How can I pre-train the model?

[hayeong0]

Hello, Thank you for providing good papers and code.

I'd like to ask a question about the ESRGAN pre-trained model. I wonder how the pre-trained model parameters provided on your github were trained. I tried synthesizing the image using the pre-trained model parameters, and I'm going to try the pre-training myself.

The question is: Q1. As mentioned in the paper, is it correct to combine DIV2K, Flickr2K, and OST as the training dataset? If you used 3 datasets, I'm curious as to how they were configured in detail. Q2. In the same context as Q1, are all three datasets configured in the same way as DIV2K sampling with script/data_preparation/extract_subimages.py? Q3. Did you train each model on x2, x3, and x4 data and combine them? I'd like to know a specific way to do this. Q4. I would like to know how much GPU usage and how long it takes to train the model according to your method.

I expect a clear answer to the above question. Thank you.