The original code for the paper "How to train your MAML" along with a replication of the original "Model Agnostic Meta Learning" (MAML) paper in Pytorch.
Per your paper and code you use a cosine-annealed learning rate in the outer loop, but it is unclear if you also use warm restarts, as described in the SGDR paper which introduced this schedule.
Have you experimented with warm restarts at all, or am I just missing something?
Hi Antreas, thanks for providing this code!
Per your paper and code you use a cosine-annealed learning rate in the outer loop, but it is unclear if you also use warm restarts, as described in the SGDR paper which introduced this schedule.
Have you experimented with warm restarts at all, or am I just missing something?