check the shape for mat, csr and csc in prediction

[guolinke]

partially improve #812

[StrikerRUS]

Close-reopen for CI.

[StrikerRUS]

Can something similar be done for predictions for data from a file?

[StrikerRUS]

Please add a simple test from the original issue:

import numpy as np
import lightgbm as lgb

x_data = np.random.rand(100, 10)
x_bad_data = np.random.rand(100, 11)
y_data =  np.random.rand(100) > .5
self.assertNotEqual(x_data.shape[-1], x_bad_data.shape[-1])
train_dataset = lgb.Dataset(x_data, y_data)
bst = lgb.train({'objective': 'binary'}, train_dataset)
with np.testing.assert_raises_regex(lgb.basic.LightGBMError,
                                    'The number of features in data*'):
    bst.predict(x_bad_data)

[guolinke]

There are more changes than I expected. So I think we should test all cases, including the the mat, libsvm file and CSR format.
@StrikerRUS could you help for the test cases?

[StrikerRUS]

@guolinke

could you help for the test cases?

Sure!

What do you think about changing the original type to avoid casting? #2464 (comment)

Does this PR already include that check?

[guolinke]

Does this PR already include that check?

Yeah, as the zero-based and one-based libsvm format have the different number of columns.

[StrikerRUS]

@guolinke I added tests in the latest commit.

As we do not support a case when data for prediction is a list of arrays, I tried to check it for validation. And it seems that validation shape is not covered by this PR, right?

bad_valid_data = train_data.create_valid(bad_X_test, label=y_test)
bst.add_valid(bad_valid_data, "valid_bad")
bst.eval_valid()  # no error risen

[guolinke]

yeah, validation is not converted for now.

[guolinke]

due the need of pr #2485, I will merge this