Merge pull request #1096 from effigies/enh/dtype_aliases

ENH: Add static and dynamic dtype aliases to NIfTI images

Author: effigies

nibabel/nifti1.py

@@ -898,26 +898,28 @@ def set_data_dtype(self, datatype):
         >>> hdr.set_data_dtype(np.dtype(np.uint8))
         >>> hdr.get_data_dtype()
         dtype('uint8')
-        >>> hdr.set_data_dtype('implausible') #doctest: +IGNORE_EXCEPTION_DETAIL
+        >>> hdr.set_data_dtype('implausible')
         Traceback (most recent call last):
            ...
-        HeaderDataError: data dtype "implausible" not recognized
-        >>> hdr.set_data_dtype('none') #doctest: +IGNORE_EXCEPTION_DETAIL
+        nibabel.spatialimages.HeaderDataError: data dtype "implausible" not recognized
+        >>> hdr.set_data_dtype('none')
         Traceback (most recent call last):
            ...
-        HeaderDataError: data dtype "none" known but not supported
-        >>> hdr.set_data_dtype(np.void) #doctest: +IGNORE_EXCEPTION_DETAIL
+        nibabel.spatialimages.HeaderDataError: data dtype "none" known but not supported
+        >>> hdr.set_data_dtype(np.void)
         Traceback (most recent call last):
            ...
-        HeaderDataError: data dtype "<type 'numpy.void'>" known but not supported
-        >>> hdr.set_data_dtype('int') #doctest: +IGNORE_EXCEPTION_DETAIL
+        nibabel.spatialimages.HeaderDataError: data dtype "<class 'numpy.void'>" known
+        but not supported
+        >>> hdr.set_data_dtype('int')
         Traceback (most recent call last):
            ...
         ValueError: Invalid data type 'int'. Specify a sized integer, e.g., 'uint8' or numpy.int16.
-        >>> hdr.set_data_dtype(int) #doctest: +IGNORE_EXCEPTION_DETAIL
+        >>> hdr.set_data_dtype(int)
         Traceback (most recent call last):
            ...
-        ValueError: Invalid data type 'int'. Specify a sized integer, e.g., 'uint8' or numpy.int16.
+        ValueError: Invalid data type <class 'int'>. Specify a sized integer, e.g., 'uint8' or
+        numpy.int16.
         >>> hdr.set_data_dtype('int64')
         >>> hdr.get_data_dtype() == np.dtype('int64')
         True
@@ -1799,6 +1801,10 @@ class Nifti1Pair(analyze.AnalyzeImage):
     _meta_sniff_len = header_class.sizeof_hdr
     rw = True
 
+    # If a _dtype_alias has been set, it can only be resolved by inspecting
+    # the data at serialization time
+    _dtype_alias = None
+
     def __init__(self, dataobj, affine, header=None,
                  extra=None, file_map=None, dtype=None):
         # Special carve-out for 64 bit integers
@@ -2043,6 +2049,137 @@ def set_sform(self, affine, code=None, **kwargs):
             else:
                 self._affine[:] = self._header.get_best_affine()
 
+    def set_data_dtype(self, datatype):
+        """ Set numpy dtype for data from code, dtype, type or alias
+
+        Using :py:class:`int` or ``"int"`` is disallowed, as these types
+        will be interpreted as ``np.int64``, which is almost never desired.
+        ``np.int64`` is permitted for those intent on making poor choices.
+
+        The following aliases are defined to allow for flexible specification:
+
+          * ``'mask'`` - Alias for ``uint8``
+          * ``'compat'`` - The nearest Analyze-compatible datatype
+            (``uint8``, ``int16``, ``int32``, ``float32``)
+          * ``'smallest'`` - The smallest Analyze-compatible integer
+            (``uint8``, ``int16``, ``int32``)
+
+        Dynamic aliases are resolved when ``get_data_dtype()`` is called
+        with a ``finalize=True`` flag. Until then, these aliases are not
+        written to the header and will not persist to new images.
+
+        Examples
+        --------
+        >>> ints = np.arange(24, dtype='i4').reshape((2,3,4))
+
+        >>> img = Nifti1Image(ints, np.eye(4))
+        >>> img.set_data_dtype(np.uint8)
+        >>> img.get_data_dtype()
+        dtype('uint8')
+        >>> img.set_data_dtype('mask')
+        >>> img.get_data_dtype()
+        dtype('uint8')
+        >>> img.set_data_dtype('compat')
+        >>> img.get_data_dtype()
+        'compat'
+        >>> img.get_data_dtype(finalize=True)
+        dtype('<i4')
+        >>> img.get_data_dtype()
+        dtype('<i4')
+        >>> img.set_data_dtype('smallest')
+        >>> img.get_data_dtype()
+        'smallest'
+        >>> img.get_data_dtype(finalize=True)
+        dtype('uint8')
+        >>> img.get_data_dtype()
+        dtype('uint8')
+
+        Note that floating point values will not be coerced to ``int``
+
+        >>> floats = np.arange(24, dtype='f4').reshape((2,3,4))
+        >>> img = Nifti1Image(floats, np.eye(4))
+        >>> img.set_data_dtype('smallest')
+        >>> img.get_data_dtype(finalize=True)
+        Traceback (most recent call last):
+           ...
+        ValueError: Cannot automatically cast array (of type float32) to an integer
+        type with fewer than 64 bits. Please set_data_dtype() to an explicit data type.
+
+        >>> arr = np.arange(1000, 1024, dtype='i4').reshape((2,3,4))
+        >>> img = Nifti1Image(arr, np.eye(4))
+        >>> img.set_data_dtype('smallest')
+        >>> img.set_data_dtype('implausible')
+        Traceback (most recent call last):
+           ...
+        nibabel.spatialimages.HeaderDataError: data dtype "implausible" not recognized
+        >>> img.set_data_dtype('none')
+        Traceback (most recent call last):
+           ...
+        nibabel.spatialimages.HeaderDataError: data dtype "none" known but not supported
+        >>> img.set_data_dtype(np.void)
+        Traceback (most recent call last):
+           ...
+        nibabel.spatialimages.HeaderDataError: data dtype "<class 'numpy.void'>" known
+        but not supported
+        >>> img.set_data_dtype('int')
+        Traceback (most recent call last):
+           ...
+        ValueError: Invalid data type 'int'. Specify a sized integer, e.g., 'uint8' or numpy.int16.
+        >>> img.set_data_dtype(int)
+        Traceback (most recent call last):
+           ...
+        ValueError: Invalid data type <class 'int'>. Specify a sized integer, e.g., 'uint8' or
+        numpy.int16.
+        >>> img.set_data_dtype('int64')
+        >>> img.get_data_dtype() == np.dtype('int64')
+        True
+        """
+        # Comparing dtypes to strings, numpy will attempt to call, e.g., dtype('mask'),
+        # so only check for aliases if the type is a string
+        # See https://github.com/numpy/numpy/issues/7242
+        if isinstance(datatype, str):
+            # Static aliases
+            if datatype == 'mask':
+                datatype = 'u1'
+            # Dynamic aliases
+            elif datatype in ('compat', 'smallest'):
+                self._dtype_alias = datatype
+                return
+
+        self._dtype_alias = None
+        super().set_data_dtype(datatype)
+
+    def get_data_dtype(self, finalize=False):
+        """ Get numpy dtype for data
+
+        If ``set_data_dtype()`` has been called with an alias
+        and ``finalize`` is ``False``, return the alias.
+        If ``finalize`` is ``True``, determine the appropriate dtype
+        from the image data object and set the final dtype in the
+        header before returning it.
+        """
+        if self._dtype_alias is None:
+            return super().get_data_dtype()
+        if not finalize:
+            return self._dtype_alias
+
+        datatype = None
+        if self._dtype_alias == 'compat':
+            datatype = _get_analyze_compat_dtype(self._dataobj)
+            descrip = "an Analyze-compatible dtype"
+        elif self._dtype_alias == 'smallest':
+            datatype = _get_smallest_dtype(self._dataobj)
+            descrip = "an integer type with fewer than 64 bits"
+        else:
+            raise ValueError(f"Unknown dtype alias {self._dtype_alias}.")
+        if datatype is None:
+            dt = get_obj_dtype(self._dataobj)
+            raise ValueError(f"Cannot automatically cast array (of type {dt}) to {descrip}."
+                             " Please set_data_dtype() to an explicit data type.")
+
+        self.set_data_dtype(datatype)  # Clears the alias
+        return super().get_data_dtype()
+
     def as_reoriented(self, ornt):
         """Apply an orientation change and return a new image
 
@@ -2136,3 +2273,141 @@ def save(img, filename):
         Nifti1Image.instance_to_filename(img, filename)
     except ImageFileError:
         Nifti1Pair.instance_to_filename(img, filename)
+
+
+def _get_smallest_dtype(
+        arr,
+        itypes=(np.uint8, np.int16, np.int32),
+        ftypes=(),
+        ):
+    """ Return the smallest "sensible" dtype that will hold the array data
+
+    The purpose of this function is to support automatic type selection
+    for serialization, so "sensible" here means well-supported in the NIfTI-1 world.
+
+    For floating point data, select between single- and double-precision.
+    For integer data, select among uint8, int16 and int32.
+
+    The test is for min/max range, so float64 is pretty unlikely to be hit.
+
+    Returns ``None`` if these dtypes do not suffice.
+
+    >>> _get_smallest_dtype(np.array([0, 1]))
+    dtype('uint8')
+    >>> _get_smallest_dtype(np.array([-1, 1]))
+    dtype('int16')
+    >>> _get_smallest_dtype(np.array([0, 256]))
+    dtype('int16')
+    >>> _get_smallest_dtype(np.array([-65536, 65536]))
+    dtype('int32')
+    >>> _get_smallest_dtype(np.array([-2147483648, 2147483648]))
+
+    By default floating point types are not searched:
+
+    >>> _get_smallest_dtype(np.array([1.]))
+    >>> _get_smallest_dtype(np.array([2. ** 1000]))
+    >>> _get_smallest_dtype(np.longdouble(2) ** 2000)
+    >>> _get_smallest_dtype(np.array([1+0j]))
+
+    However, this function can be passed "legal" floating point types, and
+    the logic works the same.
+
+    >>> _get_smallest_dtype(np.array([1.]), ftypes=('float32',))
+    dtype('float32')
+    >>> _get_smallest_dtype(np.array([2. ** 1000]), ftypes=('float32',))
+    >>> _get_smallest_dtype(np.longdouble(2) ** 2000, ftypes=('float32',))
+    >>> _get_smallest_dtype(np.array([1+0j]), ftypes=('float32',))
+    """
+    arr = np.asanyarray(arr)
+    if np.issubdtype(arr.dtype, np.floating):
+        test_dts = ftypes
+        info = np.finfo
+    elif np.issubdtype(arr.dtype, np.integer):
+        test_dts = itypes
+        info = np.iinfo
+    else:
+        return None
+
+    mn, mx = np.min(arr), np.max(arr)
+    for dt in test_dts:
+        dtinfo = info(dt)
+        if dtinfo.min <= mn and mx <= dtinfo.max:
+            return np.dtype(dt)
+
+
+def _get_analyze_compat_dtype(arr):
+    """ Return an Analyze-compatible dtype that ``arr`` can be safely cast to
+
+    Analyze-compatible types are returned without inspection:
+
+    >>> _get_analyze_compat_dtype(np.uint8([0, 1]))
+    dtype('uint8')
+    >>> _get_analyze_compat_dtype(np.int16([0, 1]))
+    dtype('int16')
+    >>> _get_analyze_compat_dtype(np.int32([0, 1]))
+    dtype('int32')
+    >>> _get_analyze_compat_dtype(np.float32([0, 1]))
+    dtype('float32')
+
+    Signed ``int8`` are cast to ``uint8`` or ``int16`` based on value ranges:
+
+    >>> _get_analyze_compat_dtype(np.int8([0, 1]))
+    dtype('uint8')
+    >>> _get_analyze_compat_dtype(np.int8([-1, 1]))
+    dtype('int16')
+
+    Unsigned ``uint16`` are cast to ``int16`` or ``int32`` based on value ranges:
+
+    >>> _get_analyze_compat_dtype(np.uint16([32767]))
+    dtype('int16')
+    >>> _get_analyze_compat_dtype(np.uint16([65535]))
+    dtype('int32')
+
+    ``int32`` is returned for integer types and ``float32`` for floating point types:
+
+    >>> _get_analyze_compat_dtype(np.array([-1, 1]))
+    dtype('int32')
+    >>> _get_analyze_compat_dtype(np.array([-1., 1.]))
+    dtype('float32')
+
+    If the value ranges exceed 4 bytes or cannot be cast, then a ``ValueError`` is raised:
+
+    >>> _get_analyze_compat_dtype(np.array([0, 4294967295]))
+    Traceback (most recent call last):
+       ...
+    ValueError: Cannot find analyze-compatible dtype for array with dtype=int64
+        (min=0, max=4294967295)
+
+    >>> _get_analyze_compat_dtype([0., 2.e40])
+    Traceback (most recent call last):
+       ...
+    ValueError: Cannot find analyze-compatible dtype for array with dtype=float64
+        (min=0.0, max=2e+40)
+
+    Note that real-valued complex arrays cannot be safely cast.
+
+    >>> _get_analyze_compat_dtype(np.array([1+0j]))
+    Traceback (most recent call last):
+       ...
+    ValueError: Cannot find analyze-compatible dtype for array with dtype=complex128
+        (min=(1+0j), max=(1+0j))
+    """
+    arr = np.asanyarray(arr)
+    dtype = arr.dtype
+    if dtype in (np.uint8, np.int16, np.int32, np.float32):
+        return dtype
+
+    if dtype == np.int8:
+        return np.dtype('uint8' if arr.min() >= 0 else 'int16')
+    elif dtype == np.uint16:
+        return np.dtype('int16' if arr.max() <= np.iinfo(np.int16).max else 'int32')
+
+    mn, mx = arr.min(), arr.max()
+    if np.can_cast(mn, np.int32) and np.can_cast(mx, np.int32):
+        return np.dtype('int32')
+    if np.can_cast(mn, np.float32) and np.can_cast(mx, np.float32):
+        return np.dtype('float32')
+
+    raise ValueError(
+        f"Cannot find analyze-compatible dtype for array with dtype={dtype} (min={mn}, max={mx})"
+    )

nibabel/tests/test_nifti1.py

@@ -1119,6 +1119,63 @@ def test_write_scaling(self):
             with np.errstate(invalid='ignore'):
                 self._check_write_scaling(slope, inter, e_slope, e_inter)
 
+    def test_dynamic_dtype_aliases(self):
+        for in_dt, mn, mx, alias, effective_dt in [
+                (np.uint8, 0, 255, 'compat', np.uint8),
+                (np.int8, 0, 127, 'compat', np.uint8),
+                (np.int8, -128, 127, 'compat', np.int16),
+                (np.int16, -32768, 32767, 'compat', np.int16),
+                (np.uint16, 0, 32767, 'compat', np.int16),
+                (np.uint16, 0, 65535, 'compat', np.int32),
+                (np.int32, -2**31, 2**31-1, 'compat', np.int32),
+                (np.uint32, 0, 2**31-1, 'compat', np.int32),
+                (np.uint32, 0, 2**32-1, 'compat', None),
+                (np.int64, -2**31, 2**31-1, 'compat', np.int32),
+                (np.uint64, 0, 2**31-1, 'compat', np.int32),
+                (np.int64, 0, 2**32-1, 'compat', None),
+                (np.uint64, 0, 2**32-1, 'compat', None),
+                (np.float32, 0, 1e30, 'compat', np.float32),
+                (np.float64, 0, 1e30, 'compat', np.float32),
+                (np.float64, 0, 1e40, 'compat', None),
+                (np.int64, 0, 255, 'smallest', np.uint8),
+                (np.int64, 0, 256, 'smallest', np.int16),
+                (np.int64, -1, 255, 'smallest', np.int16),
+                (np.int64, 0, 32768, 'smallest', np.int32),
+                (np.int64, 0, 4294967296, 'smallest', None),
+                (np.float32, 0, 1, 'smallest', None),
+                (np.float64, 0, 1, 'smallest', None)
+                ]:
+            arr = np.arange(24, dtype=in_dt).reshape((2, 3, 4))
+            arr[0, 0, :2] = [mn, mx]
+            img = self.image_class(arr, np.eye(4), dtype=alias)
+            # Stored as alias
+            assert img.get_data_dtype() == alias
+            if effective_dt is None:
+                with pytest.raises(ValueError):
+                    img.get_data_dtype(finalize=True)
+                continue
+            # Finalizing sets and clears the alias
+            assert img.get_data_dtype(finalize=True) == effective_dt
+            assert img.get_data_dtype() == effective_dt
+            # Re-set to alias
+            img.set_data_dtype(alias)
+            assert img.get_data_dtype() == alias
+            img_rt = bytesio_round_trip(img)
+            assert img_rt.get_data_dtype() == effective_dt
+            # Seralizing does not finalize the source image
+            assert img.get_data_dtype() == alias
+
+    def test_static_dtype_aliases(self):
+        for alias, effective_dt in [
+                ("mask", np.uint8),
+                ]:
+            for orig_dt in ('u1', 'i8', 'f4'):
+                arr = np.arange(24, dtype=orig_dt).reshape((2, 3, 4))
+                img = self.image_class(arr, np.eye(4), dtype=alias)
+                assert img.get_data_dtype() == effective_dt
+                img_rt = bytesio_round_trip(img)
+                assert img_rt.get_data_dtype() == effective_dt
+
 
 class TestNifti1Image(TestNifti1Pair):
     # Run analyze-flavor spatialimage tests