Awesome
MAML in pytorch
Performances are reported as best test accuracy. Converged test accuracy might be smaller than or similar to the reported performances.
Comparison to original MAML implementation for 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 | 99.4% | 99.9% | 92.8% | - |
Comparison to original MAML implementation for miniImageNet
| 5-way 1-shot | 5-way 5-shot | |
|---|---|---|
| MAML | 48.7% | 63.1% |
| Ours | 48.4% | 64.8% |
TODO
- Investigate under-performance of 20-way 1-shot omniglot experiments (This is similar behavior to that of https://github.com/katerakelly/pytorch-maml)