HAP is an advanced structured pruning library written for PyTorch. HAP prunes channels based on their second-order sensitivity. Channels are sorted based on this metric, and only insensitive channels are pruned. In the Neural Implant method, channels with moderate sensitivity levels are replaced with a trainable implant.
- PyTorch version >= 1.4.0
- Python version >= 3.6
- To install HAP and develop locally:
git clone https://github.com/yaozhewei/HAP
- Then run
bash install.sh-
Pretraining:
# for pretraining CIFAR10 $ python main_pretrain.py --learning_rate 0.1 --weight_decay 0.0002 --dataset cifar10 --epoch 200 -
Testing without Neural Implant on CIFAR10
python main_prune.py --network resnet --depth 56 --dataset cifar10 --batch-size 64 --learning-rate 0.02 --weight-decay 4e-4 --ratio 0.5 --use-decompose 0 --gpu "0"
HAP is released under the MIT license.
THIS SOFTWARE AND/OR DATA WAS DEPOSITED IN THE BAIR OPEN RESEARCH COMMONS REPOSITORY ON 02/07/23.