We explain some expandable options in our released code:
- --num-updates-for-gradient: It is set to 1 in all our experiments, but other value can also be applied.
- --meta-sgd: It is proposed in: https://arxiv.org/abs/1707.09835. We set this option to False in all our experiments.
- --second-order-grad: The first-order approximation and the full second-order implementation are proposed in: https://arxiv.org/abs/1703.03400, which shows both the two implementations give similar results.
- --first-meta-update: It is explicitly proposed in: https://arxiv.org/abs/1710.03463 and it can be viewed as a supplement to the original meta-learning loss. We set it to False in our experiments.
We provide those options for further extention.