Re: Implications of setting SIGCHLD in relation to NIFs
Max Lapshin <[email protected]>
| Newsgroups | gmane.comp.lang.erlang.general |
|---|---|
| Message-ID | <CAMxVRxCvr3nCq64-vHKeaw_ATyszneXo5ianpZy6+muK9AcdBA@mail.gmail.com> |
We use caffe + tensor-rt for video processing neural networks. They are tightly integrated with our video decoding code: it is very important to run all this in a single process to save memory. Flussonic unpacks protocols to frames, card decodes video, processes it and sends output. All this is done in a single process or a per-card process. Not easy to debug, running separate process is easier and more reliable of course. On Tue, Nov 24, 2020 at 9:07 PM Frank Muller <[email protected]> wrote: > > Hi Max > > Interesting... Can you shed some light on how you integrate Neural Network with Flussonic from a design perspective? > > Do you use an external AI library for that ? > > /Frank > > Le mar. 24 nov. 2020 à 18:40, Max Lapshin <[email protected]> a écrit : >> >> > I'm not sure I would like to have TensorFlow running inside the Erlang VM in the first place. >> >> We analyse video streams with neural networks. It is almost impossible >> and useless to run these things in different processes. >> >> Everything is running in same process because of enormous data >> streams, so it is absolutely ok to run all this inside erlang VM =) >> >> >> However, if your traffic is small, it maybe ok to split into different processes