RF: allow for run-to-run difference in SVD results

matthew-brett · matthew-brett · commit 91fddffbae25 · 2014-04-30T15:33:59.000-07:00
On 32-bit, exploring SVD seems to confirm that the results differ
between runs at around 10**-10 precision.  Allow for this.
diff --git a/nipy/algorithms/diagnostics/tests/test_screen.py b/nipy/algorithms/diagnostics/tests/test_screen.py
@@ -128,8 +128,10 @@ def test_screen_slice_axis():
         # Now the analysis works without warning
         res = screen(explicit_img)
         # And is the expected analysis
-        assert_array_equal(pca_pos(res['pca'].get_data()),
-                           pca_pos(exp_res['pca'].get_data()))
+        # Very oddly on scipy 0.9 32 bit - at least - results differ between
+        # runs, so we need assert_almost_equal
+        assert_almost_equal(pca_pos(res['pca'].get_data()),
+                            pca_pos(exp_res['pca'].get_data()))
         assert_array_equal(res['ts_res']['slice_mean_diff2'],
                            exp_res['ts_res']['slice_mean_diff2'])
         # Turn off warnings, also get expected analysis
