[FIXED] How to instantiate a scikit-learn model from a string variable?

Issue

I have a list of models as parameters in a YAML file. When I read them into my Python script, they appear there as strings:

my_model = "LinearRegression()"
print(type(my_model)) #output: <class 'str'>

I want to initialize my sklearn model as:

model = LinearRegression()

but with LinearRegression() being a string variable.

At the moment, naturally, model = my_model is a string, and can't be fitted, etc.

How to instantiate the sklearn model from a string variable?

My use case: I want to put different models in a YAML file as params in order to sweep over them in the model fine-tuning procedure (a simple AutoML).

Solution

You have few options:

1. Use eval as a quick&dirty solution:

from sklearn.linear_model import LinearRegression
model_str = 'LinearRegression()'
model = eval(model_str)
print(model)

Note: using eval is not a safe because it will execute any string so if you have no control about a string variable being executed a malicious code may be executed.

Also you need to import exact the same class before using eval. Without the first line the code will not work.

2. use pickle.dumps

from sklearn.linear_model import LinearRegression

model_class_serialized = pickle.dumps(LinearRegression)
print(model_class_serialized)
model_class = pickle.loads(model_class_serialized)
print(model_class)

In that case you do not need explicitly import classes from sklearn.

An example of saving models to YAML:

import pickle
import yaml

from sklearn.linear_model import LinearRegression
from sklearn.ensemble import RandomForestClassifier
from sklearn.svm import LinearSVC

models = [LinearRegression, RandomForestClassifier, LinearSVC]

with open('models.yml', 'w') as out_file:
    yaml.dump({m.__name__: pickle.dumps(m) for m in models}, out_file)

After that you can read YAML file like that:

import pickle
import yaml

with open('models.yml', 'r') as in_file:
    models = yaml.load(in_file, Loader=yaml.FullLoader)
    models = {name: pickle.loads(serialized_cls) for name, serialized_cls in models.items()}

print(models)

Which produces:

{'LinearRegression': <class 'sklearn.linear_model._base.LinearRegression'>, 'LinearSVC': <class 'sklearn.svm._classes.LinearSVC'>, 'RandomForestClassifier': <class 'sklearn.ensemble._forest.RandomForestClassifier'>}

so you can easily instantiate these models:

model = models['RandomForestClassifier']()
print(model)
# model.fit(...)

Answered By - u1234x1234