scikit-learn: machine learning in Python. Please feel free to ask specific questions about scikit-learn. Please try to keep the discussion focused on scikit-learn usage and immediately related open source projects from the Python ecosystem.
1 reply
As far as I can understand, this issue requires that the documentation be updated, does that indicate the docstring within the function definition only, or is that referring to another piece of documentation?
One of the commentators on the issue also mentions ensuring there are tests that break if this documentation doesn't exist, how do I go about doing that effectively?
I have a general question: If for my dataset a kneighbor classifier works well (compared to e.g. SVC and Random Forest), are there other classifiers that might also work equally well?
I think it will depend on the data set. It also depends on how you are pre-processing your data. So kinda hard to say without knowing more.
I need some help please.
I want to serve an onnx model.
Input = 144 columns ( medical records, some categoricals, some not ).
Output = classification.
Pipeline = StandardScaler + LabelEncoder + LightGBM.
I am stuck with the LabelEncoder. Any example of such configuration somewhere ? Google was not my friend.
I was able to produce an onnx model when bypassing the LabelEncoder... but I need it and want to avoid 1HE because LightGBM performs much better without 1HE.
Anyone ?
Also tree based models it's better to use OrdinalEncoder instead for categorical features
I'm not sure that's true, using OE will make the trees treat categories as ordered values, but they're not. Native categorical support (as in LightGBM) properly treats categories as un-ordered and can yield the same splits with less tree depth
OrdinalEncoder is probably the pragmatic solution. OneHotEncoder is only efficient if you use sparse output which are currently not supported by ONNX as far as I know.
@citron also you said "Pipeline = StandardScaler + LabelEncoder + LightGBM." but I assume you use a column transformer to separate to only scale the numerical features and encode the categorical feature separately: https://scikit-learn.org/stable/modules/compose.html#columntransformer-for-heterogeneous-data
BTW, StandardScaling the numerical features if often useless for tree-based models in general, and even more so for implementations such as LightGBM than bin the features.
I have tried plenty of frameworks and among them, lightGBM was the best. Now, as I need to export the model and the pipeline in ONNX/ONNX-ML format, I need to wrap lightGBM in something to keep the pipeline around.
For the categorical columns, try to use OrdinalEncoder. In 0.24+ we have better support for unknown categories at test time:
https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.OrdinalEncoder.html
Although I am not sure that sklearn-onnx has replicated that feature yet.
convert_sklearn prompts an error : 'numpy.bool' object has no attribute 'encode'StringTensorTypeStringTensorType
Any idea please ?
2 replies
Hello,
I'm trygin to use SimpleImputer(strategy='most_frequent') in Pipeline on dataframe with ~ 1.5 M samples but I take a lot of time
Is it normal ? If so, are there some alternatives to solve this issue ?
def vectorizer_df(input_data, categorical_cols, numerical_cols):
categorical_pipe = Pipeline([
('imputer', SimpleImputer(strategy='most_frequent'))
])
numerical_pipe = Pipeline([
('imputer', SimpleImputer(strategy='median')),
('bucketizer', KBinsDiscretizer(n_bins=10, strategy='uniform', encode='ordinal')) # ordinal
])
preprocessing = ColumnTransformer(
[('cat', categorical_pipe, categorical_cols),
('num', numerical_pipe, numerical_cols)
])
vectorizer_pipeline = Pipeline([
('vectorize', preprocessing)
])
return vectorizer_pipeline.fit_transform(input_data)Thanks
4 replies
5 replies