python3 main.py qconfig.yaml [path to imagenet dataset] -a [architecture of model] --pretrained
python3 main.py qconfig_bn_tuning.yaml [path to imagenet dataset] -a [architecture of model] --pretrained
- first:
download ImageNet from https://image-net.org/download-images - second:
run a script from https://github.com/pytorch/examples/blob/main/imagenet/extract_ILSVRC.sh
- Task: ImageNet
- Eval data num: 50k
- Calibration data num: 256
- Weight bit: 8
- Feature bit: 8
- Weight observer: MinMax
- Feature observers: As shown in following table
- Backend: TensorRT (symmetric feature quantization)
| Model | Float | MinMax | MSE | Percentile w/ alpha=1e-3 | Moving average w/ ema_ratio=0.9 | ACIQ | KL histogram |
|---|---|---|---|---|---|---|---|
| ResNet18 | 69.76% | 69.544% | 69.59% | 68.42% | 69.558% | 69.544% | 69.354% |
| ResNet50 | 76.146% | 75.888% | 76.042% | 75.34% | 76.026% | 76.014% | 75.634% |
| MobileNetV2 | 70.49% | 68.468% | 69.498% | 69.242% | 69.708% | 68.896% | 68.578% |
| EfficientNet-Lite0 | 75.394% | 75.116% | 75.148% | 74.862% | 75.09% | 75.116% | 31.082% |
| RegNetX-600MF | 75.034% | 74.56% | 74.776% | 72.946% | 74.692% | 74.602% | 74.458% |
- 09/01/2022: Add a batchnorm tuning algorithm. paper