MAML-Tensorflow

Tensorflow r2.1 reimplementation of Model-Agnostic Meta-Learning from this paper:

Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks

Reinforcement Learning part is adapted from MoritzTaylor/maml-rl-tf2 and tristandeleu/pytorch-maml-rl Sincerely Thanks for their fantastic works!!!

Project Requirements

1. python 3.x
2. Tensorflow r2.1
3. numpy
4. matplotlib
5. ...
6. All Scripts in image_classification folder is tested on python 3.6.5

MiniImagenet Dataset

I wrote a task generator which samples randomly from the whole dataset to set up a train batch during every training steps so it won't consume too much GPU memory.

For 5-way 1-shot tasks on the MiniImagenet, it takes at round 1.6 s to run one training steps on the GTX1070 and for each task update fast_weights 1 time, and it allocate 1.3GB GPU memory.

For 5-way 5-shot tasks on MiniImagenet, it takes at around 2.2 s to run one training steps.

If you set the --update_steps > 1, it will take more time for one training step.

1. Download the MiniImagenet, and the split files train.csv, test.csv, val.csv from here

2. Put the MiniImagenet dataset to your project folder like this

   dataset/miniimagenet
   |-- images
   	|--- n0153282900000005.jpg
   	|--- n0153282900000006.jpg
   	...
   |-- test.csv
   |-- train.csv
   |-- val.csv
   

3. Run the python script to resize and split the whole dataset

   cd scripts/image_classification
   python image_preprocess.py --dataset=miniimagenet
   

4. Modify the path to dataset in scripts/image_classification/task_generator.py

   if self.dataset == 'miniimagenet':
       ...
       META_TRAIN_DIR = '../../dataset/miniImagenet/train'
       META_VAL_DIR = '../../dataset/miniImagenet/test'
       ...
   
   if self.dataset == 'omniglot':
       ...
           DATA_FOLDER = '../../dataset/omniglot'
           ... 

5. Run the main python script

   cd scripts/image_classification
   # For 5-way 1-shot on miniimagenet
   python main.py --dataset=miniimagenet --mode=train --n_way=5 --k_shot=1 --k_query=15
   # For 5-way 5-shot on miniimagenet
   python main.py --dataset=miniimagenet --mode=train --n_way=5 --k_shot=5 --k_query=15
   

Omniglot Dataset

For Omniglot dataset, it will consume fewer computing resource and time

For 5-way 1-shot, 0.3 s for one training step

For 20-way 1-shot, 0.7 s for one training step

1. Download Omniglot dataset from [here](git clone git@github.com:brendenlake/omniglot.git) and extract the contents of python/images_background.zip and python/images_evaluation.zip to the dataset/omniglot it will looks like this:

   dataset/omniglot
   |-- Alphabet_of_the_Magi
   |-- Angelic
   ...
   

2. Run the python script to resize the images

   cd scripts/image_classification
   python image_preprocess.py --dataset=omniglot
   

3. Run the main python script

   cd scripts/image_classification
   # For 5-way 1-shot on Omniglot
   python main.py --dataset=omniglot --mode=train --n_way=5 --k_shot=1 --k_query=1 --inner_lr=0.1
   # For 20-way 1-shot on Omniglot
   python main.py --dataset=omniglot --mode=train --n_way=20 --k_shot=1 --k_query=1 --inner_lr=0.1
   

References

This project is, for the most part, a reproduction of the original implementation cbfinn/maml_rl in TensorFlow 2. The experiments are based on the paper

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]

If you want to cite this paper

@article{DBLP:journals/corr/FinnAL17,
  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}
}