search

D-X-Y / AutoDL-Projects

Automated deep learning algorithms implemented in PyTorch.
MIT License
1.56k stars 281 forks source link

How to quickly retrain network architecture from scratch after getting the best topology #117

Closed xielinzhen closed 2 years ago

xielinzhen commented 2 years ago

python ./exps/NATS-algos/search-cell.py --dataset cifar10 --data_path $TORCH_HOME/cifar.python --algo darts-v1 --rand_seed 777

With the above commands, I have the best topology(|nor_conv_1x1~0|+|none~0|none~1|+|nor_conv_3x3~0|nor_conv_1x1~1|none~2|),

How do I train this topology from scratch？

D-X-Y commented 2 years ago

For the NATS-bench, it does not need to re-train it from scratch, as you can easily query the re-training performance via NATS-Bench API, see https://github.com/D-X-Y/NATS-Bench And see examples in our code: https://github.com/D-X-Y/AutoDL-Projects/blob/main/exps/NATS-algos/search-cell.py#L727

xielinzhen commented 2 years ago

For the NATS-bench, it does not need to re-train it from scratch, as you can easily query the re-training performance via NATS-Bench API, see https://github.com/D-X-Y/NATS-Bench And see examples in our code: https://github.com/D-X-Y/AutoDL-Projects/blob/main/exps/NATS-algos/search-cell.py#L727

Thanks for your reply and your great work!