search

lmb-freiburg / Unet-Segmentation

The U-Net Segmentation plugin for Fiji (ImageJ)
https://lmb.informatik.uni-freiburg.de/resources/opensource/unet
GNU General Public License v3.0
87 stars 25 forks source link

The larger tile shape results lower speed in finetuning. #73

Open dean09131 opened 4 years ago

dean09131 commented 4 years ago

Hello, my GPU is Tesla V100-32G, when I use 508x508 tile shape as you did in the tutorial video, the speed is somewhat well, but when I use 1500x1500 tile shape, the estimated memory is about 18G, less than my GPU's limit, but the speed is quite slow. I'm not familar with caffe, so I thought larger tile shape is good for accelerating the finetuning, is it right?

ThorstenFalk commented 2 years ago

Factor 10 slower is expected with factor 10 larger input, everything above is overhead from data augmentation and transfer. The number of iterations may be affected by input shape but I would not in general say the bigger the better. I usually train with relatively small tiles and batch size one to increase randomness. Curves become wiggly but output is quite robust.