Dear reviewer,

To reproduce the 6.6X pruned DenseNet-40 result in our paper Global Sparse Momentum SGD for Pruning Very Deep Neural Networks, please install tensorflow-gpu 1.11, run

python gsm_reproduce_dc40.py $CIFAR10_TFRECORDS_DIR

where $CIFAR10_TFRECORDS_DIR should contain two files: train.tfrecords and validation.tfrecords. 

Please follow https://github.com/tensorflow/models/blob/master/research/slim/datasets/download_and_convert_cifar10.py, download the CIFAR-10 dataset, convert it to TFRecord format, and rename the two output files as train.tfrecords and validation.tfrecords.

We used Tensorflow 1.11, CUDA9.2, Ubuntu16.04.

The codes will

1.	load the pre-trained model, use GSM to prune it by 8X, i.e., achieve 12.5% non-zero ratio.

2.	evaluate the output model

The core codes 

1.	rts_gradient_handler.py

2.	rts_base.py


Best regards,

The authors of the paper.