I have a MSX i7 notebook with nVidia 1070 (8GB RAM) 32GB RAM 256 SSD on Ubuntu 18.1
I managed to run Lesson1 in cuda without any problems.
But, I failed at variant lessons lesson1-vgg and lesson1-rxt50
RuntimeError: CUDA out of memory. Tried to allocate 9.00 MiB (GPU 0; 7.93 GiB total capacity; 7.08 GiB already allocated; 13.12 MiB free; 2.96 MiB cached)
Is it normal? Is my GPU too weak?
I rebooted the machine before retrying the lessons
I played around and of course, this is quite normal.
I watched the dogbreed lesson and I run the classification now.
Everything runs perfect with sz=64 and bs=56
Then, I changed sz to 224 and the GPU RAM immediately went 100%
So, I changed bs=16
I wonder if learning quality suffers, or only time.
It’s quite fascinating
Measuring the bs effect on the learning performances of common neural network architectures is an active research topic in the field of SGD methods. If you want to learn more a “soft” start could be this paper.
however, looking at the reviewers comments, do not expect any new insight inside.
So, basically,the conclusion is that there is a critical bs. Training is not faster.
I will look into that, at the other end of the stick