Reduce numpy's import times by delaying a few time consuming imports …

…to the point of actual use and global instantiations of finfo. Thanks to David Cournapeau for tracking down and fixing the import part of the problem.

numpy/lib/_datasource.py

@@ -37,7 +37,6 @@
 import os
 import tempfile
 from shutil import rmtree
-from urllib2 import urlopen, URLError
 from urlparse import urlparse
 
 # TODO: .zip support, .tar support?
@@ -196,6 +195,9 @@ def _cache(self, path):
         Creates a copy of the file in the datasource cache.
 
         """
+        # We import these here because importing urllib2 is slow and
+        # a significant fraction of numpy's total import time.
+        from urllib2 import urlopen, URLError
 
         upath = self.abspath(path)
 
@@ -337,6 +339,9 @@ def exists(self, path):
             is accessible if it exists in either location.
 
         """
+        # We import this here because importing urllib2 is slow and
+        # a significant fraction of numpy's total import time.
+        from urllib2 import URLError
 
         # Test local path
         if os.path.exists(path):

numpy/lib/getlimits.py

@@ -21,14 +21,16 @@ def _frz(a):
     }
 
 class finfo(object):
-    """Machine limits for floating point types.
+    """ Machine limits for floating point types.
 
-    :Parameters:
-        dtype : floating point type or instance
-
-    :SeeAlso:
-      - numpy.lib.machar.MachAr
+    Parameters
+    ----------
+    dtype : floating point type, dtype, or instance
+        The kind of floating point data type to get information about.
 
+    See Also
+    --------
+    numpy.lib.machar.MachAr
     """
 
     _finfo_cache = {}

numpy/lib/polynomial.py

@@ -17,8 +17,6 @@
 from numpy.lib.function_base import trim_zeros, sort_complex
 eigvals = None
 lstsq = None
-_single_eps = finfo(NX.single).eps
-_double_eps = finfo(NX.double).eps
 
 class RankWarning(UserWarning):
     """Issued by polyfit when Vandermonde matrix is rank deficient.
@@ -301,11 +299,7 @@ def polyfit(x, y, deg, rcond=None, full=False):
 
     # set rcond
     if rcond is None :
-        xtype = x.dtype
-        if xtype == NX.single or xtype == NX.csingle :
-            rcond = len(x)*_single_eps
-        else :
-            rcond = len(x)*_double_eps
+        rcond = len(x)*finfo(x.dtype).eps
 
     # scale x to improve condition number
     scale = abs(x).max()

numpy/lib/tests/test__datasource.py

@@ -5,6 +5,7 @@
 from tempfile import mkdtemp, mkstemp, NamedTemporaryFile
 from shutil import rmtree
 from urlparse import urlparse
+from urllib2 import URLError
 
 from numpy.testing import *
 
@@ -16,7 +17,7 @@ def urlopen_stub(url, data=None):
         tmpfile = NamedTemporaryFile(prefix='urltmp_')
         return tmpfile
     else:
-        raise datasource.URLError('Name or service not known')
+        raise URLError('Name or service not known')
 
 # Rebind urlopen during testing.  For a 'real' test, uncomment the rebinding
 # below.

numpy/lib/utils.py

@@ -1,11 +1,9 @@
-import compiler
 import os
 import sys
-import inspect
 import pkgutil
 import types
 import re
-import pydoc
+
 from numpy.core.numerictypes import obj2sctype, generic
 from numpy.core.multiarray import dtype as _dtype
 from numpy.core import product, ndarray
@@ -327,7 +325,8 @@ def info(object=None,maxwidth=76,output=sys.stdout,toplevel='numpy'):
                 p[0]*(x**N-1) + p[1]*(x**N-2) + ... + p[N-2]*x + p[N-1]
     """
     global _namedict, _dictlist
-    import pydoc
+    # Local import to speed up numpy's import time.
+    import pydoc, inspect
 
     if hasattr(object,'_ppimport_importer') or \
        hasattr(object, '_ppimport_module'):
@@ -467,6 +466,8 @@ def info(object=None,maxwidth=76,output=sys.stdout,toplevel='numpy'):
 def source(object, output=sys.stdout):
     """Write source for this object to output.
     """
+    # Local import to speed up numpy's import time.
+    import inspect
     try:
         print >> output,  "In file: %s\n" % inspect.getsourcefile(object)
         print >> output,  inspect.getsource(object)
@@ -599,6 +600,8 @@ def _lookfor_generate_cache(module, import_modules, regenerate):
 
     """
     global _lookfor_caches
+    # Local import to speed up numpy's import time.
+    import inspect
 
     if module is None:
         module = "numpy"
@@ -751,6 +754,8 @@ def safe_eval(source):
       ...
     SyntaxError: Unknown name: dict
     """
+    # Local import to speed up numpy's import time.
+    import compiler
     walker = SafeEval()
     try:
         ast = compiler.parse(source, "eval")

numpy/ma/core.py

@@ -77,7 +77,6 @@
 MaskType = np.bool_
 nomask = MaskType(0)
 
-divide_tolerance = np.finfo(float).tiny
 np.seterr(all='ignore')
 
 def doc_note(note):
@@ -398,9 +397,14 @@ def __call__ (self, x):
 #............................
 class _DomainSafeDivide:
     """Define a domain for safe division."""
-    def __init__ (self, tolerance=divide_tolerance):
+    def __init__ (self, tolerance=None):
         self.tolerance = tolerance
     def __call__ (self, a, b):
+        # Delay the selection of the tolerance to here in order to reduce numpy
+        # import times. The calculation of these parameters is a substantial
+        # component of numpy's import time.
+        if self.tolerance is None:
+            self.tolerance = np.finfo(float).tiny
         return umath.absolute(a) * self.tolerance >= umath.absolute(b)
 #............................
 class _DomainGreater:

numpy/ma/extras.py

@@ -39,7 +39,7 @@
 from numpy import ndarray, array as nxarray
 import numpy.core.umath as umath
 from numpy.lib.index_tricks import AxisConcatenator
-from numpy.lib.polynomial import _lstsq, _single_eps, _double_eps
+from numpy.lib.polynomial import _lstsq
 
 #...............................................................................
 def issequence(seq):
@@ -866,10 +866,7 @@ def polyfit(x, y, deg, rcond=None, full=False):
         x[m] = y[m] = masked
     # Set rcond
     if rcond is None :
-        if x.dtype in (np.single, np.csingle):
-            rcond = len(x)*_single_eps
-        else :
-            rcond = len(x)*_double_eps
+        rcond = len(x)*np.finfo(x.dtype).eps
     # Scale x to improve condition number
     scale = abs(x).max()
     if scale != 0 :