Latency probably isn't an issue with TensorFlow model training, which can take hours or even days for a training run. Model execution latency could be an issue in some applications, but often getting the answer in an hour or so is acceptable for high-value answers which can't be obtained using other methods
Skip On May 25, 2017 11:41 AM, "Raul Miller" <[email protected]> wrote: > Their API looks.... overly complicated, right now. > > https://www.tensorflow.org/api_docs/ > > They do not even document the network API - instead, they document > wrappers for that API for a few programming languages. So you'd have > to take apart one of the wrappers (or add another layer of crud over > top one of those wrappers - python, probably, since that's the only > one they think is stable). > > My guess is that latency is high enough that they need this level of > indirection to deflect people inclined to complain about such things. > > -- > Raul > > > On Thu, May 25, 2017 at 11:37 AM, Skip Cave <[email protected]> > wrote: > > Google has released the latest version of TensorFlow, their open-source > > machine learning package at their recent Google I/O event. TensorFlow > runs > > on Android, iOS, Raspberry Pi and both Google and AWS cloud services, > using > > the same API. TensorFlow supports a wide range of CPUs, GPUs., and now > TPUs > > (Tensor Processing Units). Tensor is Google's name for vectors and > > matrices. Google provides TensorFlow API support in four languages: > Python, > > C++, Java, & Go. Other groups have ported the TensorFlow API to Haskell, > > Julia, C#, and R. TensorFlow is designed to hide the concurrent nature > of > > the underlying processes, so large numbers of parallel processes can be > > treated as a single process at the top-level TensorFlow API. > > > > Check out the TensorFlow overview video given at Google I/O last week (36 > > minutes) on YouTube at: > > https://www.youtube.com/watch?v=OzAdKMPgUt4&list= > TLGGqXCgIcW-mFUyNDA1MjAxNw > > > > Check out the basic TensorFlow machine learning formula at 18:36 in the > > video. This works as a line of J code (getting rid of the square > brackets, > > and adding parenthesis to override J's right-to-left execution). > > > > Imagine! You can debug a machine-learning application on one's PC or > > smartphone using a test dataset. Then you run the compute-intensive > > big-data training in the cloud, running on dozens (or hundreds) of GPUs, > or > > now TPUs. Finally, one can run the trained models back on a local > machine, > > or keep the model execution in the cloud, if the model still needs lots > of > > processing. > > > > This seems like a perfect fit for a J TensorFlow library. Unfortunately, > > implementation of such a library is way above my J skill level. > > > > Skip Cave > > Cave Consulting LLC > > ---------------------------------------------------------------------- > > For information about J forums see http://www.jsoftware.com/forums.htm > ---------------------------------------------------------------------- > For information about J forums see http://www.jsoftware.com/forums.htm ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
