PKTQ: Per-Kernel Thresholded Quantization With Size-Constrained Two-Stage Optimization for CNNs

Journal article, 2026

Convolutional neural networks are challenging to deploy on resource-constrained computing platforms due to their high storage and computational demands. Mixed-precision quantization can mitigate these issues by assigning different bit-widths per layer. However, finding an optimal quantization policy is challenging due to the exponential size of the policy search space and the need for extensive policy evaluations. Additionally, existing works often overlook strict model size constraints and apply per-layer quantization, which ignores intra-kernel variances. We propose a two-stage deep deterministic policy gradient (DDPG) framework. In the first stage, it optimizes accuracy; in the second, it jointly optimizes accuracy and model size with a constraint-aware reward that enforces storage limits during training. On this basis, a per-kernel clipping threshold is introduced to reduce quantization error by adapting the clipping threshold to individual kernel distributions. We perform extensive simulations based on ImageNet with MobileNet-V1/V2 and ResNet-50. The results show that our method outperforms the state-of-the-art baselines under the same compression ratios, achieving minimal accuracy drops with aggressive compression and near-lossless performance at moderate levels.