Research tools for communication-efficient deep learning, developed by Muhang Lan at USTC.

Latest version: 0.1.0

Installation

Install using pip

pip install deepcom

Check version number

import deepcom as dc
print(dc.__version__)

Function list

Basic tools

• Batch processing with enumerating given argument values: python -m deepcom config.json

Deep learning perspective

• Convert model parameters to a numpy array: model2params()

• Load a numpy array as model parameters: params2model()

Communication perspective

• Calculating mutual information: mutual_info()

Compression for training model

Compression for post-training model

• SuRP algorithm as a sparse compression for Laplacian sequence: surp_algorithm()