EPTQ is a PyTorch post-training quantization method for CV networks. It uses a Hessian-guided knowledge distillation loss to optimize the rounding error of the quantized parameters.
We support a large set of pre-trained models. The models are based on the implementation provided by BRECQ [1].
The names of the available models can be found under models/models_dict.py.
In order to run a model, you'll need to download the model's checkpoints, as explained in BRECQ's repository and in the following example:
wget https://github.com/yhhhli/BRECQ/releases/download/v1.0/resnet18_imagenet.pth.tar
The available models include:
| Model | Model usage name |
|---|---|
| ResNet18 | resnet18 |
| ResNet50 | resnet50 |
| MobileNetV2 | mobilenetv2 |
| RegNet-600M | regnet_600m |
| RegNet-3.2GF | regnet_3200m |
| MnasNet-2.0 | mnasnet |
pip install -r requirements.txt
python main.py -m resnet18 --train_dir /path/to/trainind/sampls/dir --val_dir /path/to/validation/sampls/folder --model_checkpoints /path/to/model/checkpoints.tar
This example would execute EPTQ for ResNet18 quantization with 8-bit weights and activation, using the default hyperparameters detailed in the EPTQ paper.
For faster execution you can reduce the number of optimization steps, using the flag
--eptq_iters (80K by default).
This might result in a small reduction in the final accuracy results.
[1] Li, Yuhang, et al. "BRECQ: Pushing the Limit of Post-Training Quantization by Block Reconstruction." ICLR 2021.