Skip to content

Latest commit

 

History

History
156 lines (107 loc) · 4.38 KB

training.md

File metadata and controls

156 lines (107 loc) · 4.38 KB
Raw

Training

Train DF2K

Dataset Preparation

DF2K: Combine the DIV2K dataset with the Flickr2K dataset.

Step 1: Download the DF2K dataset and extract it to ./data.

cd scripts
bash get_df2k_datasets.sh

Step 2: Download the test dataset and extract it to ./data.

cd scripts
# will download Set5
bash get_test_datasets.sh

Step 3: Crop to sub images.

The main purpose is to reduce the IO consumption of hard drives and speed up training.

python tools/misc/slice_sub_image.py -i datasets/DF2K/DF2K -o datasets/DF2K/DF2K_sub_image_384x384 --crop-size 384 --step 192 --thresh-size 384

If the machine has enough performance, you can increase the number of threads to speed up processing.

python tools/misc/slice_sub_image.py -i datasets/DF2K/DF2K -o datasets/DF2K/DF2K_sub_image_384x384 --crop-size 384 --step 192 --thresh-size 384 --num-workers 32

Step 4: Download the pretrained model weights to ./results/pretrained_models.

wget https://github.com/Lornatang/Real_ESRGAN-PyTorch/releases/download/0.1.0/realesrnet_x4-df2k_degradation.pkl -O results/pretrained_models/realesrnet_x4-df2k_degradation.pkl
wget https://github.com/Lornatang/Real_ESRGAN-PyTorch/releases/download/0.1.0/realesrgan_x4-df2k_degradation.pkl -O results/pretrained_models/realesrgan_x4-df2k_degradation.pkl

Train

Finetune (Recommended)

Step 1: Finetune the realesrnet_x4 model.

python tools/train.py configs/Real_ESRGAN/realesrnet_x4-finetune.yaml
# Results will be saved in `./results/train/realesrnet_x4_degradation-finetune`

Step 2: Finetune the realesrgan_x4 model.

python tools/train.py configs/Real_ESRGAN/realesrgan_x4-finetune.yaml
# Results will be saved in `./results/train/realesrgan_x4_degradation-finetune`

Train from scratch

Step 1: train the realesrnet_x4 model.

python tools/train.py configs/Real_ESRGAN/realesrnet_x4.yaml
# Results will be saved in `./results/train/realesrnet_x4_degradation`

Step 2: train the realesrgan_x4 model.

python tools/train.py configs/Real_ESRGAN/realesrgan_x4.yaml
# Results will be saved in `./results/train/realesrgan_x4_degradation`

Train custom dataset

Dataset Preparation

Step 1: Put the training data in the ./data directory.

The data structure is as follows:

|---datasets
    |---custom
        |---gt
            |---0001.png
            |---0002.png
            |---...
        |---gt_sub_image_384x384
            |---0001-0001_0001.png
            |---0001-021_0301.png
            |---...

Step 2: Crop to sub images.

The main purpose is to reduce the IO consumption of hard drives and speed up training.

python tools/misc/slice_sub_image.py -i datasets/custom/gt -o datasets/custom/gt_sub_image_384x384 --crop-size 384 --step 192 --thresh-size 384

If the machine has enough performance, you can increase the number of threads to speed up processing.

python tools/misc/slice_sub_image.py -i datasets/custom/gt -o datasets/custom/gt_sub_image_384x384 --crop-size 384 --step 192 --thresh-size 384 --num-workers 32

Step 3: Download the test dataset and extract it to ./data.

cd scripts
# will download Set5
bash get_test_datasets.sh

Train

Finetune (Recommended)

Step 1: Change realesrnet_x4 yaml.

Find the realesrnet_x4_degradation-finetune.yaml file in the configs/Real_ESRGAN/realesrnet_x4_degradation-finetune.yaml directory and follower modify.

# line 48
TRAIN_GT_IMAGES_DIR: "data/custom/gt_sub_image_384x384"  # 178574 images

Step 2: Finetune the realesrnet_x4 model.

python tools/train.py configs/Real_ESRGAN/realesrnet_x4-finetune.yaml
# Results will be saved in `./results/train/realesrnet_x4_degradation-finetune`

Step 3: Change realesrgan_x4 yaml.

Find the realesrgan_x4_degradation-finetune.yaml file in the configs/Real_ESRGAN/realesrgan_x4_degradation-finetune.yaml directory and follower modify.

# line 48
TRAIN_GT_IMAGES_DIR: "data/custom/gt_sub_image_384x384"  # 178574 images

Step 4: Finetune the realesrgan_x4 model.

python tools/train.py configs/Real_ESRGAN/realesrgan_x4-finetune.yaml
# Results will be saved in `./results/train/realesrgan_x4_degradation-finetune`