AM3-MAML
MAML with the initialization induced by word embeddings.
Our code is based on https://github.com/sungyubkim/GBML
- [AM3-MAML]
python3 main.py --download False
Results on miniImagenet
- Without pre-trained encoder (Use 64 channels by default. The exceptions are in parentheses)
5way 1shot | 5way 1shot (ours) | 5way 5shot | 5way 5shot (ours) | |
---|---|---|---|---|
MAML | - | - | 63.11 (64) | - |
AM3-MAML | - | - | - | 66.41 (64) |
How to run AM3-MAML
- Download Common Crawl (840B tokens, 2.2M vocab, cased, 300d vectors, 2.03 GB download) in Glove github.
- Extract files and run "mml/preprocess.py" to get word embeddings from pretrained Glove.
- Install Higher and Torchmeta (links in dependencies).
- Set "--download True" if you need to download miniimagenet and run the command above it.
Related work
[1] Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
[2] Adaptive Cross-Modal Few-Shot Learning
[3] GloVe: Global Vectors for Word Representation
[4] Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples
[5] Set Transformer: A Framework for Attention-based Permutation-Invariant Neural Networks
Reference for codes
[1] GBML for the base
[2] Torchmeta for the dataset
[3] glove_pretrain for the pretrained glove embeddings
[4] SetTransformer for SetTransformer modules
Dependencies
Acknowledgement
This work was supported by Institute of Information & Communications Technology Planning & Evaluation(IITP) grant funded by the Korea government (MSIT) (No.2019-0-01371, Development of brain-inspired AI with human-like intelligence)