PocketFlow

Why we included this project

PocketFlow handles the tedious part of getting deep learning models onto phones and other devices with limited compute, which is shrinking them and making them run faster. Instead of hand-tuning pruning or quantization parameters, you set a target compression or acceleration ratio and the framework searches for good hyper-parameters on its own. It bundles several techniques, including channel pruning, weight sparsification, and both uniform and non-uniform quantization, so you can compare how each approach behaves on the same model. Fine-tuning and distillation options help keep accuracy losses small, and distributed training support helps when you are iterating on larger networks. The framework was developed by the machine learning team at Tencent AI Lab for compacting neural networks used in industrial applications.