Machine learning in trading: theory, models, practice and algo-trading - page 3233

New comment
 
Aleksey Vyazmikin #:
The main rule is that the model must be converted to ONNX.

thanks - I will study the topic....

 
mytarmailS #:
I in turn don't understand metaquotes with their idea of implementing ONNX.

To implement my model I need to learn the onnx language as well as python.
Great simplification of life, you want to implement a model - learn only 2 new languages))) powerful!!!

And as I understood all preprocessing all generation of features should be from outside as an input... To me it looks more like a stick in the wheels, not a breakthrough.

There are docker coneiners.

Everyone in any language can implement anything, any code, use any libraries, and pack it into a container.

No, they built some crutch with a lot of restrictions, with a large entry threshold and are proud of it...


There is no need to learn the onnx language - it is an internal representation language, just like you don't need to learn the pdf format if you only save or read documents in this format.

You need to learn Python only because of the lack of full-fledged ONNX support in R. And this, IMHO, is already a serious bell about the beginning of language obsolescence.

 

just a hypothetical example...

For example I am a javascript (or any other) programmer, I am interested in the market ,

I come up with some complex trading algorithm.


1) I input OHLC.

2) then a huge 100000 lines of javascript code on how to deal with traits (preprocessing)

3) then I train a javascript neural network on TensorFlow.js ( model ).

I get my robot as output.


I package all this code in a docker container and can integrate it anywhere.

On any comp, no dependencies needed, everything is already inside the container.

If metatrader supported this, it would be a breakthrough!


But what I am offered is this :

1) To have a metatrader to get OHLC date (well, everything is good here).

2) Learn the new MQL5 language to completely rewrite the entire code for data preprocessing (I'm already excited, aren't you? But everything is still ahead of me).

3) Learn new Python, Tensorflow framework for Python, ONNX for Python, train the model, save the model in ONNX. Models have limitations, you choose only what is available, not what you need, for example in ONNX there are no associative rules, no dbscan and thousands of other algorithms MO NO .



==============================================================================================

As a result, I have to completely rewrite all my algorithms in two new languages, having learnt them beforehand....

So that I can implement my model in ONNX format, and not just any model, but only the one that this format supports!!!!

COOL!!!!!!!!!!!

and the first variant with docker solves any problem, with any algorithm, and in one favourite ANY language.

 
Maxim Dmitrievsky #:
You can convert all the preprocessing too.

Where did you read that? I don't see any such information.

 
Aleksey Vyazmikin #:

Where did you read that? I don't see any such information.

Pitorch, tensorflo, sclern. Any large framework.
 
Aleksey Vyazmikin #:

Where did you read that? I don't see any such information.

Pitorch, tensorflo, sclern. Any big framework.

Freestyle rak zer tensorflo, henh henh henh zer flo, heard that song?
 
Maxim Dmitrievsky #:
Pitorch, tensorflo, sklern. Any big framework.

Good if that's the case. I've read several articles and watched videos - everywhere it was just about the model.

 
Aleksey Vyazmikin #:

Good if that's the case. I've read a few articles and watched videos - everywhere it was just about the model.

It's an open format, you can convert anything to it
 
Maxim Dmitrievsky #:
You don't need to have a bunch of docker overhead to do this.

Yes, it's better to spend months on learning new languages, frameworks and api onnx guts....

All for the sake of rewriting something you've already written...

 
mytarmailS #:

Yes, it's better to spend months learning new languages, frameworks and api onnx guts

That's for professional developers. If you start whining, you're no longer a pro.

Everything converts in 5 minutes. But there are inconsistencies so far on the MT5 side that need to be understood.
1...322632273228322932303231323232333234323532363237323832393240...3699
New comment