Working implementation of MAML (meta-learning) for FluxML?

Specific Domains Machine Learning
,
1

Hi everyone,

I was wondering if there are any public implementations of meta-learning techniques such as MAML ([1703.03400] Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks) in Flux. The model zoo has a MAML implementation (https://github.com/FluxML/model-zoo/blob/master/contrib/meta-learning/MetaLearning.jl); however, it uses deprecated Flux functionality such as Flux.Optimise._update_params! and Flux.Tracker.gradient. I’m fairly new to Flux, so it would be neat to see what a proper and current implementation would look like

Thanks in advance!

1 Like
2

I did this implementation awhile ago. It works w/ Flux + Zygote. It was a quick and dirty port from the version found in the model zoo. I haven’t had time to write a pull request for that yet though, maybe this week.

There is almost definitely more efficient/better ways to implement this.