These are chat archives for beniz/deepdetect

11th
Oct 2017
Marc Reichman
@marcreichman
Oct 11 2017 12:59
Thanks @beniz. So to be clear on the first question, when you use a URL, is it written to disk? or just spooled in memory
Emmanuel Benazera
@beniz
Oct 11 2017 12:59
@marcreichman no it's never written to disk
Marc Reichman
@marcreichman
Oct 11 2017 12:59
ok thank you!
Emmanuel Benazera
@beniz
Oct 11 2017 13:00
URLs are pooled in parallel threads in order to best accomodate batch usage
Marc Reichman
@marcreichman
Oct 11 2017 13:01
do you have any recommendations for a performance/resource balance for batch sizes? we've typically used size 10 with CPU-only workloads, i went up to 30 yesterday with URLs and the memory really expanded, which is trouble for us since we're typically running a number of instances in docker behind load balancing
Emmanuel Benazera
@beniz
Oct 11 2017 13:02
Look at https://github.com/jolibrain/dd_performances though that's for GPUs
Also, you can use https://github.com/beniz/deepdetect/blob/master/clients/python/dd_bench.py to find the exact sweet spot for your setup (CPU or GPU)
Marc Reichman
@marcreichman
Oct 11 2017 13:02
ok, thanks for the pointer to that
Emmanuel Benazera
@beniz
Oct 11 2017 13:04
We have versions optimized for AWS GPUs, but Amazon marketplace is taking a long time to release them at the moment, so they're in the queue... should get live whenever they decide to...