Extract subset of ImageNet
python subImageNet.py --data-path /your_data_path/Searching with 8 GPUs
n_doe is the initial population size in the evolutionary search, while gene_num is the generation number. pruned_num and flop_num are numbers of best-performing Lottery Tickets and the largest computation models when formulating parent populations, respectively. shrink_num is the number of generated pruning proposals for each trained network. actual_prune is the number of Lottery Tickets selected from pruning proposals. data_dir is the path of 10% ImageNet training set, while sub_val_set is the path of 10% ImageNet val set.
python search.py --n_doe 60 --gene_num 5 --pruned_num 24 --flop_num 6 --shrink_num 40 --actual_prune 2 --data_dir /your_10%_train_path/ --sub_val_set /your_10%_val_path/ Re-training with 8 GPUs
we present the searched model dictionary in searched/net.dict
sh scripts/retrain.sh 8 /your_imagenet_path/ --model_cfg ./searched/net.dict --model /name_for_save/ -b 128 --sched step --epochs 300 --decay-epochs 2.4 --decay-rate .97 --opt rmsproptf --opt-eps .001 -j 8 --warmup-lr 1e-6 --weight-decay 1e-5 --drop 0.2 --drop-connect 0.2 --model-ema --model-ema-decay 0.9999 --aa rand-m9-mstd0.5 --remode pixel --reprob 0.2 --amp --lr .064 Throughput
To measure the throughput, run:
sh scripts/retrain.sh 1 /home/pdl/datasets/ImageNet/ --model_cfg ./searched/net.dict --model tttest -b 128 --amp --throughput