Official Github repo for paper titled "GeneCAI: Genetic evolution for acquiring Compact AI" published in 2020 GECCO conference. The repo contains the source codes for applying structured pruning and decomposition on pre-trained models trained with the CIFAR10 dataset.
To compress the models, run main.py which expects to receive the checkpoint of a pretrained model in the following path: dataset_name/architecutre_name/checkpoint/ckpt.t7. For a complete list of parameters required bymain.py run python main.py --help. In the following, we provide example commands for compressing a ResNet56 architecture:
- Place the pretrained checkpoint for the non-compressed model in
CIFAR10/ResNet56/checkpoint/ckpt.t7 - Find optimized per-layer decomposition paramaters for the pretrained model:
python main.py --phase D --arch ResNet56 --num_population 100 --acc_threshold 90 - Select one config from path
CIFAR10/ResNet56/best_configs/decomposition(e.g.,iter_xxx_acc_yyy_flops_zzz.pkl) and perform fine-tuning to recover the accuracy opf the decomposed model:python main.py --phase D_FT --arch ResNet56 --config iter_xxx_acc_yyy_flops_zzz.pklThis saves the pretrained checkpoint toCIFAR10/ResNet56/checkpoint/decomposition/flops_xxx_acc_yyy.t7 - Find optimized per-layer pruning paramaters for the decomposed model:
python main.py --phase CP --arch ResNet56 --compressed_ckpt 'CIFAR10/ResNet56/checkpoint/decomposition/flops_xxx_acc_yyy.t7' --num_population 100 --acc_threshold 60 - Select one config from path
CIFAR10/ResNet56/best_configs/channel_pruning(e.g.,iter_xxx_acc_yyy_flops_zzz.pkl) and perform fine-tuning to recover the accuracy of the decomposed+pruned model:python main.py --phase CP_FT --arch ResNet56 -compressed_ckpt 'CIFAR10/ResNet56/checkpoint/decomposition/flops_xxx_acc_yyy.t7' --config iter_xxx_acc_yyy_flops_zzz.pklThis saves the pretrained checkpoint toCIFAR10/ResNet56/checkpoint/pruned/flops_xxx_acc_yyy.t7