How to train your MAML in Pytorch

A replication of the paper "How to train your MAML", along with a replication of the original "Model Agnostic Meta Learning" (MAML) paper.

Introduction

Welcome to the code repository of How to train your MAML. This repository includes code for training both MAML and MAML++ models, as well as data providers and the datasets for both. By using this codebase you agree to the terms and conditions in the LICENSE file. If you choose to use the Mini-Imagenet dataset, you must abide by the terms and conditions in the ImageNet LICENSE

Installation

The code uses Pytorch to run, along with many other smaller packages. To take care of everything at once, we recommend using the conda package management library. More specifically, miniconda3, as it is lightweight and fast to install. If you have an existing miniconda3 installation please start at step 3. If you want to install both conda and the required packages, please run:

wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
Go through the installation.
Activate conda
conda create -n meta_learning_pytorch_env python=3.6.
conda activate meta_learning_pytorch_env
At this stage you need to choose which version of pytorch you need by visiting here
Choose and install the pytorch variant of your choice using the conda commands.
Then run bash install.sh

To execute an installation script simply run: bash <installation_file_name>

To activate your conda installations simply run: conda activate

Datasets

We provide the omniglot dataset in the datasets folder directly in this repo. However, due to Mini-ImageNet being substantially larger than github's limit, we chose to upload it on gdrive using pbzip (parallel zip) compression, which is one of the fastest compressions available as we are writing this (really helps if you are zipping something as big as today's large scale datasets). We have automated the unzipping and usage of the dataset, all one needs to do is download it from our mini_imagenet gdrive folder. Once download, please place it in the datasets folder in this repo. The rest will be done automagically when you run your mini-imagenet experiment.

Note: By downloading and using the mini-imagenet datasets, you accept terms and conditions found in imagenet_license.md

Overview of code:

datasets folder: Contains the dataset pbzip files and folders containing the images in a structure readable by the custom data provider.
experiments_config: Contains configuration files for each and every experiment listed in the experiments script folder. All of the scripts are automatically generated using the script_generation_tools/generate_configs.py script.
experiment_scripts: Contains scripts that can reproduce every results in the paper. Each script is easily runnable simply by executing: bash <experiment-script.sh>
experiment_template_config: Contains the template configuration files. These files have variables declared in their files indicated as $variable_name$, which are then filled automatically by the generate_configs.py script. This way one can vary various hyperparameters automatically.
script_generation_tools: Contains scripts and template files for the automatic generation of experiment scripts.
utils: Contains utilities for dataset extraction, parser argument extraction and storage of statistics and others.
data.py: Contains the data providers for the few shot meta learning task generation. The data provider is agnostic to dataset, which means it can be used with any dataset. Most importantly, it can only scan and use datasets when they are presented in a specific format. The two formats that the data provider can read are:

A folder structure where the top level folders are the classes and the contained images of each folders, the images of that class, as illustrated below:

Dataset
    ||______
    |       |
 class_0 class_1 ... class_N
    |       |___________________
    |                           |
samples for class_0    samples for class_1

In this case the data provider will split the data into 3 sets, train, val and test using the train_val_test_split variable found in the experiment_config files. However, in the case where you have a pre-split dataset, such as mini_imagenet, you can instead use: 2. A folder structure where the higher level folders indicate the set (i.e. train, val, test), the mid level folders (i.e. the folders within a particular set) indicate the class and the images within each class indicate the images of that class.

Dataset
    ||
 ___||_________
|       |     |
Train   Val  Test
|_________________________
    |       |            |
 class_0 class_1 ... class_N
    |       |___________________
    |                           |
samples for class_0    samples for class_1

experiment_builder.py: Builds an experiment ready to train and evaluate your meta learning models. It supports automatic checkpoining and even fault-tolerant code. If your script is killed for whatever reason, you can simply rerun the script. It will find where it was before it was killed and continue onwards towards convergence!
few_shot_learning_system.py: Contains the meta_learning_system class which is where most of MAML and MAML++ are actually implemented. It takes care of inner and outer loop optimization, checkpointing, reloading and statistics generation, as well as setting the rng seeds in pytorch.
meta_neural_network_architectures: Contains new pytorch layers which are capable of utilizing either internal parameter or externally passed parameters. This is very useful in a meta-learning setting where inner-loop update steps are applied on the internal parameters. By allowing layers to receive weight which they will only use for the current inference phase, one can easily build various meta-learning models, which require inner_loop optimization without having to reload the internal parameters at every step. Essentially at the technical level, the meta-layers forward prop looks like:

def forward(x, weights=None):

    if weights is not None:
        out = layer(x, weights)
    else:
        out = layer(x, self.parameters)
    return out

If we pass weights to it, then the layer/model will use those to do inference, otherwise it will use its internal parameters. Doing so allows a model like MAML to be build very easily. At the first step, use weights=None and for any subsequent step just pass the new inner loop/dynamic weights to the network.

train_maml_system.py: A very minimal script that combines the data provider with a meta learning system and sends them to the experiment builder to run an experiment. Also takes care of automated extraction of data if they are not available in a folder structure.

Running an experiment

To run an experiment from the paper on Omniglot:

Activate your conda environment conda activate pytorch_meta_learning_env
cd experiment_scripts
Find which experiment you want to run.
bash experiment_script.sh

Note: By downloading and using the mini-imagenet datasets, you accept terms and conditions found in imagenet_license.md

To run an experiment from the paper on Mini-Imagenet:

Activate your conda environment conda activate pytorch_meta_learning_env
Download the mini_imagenet dataset from the mini_imagenet gdrive folder
copy the .pbzip file in datasets
cd experiment_scripts
Find which experiment you want to run.
bash experiment_script.sh

To run a custom/new experiment on any dataset:

Activate your conda environment conda activate pytorch_meta_learning_env
Make sure your data is in datasets/ in a folder structure the data provider can read.
cd experiment_template_config
Find an experiment close to what you want to do and open its config file.
For example let's take an omniglot experiment on maml++. Make changes to the hyperparameters such that your experiment takes form. Note that the variables in $$ are hyperparameters automatically filled by the config generation script. If you add any new of those, you'll have to change the generate_configs.py file in order to tell it what to fill those with.

{
  "batch_size":16,
  "image_height":28,
  "image_width":28,
  "image_channels":1,
  "gpu_to_use":0,
  "num_dataprovider_workers":8,
  "max_models_to_save":5,
  "dataset_name":"omniglot_dataset",
  "dataset_path":"omniglot_dataset",
  "reset_stored_paths":false,
  "experiment_name":"MAML++_Omniglot_$num_classes$_way_$samples_per_class$_shot_$train_update_steps$_filter_multi_step_loss_with_max_pooling_seed_$train_seed$",
  "train_seed": $train_seed$, "val_seed": $val_seed$,
  "train_val_test_split": [0.70918052988, 0.03080714725, 0.2606284658],
  "indexes_of_folders_indicating_class": [-3, -2],
  "sets_are_pre_split": false,

  "total_epochs": 150,
  "total_iter_per_epoch":500, "continue_from_epoch": -2,

  "max_pooling": true,
  "per_step_bn_statistics": true,
  "learnable_batch_norm_momentum": false,

  "dropout_rate_value":0.0,
  "min_learning_rate":0.00001,
  "meta_learning_rate":0.001,   "total_epochs_before_pause": 150,
  "task_learning_rate":-1,
  "init_task_learning_rate":0.4,
  "first_order_to_second_order_epoch":80,

  "norm_layer":"batch_norm",
  "cnn_num_filters":64,
  "num_stages":4,
  "number_of_training_steps_per_iter":$train_update_steps$,
  "number_of_evaluation_steps_per_iter":$val_update_steps$,
  "cnn_blocks_per_stage":1,
  "num_classes_per_set":$num_classes$,
  "num_samples_per_class":$samples_per_class$,
  "num_target_samples": $target_samples_per_class$,

  "second_order": true,
  "use_multi_step_loss_optimization":true,
  "use_gdrive":false
}

cd script_generation_tools
python generate_configs.py; python generate_scripts.py
Your new scripts can be found in the experiment_scripts, ready to be run.

Acknowledgments

Thanks to the University of Edinburgh and EPSRC research council for funding this research.