Meta-Learner LSTM for few Shot Learning

Authors: Stanislas Furrer & Yiğit Efe Erginbaş & Mert Kayaalp

Date : 01.06.2020

Ravi & Larochelle have addressed the weakness of neural networks trained with gradient-based optimization on the few-shot learning problem with an LSTM-based meta-learner. Our paper expands the performance analysis of their algorithm.

TensorFlow implementation of Optimization as a Model for Few-shot Learning together with extended experiments.

Our Paper : Meta-Learner LSTM for few Shot Learning

Getting Started

Requirements

• Python3
• Python packages in requirements.txt

Data preparation

• Mini-Imagenet data
  • You can download it from here (~2.7GB, google drive link)

The data should placed and split properly as follows:

- data/
  - miniImagenet/
    - train/
    - val/
    - test/
- main.py
- ...

• Check out the configuration in main.py to make sure --data-root is properly set

Running with Default Settings

Training a Model

main.py is configured to train a default model on mini-ImageNet data set using a random seed.

For training with default settings run:

python main.py

Testing a Pre-Trained Model

This project also includes pre-trained models that can be used for testing.

For testing a model that has been trained on miniImagenet using the default settings:

python main.py --mode test --seed 100 --ID 1

Changing the hyper-parameters

In order to change the hyper-parameters you can pass the following command line arguments together with desired values.

• --n-shot : How many examples per class for episode-training (k)
• --n-eval : How many examples per class for episode-evaluation
• --n-class : How many classes (N)
• --hidden-size : Hidden size for the meta-learner
• --lr : Learning rate
• --episode : Episodes for meta-train
• --episode-test : Episodes for meta-test
• --epoch : Epoch to train for an episode
• --meta-learner : Type of the meta-learner to be used (choices=['lstm', 'full-lstm', 'gru', 'sgd'])

This is not an exhaustive list of the all possible arguments. For further information, please check main.py.

Testing with Other Pre-trained Models

The pre-trained model library also includes models trained with different hyper-parameters.

For example, in order to test the model trained with "hidden-size = 1", you can run

python main.py --mode test --seed 100 --ID 0 --hidden-size 1

You can check models.md for a list of all pre-trained models together with their arguments.

You can check run.ipynb for a notebook consisting of the experiments in the paper.

Other data sets

Directory for other data sets should be also set up in the same way as miniImagenet.

In order to train/test with Omniglot/CUB data, you can run main_omniglot.py or main_CUB.py in the same way descibed for main.py.

Cross-Domain Scenario

For these tests you do not require the Omniglot/CUB data.

For testing a model trained with CUB data on miniImagenet data set:

python main.py --mode test --seed 75 --ID 4 --hidden-size 1

For testing a model trained with Omniglot data on miniImagenet data set:

python main.py --mode test --seed 150 --ID 1 --hidden-size 1