toologicbv / meta_learner Goto Github PK

Fix Line 112 in rnn_optimizer.py from self.cuda to self.use_cuda

Not entirely sure about this but it seems to me as if you are computing the nll twice for every set of optimizee parameters:
First, you compute
loss = reg_funcs.compute_neg_ll(average_over_funcs=False, size_average=False)
around line 266 in train_optimizer.py
Then you compute the parameter delta with the meta optimizer and based on the updated optimizee parameters you compute the new nll:
loss_step = meta_optimizer.step_loss(reg_funcs, par_new, average_batch=True)
around line 292 in train_optimizer.py.

Then, in the next iteration you compute the loss
loss = reg_funcs.compute_neg_ll(average_over_funcs=False, size_average=False) again, which is, as far as I understand, the same as what happened in line 292 in the previous iteration.

It seems to me you could gain efficiency here if I understand correctly.

I'm not 100% clear about pytorch syntax. Should the two following ways to compute the gradients df/dtheta be equivalent? Why are they not? :) I'm not entirely sure what loss.backward(backward_ones) does. Is this df/d1 ?

loss.mean().backward(retain_variables=True)
print(reg_funcs.params.grad.data)
reg_funcs.params.data.zero_()
loss.backward(backward_ones)
print(reg_funcs.params.grad.data)