FFTFormer Documentation

To start the training process, execute the provided shell script from the root directory.

bash train.sh

To run inference/testing, follow these steps:

1. Download Pre-trained Model: Download the Realblur pre-trained model from this Google Drive link.
2. Run Test Script: Execute the test shell script.
3. Locate Results: Output images are saved in the results/model_name/dataset_name/ directory.

Note: PSNR measurements are performed using the official RealBlur test code.

bash test.sh

To use FFTFormer, ensure you have the following Python packages installed:

• Python
• Pytorch (version 1.11)
• scikit-image
• opencv-python
• Tensorboard
• einops

The repository provides source code for generating representative figures used in the project. These scripts can be modified to suit specific visualization needs.

Note that these scripts rely on utility functions defined in the basicsr/utils/plot_util.py module (from the BasicSR library), such as:

• read_data_from_tensorboard: To ingest data from TensorBoard logs.
• smooth_data: To apply smoothing to data series.
• read_data_from_txt_2v: To read data from text files.

One available script is model_complexity_cmp_bsrn.py, which compares model complexity (based on the BSRN paper).