Exploiting the Relationship between Pruning Ratio and Compression Effect for Neural Network Model Based on TensorFlow

Table 1

Descriptions of parameters of the pruning command.

Parameter	Description	Default value
−1	Most basic model training
−2	Pruning is performed on the most basic model
−3	Pruning is performed, and network parameters of sparsification are stored
--train_data_dir	Folder path containing training data	/tmp/mnist_data
--train_dir	Log data during training and folder path containing model data	/tmp/mnist_train
--variables_dir	Folder containing data of the last trained model	/tmp/mnist_variables
--max_steps	Number of iterations of the basic model	10,000
--batch_size	Sample data size of each batch during model training	32
--sparse_ratio	Percentage of compression parameter of pruning/volume of parameters set to 0	0.9
--pruning_variable_names	Parameters on which pruning could be performed. Optional parameters include w_conv1, w_conv2, w_fc1, and w_fc2	w_fc1, w_fc2