How to productionize the code | Any recommendation

[jaiswalvineet]

Can you please help me to productionize the model, here the difficulty I have faced

1. TF serving with ngnix (recommended) : Because it store model in pickle file so not sure how to provide pickle file path as model, can you please help me
2. Flask App with ngnix : It loose graph reference after few calls, random behavior and seems its too hard to provide same offering as TF serving

please help us with sample code

[arshabhi]

Hi, Thanks for pointing out the issue @jaiswalvineet , I am also having trouble with tensorflow serving of the model. Any help would be appreciated.
Thanks in advance

[gallmerci]

Regarding question two, I responded to your problem in #120. I hope this fixes your problem. But I am also very interested in running TensorRec in production. Especially in performance optimizations. Currently solution with flask and TensorRec using predict_ranks method gives me a latency of 150-300ms even on larger compute instances on AWS.
Use case is computing a recommendation for a single user and roughly 100 items with 12 item features.

If anybody can recommend some performance tuning tipps for model serving, it would be awesome :)

[gallmerci]

I did some load testing with TensorRec for serving using TensorFlow 1.12. and TensorFlow < 1.12 and the latter was way faster (up to factor of 10!) than 1.12. Is there anything known why 1.12. is so slow with TensorRec? @jfkirk also noticed that the tests are running much slower with this version...