Trading Strategy with Cross Validation - Runtime error

AI & Machine learning
1

I was trying our the code provided on this particular notebook and man you guys skip a LOT of informations. For example the importing a local python script which is designed to make a new Neural Network was skipped, I had to search around a bit to understand why 

from Keras_CV import create_new_model

was causing an error.

Immediately after fixing that issue, I am facing this:

grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=1,verbose=2)

grid_result = grid.fit(X_train, y_train)

RETURNS:
RuntimeError: Cannot clone object <keras.wrappers.scikit_learn.KerasClassifier object at 0x0000014EBE7032E8>, as the constructor either does not set or modifies parameter class_weight

Apparently this error is because my code is trying to clone something.
How do I fix this?
Pretty disappointing that I have to come here and ask for help even for the code which is provided to us as an example.
Now I have to wait for hours for someone to actually reply to this so that I can move ahead with my learning/experimenting because you guys won't provide an actual working strategy.
What a waste of time and money. Thank you Quantra.
 

 

2

Hi Abinash,



Please replace the following line of code:



model = KerasClassifier(build_fn=create_new_model, epochs=20, batch_size=128, verbose=2, 

                       validation_split=0.2,class_weight=class_weight)



by



model = KerasClassifier(build_fn=create_new_model, epochs=20, batch_size=128, verbose=2, 

                       validation_split=0.2)



by removing class_weight=class_weight from the call. And the code shall work.



Please let us know if you need more help.



Thanks,

Ashish Garg

3

The above fix worked but that defeats the purpose of balancing the input data based on their weights. Doesn’t that ruin the purpose of pre-processing ?

4

Plus in the same example of Trading strategy with cross-validation, why the final output from the best_model.predict() is a decimal number instead of 0 or 1 , which it was trained to predict. y_train consists of 0s and 1s

5

Hi Abhinash, The final output is a probability of the data belonging to a particular class. This is explained in the strategy notebook used in the Section 2 Unit 11 of the course. We are currently working on the runtime error issue that you have posted earlier and will get back to you as soon as we find a solution to this problem.

6

Hi Abhinash, There is solution to the class_weights problem. You can pass them as parameter when you fit the data to the grid search object. Please check the following code for clarity:

7

from sklearn.pipeline import Pipeline neurons_params = [ 225,150,175] act_1_params=[‘tanh’,‘sigmoid’,‘relu’] dropout_ratio_params=[0.18,0.30,0.23] pipeline = Pipeline(steps=[(‘clf’, model )]) param_grid = dict(clf__neurons=neurons_params,clf__act_1=act_1_params,clf__dropout_ratio=dropout_ratio_params) grid = GridSearchCV(estimator=pipeline, param_grid=param_grid, n_jobs=1,verbose=2) grid_result = grid.fit(X_train, y_train,clf__class_weight=class_weight)

8

The class_weights feature is facing some issues in the latest version of Keras wrapper and is a known issue. You can track it on the github: Using callbacks with the SciKit-Learn API? · Issue #4081 · keras-team/keras · GitHub . Meanwhile, we will update the current solution file and downloadable code in the course so that you can try the solution at your end and confirm it.

9

Hi Abinash, Notebook and downloadables section have been updated. Please take a look.