Solution to CVPR2021 NAS track1

• Stage 2, Leaderboard B, 3rd (tied)

Please follow readme.ipynb to reproduce the results.

You can also reproduce results by running following commands.

First, please download dataset and architectures and put them under ./data. Then, download trained supernet models from released package and put them under ./saved_models. The files should in the following organization:

• data
  • cifar-100-python.tar.gz
  • Track1_final_archs.json
• saved_models
  • model_97122.pdparams
  • model_97131.pdparams

Train

Run following command to train the supernet.

python -m supernet.scripts.train --n 18 --kd --sandwich

Evaluate

Run following command to reproduce the 0.97122 and 0.97131 submission.

python -m supernet.scripts.evaluate --path ./saved_models/model_97122.pdparams --output ./saved_models/submit_97122.json
python -m supernet.scripts.evaluate --path ./saved_models/model_97131.pdparams --output ./saved_models/submit_97131.json