how to train own datasets

[xingjunhong]

how to train own datasets

[zhengchen1999]

You can modify the training configs in .yml files, like train_DAT_light_x4.yml. Replace your own data (LR and HR) path in dataroot_gt and dataroot_lq.

    dataroot_gt: datasets/DF2K/HR
    dataroot_lq: datasets/DF2K/LR_bicubic/X4

PS: Our code is based on BasicSR. It provide detail commands tutorials in doc. Maybe you refer to it.

[xingjunhong]

How much data does it take to train a demo model

[zhengchen1999]

Our training dataset is DK2K, which contains 3550 images. The training setting of our models: batch size = 32, patch size = 64 $\times$ 64, training iteration = 500,000.

[xingjunhong]

Are we training with high-definition images? Or use low resolution images for training?

[zhengchen1999]

Training needs high-resolution and low-resolution image pairs. This is a supervised task, where high-resolution images are ground truth.

[xingjunhong]

Isn't it using low resolution images for training and then performing loss calculations with high-resolution images?

[zhengchen1999]

Yes.

[xingjunhong]

Is the original image of the dataset we collected high-resolution? Then use the script to turn it into a low resolution image and train it with a low resolution image?

[zhengchen1999]

Yes, in our approach, we use bicubic downsampling for high-resolution images (x2, x3, x4). Of course, you can also utilize other methods, such as the degradation model in Real-ESRGAN, which is suitable for real-world SR.