Overview
A systematic evaluation of FP32, INT8 Post-Training Quantization (PTQ), and INT8 Quantization-Aware Training (QAT) using custom TinyCNN and MobileNetV2 on CIFAR-10 and CIFAR-10-C. Proposes INT8-RPTQ, a degradation-aware PTQ calibration strategy that improves robustness to noise, blur, and compression without full QAT retraining, and analyses deployment trade-offs across latency, peak memory, throughput, power consumption, and energy per inference.
Methods
- Degradation-aware calibration
- Robustness evaluation
- Hardware performance evaluation