Neural Networks Quantization

Linear Quantization

Post Training Quantization

• Same, Same But Different - Recovering Neural Network Quantization Error Through Weight Factorization
• Data-Free Quantization through Weight Equalization and Bias Correction
• Fighting Quantization Bias With Bias
• Improving Neural Network Quantization without Retraining using Outlier Channel Splitting
• Cell Division: Weight Bit-Width Reduction Technique for Convolutional Neural Network Hardware Accelerators
• Figting bias with bias

Quantization-Aware Training

• PACT: Parameterized Clipping Activation for Quantized Neural Networks

  • Relu-α where α is a learnable parameter per layer with L2 regularization

• Accurate and Efficient 2-bit Quantized Neural Networks

  • PACT quantization aware training
  • Statistic-Aware Weight Binning
  • Full precision for shortcut connections in resnet

• Fully Quantized Network for Object Detection

  • 4bits quantization
  • Freeze BN statistic
  • clamped activation from calibration statistic
  • channel-wise quantization scale

• Trained Uniform Quantization for Accurate and Efficient Neural Network Inference on Fixed-Point Hardware

• Simultaneously Optimizing Weight and Quantizer of Ternary Neural Network using Truncated Gaussian Approximation

state of art

4-bits

• DISCOVERING LOW-PRECISION NETWORKS CLOSE TO FULL-PRECISION NETWORKS FOR EFFICIENT EMBEDDED INFERENCE
• Learned Step Size Quantization
• (https://arxiv.org/pdf/1810.05723.pdf)

2-bits

• Accurate and Efficient 2-bit Quantized Neural Networks
• Simultaneously Optimizing Weight and Quantizer of Ternary Neural Network using Truncated Gaussian Approximation

Anealing from Continue to Discrete

• Quantization Networks
• Learning low-precision neural networks without Straight-Through Estimator (STE)

Non-Uniform Quantization

• LQ-Nets: Learned Quantization for Highly Accurate and Compact Deep Neural Networks
• Quantization Networks

Sparsity and Quantization

• SeerNet: Predicting Convolutional Neural Network Feature-Map Sparsity Through Low-Bit Quantization
• Value-aware Quantization for Training and Inference of Neural Networks
• Efficient and Effective Quantization for Sparse DNNs
• BiScaled-DNN: Quantizing Long-tailed Datastructures with Two Scale Factors for Deep Neural Networks

Quantization Support in Libraries

Post Training Quantization

• TensorRT
  • Per channel weight scale
  • Calibration: minimize KL Divergence
• Tensorflow lite
  • Per channel weight scale
  • Calibration: min max??
• TVM
  • Per channel weight scale
  • Calibration: minimize MSE

LUT-based

https://arxiv.org/pdf/1906.04798.pdf