[python package]: suggestion: lgb.Booster.predict() should check that the input X data makes sense

[j-mark-hou]

In particular, I'm thinking about these things:

1. if the input is an np.array, check that the columns is the same as the number of features the lgb.Booster object uses. if not, throw a warning.
2. if the input is a pd.Dataframe object, should check that the feature_names of the lgb.Booster object is a superset of the columns of the pd.Dataframe
3. if feature names in the booster object are repeated, or if column names in the pd.Dataframe are repeated, fall back to 0.

if these things sound reasonable I'd be happy to add these checks to the lgb.Booster.predict() function prior to calling the lgb._InnerPredictor.predict()

[guolinke]

@j-mark-hou sure, very happy to see this feature.

[guolinke]

@j-mark-hou any updates ?

[j-mark-hou]

I think maybe this should be rolled into a more systematic rewrite of the pandas api. There's some current things about the implementation that I don't quite understand, and unfortunately I don't currently have the time to dig deeper into this. Sorry, I'll let you know if I revisit this at some point in the future.

[arsenyinfo]

Currently, it leads to some kind of inconsistency:

In [1]: from lightgbm import LGBMClassifier, Dataset, train
   ...: import numpy as np
   ...:
   ...: x_data, y_data = np.random.rand(1000, 100), np.random.rand(1000) > .5
   ...: x_bad = np.random.rand(1000, 101)
   ...:
   ...:
   ...: def sklearn_style():
   ...:     clf = LGBMClassifier()
   ...:     clf.fit(x_data, y_data)
   ...:     return clf.predict(x_bad)
   ...:
   ...:
   ...: def xgboost_style():
   ...:     dataset = Dataset(x_data, y_data)
   ...:     params = {'application': 'binary'}
   ...:     booster = train(params, dataset)
   ...:     return booster.predict(x_data)
   ...:

In [2]: sklearn_style().shape
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-2-281be6aeee0d> in <module>()
----> 1 sklearn_style().shape

<ipython-input-1-adc7de9c0c86> in sklearn_style()
      9     clf = LGBMClassifier()
     10     clf.fit(x_data, y_data)
---> 11     return clf.predict(x_bad)
     12
     13

~/.pyenv/versions/3.6.2/lib/python3.6/site-packages/lightgbm/sklearn.py in predict(self, X, raw_score, num_iteration)
    674
    675     def predict(self, X, raw_score=False, num_iteration=0):
--> 676         class_probs = self.predict_proba(X, raw_score, num_iteration)
    677         class_index = np.argmax(class_probs, axis=1)
    678         return self._le.inverse_transform(class_index)

~/.pyenv/versions/3.6.2/lib/python3.6/site-packages/lightgbm/sklearn.py in predict_proba(self, X, raw_score, num_iteration)
    704                              "match the input. Model n_features_ is %s and "
    705                              "input n_features is %s "
--> 706                              % (self._n_features, n_features))
    707         class_probs = self.booster_.predict(X, raw_score=raw_score, num_iteration=num_iteration)
    708         if self._n_classes > 2:

ValueError: Number of features of the model must match the input. Model n_features_ is 100 and input n_features is 101

In [3]: xgboost_style().shape
[LightGBM] [Info] Number of positive: 506, number of negative: 494
[LightGBM] [Info] Total Bins 25500
[LightGBM] [Info] Number of data: 1000, number of used features: 100
[LightGBM] [Info] Finished loading 100 models
Out[3]: (1000,)

Imho, check on feature count is really important, and this lack of assertion may lead to harsh, not so easy to debug issues.

[StrikerRUS]

Closed in favor of being in #2302. We decided to keep all feature requests in one place.

Welcome to contribute this feature! Please re-open this issue (or post a comment if you are not a topic starter) if you are actively working on implementing this feature.

[jsh9]

I may be interested helping improve this. What would be the ideal behavior here? Does LightGBM make predictions only based on column order (like sklearn), or based on column names?

[StrikerRUS]

The shape of data for prediction is now checked at cpp side, thanks to #2464.

The next steps may be to check the type and order of features, and their names in case of pandas DataFrame.