Research map / Architectures and efficiency / Quantization
Lossy and Lossless (L2) Post-training Model Size Compression
Yumeng Shi, Shihao Bai, Xiuying Wei, Ruihao Gong, Jianlei Yang
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 |
|---|---|---|---|
| add pruning to quantization (joint pruning and quantization transformation) at 20x compression (ablation) | L2 compression · ImageNet-1k val, ResNet-18 at ~20x CR | Top-1 accuracy (%): 68.34% (linear quant) -> 69.15% | Improved |
| use joint pruning and quantization instead of single quantization at ~15x CR (ablation) | L2 compression · ImageNet-1k val, ResNet-18 at ~15x CR | Top-1 accuracy (%): 70.37% (linear) vs 70.20% (joint) | No clear effect |
| vary kernel function (cosine, linear, triangle) and resolution (16-128) of the differentiable counter (ablation) | L2 compression · ImageNet-1k val, ResNet-18 at 15x CR | Top-1 accuracy (%): 70.04% to 70.25% | No clear effect |
| use range coding instead of Huffman coding (ablation) | L2 compression · ResNet-18 compressed size | model size (KB): 2500.24KB -> 2288.99KB (smallest target) | Improved |
| increase calibration set from 256 to 768+ images (ablation) | L2 compression · ImageNet-1k val, ResNet-18 at 12x CR | Top-1 accuracy (%): 67.93 -> 70.75 (768 images) | Improved |
| increase calibration set from 1000 to 2000 images (ablation) | L2 compression · ImageNet-1k val, ResNet-18 at 12x CR | Top-1 accuracy (%): 70.79 -> 70.78 | No clear effect |
| add weight fine-tuning after transformation training (ablation) | L2 compression · ImageNet-1k val, ResNet-18 at ~10x CR | Top-1 accuracy (%): 70.47% (no fine-tune, 10.05x) vs 70.79% (12x with fine-tune) | No clear effect |
| apply L2 compression versus AdaRound at similar/higher CR | MobileNetV2 · ImageNet-1k val | Top-1 accuracy (%): 69.01% -> 71.53% | Improved |
| increase target CR on ViT-base from 6x to 14x (ablation) | L2 compression · ImageNet-1k val, ViT-base-patch16 | Top-1 accuracy (%): 84.97% -> 81.86% | Hurt |
Closest papers by meaning
- Weightless: Lossy weight encoding for deep neural network compression (ICML 2018 · similarity 0.90)
- On Compressing Deep Models by Low Rank and Sparse Decomposition (CVPR 2017 · similarity 0.89)
- Compressing Neural Networks: Towards Determining the Optimal Layer-wise Decomposition (NeurIPS 2021 · similarity 0.88)
- Rate Distortion For Model Compression:From Theory To Practice (ICML 2019 · similarity 0.88)
- Model Preserving Compression for Neural Networks (NeurIPS 2022 · similarity 0.88)
- Optimal Brain Compression: A Framework for Accurate Post-Training Quantization and Pruning (NeurIPS 2022 · similarity 0.88)
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.
