search

dragen1860 / MAML-Pytorch

Elegant PyTorch implementation of paper Model-Agnostic Meta-Learning (MAML)
MIT License
2.29k stars 420 forks source link

question about Hessian-vector products #18

Open shawnkx opened 5 years ago

shawnkx commented 5 years ago

In the original paper, the authors claimed that MAML needs second gradient and Hessian-vector products. Could you explain how do you implement this or Pytorch just do this automatically? Thanks!

dragen1860 commented 5 years ago

@shawnkx Does the official MAML git repo. implement hessian vector product ? I forgot already.

wechto commented 5 years ago

The results in the original paper claim K-shot, which comes to a doubt that the K means the number on the stage of meta-training or meta-testing (fine-tuning) with regard to the different shot number in the code.

seanie12 commented 5 years ago

@dragen1860 In the official MAML you can choose whether to use second order deriative as follows https://github.com/cbfinn/maml/blob/master/main.py#L60 https://github.com/cbfinn/maml/blob/master/maml.py#L92