Research map / Architectures and efficiency / Quantization
HMQ: Hardware Friendly Mixed Precision Quantization Block for CNNs
Hai Victor Habi, Roy H. Jennings, Arnon Netzer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HMQ instead of HAQ at ~16x weight compression | MobileNetV1 quantization · ImageNet (WCR ~16) | top-1 accuracy (%): 57.14 -> 68.36 | Improved |
| use HMQ instead of HAQ at ~16x weight compression | ResNet-50 quantization · ImageNet (WCR ~16) | top-1 accuracy (%): 70.63 -> 75 | Improved |
| use HMQ instead of HAQ at ~16x weight compression | MobileNetV2 quantization · ImageNet (WCR ~16) | top-1 accuracy (%): 66.75 -> 65.7 | Hurt |
| use HMQ instead of HAQ at ~10x weight compression | MobileNetV1 quantization · ImageNet (WCR ~10) | top-1 accuracy (%): 67.66 -> 69.88 | Improved |
| use HMQ instead of HAQ at ~10x weight compression | MobileNetV2 quantization · ImageNet (WCR ~10) | top-1 accuracy (%): 70.9 -> 70.12 | Hurt |
| use HMQ instead of HAQ at ~10x weight compression | ResNet-50 quantization · ImageNet (WCR ~10) | top-1 accuracy (%): 75.30 -> 76.1 | Improved |
| use HMQ instead of HAQ at ~8x weight compression | MobileNetV1 quantization · ImageNet (WCR ~8) | top-1 accuracy (%): 71.74 -> 70.912 | Hurt |
| use HMQ instead of HAQ at ~8x weight compression | ResNet-50 quantization · ImageNet (WCR ~8) | top-1 accuracy (%): 76.14 -> 76.3 | Improved |
| use HMQ instead of DQ with ACR 8 | MobileNetV2 quantization · ImageNet (ACR 8) | top-1 accuracy (%): 69.74 -> 70.9 | Improved |
| use HMQ instead of HAWQ-V2 with ACR 8 | ResNet-50 quantization · ImageNet (ACR 8) | top-1 accuracy (%): 75.7 -> 75.45 | Mixed |
| raise target weight compression Rw from 4 to 8 | HMQ on EfficientNet-B0 · ImageNet (ACR 4) | top-1 accuracy (%): 76.4 -> 76 | Hurt |
| raise target weight compression Rw from 8 to 12 | HMQ on EfficientNet-B0 · ImageNet (ACR 4) | top-1 accuracy (%): 76 -> 74.6 | Hurt |
Closest papers by meaning
- HAWQ: Hessian AWare Quantization of Neural Networks With Mixed-Precision (ICCV 2019 · similarity 0.90)
- InfoQ: Mixed-Precision Quantization via Global Information Flow (AAAI 2026 · similarity 0.90)
- HAQ: Hardware-Aware Automated Quantization With Mixed Precision (CVPR 2019 · similarity 0.89)
- FracBits: Mixed Precision Quantization via Fractional Bit-Widths (AAAI 2021 · similarity 0.89)
- HAWQ-V3: Dyadic Neural Network Quantization (ICML 2021 · similarity 0.89)
- RMSMP: A Novel Deep Neural Network Quantization Framework With Row-Wise Mixed Schemes and Multiple Precisions (ICCV 2021 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
