Forums

Hi, <p>I'm trying to host a prototype Keras ML predictor in my pythonanywhere Hacker-level account. I'm using flask and when I try to call the endpoint, it takes ages to load and finally returns a pythonanywhere error, which I guess is a timeout error. <p> <p>When testing this code locally (oldish macbook pro), the Flask server runs normally and my predictions are pretty much instant, since I load the model when starting the server. So it can't be that the code execution is just too slow. Also I tracked that the method uses around 480mb of memory. I think I saw that pythonanywhere instances should have around 3 gigs of memory, so I don't see how even that would be the problem? Shouldn't the worker be killed instantly if it exceeds the memory limit? I do load the model in memory, but it's only about 150 mb.<p> Any help?

2018-05-13 08:39:25 Sun May 13 08:39:24 2018 - *** HARAKIRI ON WORKER 2 (pid: 97, try: 1) ***
2018-05-13 08:39:25 Sun May 13 08:39:24 2018 - HARAKIRI !!! worker 2 status !!!
2018-05-13 08:39:25 Sun May 13 08:39:24 2018 - HARAKIRI [core 0] 10.0.0.31 - POST //predict since 1526200163
2018-05-13 08:39:25 Sun May 13 08:39:24 2018 - HARAKIRI !!! end of worker 2 status !!!
2018-05-13 08:39:25 DAMN ! worker 2 (pid: 97) died, killed by signal 9 :( trying respawn ...

There are a number of threads on our forums where people are discussing the use of Keras in web apps. It seems that some uses of Keras are incompatible with web apps as they are now. But there were some workarounds in some of the threads.