We develop a training tool for Hybrid Fixed Point CNN Model. The tool quantize the each layers of AI Model from float32 to fixed point 1, 2, 4 or 8bit with reducing few accuracy. For the different applications, we can provides 1 bit fixed point model, 8 bit fixed point model or the hybrid fixed point model.

ezHybrid-M, a world-first hybrid fixed point CNN model training and inferencing framework, which can maintain the object detection quality while reducing over 90% model size.  
For hardware accelerator, we can reduce memory cost and 75% bandwidth and increase hardware computing efficiency.
Nowadays,  ezHybrid-M provides a world first novel method for training and inferencing Hybrid fixed point/binary CNN Model.

The ezHybrid-M: a Hybrid CNN Model training tool is suitable for low power/high efficiency Convolutional Neural Networks applications, including image classification, multi-object detection and semantic segmentation models. Take the mobilenet SSD as an example, using the proposed ezHybrid-M can generate a light weight fixed point CNN model that requires only 9% model size as compared to its floating point model with only 1% quality degradation.