Meta-Learning & Multimodal Learning Library for TensorFlow 2.0

metaX library is a python library with deep neural networks and datasets for meta learning and multi-view learning base on Tensorflow 2.0.

We provide...

• Deep neural networks for meta learning that can solve one-shot or few-shot problems.
• Deep neural networks for multi-view learning
• Various types of experimental datasets that can be used for experiments using the provided model
• Load codes for provided dataset with few-shot learning settings

• metaX.dataset
• metaX.model

dataset/
	data_generator.py (Omniglot, mini-ImageNet) (Completed)
        KTS_data_generator.py                       (Completed)
	FLOWER_data_generator.py                    (In progress)
	KMSCOCO_data_generator.py                   (In progress)
	KVQA_data_generator.py                      (In progress)
	CropDisease.py                              (Completed)
	EuroSAT.py                                  (Completed)
	ISIC.py                                     (Completed)
 	ChestX.py                                   (Completed)
  data/
  raw_data/
  
model/
	LearningType.py 
	metric_based/
		Relation_network.py                 (In progress)
		Prototypical_network.py             (In progress)
		Siamese_network.py                  (Completed)
	model_based/
		MANN.py                             (Completed)
		SNAIL.py
	optimization_based/
		MAML.py                             (Completed)
		MetaSGD.py
		Reptile.py                          (In progress)
	heterogeneous_data_analysis/
		image_text_embeding.py              (In progress)
		Vis_LSTM.py                         (In progress)
                Modified_mCNN.py                    (In progress)
		
train.py
utils.py (accuracy, mse)

• Website
• Documentation
• Tutorials
• Examples
• GitHub

pip install metax

Results Among various optimization based meta-learning algorithms for few-shot learning, MAML(model-agnostic meta-learning) has been highly popular due to its great performance on several benchmaks. This idea is to establish a meta-learner that seeks an initialization useful for fast learning of different tasks, then adapt to specific tasks quickly and efficiently.

Usage

from metaX.model.optimization_based.MAML import ModelAgnosticMetaLearning
from metaX.model.optimization_based.MAML import OmniglotModel
from metaX.datasets import OmniglotDatabase

# 1. Preprocess the Dataset
database = OmniglotDatabase(
    raw_data_address="dataset\raw_data\omniglot",
    random_seed=47,
    num_train_classes=1200,
    num_val_classes=100)

# 2. Create the learner model
network_cls=OmniglotModel

# 3. Wrap the meta-learning method(MAML) on the learner model and dataset
maml = ModelAgnosticMetaLearning(args, database, network_cls)

# 4. Meta-Train
maml.meta_train(epochs = args.epochs)

# 5. Meta-Test
maml.meta_test(iterations = args.iterations)

# 6. Load the trained model
maml.load_model(epochs = args.epochs)

# 7. Predict with support set
print(maml.predict_with_support(meta_test_path='/dataset/data/omniglot/test'))

• This library was supported by Next-Generation Information Computing Development Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT (NRF-2017M3C4A7083279).

• Department of Computer Engineering, Dongguk University.
• Department of Computer Engineering, Duksung Women's University.
• Department of Statistics, Chungang University.
• Department of Statistics, Dongguk University.
• BI MATRIX CO., LTD.