This software package deploys an ultra-low loss quantization method that provides Convolutional Neural Network (CNN) quantization schemes based on comprehensive quantitative weight data analysis. This software package compresses the large size of CNN models to be friendly to smaller devices such as low-end edge-based systems or embedded systems. It is attractive to companies that provide computer vision and/or Artificial Intelligence (AI) full stack products.
This software package achieves the ultra-low loss quantization through data distribution transformation and optimal parameter enumeration during the training of a targeted CNN model.
This software package is built on-top of the open-source framework called CAFFE, with additional layer functions. It provides flexible 1, 2, 4, 8-bit width weight data quantization and maintains less than 1% accuracy loss.
We have witnessed the increasing requirement of AI and computer vision-based solutions in the recent years. However, the exponentially growing model size prevents further adoption of the technology in a number of applications, especially that in consumer electronics.
This software package enables ultra-low loss training quantization, which compresses the original huge CNN models to the target bitwidth to achieve multiple times of memory size reduction while maintaining the original model accuracy. It enables the deployment of the CNN models with huge memory requirement to the low-end devices while improving the potential computational efficiency optimization with low-bitwidth arithmetic that is supported by the modern platforms.