Modified PyTorch implementation of Model-Agnostic Meta-Learning (MAML) from this repo.
- Re-write meta (outer) learner to be consistent with the Pytorch nn.Module
- Support all three datasets in the original paper
- Minor bug fixes
- python: 3.x
- Pytorch: 1.0+
10-shot sinusoid:
python main.py --data_source sinusoid --task_num 25 --class_num 5 --train_sample_size_per_class 10 --inner_lr 0.001 --inner_step 15-way, 1-shot omniglot:
python main.py --datasource omniglot --task_num 32 --class_num 5 --train_sample_size_per_class 1 --inner_lr 0.4 --inner_step 1 --img_size [28, 28] --data_folder path/to/omniglot20-way, 1-shot omniglot:
python main.py --datasource omniglot --task_num 16 --class_num 20 --train_sample_size_per_class 1 --inner_lr 0.1 --inner_step 5 --img_size [28, 28] --data_folder path/to/omniglotFirst download MiniImagenet dataset from here and the label files train/val/test.csv from here. Extract it like:
miniimagenet/
├── images
├── n0210891500001298.jpg
├── n0287152500001298.jpg
...
├── test.csv
├── val.csv
└── train.csv5-way 1-shot mini imagenet:
python main.py --datasource miniimagenet --epoch 60000 --task_num 4 --class_num 5 --update_batch_size 1 --inner_lr 0.01 --inner_step 5 --img_size [84, 84] --data_folder path/to/miniimagenet5-way 5-shot mini imagenet:
python main.py --datasource miniimagenet --epoch 60000 --task_num 4 --class_num 5 --update_batch_size 5 --inner_lr 0.01 --inner_step 5 --img_size [84, 84] --data_folder path/to/miniimagenet