Tencent/PocketFlow

Supports multiple compression techniques—channel pruning, weight sparsification, and quantization—with automatic hyperparameter tuning via reinforcement learning (DDPG), Gaussian Processes, or Tree-structured Parzen Estimators to optimize compression ratios without manual tuning. The framework employs a learner-optimizer loop where compression algorithms generate candidate models that are evaluated and fed back to guide the search space exploration. Includes training enhancements like knowledge distillation, multi-GPU distributed training, and group fine-tuning to minimize accuracy degradation on deep learning models for mobile and resource-constrained deployment.

2,914 stars. No commits in the last 6 months.