pix2pix: ideal training images dimensions

[Armand Collin]

First, thank you all for this amazing project.

I would like to train a pix2pix model with my own dataset. I have rectangular images, which I understand will require preprocessing flags. The dataset is quite large and is almost 2x bigger in total volume than the map <-> aerial dataset used in the original pix2pix paper. Images are 3762x2286 pairs, 8-bit grayscale. My guess is that the images are too large and would cause memory issues during training. For GPUs, I have access to GTX TITAN X (12 Gb) or possibly Tesla P100 (16 Gb).

Do you think it would be best to first crop these images in 4 or maybe 8 parts before feeding them to the combine_A_and_B script?

[Jun-Yan Zhu]

For grayscale images, you need to use the flag -input_nc 1 --output_nc.
You can directly use combine_A_and_B for the original images.
During training time, you can use different --preprocessing flags to crop patches on the fly. See Training & Test tips for more details.