Cuda out of memory #68
|
I make the SR 1x model for removing water models, of course only for learning :) What to do not to turn off the computer after finishing training the model each time, to reset the memory "Cuda out of memory". I can start training for the night, by the morning the model was ready, is it possible? |
Replies: 2 comments
|
Hello! Typically 1x models will consume more VRAM than 4x models if the same configuration is used (specially batch size and crop size). If using the standard ESRGAN architecture, the recommendation here is to reduce either the batch size, the crop size or both, so the tensors fit in memory. Alternatively, you can try to use the pixel-unshuffle wrapper, but that depends on how you plan to use the model after it's trained and if it that code supports using the models with the wrappers. |
|
Thanks to Victoria also gives her Train_sr.yml maybe you will come up with an idea what can be improved. Thanks again!
|
Hello! Typically 1x models will consume more VRAM than 4x models if the same configuration is used (specially batch size and crop size).
If using the standard ESRGAN architecture, the recommendation here is to reduce either the batch size, the crop size or both, so the tensors fit in memory.
Alternatively, you can try to use the pixel-unshuffle wrapper, but that depends on how you plan to use the model after it's trained and if it that code supports using the models with the wrappers.