CADyQ : Content-Aware Dynamic Quantization for Image Super Resolution

This respository is the official implementation of our ECCV2022 paper.

The overview of the proposed quantization framework CADyQ for SR network, which we illustrate with a residual block based backbone. For each given patch and each layer, our CADyQ module introduces a light-weight bit selector that dynamically selects the bit-width and its corresponding quantization function $Q_{b^{k}}$ among the candidate quantization functions with distinct bit-widths. The bit selector is conditioned on the estimated quantization sensitivity (the average gradient magnitude ${|\nabla{}|}$ of the given patch and the standard deviation ${\sigma}$ of the layer feature). Qconv denotes the convolution layer of the quantized features and weights.

Our implementation is based on EDSR(PyTorch) and PAMS(PyTorch).

Conda Environment setting

conda env create -f environment.yml --name CADyQ
conda activate CADyQ

Dependencies

• Python 3.6
• PyTorch == 1.1.0
• coloredlogs >= 14.0
• scikit-image

Datasets

• For training, we use DIV2K datasets.

• For testing, we use benchmark datasets and Test2K,4K.8K. Test8K contains the images (index 1401-1500) from DIV8K. Test2K/4K contain the images (index 1201-1300/1301-1400) from DIV8K which are downsampled to 2K and 4K resolution.

  # for training
  DIV2K 

  # for testing
  benchmark
  Test2K
  Test4K
  Test8K

How to train CADyQ

sh train_carn_cadyq.sh
sh train_idn_cadyq.sh
sh train_edsrbaseline_cadyq.sh
sh train_srresnet_cadyq.sh

Pretrained model for stduent and teacher model to start training from can be accessed from Google Drive.

How to test CADyQ

sh test_cadyq_patch.sh # for patch-wise inference
sh test_cadyq_image.sh # for image-wise inference

Citation

Contact

Email : [email protected]