How to choose an ML framework for sales?

Question

Neural networks

For a long time I delved into the world of machine learning and wondered how to choose a machine learning framework for production, where the model will be built exclusively on neural networks?
The bottom line is that there are many different frameworks and they are all +- similar, but they all have their jambs and nedopili.
For example, the very popular tensorflow. It was written by Google, and judging by the android, Google is not very fast at closing bugs that are important for third-party developers, and even in the data-flow of this framework, the value vector is copied in each "computing node". But on the other hand, tf has a very large community and even has tfhub.com, which promises to be something interesting. Yes, and it is written in python, which is not bad for ml tasks.
There is deeplearning4j, which, right out of the box, can integrate with Java microservices and Java big data projects, and this is very cool, because mostly large "sites" are written in Java (well, or in Sharp), but (suddenly) dl4j itself is written in java, although in the world of ml the de facto main language is python => java can limit in terms of available libraries. Oh yes, you will need to allocate a couple of gigabytes of RAM to train a neural network of 20 neurons.
Well, and so on ..
In general, I would like someone to suggest which points are critical in ML frameworks from the point of view of sales, which are not. It will be good if someone shares their experience of using this or that tool, lists stuffed cones, etc.

Answer 1

Wow, what a general question. And it is strange that after a "long enough deepening" it is still not possible to decide what is the best way to train a network for 20 neurons.
Comparison of various frameworks (programming languages, operating systems, database management systems, etc.) is akin to religious wars: there are adherents of each of the religions who, in fact, only know it. There are those who firmly believe in what is written in all sorts of catechisms, reviews, comparisons. There are followers, preachers, enthusiasts, apostates. There are quiet believers and developers who calmly write on what they were told from above. Only now there are no right and wrong in such wars. For there is no "best framework" and the search for it is meaningless. And the experience of some in the working conditions of others may be, to put it mildly, inapplicable.
Well, if you really need it, you can easily find comparisons of different tools, for example
https://www.netguru.co/blog/deep-learning-framewor...
https://medium.com/the-mission/8-best- deep-learnin...
https://www.datanyze.com/market-share/machine-lear...
But thank God engineering is not a religion. Here for the fact that you change one tool for another - they don’t burn at the stake. You can spend a lot of energy choosing the "best framework" and still not guess somewhere. And you can start doing your project on what you know, understand, own. Proud of the fact that you learned how to hammer nails with a hammer and not beat off your fingers at the same time, and not that you hold in your hand that Steanly hammer, and not a Toptul hammer. And then specifically to find out in what way your tool turned out to be weak exactly for your task, and already purposefully, and not abstractly, look for the best one in terms of a specific indicator.