How to train the synthetic NN without apply its outputs to the base NN?

[parap1uie-s]

As the issue title.

My base module is a 3-layers GRU, and the synthetic module is another RNN.

I want to training base module in BPTT mode without synthetic gradients, for a few starting epochs.

From my understand, the make_trigger() function makes a trigger, so the synthetic module can be trained from true gradients. And backward_interface.backward() function apply the synthetic gradients to final hidden states and base module, at same time.

So I tried to forward the base module with make_trigger() and without interface.backward(), but the gradients of base module is still different from the model without make_trigger().

Any help?

Thanks.