Meta-Learning: Learning to learn

This repo accompanies the code from our Meta-Learning report discussed in the Experiment section.

Setup environment

Step 1: Install Torchmeta from source

    git clone https://github.com/nishamuktewar/pytorch-meta.git
    conda create --name meta_env python=3.7 ipykernel
    conda activate meta_env
    cd pytorch-meta
    pip install -e .

Step 2: Install other libraries

    conda install -c anaconda scikit-learn

Repo structure

.
├── figures
│   ├── 10.png
│   ├── 11.png
│   ├── 12.png
│   └── 13.png
├── maml
│   ├── datasets.py
│   ├── __init__.py
│   ├── metalearners
│   │   ├── __init__.py
│   │   ├── maml.py
│   │   ├── meta_sgd.py
│   ├── model.py
│   └── utils.py
├── models
│   ├── 2020-09-22_164258
│   │   ├── config.json
│   │   ├── model_results.json
│   │   ├── model.th
│   │   ├── results.json
│   │   └── run.txt
.
.
.
├── notebooks
│   ├── visualize_runs_100samples.ipynb
│   └── visualize_runs_20samples.ipynb
├── README.md
├── test.py
└── train.py

Run experiments

Meta-training

• Run the following command to train a meta-learning model on the meta-training dataset.
  Note that this step will take a few hours if run for the first time. That is because it downloads the entire Quick Draw! dataset and transforms it into train/validation/test meta-datasets to be consumed during training and evaluation.

python train.py <<dataset path>> \
    --use-cuda \
    --num-training-samples 20 \
    --num-ways 5 \
    --num-shots 1 \
    --num-shots-test 1 \
    --num-steps 5 \
    --hidden-size 20 \
    --batch-size 10 \
    --num-batches 100 \
    --num-epochs 50 \
    --step-size 0.005 \
    --meta-lr 0.0005 

• The configuration, model and result files are saved in the ./models folder using the date-timestamp as the foldername

Meta-inference

• Test the meta learning model

python test.py ./models/<<date-timestamp>>/config.json

Results

5-way, 1/5/10-shot results based on 100 random sampled images

10-way, 1/5/10-shot results based on 100 random sampled images

5-way, 1/5/10-shot results based on 20 random sampled images

10-way, 1/5/10-shot results based on 20 random sampled images

References

Leveraged source code from the following repos:

Tristan Deleu, Tobias Würfl, Mandana Samiei, Joseph Paul Cohen, and Yoshua Bengio. Torchmeta: A Meta-Learning library for PyTorch, 2019 [ArXiv]

@misc{deleu2019torchmeta,
  title={{Torchmeta: A Meta-Learning library for PyTorch}},
  author={Deleu, Tristan and W\"urfl, Tobias and Samiei, Mandana and Cohen, Joseph Paul and Bengio, Yoshua},
  year={2019},
  url={https://arxiv.org/abs/1909.06576},
  note={Available at: https://github.com/tristandeleu/pytorch-meta, https://github.com/tristandeleu/pytorch-maml}
}

Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. International Conference on Machine Learning (ICML), 2017 [ArXiv]

@article{finn17maml,
  author={Chelsea Finn and Pieter Abbeel and Sergey Levine},
  title={Model-{A}gnostic {M}eta-{L}earning for {F}ast {A}daptation of {D}eep {N}etworks},
  journal={International Conference on Machine Learning (ICML)},
  year={2017},
  url={http://arxiv.org/abs/1703.03400},
  note={Available at: }
}