Machine learning in trading: theory, models, practice and algo-trading - page 1947

New comment
 
Evgeny Dyuka:
I didn't come up with anything effective, at first I just tried to see what has the best effect on the result, but then I gave up, it was too tedious. It seems like TensorBoard can help. I haven't mastered it yet, if you'll get into it, please let me know if you set it up.

There's nothing interesting there. 

%load_ext tensorboard
import datetime, os
logdir = os.path.join("logs", datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))
tensorboard_callback = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
history = model.fit(InTrain, OutTrain, epochs=10, batch_size=32, callbacks=[tensorboard_callback])
%tensorboard --logdir logs

I want to try to pull the values of the weights of the input layer, I need to find out how.

 
Rorschach:

Nothing interesting there.

I want to try to pull the values of the input layer weights, I need to find out how.

Thanks for the link.
Is it necessary to bother with this at all? If there are not thousands of features, but dozens, then neuron will figure out what it needs, the main thing is to play with dropout. When I feed a lot to input, I set the dropout to 0.5 and let it think for itself what it needs.
 
Evgeny Dyuka:
Thanks for the link.
Is it necessary to bother with it at all? If there are dozens of features, neuron will figure out what it needs, just play with the dropout. When I feed a lot of stuff to input, I set the dropout to 0.5 and let it figure out for itself what it needs.

I think it needs it. I feed 10 lag traine and validation show similar numbers, I feed 100 traine begins to retrain.

 
Rorschach:

I think we should. I feed 10 lag traine and validation show similar numbers, I feed 100 traine starts retraining.

I decided the problem of retraining once and for all, when I started to podvat from 5 to 10 thousand for the feature, and epochs 100-150. In general, there is no problem retraining.
 
It seems to be Saturday...
 
Rorschach:

Everybody curses C++ for links, but python decided to go further and cram them everywhere.

You just have to learn how to use them properly. Getting a slice of data is one thing, but assigning back and forth until you don't understand where it's coming from is another :)

 
mytarmailS:

I'm studying the tsmp package.

Interesting stuff, sort of state recognition in a hidden Markov model

I don't know how to use it, but I'll keep it in mind...

function

https://sites.google.com/site/snippetfinderinfo/

oh I can't find any patterns, I guess there are none.

I can, but they run out quickly on new data.

 
Mihail Marchukajtes:
It seems to be Saturday...

In **** came.

It happens...
 
mytarmailS:

Alexey you make me nervous again)

I write a dozen codes every day, and I have to remember the code I wrote especially for you? I wrote it for you to learn something, and I should know if you changed the code or not?

And you haven't even learned how to look at a variable? You just type "X" into the console and press enter!

And I'm asking strange questions? Aren't you ashamed, Alexei?

No need to be nervous - practice - when the kids will be useful :)

So, what exactly is this function - the translator gives out:

predict-it's a universal function for predicting from the results of various model fitting functions. The function calls certain methods that depend on the class of the first argument.

As I understand it, it's essentially a function to apply the model to new data.

I read the UMAP help from which I concluded that the resulting model is essentially a matrix.

That's the matrix I was wondering how to get. In other methods of model creation it could be something else - mathematical formulas or a set of logical rules.

But, why isn't the algorithm for applying the model to new data described - how to use this matrix to relate a row of test sample to a specific coordinate? Without that, this whole direction is rubbish.

 
mytarmailS:

These are not returnees, there are no patterns in returnees ( verified by 7 years of experience) This is an abbreviated dimension, there are 2.5k features in these two curves Te is searching for patterns on steroids )

How did you get these curves? The main components?

1...194019411942194319441945194619471948194919501951195219531954...3494
New comment
Reason: