How to train and test 2560*1440 pairs of images?

[songyn95]

Hello, I have pairs of 2560*1440 traffic images. How can I train 2560*1440 pairs of images? GPU size is 8*8=64gb，160 training set images and 40 test set images

python train.py --name r2560 --label_nc 0 --no_instance --resize_or_crop none --gpu_ids 0,1,2,3,4,5,6,7 --dataroot /data1/pix2pixHD/datasets/testIR --batchSize 8 --ngf 32 In this case, there will be an out of memory error（RuntimeError: CUDA out of memory. Tried to allocate 86.00 MiB (GPU 0; 7.43 GiB total capacity; 6.80 GiB already allocated; 18.94 MiB free; 123.67 MiB cached)）.

1. how can I modify the training command to make my effect better? 2. how to set the test command? 3. how many iterations are required for high resolution (2560*1440)?

Looking forward to your reply！

[mahaveer220]

Hi songyn95,

I also have the same issue it seems,I am facing errors when inferencing on high-resolution images. If you find a way please let me know as well. In my case I am using different sized images for each inference.

[songyn95]

@mahaveer220 ，The following are my training and test commands. You can try them.

• python train.py --name test-mix --label_nc 0 --no_instance --loadSize 1024 --fineSize 1024 --resize_or_crop scale_width --gpu_ids 0,1,2,3,4,5,6,7 --dataroot /data1/pix2pixHD/datasets/test-mix --batchSize 8
• python test.py --name test-mix --resize_or_crop scale_width --dataroot /data1/pix2pixHD/datasets/test-mix --label_nc 0 --no_instance --gpu_ids 0,1,2,3,4,5,6,7