Support for float32 in KDTree and BallTree #15474
Description
Description
The conversion from float32 to float64 in Isomap leads to cython class BinaryTree
Steps/Code to Reproduce
Debug example:
from sklearn.datasets import load_digits
from sklearn.manifold import Isomap
X, _ = load_digits(return_X_y=True)
X = X.astype('float32')
embedding = Isomap(n_components=2)
X_transformed = embedding.fit_transform(X[:100])The change of data type occurs in line 157 of isomap when calling function:
kng = kneighbors_graph(self.nbrs_, self.n_neighbors,
metric=self.metric, p=self.p,
metric_params=self.metric_params,
mode='distance', n_jobs=self.n_jobs)which leads to class KNeighborsMixin in _base.py, method kneighbors (line 531):
chunked_results = Parallel(n_jobs, **parallel_kwargs)(
delayed_query(
self._tree, X[s], n_neighbors, return_distance)
for s in gen_even_slices(X.shape[0], n_jobs)
)here, X is dtype float32, chunked_results is a list of arrays dtype float64
The problem arises in the method query of class BinaryTree in _binary_tree.pxi (line 1271). Data are casted into either DTYPE or DTYPE_t, which are defined as np.float64.
class BinaryTree should be changed to allow computations in both float32 and float64 data types