Is it possible to pickle only trained model and call it back for test and/or forecast without any training process in another .py folder? about scalecast HOT 5 CLOSED

After calling f.auto_forecast() and before changing the estimator, the scikit-learn trained model is available in the regr attribute. So, you can use code like this to do what you are asking:

import pickle

xgboost_grid = {
'max_depth': [15],
'tree_method': 'gpu_hist'
}

f.set_estimator('xgboost')
f.ingest_grid(xgboost_grid)
f.tune()
f.auto_forecast()

with open('xgboost_regr.pckl','wb') as pckl:
  pickle.dump(f.regr,pckl)

In a separate issue, I think you need brackets around 'gpu_hist' in the grid. If there is an error with all hyperparam combinations during grid evaluation, the model is trained with default parameters.

from scalecast.

Thanks for the response. I tried your suggested solution but I am having error in the process call it back and use it for predictions.

Error;
ValueError: DataFrame.dtypes for data must be int, float, bool or category. When
categorical type is supplied, DMatrix parameter enable_categorical must
be set to True.

Code;
`with open('xgboost_regr.pckl','wb') as pckl:
pickle.dump(f.regr,pckl)

filename = open("xgboost_regr.pckl", "rb")
xgb_new = pickle.load(filename)
filename.close()

pickle_pred = xgb_new.predict(X_test)`

X_test concludes only date time series that i want to make forecast.

from scalecast.

In scalecast, the model is using a whole dataframe of predictors to make forecasts, not just the date-type column. To see the dataframe that is generated for your model, you can run:

f.export_Xvars_df()

You will only be able to use this dataframe to predict one step into the future if you are using any AR terms. You will need to create a process that predicts future values of your AR terms to go more than one step out. But, I think the easiest way to get where you are going is to not pickle out the model and use the forecasts that scalecast produces after running auto_forecast() by default:

f.auto_forecast()
f.export("lvl_fcsts")

The forecast horizon you specified in your analysis when calling f.generate_future_dates() will determine how many forecast steps are exported this way. The reason the package both tests and forecasts with the model in the same line of code is to simplify the dynamic process used to make predictions.

from scalecast.

thanks for your response, it helped a lot. 💯

from scalecast.

Glad to help! Closing the issue.

from scalecast.