[Accounts pit] mixed programming related

[Accounts pit] mixed programming related

This article is my summary of the data access order, for their future intentions to learn to use, the contents of which probably did not help

1. Multi-language interactive mode: https://www.zhihu.com/question/270065760

2. C ++ and C # programming mix: https://zhuanlan.zhihu.com/p/81924996

3. C ++ mixed programming with Python: https://www.zhihu.com/question/23003213

   The Q high praise answers are worth a visit.

   There are geese plant this:

   Author: Tencent Technology Engineering

   With the extensive use of deep learning, a search engine / recommendation system / machine vision and other business systems, more and more deep learning model to deploy online services. The main use python implementation language, after completion of the offline training machine learning models on the line to deploy to re-implement in C ++ feature of these engineering logic code of python.

   We found that re-implemented in C ++ this step, the actual business to bring a lot of problems:

   1. Tedious, time-consuming, very prone to python and C ++ code is inconsistent
   2. Inconsistencies will directly affect the results of the model line, leading to broader market index of business as expected, produce a variety of bad case
   3. Inconsistency difficult to find, can not test, not monitor, often rely on user feedback complaints, even the market data anomalies can be found

   Currently the industry has so few Solutions:

   1. Online features stored for offline use
   2. Online C ++ code compiled into so export to use offline
   3. According to a configuration code is generated offline and online
   4. Extraction of common code, strengthen code reuse, software engineering tools, reduce inconsistencies

   But these ideas have various disadvantages:

   1. All features of all online requests, this large amount of data storage capacity
   2. Algorithms change the code to wait for background development, reducing the efficiency of the algorithm students
   3. Complexity feature processing proceeds to code the configuration file may not be able to fully express, but also increases the cost of learning configuration format
   4. It is true here offline feature processing code, most of the code do not pull out the common code reuse.

   Back to the starting point to consider the issue, it is obvious that the question boils down to is the need for a "python to c ++ translator"