Training.md

Pretrained SR models

We initialize the SR network from the network weights trained on the DIV2K dataset.

You can download those pre-trained SR network weights from here, then please locate them in experiments/.

To be precise, the weights should be located as experiments/pretrained_models/*.

Training command

Single GPU

CUDA_VISIBLE_DEVICES=0 python src/main.py -opt path/to/config
# e.g., CUDA_VISIBLE_DEVICES=0 python src/main.py -opt options/seg/000_H2T.yml

Multi GPU (DDP)

CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node 2 src/main.py -opt path/to/config      # 2 GPU
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --nproc_per_node 4 src/main.py -opt path/to/config  # 4 GPU

Configuration files

We offer all training configuration files used in our experimental section (Table 1, 2, and 3 of our main manuscript).

You can check detailed training configurations in options/ at the following directories:

 ${ROOT}
 ├──options
    ├──seg
    ├──det
    └──cls
       ├── StanfordCars
       └── CUB200

For each training configuration, the required number of GPU differs, and you can find it on the commented note beside training batch_size.

By default, we assume that the memory size of the single GPU is 48GB.

However, if your GPU memory size is smaller, so the OOM issue occurs, then you can consider reducing the training batch_size and using multi-GPU.

Link to Testing.

Return to main: https://github.com/JaehaKim97/SR4IR