implement my own train function based on the available template in Flux:
for d in training_set
# Our super logic
gs = gradient(params(m)) do
l = loss(d...)
end
update!(opt, params(m), gs)
end
What I want to do, using SimpleChains, is to incorporate a simple logic that allows me to discard certain examples during training. I want to avoid computing the gradient for these examples. Is there a specific method to do this?
I have already implemented my own logic by modifying the code in the train_unbatched function, but I would like to achieve the same result using train_batched. However, train_batched appears to be more complex, and I’m finding it challenging to make the necessary modifications.
Additionally, I’ve observed that the performance (in terms of loss and accuracy) of train_batched is significantly better than train_unbatched. If it’s not possible to create an explicit train function, is there a way to improve train_unbatched so that it is comparable to train_batched?
Any guidance or suggestions would be greatly appreciated!
Thank you 