Meta-SGD-pytorch
Meta-SGD-pytorch copied to clipboard
Published
20 hours ago •
jik0730
jik0730
Neat implementation of Meta-SGD in pytorch: https://arxiv.org/abs/1707.09835
Meta-SGD in pytorch
The only difference compared to MAML is to parametrize task learning rate in vector form when meta-training. As the authors said, we could see fast convergence and higher performance than naive MAML. For our version of Meta-SGD, we did not use other tricks to improve performances such as regularization or 1-shot meta-training. Our pytorch version of MAML is at https://github.com/jik0730/MAML-in-pytorch.
Performance comparisions to MAML
The reported performance are refered to the ones in Meta-SGD paper.
| Omniglot | 5-way 1-shot | 5-way 5-shot | 20-way 1-shot | 20-way 5-shot |
|---|---|---|---|---|
| MAML | 98.7% | 99.9% | 95.8% | 98.9% |
| Ours MAML | 99.4% | 99.9% | 92.8% | - |
| Meta-SGD | 99.5% | 99.9% | 95.9% | 99.0% |
| Ours Meta-SGD | 99.3% | 99.8% | 95.4% | 97.8% |
| miniImageNet | 5-way 1-shot | 5-way 5-shot | 20-way 1-shot | 20-way 5-shot |
|---|---|---|---|---|
| MAML | 48.7% | 63.1% | 16.5% | 19.3% |
| Ours MAML | 48.4% | 64.8% | - | - |
| Meta-SGD | 50.5% | 64.0% | 17.6% | 28.9% |
| Ours Meta-SGD | 49.1% | 66.0% | 17.0% | 29.9% |
TODO
- TBD
jik0730