Abstract: The nonuniform quantization strategy for compressing neural networks usually achieves better performance than its counterpart, i.e., uniform strategy, due to its superior representational capacity. However, many nonuniform quantization methods overlook the complicated projection process in implementing the nonuniformly quantized weights/activations, which incurs non-negligible time and space overhead in hardware deployment. In this study, we propose Nonuniform-to-Uniform Quantization (N2UQ), a method that can maintain the strong representation ability of nonuniform methods while being hardware-friendly and efficient as the uniform quantization for model inference. We achieve this through learning the flexible in-equidistant input thresholds to better fit the underlying distribution while quantizing these real-valued inputs into equidistant output levels. To train the quantized network with learnable input thresholds, we introduce a generalized straight-through estimator (G-STE) for intractable backward derivative calculation w.r.t. threshold parameters. Additionally, we consider entropy preserving regularization to further reduce information loss in weight quantization. Even under this adverse constraint of imposing uniformly quantized weights and activations, our N2UQ outperforms state-of-the-art nonuniform quantization methods by 0.5~1.7 on ImageNet, demonstrating the contribution of N2UQ design. Code and models are available at: this https URL.

Nonuniform-to-Uniform Quantization

This repository contains the training code of N2UQ introduced in our CVPR 2022 paper: "Nonuniform-to-Uniform Quantization: Towards Accurate Quantization via Generalized Straight-Through Estimation"

In this study, we propose a quantization method that can learn the non-uniform input thresholds to maintain the strong representation ability of nonuniform methods, while output uniform quantized levels to be hardware-friendly and efficient as the uniform quantization for model inference.

To train the quantized network with learnable input thresholds, we introduce a generalized straight-through estimator (G-STE) for intractable backward derivative calculation w.r.t. threshold parameters.

The formula for N2UQ is simply as follows,

Forward pass:

Backward pass:

Moreover, we proposed L1 norm based entropy preserving weight regularization for weight quantization.

Citation

If you find our code useful for your research, please consider citing:

@inproceedings{liu2022nonuniform,
  title={Nonuniform-to-Uniform Quantization: Towards Accurate Quantization via Generalized Straight-Through Estimation},
  author={Liu, Zechun and Cheng, Kwang-Ting and Huang, Dong and Xing, Eric and Shen, Zhiqiang},
  journal={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2022}
}

Run

1. Requirements:

• python 3.6, pytorch 1.7.1, torchvision 0.8.2
• gdown

2. Data:

• Download ImageNet dataset

3. Pretrained Models:

• pip install gdown # gdown will automatically download the models
• If gdown doesn't work, you may need to manually download the pretrained models and put them in the correponding ./models/ folder.

4. Steps to run:

(1) For ResNet architectures:

• Change directory to ./resnet/
• Run bash run.sh architecture n_bits quantize_downsampling
• E.g., bash run.sh resnet18 2 0 for quantize resnet18 to 2-bit without quantizing downsampling layers

(2) For MobileNet architectures:

• Change directory to ./mobilenetv2/
• Run bash run.sh

Models

1. ResNet

Note that N2UQ without * denotes quantizing all the convolutional layers except the first input convolutional layer.

N2UQ with * denotes quantizing all the convolutional layers except the first input convolutional layer and three downsampling layers.

W2/A2, W3/A3, W4/A4 denote the cases where the weights and activations are both quantized to 2 bits, 3 bits, and 4 bits, respectively.

2. MobileNet

Contact

Zechun Liu, HKUST (zliubq at connect.ust.hk)

Download Source Code

Paper Preview