I’m having trouble with CNN : I implemented an encoder-decoder (UNET) for image segmentation purpose. I made a custom dataset wich return an image with its associated ground truth loaded (thanks to getitem() method).
The problem is I cannot load more than 200 images (size : 480x640x3) otherwise my GPU is OOM .
Here is a screenshot of what’s happening inside my RAM during training (each drop correspond to the end of an epoch and the beginning of another)
How can I increase the number of image to train on ?
Thank you very much