This is an experimental code base

The biggest difference between MAML and pre-training weights：Pre-training weights minimize only for original task loss. MAML can minimize all task loss with a few steps of training.

1. Pull repository.

git clone https://github.com/verages/MAML.git

2. You need to install some dependency package.

cd MAML
pip installl -r requirements.txt

3. Download the Omiglot dataset and maml weights.

wget https://github.com/verages/MAML/releases/download/v0.1/Omniglot.tar
wget https://github.com/verages/MAML/releases/download/v0.1/maml.h5
tar -xvf Omniglot.tar

4. Run evaluate.py, you'll see the difference between MAML and random initialization weights.

python evaluate.py

1. You should set same parameters in config.py.

n_way = "number of classes"
k_shot = "number of support set"
q_query = "number of query set"

2. Start training.

python train.py

3. Running tensorboard to monitor the training process.

tensorboard --logdir=./summary

• Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
• cbfinn/maml
• dragen1860/MAML-Pytorch