Fit_transform() vs. transform()

[EngWiPy]

Hi,

I noticed that in some cases we first call fit_transform(), and afterwards we call transform(). Like in the following example:

from sklearn.preprocessing import PolynomialFeatures

X_train = np.array([6, 8, 10, 14, 18]).reshape(-1, 1)
X_test = np.array([6, 8, 11, 16]).reshape(-1, 1)

quadratic_featurizer = PolynomialFeatures(degree = 2)

X_train_qudratic = quadratic_featurizer.fit_transform(X_train)
X_test_qudratic = quadratic_featurizer.transform(X_test)

Why? What is the difference between the two methods?

Thanks

 

[Smile Say Hello]

I haven't used PolynomialFeatures before but fit(), fit_transform(), and transform() are standard methods is scikit-learn.

fit_transform() is essentially the same as calling fit() and then transform() - so is like a shortcut for two commands in one if you wish.

So when you do X_train_qudratic = quadratic_featurizer.fit_transform(X_train) what you are doing is fitting quadratic_featurizer on X_train and using it to transform X_train itself. This should be equal to (and is a shorthand for):

quadratic_featurizer.fit(X_train)
X_train_qudratic = quadratic_featurizer.transform(X_train)

On the other hand, when you do X_test_qudratic = quadratic_featurizer.transform(X_test) you are using a previously fitted quadratic_featurizer to transform X_test. This should fail unless you have previously called either .fit() or .fit_transform on quadratic_featurizer.

Hope it makes sense.

I am guessing what you are trying to do is actually:
quadratic_featurizer.fit(X_train)
X_test_qudratic = quadratic_featurizer.transform(X_test)

Although what you did, e.g.:
X_train_qudratic = quadratic_featurizer.fit_transform(X_train)
X_test_qudratic = quadratic_featurizer.transform(X_test)

will also work but you are unnecessarily transforming X_train by calling fit_transform() instead of fit().

 

[EngWiPy]

I haven't used PolynomialFeatures before but fit(), fit_transform(), and transform() are standard methods is scikit-learn.

fit_transform() is essentially the same as calling fit() and then transform() - so is like a shortcut for two commands in one if you wish.

So when you do X_train_qudratic = quadratic_featurizer.fit_transform(X_train) what you are doing is fitting quadratic_featurizer on X_train and using it to transform X_train itself. This should be equal to (and is a shorthand for):

quadratic_featurizer.fit(X_train)
X_train_qudratic = quadratic_featurizer.transform(X_train)

On the other hand, when you do X_test_qudratic = quadratic_featurizer.transform(X_test) you are using a previously fitted quadratic_featurizer to transform X_test. This should fail unless you have previously called either .fit() or .fit_transform on quadratic_featurizer.

Hope it makes sense.

I am guessing what you are trying to do is actually:
quadratic_featurizer.fit(X_train)
X_test_qudratic = quadratic_featurizer.transform(X_test)

Although what you did, e.g.:
X_train_qudratic = quadratic_featurizer.fit_transform(X_train)
X_test_qudratic = quadratic_featurizer.transform(X_test)

will also work but you are unnecessarily transforming X_train by calling fit_transform() instead of fit().

It makes sense. I did X_train_qudratic = quadratic_featurizer.fit_transform(X_train) because later in my code I use X_train_quadratic to train a model using .fit() and then test the performance of the model on X_test_quadratic.

I have one question: the method .fit_transform() fits what to the training data X_train? For example if
X_train = [1
2
3
4]
the quadratic_featurizer.fit_transform(X_train) will result in
[ 1 1 1
1 2 4
1 3 9
1 4 16]
which is basically the value of the independent variable x_1 in the polynomial equation
$$y = \beta_0 + \beta_1x_1 + \beta_2x_1^2$$

In this case .fit_transform() fits what to X_train?

Thanks

 

[Smile Say Hello]

It makes sense. I did X_train_qudratic = quadratic_featurizer.fit_transform(X_train) because later in my code I use X_train_quadratic to train a model using .fit() and then test the performance of the model on X_test_quadratic.

I have one question: the method .fit_transform() fits what to the training data X_train? For example if
X_train = [1
2
3
4]
the quadratic_featurizer.fit_transform(X_train) will result in
[ 1 1 1
1 2 4
1 3 9
1 4 16]
which is basically the value of the independent variable x_1 in the polynomial equation
$$y = \beta_0 + \beta_1x_1 + \beta_2x_1^2$$

In this case .fit_transform() fits what to X_train?

Thanks

 

[Smile Say Hello]

It looks like it is not actually fitting anything-- I think it is called fit_transform() simply because scikit-learn tries to provide a uniform interface and a lot of other modules in scikit-learn use the same terminology. What fit() and the fit part of fit_transform() seems to do is simply determine the combinations of features it needs to return for the given input shape. So when you later call transform many times, it can skip that part and simply return the values.

So, in this case the fit() part figures that it is a single feature and x^0,x^1, and x^2 need to be returned and the transform() part simply returns them for each sample on that basis.

 

[EngWiPy]

It looks like it is not actually fitting anything-- I think it is called fit_transform() simply because scikit-learn tries to provide a uniform interface and a lot of other modules in scikit-learn use the same terminology. What fit() and the fit part of fit_transform() seems to do is simply determine the combinations of features it needs to return for the given input shape. So when you later call transform many times, it can skip that part and simply return the values.

So, in this case the fit() part figures that it is a single feature and x^0,x^1, and x^2 need to be returned and the transform() part simply returns them for each sample on that basis.

Thanks for your replies. It is more clear now.