Data-Free Network Quantization With Adversarial Knowledge Distillation

Data-Free Network Quantization With Adversarial Knowledge Distillation PyTorch (Reproduced)

Requirements

• Pytorch 1.4.0
• Python 3.6
• Torchvision 0.5.0
• tensorboard
• tensorboardX

Running the code

cifar10 dataset

CUDA_VISIBLE_DEVICES=0 python main.py --dataset=cifar10 --alpha=0.01 --do_warmup=True --do_Ttrain=True

The generated images and a trained student network from Knowledge distillation will be saved in ./outputs (default) folder.

cifar100 dataset

if you did train the teacher network, let argument "do_Ttrain" be False like as belows:

CUDA_VISIBLE_DEVICES=0 python main.py --dataset=cifar10 --alpha=0.01 --do_warmup=True --do_Ttrain=False

Arguments:

• dataset - Choose a dataset name
  • [cifar10, cifar100]
• data - dataset path
• teacher_dir - save path for teacher
• n_epochs - Epochs
• iter - Iterations
• batch_size - Size of the batches
• lr_G - learning rate for generator
• lr_S - learning rate for student
• alpha - Alpha value
• latent_dim - Dimensionality of the latent space
• img_size - Size of each image dimension
• channels - Number of image channels
• saved_img_path - Save path for generated images
• saved_model_path - Save path for trained stduent
• do_warmup - Do warm-up??
• do_Ttrain - Do train teacher network??

Result examples

Reference

Choi, Yoojin, et al. "Data-free network quantization with adversarial knowledge distillation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2020.