How to fine-tune existing model? about ddcolor

[piddnad]

Hi, you can refer to #3 (comment) and set pretrain_network_g as an existing model in the configuration file.

[mycomedico]

Do i need both the black and white and color images? or does the training automatically generate black and white images? do i need to specify seperate folders for the images if i need to include both? my images have resolution of 720height x 992 width. is this the correct format to specify that: gt_size: [720, 992]

[piddnad]

Do i need both the black and white and color images? or does the training automatically generate black and white images? do i need to specify seperate folders for the images if i need to include both? my images have resolution of 720height x 992 width. is this the correct format to specify that: gt_size: [720, 992]

1. Only color images are needed. During training, color images will automatically converted to black and white images. So just set your color image path list file in "meta_info_file" (e.g. 'data_list/imagenet.txt').
2. No, there is no need to change "gt_size". "gt_size" represents the resolution of GT samples and 256 is usually sufficient. When training, dataloader will automatically resize the input to the this size (see here) .

Hope this helps!