MLPredictor runs models

v1.84. MLPredictor lets you run ML models on dynamic datasets and render the output in a FormHandler-like interface.

You should use MLPredictor if

• You’ve already built a model. You can use any library that supports a scikit-learn compatible .predict() function. You must save it as model.pkl
• You want to apply it to a dynamic data source, e.g. a database or file that’s refreshed daily

Let’s create a sample classifier model with this iris.csv and save it in model.pkl:

import joblib
import pandas as pd
from sklearn.neighbors import KNeighborsClassifier

data = pd.read_csv('iris.csv')
model = KNeighborsClassifier()
model.fit(data.drop(columns=['species']), data['species'])
joblib.dump(model, 'model.pkl')

This model.pkl can be used to predict data from any file (including the original iris.csv) with this gramex.yaml config:

url:
  mlpredictor-iris:
    pattern: /iris
    handler: MLPredictor
    kwargs:
      config_dir: $YAMLPATH
      data:
        url: iris.csv
      columns: [sepal_length, sepal_width, petal_length, petal_width]
      target_col: species

Now, visit /iris?petal_length>=6&_format=html. This will run the predictions for all petal lengths > 6, and show:

The kwargs: MLPredictor accepts are:

• config_dir: The directory containing a model.pkl file. The file MUST be called model.pkl. This is consistent with MLHandler
• data: The location of the data to predict from. This can contain any dynamic database dynamic file. E.g.:
  • data: {url: /path/to/file.xlsx}
  • data: {url: 'postgresql://$USER:$PASS@server/db'; table: public.sales}
• columns: List of input columns from data to pass to the model. If the model was created with MLHandler, this is optional.
• target_col: Optional name of the predicted column. If the model was created with MLHandler, it picks it from the training dataset. Else defaults to prediction.

All FormHandler filters are applicable, making MLPredictor an good way to predict for different subsets. For example:

• /mlpredictor?date=2022-22-02 – predict all results for a specific date
• /mlpredictor?sort=-date&_limit=30 – predict the 30 most recent results
• /mlpredictor?city=Rome – predict only for the city of Rome