Awesome

Learning to Learn with Smooth Regularization

This is our PyTorch implementation of learned RNN optimizers with regularzation to enforce smoothness. Experiments are conducted on one GeForce GTX 1080Ti.

Prerequisites

You can find all prerequisites in requirements.txt and install them by

pip install -r requirements.txt

Data

• MNIST and CIFAR-10 can be downloaded automatically by torchvision.
• Mini-Imagenet as described here
  • You can download it from here (~2.7GB, google drive link)

Run

Optimization of neural networks

Codes for this task can be found in the folder optimization. Default values for both lambda and epsilon are 1.0. You can train the RNN optimizer with your own lambda and epsilon by

python main.py --lamb [your lambda] --eps [your epsilon]

For more details about hyperparameters, you can check main.py. You can tune these hyperparemeters to obtain better performance.

In addition, to evaluate the learned optimizer, you can run

python test.py --batch_size [your batch size] --optimizer_steps [your optimizer steps]

Few-Shot Learning

Codes for few-shot learning can be found in the folder few-shot-learning. Here are details for running experiments.

For 5-shot, 5-way training, run

bash scripts/train_5s_5c.sh

Hyper-parameters are referred to this repo.

For 5-shot, 5-way evaluation, run (remember to change --resume and --seed arguments)

bash scripts/eval_5s_5c.sh

In addition, for 1-shot, 5-way training and evaluation, run respectively

bash scripts/train_1s_5c.sh

and

bash scripts/eval_1s_5c.sh