AI inference can often be a slow, memory-crushing process due to the need for precision coupled with model computational complexity.
This session looks at a way to solve these issues using quantization: the process of converting data in FP32 to a smaller precision (like int8) while maintaining accuracy and performance and saving memory bandwidth.
AI software engineers Neo Zhang and Severine Habert introduce the tools and techniques to quantize your AI models easily and quickly, including:
- An overview of Intel® Neural Compressor and Intel® Deep Learning Boost
- A demonstration showcasing an end-to-end pipeline to train a TensorFlow* model with a small Keras* dataset, followed by speeding it up using quantization
- Performance comparisons of FP32 and int8 models by the same script