I am trying to do grid search over a sklearn pipeline that uses a custom transformer in a pipeline with FeatureUnion. It works fine when the pipeline uses the custom transformer class in FeatureUnion; however, it fails when the custom class is ignored in the pipeline by setting passthrough
in the grid search parameters.
The full pipeline is defined as follows:
from sklearn.feature_extraction.text import CountVectorizer, TfidfTransformer
from sklearn.naive_bayes import MultinomialNB
from sklearn.pipeline import Pipeline, FeatureUnion
ngram_vectorizer = Pipeline([
("vectorizer", CountVectorizer(analyzer="char_wb", ngram_range=(1,3))),
("tfidf", TfidfTransformer())
])
pipe_full = Pipeline(
[
("features", FeatureUnion(
[
("ngrams", ngram_vectorizer),
("lengths", TextLengthExtractor())
]
)
),
("classifier", MultinomialNB())
]
)
The custom transformer class TextLengthExtractor
simply computes the number of characters from an input string:
from sklearn.base import BaseEstimator, TransformerMixin
class TextLengthExtractor(BaseEstimator, TransformerMixin):
def fit(self, X, y = None):
return self
def transform(self, X, y = None):
string_lengths = np.array([len(doc) for doc in X])
return string_lengths.reshape(-1,1)
The tuning parameters for grid search are defined through a dictionary params
. Importantly, the parameters for the custom TextLengthExtractor
contain the passthrough
option to ignore the entire features__lengths
step from the pipeline (see also the sklearn's documentation on pipelines):
params = {
"features__lengths": [TextLengthExtractor(), "passthrough"],
"features__ngrams__vectorizer__ngram_range" : [(1,3), (2,6)],
}
When the pipeline is fit on the following dummy data
X_train_dummy = ["a", "ab", "a bc", "aaaaa", "b ab cc b", "ba", "baba", "cc bb aa", "c", "bca"]
y_train_dummy = [1,0,1, 1, 0, 1, 0, 1, 0, 0]
pipe_full.fit(X_train_dummy, y_train_dummy)
it can be seen that the lengths
step of the FeatureUnion
pipeline works as expected:
pipe_full["features"].get_params()["lengths"].transform(X_train_dummy)
# gives the following output of shape (10,1)
# array([[1], [2], [4], [5], [9], [2], [4], [8], [1], [3]])
However - and now comes the problem - when grid search is performed as follows:
from sklearn.model_selection import GridSearchCV
grid_search = GridSearchCV(pipe_full, params, cv=5, n_jobs=-1, verbose=10)
grid_search.fit(X_train_dummy, y_train_dummy)
all fits that ignore the lengths
step (as defined by the passthrough
option from params["features__lengths"]
throw the following error:
5 fits failed out of a total of 10.
The score on these train-test partitions for these parameters will be set to nan.
If these failures are not expected, you can try to debug them by setting error_score='raise'.
Below are more details about the failures:
--------------------------------------------------------------------------------
5 fits failed with the following error:
Traceback (most recent call last):
File "C:\dev\NameClassification\venv\lib\site-packages\sklearn\model_selection\_validation.py", line 686, in _fit_and_score
estimator.fit(X_train, y_train, **fit_params)
File "C:\dev\NameClassification\venv\lib\site-packages\sklearn\pipeline.py", line 378, in fit
Xt = self._fit(X, y, **fit_params_steps)
File "C:\dev\NameClassification\venv\lib\site-packages\sklearn\pipeline.py", line 336, in _fit
X, fitted_transformer = fit_transform_one_cached(
File "C:\dev\NameClassification\venv\lib\site-packages\joblib\memory.py", line 349, in __call__
return self.func(*args, **kwargs)
File "C:\dev\NameClassification\venv\lib\site-packages\sklearn\pipeline.py", line 870, in _fit_transform_one
res = transformer.fit_transform(X, y, **fit_params)
File "C:\dev\NameClassification\venv\lib\site-packages\sklearn\pipeline.py", line 1162, in fit_transform
return self._hstack(Xs)
File "C:\dev\NameClassification\venv\lib\site-packages\sklearn\pipeline.py", line 1216, in _hstack
Xs = sparse.hstack(Xs).tocsr()
File "C:\dev\NameClassification\venv\lib\site-packages\scipy\sparse\_construct.py", line 532, in hstack
return bmat([blocks], format=format, dtype=dtype)
File "C:\dev\NameClassification\venv\lib\site-packages\scipy\sparse\_construct.py", line 665, in bmat
raise ValueError(msg)
ValueError: blocks[0,:] has incompatible row dimensions. Got blocks[0,1].shape[0] == 1, expected 8.
I do understand that both steps require identical row dimensions for both ngrams
and lengths
in the FeatureUnion
, where the number of rows in the extracted feature matrices must equal the number of samples in the respective split. However, I have no idea how to control the shape of matrices when ignoring the lengths
part of FeatureUnion
using the passthrough
option in the gird search params.
I have found any solution to the problem on SE or any other sklearn related resource. Does anyone have an idea on how to solve the issue?
I think I found the solution to the problem: To ignore an individual step in a FeatureUnion
, the string drop
rather than passthrough
must be used. According to sklearn's documentation of FeatureUnion:
Parameters of the transformers may be set using its name and the parameter name separated by a '__'. A transformer may be replaced entirely by setting the parameter with its name to another transformer, removed by setting to 'drop' or disabled by setting to 'passthrough' (features are passed without transformation).
An example of dropping an entire transformer in FeatureUnion
is also shown in sklearn's user guide on pipelines.
In conclusion, to solve my problem, I had to replace passthrough
with drop
in the grid search parameter dictionary as follows
Change from
params = {
"features__lengths": [TextLengthExtractor(), "passthrough"],
"features__ngrams__vectorizer__ngram_range" : [(1,3), (2,6)],
}
to
params = {
"features__lengths": [TextLengthExtractor(), "drop"],
"features__ngrams__vectorizer__ngram_range" : [(1,3), (2,6)],
}