Merge pull request #431 from matthew-brett/fix-numpy-1.13-breakage

matthew-brett · web-flow · commit f0f35adc803f · 2018-01-15T09:35:51.000Z
MRG: new numpy protects polynomial coefficients New numpy returns a copy of the polynomial coefficients from poly.c, rather than the actual coefficent array, so when we were modifying the polynomial coefficients, the polynomial object wasn't seeing our changes, leading to errors in the RFT module - see e.g. https://nipy.bic.berkeley.edu/builders/nipy-py2.7-osx-10.10/builds/52/steps/shell_9/logs/stdio Numpy issue raised on numpy-discussion mailing list.
diff --git a/nipy/algorithms/statistics/rft.py b/nipy/algorithms/statistics/rft.py
@@ -87,14 +87,14 @@ def Q(dim, dfd=np.inf):
     j = dim
     if j <= 0:
         raise ValueError('Q defined only for dim > 0')
-    poly = hermitenorm(j-1)
-    poly = np.poly1d(np.around(poly.c))
+    coeffs = np.around(hermitenorm(j - 1).c)
     if np.isfinite(m):
-        for l in range((j-1)//2+1):
-            f = np.exp(gammaln((m+1)/2.) - gammaln((m+2-j+2*l)/2.)
-                                - 0.5*(j-1-2*l)*(np.log(m/2.)))
-            poly.c[2*l] *= f
-    return np.poly1d(poly.c)
+        for L in range((j - 1) // 2 + 1):
+            f = np.exp(gammaln((m + 1) / 2.)
+                       - gammaln((m + 2 - j + 2 * L) / 2.)
+                       - 0.5 * (j - 1 - 2 * L) * (np.log(m / 2.)))
+            coeffs[2 * L] *= f
+    return np.poly1d(coeffs)
 
 
 class ECquasi(np.poly1d):
diff --git a/nipy/algorithms/statistics/tests/test_rft.py b/nipy/algorithms/statistics/tests/test_rft.py
@@ -400,16 +400,16 @@ def test_hotelling2():
         chi = rft.ChiSquared(dfn=dfn)(x)
         assert_almost_equal(h, chi)
         chi2 = scipy.stats.chi2.sf(x, dfn)
-        yield assert_almost_equal, h, chi2
+        assert_almost_equal(h, chi2)
         # XXX - p appears to be unused
         p = rft.spherical_search(dfn)
         for dfd in [40,50]:
             fac = (dfd-dfn+1.)/(dfd*dfn)
             h = rft.Hotelling(dfd=dfd,k=dfn)(x)
             f = scipy.stats.f.sf(x*fac, dfn, dfd-dfn+1)
             f2 = rft.FStat(dfd=dfd-dfn+1,dfn=dfn)(x*fac)
-            yield assert_almost_equal, f2, f
-            yield assert_almost_equal, h, f
+            assert_almost_equal(f2, f)
+            assert_almost_equal(h, f)
 
 
 @dec.slow
