Training related issues

[songyn95]

Hello, I'm currently conducting training experiments. I have 17000 infrared images (trainA) and corresponding daytime results (trainB). The data resolution is 256 256, 8 Tesla P4 graphics cards (GPU Memory: 88GB = 64GB ) and batch_ Size = 16. I have the following questions to ask you:

1. How many iterations(epoch) does it take?
2. In the test results, it is obvious that I feel fake B is more effective than real B fuzzy, because real_ A is it the reason for the infrared image?
3. I am going to use the resolution of *2560 1440 to conduct the experiment. Is the following training command correct，the input and output resolutions shall be consistent during the test: (1): python train. py --dataroot ./ traffic dataset --name train traffic --model cycle gan --batch size 16 --load size 2560 --crop size 360 --gpuids 0,1,2,3 (2): python test. py --dataroot ./test jpg --name train traffic --model cycle gan --preprocess none --gpu_ids 0,1,2,3**
4. If I use the above command (train.py) to train images with a resolution of 2560 * 1440, how many iterations do I need? At present, the default is 200.

Look forward to your reply！

[junyanz]

If you have paired datasets of infrared images -> RGB images, you should consider using paired image-to-image translation (e.g., pix2pix, pix2pixHD, GauGAN). CycleGAN might not work for your application.

[songyn95]

@junyanz I use the pix2pixhd method to verify that the effect is very good, but the traffic scenarios are limited. How can I use it to predict so many different traffic scenarios? Is expanding datasets a viable approach?

[junyanz]

Answered here.