Merge pull request #86 from matthew-brett/uncontroversial-snippets

Uncontroversial snippets

Some docstring updates

Adding a test for image array shapes too large to fit within header dimension dtypes (to avoid silent overflow)

Adding tests to confirm safe behavior when array scaling intercept or slope causes overflow when testing for clip values (test only)

Author: matthew-brett

nibabel/analyze.py

@@ -635,6 +635,10 @@ def set_data_shape(self, shape):
         dims[:] = 1
         dims[0] = ndims
         dims[1:ndims+1] = shape
+        # Check that dimensions fit
+        if not np.all(dims[1:ndims+1] == shape):
+            raise HeaderDataError('shape %s does not fit in dim datatype' %
+                                   shape)
         self._structarr['pixdim'][ndims+1:] = 1.0
 
     def get_base_affine(self):

nibabel/tests/test_analyze.py

@@ -25,6 +25,7 @@
 from ..nifti1 import Nifti1Header
 from ..loadsave import read_img_data
 from .. import imageglobals
+from ..casting import as_int
 
 from numpy.testing import (assert_array_equal,
                            assert_array_almost_equal)
@@ -217,6 +218,22 @@ def test_data_dtype(self):
                                 hdr.set_data_dtype,
                                 inp)
 
+    def test_shapes(self):
+        # Test that shape checks work
+        hdr = self.header_class()
+        for shape in ((2, 3, 4), (2, 3, 4, 5), (2, 3), (2,)):
+            hdr.set_data_shape(shape)
+            assert_equal(hdr.get_data_shape(), shape)
+        # Check max works, but max+1 raises error
+        dim_dtype = hdr.structarr['dim'].dtype
+        # as_int for safety to deal with numpy 1.4.1 int conversion errors
+        mx = as_int(np.iinfo(dim_dtype).max)
+        shape = (mx,)
+        hdr.set_data_shape(shape)
+        assert_equal(hdr.get_data_shape(), shape)
+        shape = (mx+1,)
+        assert_raises(HeaderDataError, hdr.set_data_shape, shape)
+
     def test_read_write_data(self):
         # Check reading and writing of data
         hdr = self.header_class()

nibabel/tests/test_utils.py

@@ -160,6 +160,34 @@ def test_array_to_file():
     assert_array_equal(data_back, np.zeros(arr.shape))
 
 
+def test_a2f_big_scalers():
+    # Check that clip works even for overflowing scalers / data
+    info = np.finfo(np.float32)
+    arr = np.array([info.min, np.nan, info.max], dtype=np.float32)
+    str_io = BytesIO()
+    # Intercept causes overflow - does routine scale correctly?
+    array_to_file(arr, str_io, np.int8, intercept=np.float32(2**120))
+    data_back = array_from_file(arr.shape, np.int8, str_io)
+    assert_array_equal(data_back, [-128, 0, 127])
+    # Scales also if mx, mn specified?
+    str_io.seek(0)
+    array_to_file(arr, str_io, np.int8, mn=info.min, mx=info.max,
+                  intercept=np.float32(2**120))
+    data_back = array_from_file(arr.shape, np.int8, str_io)
+    assert_array_equal(data_back, [-128, 0, 127])
+    # And if slope causes overflow?
+    str_io.seek(0)
+    array_to_file(arr, str_io, np.int8, divslope=np.float32(0.5))
+    data_back = array_from_file(arr.shape, np.int8, str_io)
+    assert_array_equal(data_back, [-128, 0, 127])
+    # with mn, mx specified?
+    str_io.seek(0)
+    array_to_file(arr, str_io, np.int8, mn=info.min, mx=info.max,
+                  divslope=np.float32(0.5))
+    data_back = array_from_file(arr.shape, np.int8, str_io)
+    assert_array_equal(data_back, [-128, 0, 127])
+
+
 def write_return(data, fileobj, out_dtype, *args, **kwargs):
     fileobj.truncate(0)
     array_to_file(data, fileobj, out_dtype, *args, **kwargs)

nibabel/volumeutils.py

@@ -696,6 +696,8 @@ def scale_min_max(mn, mx, out_type, allow_intercept):
     intercept : numpy scalar, dtype=np.maximum_sctype(np.float)
        value to subtract from data before dividing by scalefactor
 
+    Examples
+    --------
     >>> scale_min_max(0, 255, np.uint8, False)
     (1.0, 0.0)
     >>> scale_min_max(-128, 127, np.int8, False)
@@ -720,6 +722,9 @@ def scale_min_max(mn, mx, out_type, allow_intercept):
 
     Notes
     -----
+    We don't use this function anywhere in nibabel now, it's here for API
+    compatibility only.
+
     The large integers lead to python long types as max / min for type.
     To contain the rounding error, we need to use the maximum numpy
     float types when casting to float.