Reading through and trying to refactor code for the three MAML models I might've come across a bug in the logic - when warmup is used for HyperMAML, and gradients are computed for the fast_parameters list in _update_network_weights, it isn't updated with the gradient values in the same way it is in classic MAML's set_forward method, after each gradient step.
I'm 90% sure this isn't the intended behaviour, hope to hear your thoughts on this.
Reading through and trying to refactor code for the three MAML models I might've come across a bug in the logic - when warmup is used for HyperMAML, and gradients are computed for the
fast_parameterslist in_update_network_weights, it isn't updated with the gradient values in the same way it is in classic MAML'sset_forwardmethod, after each gradient step.I'm 90% sure this isn't the intended behaviour, hope to hear your thoughts on this.