Qimera: Data-free Quantization with Synthetic Boundary Supporting Samples

This repository is the official implementation of paper [Qimera: Data-free Quantization with Synthetic Boundary Supporting Samples].

Requirements

• Python 3.6
• PyTorch 1.8.1
• Refer requirements.txt for other requirements

To install requirements:

pip install -r requirements.txt

Training

For Imagenet training, change the path of the validation set in .hocon file. To train the model described in the paper, run one of this command:

./run_cifar10_4bit.sh
./run_cifar100_4bit.sh
./run_imgnet_resnet18_4bit.sh
./run_imgnet_resnet50_4bit.sh
./run_imgnet_mobilenet_v2_4bit.sh

Refer other_train_scripts folder for 5bit settings.

Evaluation

To evaluate trained model, run the command below after training:

./eval_cifar10_4bit.sh
./eval_cifar100_4bit.sh
./eval_imgnet_resnet18_4bit.sh
./eval_imgnet_resnet50_4bit.sh
./eval_imgnet_mobilenet_v2_4bit.sh

Visualizing Feature Space

Feature space visualization of real or synthetic images described in Figure 3.

python experiments.py --pca_source
python experiments.py --gdfq_generator_path GENERATOR_WEIGHT_PATH --pca_gdfq --image_gdfq
python experiments.py --qimera_generator_path GENERATOR_WEIGHT_PATH --pca_qimera --pca_mix --pca_path --image_gdfq --image_mix

Results

Our model achieves the following performance on :

Generated Synthetic Images for Cifar10 :

License

This project is licensed under the terms of the GNU General Public License v3.0