Changes and configuration made for Tencent/Real-SR to work on HC18 on Docker or Colab

Forked version of https://github.com/Tencent/Real-SR, please refer to the official repository. This repository is only made for research pourpose regarding Ultra-sound image Super Resolution.

INDEX

0. Path preparation and dependencies using Google Colab
1. Dataset preparation
2. Training Phase
3. Inference Phase
4. Execution using Docker

---

0. Path preparation and dependencies using Google Colab

Install the following libraries required for script execution::

!pip install numpy opencv-python lmdb pyyaml
!pip install tb-nightly future

Then download the code from the following github:

!git clone https://github.com/Domics10/Real-SR

1. Dataset preparation

Download dataset HC18 at :https://zenodo.org/record/1327317#.YpSdTu7P2iM however we already provided a clean copy of the dataset.

Then will be necessary to download the github for the generation of the kernels:

!cd preprocess && git clone https://github.com/sefibk/KernelGAN

Let's proceed to the kernel generation:

!cd ./preprocess/KernelGAN && CUDA_VISIBLE_DEVICES=0 python3 train.py --X4 --input-dir ../HC18/train_set

Training dataset generation

Run the following code to create the HR and LR images for the selected dataset:

!python3 ./preprocess/create_kernel_dataset.py --dataset hc18 --artifacts clean --kernel_path ./preprocess/KernelGAN/results

After to inject noise in the generated LR images:

!python3 ./preprocess/collect_noise.py --dataset hc18 --artifacts clean

If missing, download the file : RRDB_PSNR_x4 downloaded from: https://drive.google.com/drive/u/0/folders/17VYV_SoZZesU6mbxz2dMAIccSSlqLecY in the pretrained_models folder.

2. Training Phase

To run the training:

!CUDA_VISIBLE_DEVICES=0 python3 train.py -opt options/dped/train_kernel_noise.yml

2.0 Resume training

To resume an interrupted training, make the following changes; go to the codes / options / deped / train_kernel_noise.yml folder and edit "resume_state" as for example :

resume_state: ../experiments/hc18_kernel_noise/training_state/5000.state

Then run the training command.

3. Inference Phase

You can edit the file test_dped.yml in the folder options/dped

After the training go in the folder experiments/hc18_kernel_noise/models: move the file current_step_G.pth, where "current_step" è is the last step of the training phase (5000, 10000, 20000, etc.), then move the file in pretrained_models and change the current_step_G.pth in the file codes/option/dped/test_dped.yml, as for example:

pretrained_model_G: ./pretrained_model/current_step_G.pth

Inference results will be saved in the folder results/hc18_results .

To run the test, execute the following line:

CUDA_VISIBLE_DEVICES=1,0 python3 test.py -opt options/dped/test_dped.yml

4. Execution using Docker

To use Docker and entrypoint.sh, simply run the following commands:

docker build -t image_name .

Then run the following script to launch the container execution:

docker run --rm --gpus all --env MODE=mode -it -v volume --name container_name image_name

mode can be:

• kc: kernel creation
• ka: kernel application to images
• na: collecting noise
• tr: Training phase
• te: Testing phase

Depending on the value of MODE, the volumes may change, these are working examples:

• kc: /mnt/disk1/vrai/CVDL2022/Tencent/Real-SR/codes/preprocess/KernelGAN/results:/home/Tencent/Real-SR/codes/preprocess/KernelGAN/results
• ka: /mnt/disk1/vrai/CVDL2022/Tencent/Real-SR/datasets:/home/Tencent/Real-SR/datasets
• na: /mnt/disk1/vrai/CVDL2022/Tencent/Real-SR/datasets:/home/Tencent/Real-SR/datasets
• tr: /mnt/disk1/vrai/CVDL2022/Tencent/Real-SR/experiments:/home/Tencent/Real-SR/experiments
• te: /mnt/disk1/vrai/CVDL2022/Tencent/Real-SR/codes/results:/home/Tencent/Real-SR/codes/results