fix for np.count_nonzero not present in Numpy < 1.6

benchmarks/bench_sparsify.py

@@ -45,14 +45,15 @@
 
 from scipy.sparse.csr import csr_matrix
 import numpy as np
+from sklearn.utils.fixes import count_nonzero
 from sklearn.linear_model.stochastic_gradient import SGDRegressor
 from sklearn.metrics import r2_score
 
 np.random.seed(42)
 
 
 def sparsity_ratio(X):
- return np.count_nonzero(X) / float(n_samples * n_features)
+ return count_nonzero(X) / float(n_samples * n_features)
 
 n_samples, n_features = 5000, 300
 X = np.random.randn(n_samples, n_features)

doc/modules/performance.rst

@@ -95,9 +95,9 @@ an optimized BLAS implementation.
 
 Here is a sample code to test the sparsity of your input:
 
- >>> import numpy as np
+ >>> from sklearn.utils.fixes import count_nonzero
  >>> def sparsity_ratio(X):
- >>> return 1.0 - np.count_nonzero(X) / float(X.shape[0] * X.shape[1])
+ >>> return 1.0 - count_nonzero(X) / float(X.shape[0] * X.shape[1])
  >>> print("input sparsity ratio:", sparsity_ratio(X))
 
 As a rule of thumb you can consider that if the sparsity ratio is greater
@@ -219,7 +219,7 @@ compromise between model compactness and prediction power. One can also
 further tune the ``l1_ratio`` parameter (in combination with the
 regularization strength ``alpha``) to control this tradeoff.
 
-A typical `benchmark <https://github.com/scikit-learn/scikit-learn/tree/master/benchmarks/bench_sparsify.py>`_
+A typical `benchmark <https://github.com/scikit-learn/scikit-learn/tree/masternchmarks/bench_sparsify.py>`_
 on synthetic data yields a >30% decrease in latency when both the model and
 input are sparsed (with 0.000024 and 0.027400 non-zero coefficients ratio
 respectively). Your mileage may vary depending on the sparsity and size of