Question about the inner gradient step

[jingweiz]

Question about the inner gradient step

[katerakelly]

Hi Jingwei, The lines you reference are doing the meta-evaluation of the network. Before fine-tuning for each test task, the parameters of the test network are reset and the optimizer is re-created, in these lines <https://github.com/katerakelly/pytorch-maml/blob/master/src/maml.py#L100-L102> . As for the meta-training, you are correct that a batch of meta-updates is collected by fine-tuning independently on separate tasks, all starting from the same parameters, see here <https://github.com/katerakelly/pytorch-maml/blob/master/src/maml.py#L159-L164> . Let me know if you still see a problem!

…

On Sun, Feb 25, 2018 at 11:00 AM, Jingwei Zhang ***@***.***> wrote: Hi, Nice repo! One question about src/maml.py Line 107~111: as far as I understand, when doing the inner gradient steps, maml uses the same /theta for computing the loss for all the training demos, whereas in your implementation, the parameters of test_net is updated after every training demo, then the loss for the next training demo is computed using the updated parameters. Am I misunderstanding sth here? Thanks a lot in advance! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#3>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADplVuNA-IaMuuiVPJCL0xgJ_ggEUEg4ks5tYa3EgaJpZM4SSaZE> .

-- Kate Rakelly UC Berkeley EECS PhD Student rakelly@eecs.berkeley.edu

[jingweiz]

Hey,
yes I found out I was reading it wrong so I deleted my question lol, but still thank you so much for the explanation:)