classification

Test

1. pip install -r ../requirements.txt

2. Download Tiny-ImageNet from Google Drive or Dropbox. CIFAR dataset is automatically downloaded the first time the code is run. Place the dataset at your --dir_data directory.

3. Download the model zoo from Google Drive or Dropbox. This contains the compressed models. Place the models in ./model_zoo.

4. cd ./scripts/dhp_camera_ready.

5. Use the following scripts in ./scripts/dhp_camera_ready/demo_test_dhp.sh to test the compressed models.

   Be sure the change the directories --pretrain, --dir_data, and --dir_save.

   --pretrain: where the pretrained models are placed.

   --dir_data: where the dataset is stored.

   --dir_save: where you want to save the results.

6. Demo: test ResNet56 with target compression ratio at about 50%.

	# ResNet56, Ratio=0.5
	python ../../main_dhp.py --save ResNet_DHP_SHARE_L56_Ratio50 --template CIFAR10_ResNet --model ResNet_DHP_SHARE --depth 56 --test_only \
	--pretrain XXX --dir_data  XXX --dir_save XXX

Train

1. cd ./scripts/dhp_camera_ready.

2. Run the scripts dhp_XXX.sh to reproduce the results in our paper, where XXX may be replaced by mobilenet, mobilenetv2, resnet20, resnet56, resnet110 and resnet164 depending on which network you want to compress.

3. Be sure the change the directories --dir_data and --dir_save.

4. Demo: compress ResNet56 with target compression ratio 50%.

	# ResNet56, Ratio=0.50
	MODEL=ResNet_DHP_SHARE
	LAYER=56
	BATCH=64
	TEMPLATE=CIFAR10
	REG=3e-4
	T=5e-3
	LIMIT=0.01
	RATIO=0.5
	CHECKPOINT=${MODEL}_${TEMPLATE}_L${LAYER}_B${BATCH}_Reg${REG}_T${T}_Limit${LIMIT}_Ratio${RATIO}
	python ../../main_dhp.py --save $CHECKPOINT --template "${TEMPLATE}_ResNet" --model ${MODEL} --batch_size ${BATCH} --epochs 300 --decay step-20-50+step-150-225 \
	--depth ${LAYER} --prune_threshold ${T} --regularization_factor ${REG} --ratio ${RATIO} --stop_limit ${LIMIT} --print_model \
	--dir_save XXX --dir_data XXX