Committee of NAS-based models

The algorithm is based on continuous relaxation and gradient descent in the architecture space. It is able to efficiently design high-performance convolutional architectures for image classification (on CIFAR-10 and Imagenette).

Requirements

Python >= 3.5.5, PyTorch == 0.3.1, torchvision == 0.2.0

NOTE: PyTorch 0.4 is not supported at this moment and would lead to OOM.

Datasets

Instructions for acquiring imagenette can be found here. CIFAR-10 can be automatically downloaded by torchvision.

Architecture search (using small proxy models)

To carry out architecture search using 2nd-order approximation, run

cd cnn_space1 && python train_search.py      # for conv cells on CIFAR-10
cd cnn_space2 && python train_search.py      # for conv cells on CIFAR-10
cd cnn_space3 && python train_search.py      # for conv cells on CIFAR-10

Test all models

The easist way to get started is to evaluate our DARTS models.

cd cnn_space1 && python test_all.py      # for conv cells on CIFAR-10
cd cnn_space2 && python test_all.py      # for conv cells on CIFAR-10
cd cnn_space3 && python test_all.py     # for conv cells on CIFAR-10

To run the ensemble

python test_ensemble.py

To run transfer learning in imagenette

cd darts_imagenette/cnn && python train_imagenette.py      # for conv cells on CIFAR-10
cd darts_imagenette/cnn_space1 && python train_imagenette.py      # for conv cells on CIFAR-10
cd darts_imagenette/cnn_space2 && python train_imagenette.py      # for conv cells on CIFAR-10
cd darts_imagenette/cnn_space3 && python train_imagenette.py      # for conv cells on CIFAR-10

To run the ensemble in imagenette

cd darts_imagenette && python test_ensemble_imagenette.py

Sette / CoNAS