###To run Lenet compression;

1. First train the model

python lenet_network_pruning.py --early_stopping 30

Then provide the name of the trained mode

2. To prune:

Example (choose your own model from step 1):

python lenet_network_pruning.py --resume=True --prune_bool=True --path mnist_trainval0.8_epo8_acc98.70

3. To prune and retrain:

python lenet_network_pruning.py --resume=True --prune_bool=True --path mnist_trainval0.8_epo8_acc98.70 --retrain=True

or

python lenet_network_pruning.py --resume=True --prune_bool=True --path mnist_trainval0.8_epo8_acc98.70 --switch_comb load --retrain=True

Notes:

The default early-stopping value is 500, for trying the code choose a lower value. --early_stopping 30 Once you run the switch training once you can load the parameters --switch_comb load Choose an architecture to prune: --arch 8,9,36,17 Select between mnist and fashionmnist --dataset fashionmnist

WideResNet

Along with other settings, the default value for num_epoch=200 can be changed in the config.py

1. To train the base network run:

python main.py

2. Then compute the switch vectors

python main_switch.py

It is enough to run it for even 1 iteration, 3-5 are recommended.

3. To prune and retrain the previously trained model:

python main_prune.py --arch 75,85,80,80,159,159,154,159,315,315,314,316

VGG

1. To train the base network run:

python vgg_main2_tomodule.py

2. Then compute the switch vectors

python main2vgg_switch_point.py --layer conv1 --path_switch_checkpoint <path>

The above file will run the switches only for one layer. If you would like to compute it for all the layers you may run

python script_vgg_vggswitch.py

However, you need to specify first the checkpoint trained in the step 1 in the main2vgg_switch_point.py in line 394

3. Prune and retrain

python vgg_main2_tomodule.py --model <path_to_model>

You may specify the architecture and training parameters with the arguments provided in the file.