Following papers are implemented in PyTorch (DFQ)
Quantizing deep convolutional networks forefficient inference: A whitepaper
Data-Free Quantization Through Weight Equalization and Bias Correction
| Quantization with min-max | Reproduced result | Paper result |
|---|---|---|
| 8 bitwidth - Batch Norm folding | 0.1% | 0.1% |
| 8 bitwidth - Cross Layer Scaling | 69.59% | 69.91% |
| 8 bitwidth - CLS + High bias Absortion | 70.02% | 70.92% |
| Quantization with Mean Squared Error | Reproduced result | Paper result |
|---|---|---|
| 8 bitwidth - Batch Norm folding | 0.1% | 0.11% |
| 8 bitwidth - Cross Layer Scaling | 70.12% | 69.91% |
| 8 bitwidth - CLS + High bias Absortion | 70.36% | 70.92% |
BatchNorm layer has been folded
Convolution layer and Activations are fused as one operation
Cross Layer scaling & High bias fold are implemented
python main.py --images-dir <data-path> --ptq cle hba
torch : 1.4.0+cu100
torchvision : 0.5.0+cu100
MobileNet pretrained model : https://github.com/tonylins/pytorch-mobilenet-v2