Closed asmagen closed 4 years ago
Also, using a GPU accelerated instance on AWS I got the ~6k x 6k pixel image to run in about 40 minutes using the code I extracted from your package and air lab. Not sure about the accuracy yet. But isn't it much more time than what we estimated? (and this size is for 5x while I want to do 10x)
Thanks, I just pushed another update
Hmm how many iterations did that run for? That does not sound right. What transformation did you use?
The time it took to run the algo is very surprising though. I think you did not specify a GPU to use and tried to fit using CPU. Try changing gpu_device to 0.
The number of iterations is far to small to achieve a good fit
I used 1000 (not 10) iterations on that medium sized image (~6k). Is that too much? I was wondering why airlab doesn't have a termination strategy rather than a fixed # iterations. Not sure how big is the impact of additional 500 iterations. How long does it take you for such image? And what resources are you allocating? Do I still need a lot of RAM (>30GB) with a GPU instance? And does it utilize multiple GPUs? The idea is to run everything in parallel on about 100 images at a time so I hope this is all scalable and the resources allocation won't be a huge bottleneck.
I just checked out some of the stats from a previous run. I was using less than 11GB RAM, fitting images of size ~3-10k by 3-10k, with run times of 1-5 minutes per image with similar number of iterations, though I think it really depends on the task how long you want to iterate for and something I still need to fully explore for histopath image registration.
You'd want to deploy each image to separate GPU instances. But not sure why you were getting those runtimes.
I would check nvidia-smi -l
while running a job.
Going to close this since we patched this bug.
Error/bug with device: