Automated Mixed-Precision Quantization Search for Efficient Deep Neural Networks

FLIQS is the first one-shot mixed-precision quantization search framework that efficiently discovers high-accuracy models without retraining, by jointly optimizing integer and low-precision floating-point formats.