These are chat archives for beniz/deepdetect

2nd
May 2017
cchadowitz-pf
@cchadowitz-pf
May 02 2017 22:47
Hi @beniz, me again :) I'm finding that DeepDetect is occasionally segfault-ing. I initially thought it was because I was batching too many input images at once, but now I don't think that's necessarily the case. I'm limiting the number of images per predict call to 25, and sometimes it works perfectly fine, other times it crashes with a segmentation fault. I saved my server log to the gist linked below, I can also paste a sample predict call that I'm using (I'm using a python client in this case) if need be. This is running built with both Caffe and TF support, TF on GPU, Caffe on CPU, and while I have models loaded for both mllibs and am hitting all of them occasionally as a health check, the script I'm running is only hitting a Caffe model, so I would expect that to remove any possible GPU-associated complications. This is running in a docker container.
Server log gist here: https://gist.github.com/cchadowitz-pf/fa38f7ebfa9108a2f78e1c1e7f5ec81b
If you'd like, I can open an issue for this instead, just seeing if you have any initial thoughts. Thanks!