Python numpy.iinfo() Examples

def hparams(self, defaults, model_hparams):
    p = model_hparams
    # Filterbank extraction in bottom instead of preprocess_example is faster.
    p.add_hparam("audio_preproc_in_bottom", False)
    # The trainer seems to reserve memory for all members of the input dict
    p.add_hparam("audio_keep_example_waveforms", False)
    p.add_hparam("audio_sample_rate", 16000)
    p.add_hparam("audio_preemphasis", 0.97)
    p.add_hparam("audio_dither", 1.0 / np.iinfo(np.int16).max)
    p.add_hparam("audio_frame_length", 25.0)
    p.add_hparam("audio_frame_step", 10.0)
    p.add_hparam("audio_lower_edge_hertz", 20.0)
    p.add_hparam("audio_upper_edge_hertz", 8000.0)
    p.add_hparam("audio_num_mel_bins", 80)
    p.add_hparam("audio_add_delta_deltas", True)
    p.add_hparam("num_zeropad_frames", 250)

    p = defaults
    # p.stop_at_eos = int(False)
    p.input_modality = {"inputs": ("audio:speech_recognition_modality", None)}
    p.target_modality = (registry.Modalities.SYMBOL, 256) 

def fit(self, X, y):
        if X.shape[0] == 0:
            return self
        elif np.unique(y).shape[0] <= 1:
            self.update_aux(y)
            return self

        seed = self.random_state.integers(np.iinfo(np.int32).max)
        self.model.set_params(random_state = seed)
        self.model.fit(X, y)
        n_nodes = self.model.tree_.node_count
        self.pos = np.zeros(n_nodes, dtype=ctypes.c_long)
        self.neg = np.zeros(n_nodes, dtype=ctypes.c_long)
        pred_node = self.model.apply(X).astype(ctypes.c_long)
        _create_node_counters(self.pos, self.neg, pred_node, y.astype(ctypes.c_double))
        self.pos = self.pos.astype(ctypes.c_double) + self.beta_prior[0]
        self.neg = self.neg.astype(ctypes.c_double) + self.beta_prior[1]

        self.is_fitted = True
        return self 

def _do_scaling(self):
        arr = self._array
        out_dtype = self._out_dtype
        assert out_dtype.kind in 'iu'
        mn, mx = self.finite_range()
        if arr.dtype.kind == 'f':
            # Float to (u)int scaling
            self._range_scale()
            return
        # (u)int to (u)int
        info = np.iinfo(out_dtype)
        out_max, out_min = info.max, info.min
        # If left as int64, uint64, comparisons will default to floats, and
        # these are inexact for > 2**53 - so convert to int
        if (as_int(mx) <= as_int(out_max) and
            as_int(mn) >= as_int(out_min)):
            # already in range
            return
        # (u)int to (u)int scaling
        self._iu2iu() 

def _get_valid_range(self):
        ''' Return valid range for image data

        The valid range can come from the image 'valid_range' or
        image 'valid_min' and 'valid_max', or, failing that, from the
        data type range
        '''
        ddt = self.get_data_dtype()
        info = np.iinfo(ddt.type)
        try:
            valid_range = self._image.valid_range
        except AttributeError:
            try:
                valid_range = [self._image.valid_min,
                               self._image.valid_max]
            except AttributeError:
                valid_range = [info.min, info.max]
        if valid_range[0] < info.min or valid_range[1] > info.max:
            raise ValueError('Valid range outside input '
                             'data type range')
        return np.asarray(valid_range, dtype=np.float) 

def test_int_int_min_max():
    # Conversion between (u)int and (u)int
    eps = np.finfo(np.float64).eps
    rtol = 1e-6
    for in_dt in IUINT_TYPES:
        iinf = np.iinfo(in_dt)
        arr = np.array([iinf.min, iinf.max], dtype=in_dt)
        for out_dt in IUINT_TYPES:
            try:
                aw = SlopeInterArrayWriter(arr, out_dt)
            except ScalingError:
                continue
            arr_back_sc = round_trip(aw)
            # integer allclose
            adiff = int_abs(arr - arr_back_sc)
            rdiff = adiff / (arr + eps)
            assert_true(np.all(rdiff < rtol)) 

def test_int_int_slope():
    # Conversion between (u)int and (u)int for slopes only
    eps = np.finfo(np.float64).eps
    rtol = 1e-7
    for in_dt in IUINT_TYPES:
        iinf = np.iinfo(in_dt)
        for out_dt in IUINT_TYPES:
            kinds = np.dtype(in_dt).kind + np.dtype(out_dt).kind
            if kinds in ('ii', 'uu', 'ui'):
                arrs = (np.array([iinf.min, iinf.max], dtype=in_dt),)
            elif kinds == 'iu':
                arrs = (np.array([iinf.min, 0], dtype=in_dt),
                        np.array([0, iinf.max], dtype=in_dt))
            for arr in arrs:
                try:
                    aw = SlopeArrayWriter(arr, out_dt)
                except ScalingError:
                    continue
                assert_false(aw.slope == 0)
                arr_back_sc = round_trip(aw)
                # integer allclose
                adiff = int_abs(arr - arr_back_sc)
                rdiff = adiff / (arr + eps)
                assert_true(np.all(rdiff < rtol)) 

def test_able_casting():
    # Check the able_int_type function guesses numpy out type
    types = np.sctypes['int'] + np.sctypes['uint']
    for in_type in types:
        in_info = np.iinfo(in_type)
        in_mn, in_mx = in_info.min, in_info.max
        A = np.zeros((1,), dtype=in_type)
        for out_type in types:
            out_info = np.iinfo(out_type)
            out_mn, out_mx = out_info.min, out_info.max
            B = np.zeros((1,), dtype=out_type)
            ApBt = (A + B).dtype.type
            able_type = able_int_type([in_mn, in_mx, out_mn, out_mx])
            if able_type is None:
                assert_equal(ApBt, np.float64)
                continue
            # Use str for comparison to avoid int32/64 vs intp comparison
            # failures
            assert_equal(np.dtype(ApBt).str, np.dtype(able_type).str) 

def _linear_transformation_to_lut(linear_transformation,
                                  max_value=None, dtype=numpy.uint16):
    min_value = 0
    if max_value is None:
        max_value = numpy.iinfo(dtype).max

    def gain_offset_to_lut(gain, offset):
        logging.debug(
            'Normalize: Calculating lut values for gain '
            '{} and offset {}'.format(gain, offset))
        lut = numpy.arange(min_value, max_value + 1, dtype=numpy.float)
        return gain * lut + offset

    lut = gain_offset_to_lut(linear_transformation.gain,
                             linear_transformation.offset)

    logging.debug('Normalize: Clipping lut from [{}, {}] to [{},{}]'.format(
        min(lut), max(lut), min_value, max_value))
    numpy.clip(lut, min_value, max_value, lut)

    return lut.astype(dtype) 

def _uniform_weight_alpha(sum_masked_arrays, output_datatype):
    '''Calculates the cumulative mask of a list of masked array

    Input:
        sum_masked_arrays (list of numpy masked arrays): The list of
            masked arrays to find the cumulative mask of, each element
            represents one band.
            (sums_masked_array.mask has a 1 for a no data pixel and
            a 0 otherwise)
        output_datatype (numpy datatype): The output datatype

    Output:
        output_alpha (numpy uint16 array): The output mask
            (0 for a no data pixel, uint16 max value otherwise)
    '''

    output_alpha = numpy.ones(sum_masked_arrays[0].shape)
    for band_sum_masked_array in sum_masked_arrays:
        output_alpha[numpy.nonzero(band_sum_masked_array.mask == 1)] = 0

    output_alpha = output_alpha.astype(output_datatype) * \
        numpy.iinfo(output_datatype).max

    return output_alpha 

def _parallel_predict(self, contexts: np.ndarray, is_predict: bool):

        # Total number of contexts to predict
        n_contexts = len(contexts)

        # Partition contexts by job
        n_jobs, n_contexts, starts = self._partition_contexts(n_contexts)
        total_contexts = sum(n_contexts)

        # Get seed value for each context
        seeds = self.rng.randint(np.iinfo(np.int32).max, size=total_contexts)

        # Perform parallel predictions
        predictions = Parallel(n_jobs=n_jobs, backend=self.backend)(
                          delayed(self._predict_contexts)(
                              contexts[starts[i]:starts[i + 1]],
                              is_predict,
                              seeds[starts[i]:starts[i + 1]],
                              starts[i])
                          for i in range(n_jobs))

        # Reduce
        predictions = list(chain.from_iterable(t for t in predictions))

        return predictions if len(predictions) > 1 else predictions[0] 

def test_big_indices(self):
        # ravel_multi_index for big indices (issue #7546)
        if np.intp == np.int64:
            arr = ([1, 29], [3, 5], [3, 117], [19, 2],
                   [2379, 1284], [2, 2], [0, 1])
            assert_equal(
                np.ravel_multi_index(arr, (41, 7, 120, 36, 2706, 8, 6)),
                [5627771580, 117259570957])

        # test overflow checking for too big array (issue #7546)
        dummy_arr = ([0],[0])
        half_max = np.iinfo(np.intp).max // 2
        assert_equal(
            np.ravel_multi_index(dummy_arr, (half_max, 2)), [0])
        assert_raises(ValueError,
            np.ravel_multi_index, dummy_arr, (half_max+1, 2))
        assert_equal(
            np.ravel_multi_index(dummy_arr, (half_max, 2), order='F'), [0])
        assert_raises(ValueError,
            np.ravel_multi_index, dummy_arr, (half_max+1, 2), order='F') 

def test_allclose(self):
        # Tests allclose on arrays
        a = np.random.rand(10)
        b = a + np.random.rand(10) * 1e-8
        assert_(allclose(a, b))
        # Test allclose w/ infs
        a[0] = np.inf
        assert_(not allclose(a, b))
        b[0] = np.inf
        assert_(allclose(a, b))
        # Test allclose w/ masked
        a = masked_array(a)
        a[-1] = masked
        assert_(allclose(a, b, masked_equal=True))
        assert_(not allclose(a, b, masked_equal=False))
        # Test comparison w/ scalar
        a *= 1e-8
        a[0] = 0
        assert_(allclose(a, 0, masked_equal=True))

        # Test that the function works for MIN_INT integer typed arrays
        a = masked_array([np.iinfo(np.int_).min], dtype=np.int_)
        assert_(allclose(a, a)) 

def test_respect_dtype_singleton(self):
        # See gh-7203
        for dt in self.itype:
            lbnd = 0 if dt is np.bool_ else np.iinfo(dt).min
            ubnd = 2 if dt is np.bool_ else np.iinfo(dt).max + 1

            sample = self.rfunc(lbnd, ubnd, dtype=dt)
            assert_equal(sample.dtype, np.dtype(dt))

        for dt in (bool, int, np.long):
            lbnd = 0 if dt is bool else np.iinfo(dt).min
            ubnd = 2 if dt is bool else np.iinfo(dt).max + 1

            # gh-7284: Ensure that we get Python data types
            sample = self.rfunc(lbnd, ubnd, dtype=dt)
            assert_(not hasattr(sample, 'dtype'))
            assert_equal(type(sample), dt) 

def test_can_cast_values(self):
        # gh-5917
        for dt in np.sctypes['int'] + np.sctypes['uint']:
            ii = np.iinfo(dt)
            assert_(np.can_cast(ii.min, dt))
            assert_(np.can_cast(ii.max, dt))
            assert_(not np.can_cast(ii.min - 1, dt))
            assert_(not np.can_cast(ii.max + 1, dt))

        for dt in np.sctypes['float']:
            fi = np.finfo(dt)
            assert_(np.can_cast(fi.min, dt))
            assert_(np.can_cast(fi.max, dt))


# Custom exception class to test exception propagation in fromiter 

def argmin(self, axis=None, skipna=True, *args, **kwargs):
        """
        Returns the indices of the minimum values along an axis.

        See `numpy.ndarray.argmin` for more information on the
        `axis` parameter.

        See Also
        --------
        numpy.ndarray.argmin
        """
        nv.validate_argmin(args, kwargs)
        nv.validate_minmax_axis(axis)

        i8 = self.asi8
        if self.hasnans:
            mask = self._isnan
            if mask.all() or not skipna:
                return -1
            i8 = i8.copy()
            i8[mask] = np.iinfo('int64').max
        return i8.argmin() 

def reduce_id_bits(segmentation):
  """Reduces the number of bits used for IDs.

  Assumes that one additional ID beyond the max of 'segmentation' is necessary
  (used by GALA to mark boundary areas).

  Args:
    segmentation: ndarray of int type

  Returns:
    segmentation ndarray converted to minimal uint type large enough to keep
    all the IDs.
  """
  max_id = segmentation.max()
  if max_id <= np.iinfo(np.uint8).max:
    return segmentation.astype(np.uint8)
  elif max_id <= np.iinfo(np.uint16).max:
    return segmentation.astype(np.uint16)
  elif max_id <= np.iinfo(np.uint32).max:
    return segmentation.astype(np.uint32) 

def random_seed(seed=None):
    """
    Runs a code block with a new seed for np, mx and python's random.

    Parameters
    ----------

    seed : the seed to pass to np.random, mx.random and python's random.

    To impose rng determinism, invoke e.g. as in:

    with random_seed(1234):
        ...

    To impose rng non-determinism, invoke as in:

    with random_seed():
        ...

    Upon conclusion of the block, the rng's are returned to
    a state that is a function of their pre-block state, so
    any prior non-determinism is preserved.

    """

    try:
        next_seed = np.random.randint(0, np.iinfo(np.int32).max)
        if seed is None:
            np.random.seed()
            seed = np.random.randint(0, np.iinfo(np.int32).max)
        logger = default_logger()
        logger.debug('Setting np, mx and python random seeds = %s', seed)
        np.random.seed(seed)
        mx.random.seed(seed)
        random.seed(seed)
        yield
    finally:
        # Reinstate prior state of np.random and other generators
        np.random.seed(next_seed)
        mx.random.seed(next_seed)
        random.seed(next_seed) 

def random_seed(seed=None):
    """
    Runs a code block with a new seed for np, mx and python's random.

    Parameters
    ----------

    seed : the seed to pass to np.random, mx.random and python's random.

    To impose rng determinism, invoke e.g. as in:

    with random_seed(1234):
        ...

    To impose rng non-determinism, invoke as in:

    with random_seed():
        ...

    Upon conclusion of the block, the rng's are returned to
    a state that is a function of their pre-block state, so
    any prior non-determinism is preserved.

    """

    try:
        next_seed = np.random.randint(0, np.iinfo(np.int32).max)
        if seed is None:
            np.random.seed()
            seed = np.random.randint(0, np.iinfo(np.int32).max)
        logger = default_logger()
        logger.debug('Setting np, mx and python random seeds = %s', seed)
        np.random.seed(seed)
        mx.random.seed(seed)
        random.seed(seed)
        yield
    finally:
        # Reinstate prior state of np.random and other generators
        np.random.seed(next_seed)
        mx.random.seed(next_seed)
        random.seed(next_seed) 

def encode(self, s):
    """Transform a string with a filename into a list of float32.

    Args:
      s: path to the file with a waveform.

    Returns:
      samples: list of int16s
    """
    # Make sure that the data is a single channel, 16bit, 16kHz wave.
    # TODO(chorowski): the directory may not be writable, this should fallback
    # to a temp path, and provide instructions for installing sox.
    if s.endswith(".mp3"):
      # TODO(dliebling) On Linux, check if libsox-fmt-mp3 is installed.
      out_filepath = s[:-4] + ".wav"
      call([
          "sox", "--guard", s, "-r", "16k", "-b", "16", "-c", "1", out_filepath
      ])
      s = out_filepath
    elif not s.endswith(".wav"):
      out_filepath = s + ".wav"
      if not os.path.exists(out_filepath):
        call(["sox", "-r", "16k", "-b", "16", "-c", "1", s, out_filepath])
      s = out_filepath
    rate, data = wavfile.read(s)
    assert rate == self._sample_rate
    assert len(data.shape) == 1
    if data.dtype not in [np.float32, np.float64]:
      data = data.astype(np.float32) / np.iinfo(data.dtype).max
    return data.tolist() 

def _check_random_state(random_state):
    if random_state is None:
        return np.random.Generator(np.random.MT19937())
    if isinstance(random_state, np.random.Generator):
        return random_state
    elif isinstance(random_state, np.random.RandomState) or (random_state == np.random):
        random_state = int(random_state.randint(np.iinfo(np.int32).max) + 1)
    if isinstance(random_state, float):
        random_state = int(random_state)
    assert random_state > 0
    return np.random.Generator(np.random.MT19937(seed = random_state)) 

def _spawn_arm(self, fitted_classifier = None, n_w_rew = 0, n_wo_rew = 0,
                   buffer_X = None, buffer_y = None):
        self.n += 1
        self.rng_arm.append(self.random_state if (self.random_state == np.random) else \
                            _check_random_state(
                                self.random_state.integers(np.iinfo(np.int32).max) + 1))
        if self.smooth is not None:
            self.counters = np.c_[self.counters, np.array([n_w_rew + n_wo_rew]).reshape((1, 1)).astype(self.counters.dtype)]
        if (self.force_counters) or (self.thr > 0 and not self.force_fit):
            new_beta_col = np.array([0 if (n_w_rew + n_wo_rew) < self.thr else 1, self.alpha + n_w_rew, self.beta + n_wo_rew]).reshape((3, 1)).astype(self.beta_counters.dtype)
            self.beta_counters = np.c_[self.beta_counters, new_beta_col]
        if fitted_classifier is not None:
            if 'predict_proba' not in dir(fitted_classifier):
                fitted_classifier = _convert_decision_function_w_sigmoid(fitted_classifier)
            if partialfit:
                fitted_classifier = _add_method_predict_robust(fitted_classifier)
            self.algos.append(fitted_classifier)
        else:
            if self.force_fit or self.partialfit:
                if self.base is None:
                    raise ValueError("Must provide a classifier when initializing with different classifiers per arm.")
                self.algos.append( deepcopy(self.base) )
            else:
                if (self.force_counters) or (self.thr > 0 and not self.force_fit):
                    self.algos.append(_BetaPredictor(self.beta_counters[:, -1][1],
                                                     self.beta_counters[:, -1][2],
                                                     self.rng_arm[-1]))
                else:
                    self.algos.append(_ZeroPredictor())
        if (self.buffer is not None):
            self.buffer.append(_RefitBuffer(self.refit_buffer, self.deep_copy,
                                            self.rng_arm[-1]))
            if (buffer_X is not None):
                self.buffer[-1].add_obs(bufferX, buffer_y) 

def fit(self, X, a, r, p):
        """
        Fits the Offset Tree estimator to partially-labeled data collected from a different policy.
        
        Parameters
        ----------
        X : array (n_samples, n_features)
            Matrix of covariates for the available data.
        a : array (n_samples), int type
            Arms or actions that were chosen for each observations.
        r : array (n_samples), {0,1}
            Rewards that were observed for the chosen actions. Must be binary rewards 0/1.
        p : array (n_samples)
            Reward estimates for the actions that were chosen by the policy.
        """
        X, a, r = _check_fit_input(X, a, r)
        p = _check_1d_inp(p)
        assert p.shape[0] == X.shape[0]
        
        if self.c is not None:
            p = self.c * p
        if self.pmin is not None:
            p = np.clip(p, a_min = self.pmin, a_max = None)
        
        self._oracles = [deepcopy(self.base_algorithm) for c in range(self.nchoices - 1)]
        rs = self.random_state.integers(np.iinfo(np.int32).max, size=self.nchoices)
        Parallel(n_jobs=self.njobs, verbose=0, require="sharedmem")\
                (delayed(self._fit)(classif, X, a, r, p, rs) \
                    for classif in range(len(self._oracles)))
        self.is_fitted = True 

def __init__(self,
                 agent: 'ParallelDQNAgent',
                 create_env: Callable[[None], 'TrainingEnvironment'],
                 memory_queue: Queue,
                 model_update_queue: Queue,
                 done_queue: Queue,
                 n_steps: int,
                 n_episodes: int,
                 eps_end: int = 0.05,
                 eps_start: int = 0.99,
                 eps_decay_steps: int = 2000,
                 update_target_every: int = 2):
        super().__init__()

        self.agent = agent
        self.env = create_env()
        self.memory_queue = memory_queue
        self.model_update_queue = model_update_queue
        self.done_queue = done_queue
        self.n_steps = n_steps or np.iinfo(np.int32).max
        self.n_episodes = n_episodes or np.iinfo(np.int32).max
        self.eps_start = eps_start
        self.eps_end = eps_end
        self.eps_decay_steps = eps_decay_steps
        self.update_target_every = update_target_every

        self.env.agent_id = self.agent.id 

def set_data_shape(self, shape):
        ''' Set shape of data

        If ``ndims == len(shape)`` then we set zooms for dimensions higher than
        ``ndims`` to 1.0

        Parameters
        ----------
        shape : sequence
           sequence of integers specifying data array shape

        Notes
        -----
        Applies freesurfer hack for large vectors described in
        https://github.com/nipy/nibabel/issues/100 and
        http://code.google.com/p/fieldtrip/source/browse/trunk/external/freesurfer/save_nifti.m?spec=svn5022&r=5022#77
        '''
        # Apply freesurfer hack for vector
        hdr = self._structarr
        shape = tuple(shape)
        if (len(shape) == 3 and shape[1:] == (1, 1) and
            shape[0] > np.iinfo(hdr['dim'].dtype.base).max): # Freesurfer case
            try:
                hdr['glmin'] = shape[0]
            except OverflowError:
                overflow = True
            else:
                overflow = hdr['glmin'] != shape[0]
            if overflow:
                raise HeaderDataError('shape[0] %s does not fit in glmax datatype' %
                                      shape[0])
            warnings.warn('Using large vector Freesurfer hack; header will '
                          'not be compatible with SPM or FSL', stacklevel=2)
            shape = (-1, 1, 1)
        super(Nifti1Header, self).set_data_shape(shape) 

def able_int_type(values):
    """ Find the smallest integer numpy type to contain sequence `values`

    Prefers uint to int if minimum is >= 0

    Parameters
    ----------
    values : sequence
        sequence of integer values

    Returns
    -------
    itype : None or numpy type
        numpy integer type or None if no integer type holds all `values`

    Examples
    --------
    >>> able_int_type([0, 1]) == np.uint8
    True
    >>> able_int_type([-1, 1]) == np.int8
    True
    """
    if any([v % 1 for v in values]):
        return None
    mn = min(values)
    mx = max(values)
    if mn >= 0:
        for ityp in np.sctypes['uint']:
            if mx <= np.iinfo(ityp).max:
                return ityp
    for ityp in np.sctypes['int']:
        info = np.iinfo(ityp)
        if mn >= info.min and mx <= info.max:
            return ityp
    return None 

def _iu2iu(self):
        # (u)int to (u)int
        mn, mx = [as_int(v) for v in self.finite_range()]
        # range may be greater than the largest integer for this type.
        # as_int needed to work round numpy 1.4.1 int casting bug
        out_dtype = self._out_dtype
        t_min, t_max = np.iinfo(out_dtype).min, np.iinfo(out_dtype).max
        type_range = as_int(t_max) - as_int(t_min)
        mn2mx = mx - mn
        if mn2mx <= type_range: # might offset be enough?
            if t_min == 0: # uint output - take min to 0
                # decrease offset with floor_exact, meaning mn >= t_min after
                # subtraction.  But we may have pushed the data over t_max,
                # which we check below
                inter = floor_exact(mn - t_min, self.scaler_dtype)
            else: # int output - take midpoint to 0
                # ceil below increases inter, pushing scale up to 0.5 towards
                # -inf, because ints have abs min == abs max + 1
                midpoint = mn + as_int(np.ceil(mn2mx / 2.0))
                # Floor exact decreases inter, so pulling scaled values more
                # positive. This may make mx - inter > t_max
                inter = floor_exact(midpoint, self.scaler_dtype)
            # Need to check still in range after floor_exact-ing
            int_inter = as_int(inter)
            assert mn - int_inter >= t_min
            if mx - int_inter <= t_max:
                self.inter = inter
                return
        # Try slope options (sign flip) and then range scaling
        super(SlopeInterArrayWriter, self)._iu2iu() 

def _range_scale(self):
        """ Calculate scaling, intercept based on data range and output type """
        mn, mx = self.finite_range() # Values of self.array.dtype type
        out_dtype = self._out_dtype
        if mx == mn: # Only one number in array
            self.inter = mn
            return
        # Straight mx-mn can overflow.
        big_float = best_float() # usually longdouble except in win 32
        if mn.dtype.kind == 'f': # Already floats
            # float64 and below cast correctly to longdouble.  Longdouble needs
            # no casting
            mn2mx = np.diff(np.array([mn, mx], dtype=big_float))
        else: # max possible (u)int range is 2**64-1 (int64, uint64)
            # int_to_float covers this range.  On windows longdouble is the same
            # as double so mn2mx will be 2**64 - thus overestimating slope
            # slightly.  Casting to int needed to allow mx-mn to be larger than
            # the largest (u)int value
            mn2mx = int_to_float(as_int(mx) - as_int(mn), big_float)
        if out_dtype.kind == 'f':
            # Type range, these are also floats
            info = type_info(out_dtype)
            t_mn_mx = info['min'], info['max']
        else:
            t_mn_mx = np.iinfo(out_dtype).min, np.iinfo(out_dtype).max
            t_mn_mx= [int_to_float(v, big_float) for v in t_mn_mx]
        # We want maximum precision for the calculations. Casting will
        # not lose precision because min/max are of fp type.
        assert [v.dtype.kind for v in t_mn_mx] == ['f', 'f']
        scaled_mn2mx = np.diff(np.array(t_mn_mx, dtype = big_float))
        slope = mn2mx / scaled_mn2mx
        self.inter = mn - t_mn_mx[0] * slope
        self.slope = slope
        if not np.all(np.isfinite([self.slope, self.inter])):
            raise ScalingError("Slope / inter not both finite") 

def test_int_abs():
    for itype in np.sctypes['int']:
        info = np.iinfo(itype)
        in_arr = np.array([info.min, info.max], dtype=itype)
        idtype = np.dtype(itype)
        udtype = np.dtype(idtype.str.replace('i', 'u'))
        assert_equal(udtype.kind, 'u')
        assert_equal(idtype.itemsize, udtype.itemsize)
        mn, mx = in_arr
        e_mn = as_int(mx) + 1 # as_int needed for numpy 1.4.1 casting
        assert_equal(int_abs(mx), mx)
        assert_equal(int_abs(mn), e_mn)
        assert_array_equal(int_abs(in_arr), [e_mn, mx]) 

def test_round_trip():
    scaling_type = np.float32
    rng = np.random.RandomState(20111121)
    N = 10000
    sd_10s = range(-20, 51, 5)
    iuint_types = np.sctypes['int'] + np.sctypes['uint']
    # Remove intp types, which cannot be set into nifti header datatype
    iuint_types.remove(np.intp)
    iuint_types.remove(np.uintp)
    f_types = [np.float32, np.float64]
    # Expanding standard deviations
    for i, sd_10 in enumerate(sd_10s):
        sd = 10.0**sd_10
        V_in = rng.normal(0, sd, size=(N,1))
        for j, in_type in enumerate(f_types):
            for k, out_type in enumerate(iuint_types):
                check_arr(sd_10, V_in, in_type, out_type, scaling_type)
    # Spread integers across range
    for i, sd in enumerate(np.linspace(0.05, 0.5, 5)):
        for j, in_type in enumerate(iuint_types):
            info = np.iinfo(in_type)
            mn, mx = info.min, info.max
            type_range = mx - mn
            center = type_range / 2.0 + mn
            # float(sd) because type_range can be type 'long'
            width = type_range * float(sd)
            V_in = rng.normal(center, width, size=(N,1))
            for k, out_type in enumerate(iuint_types):
                check_arr(sd, V_in, in_type, out_type, scaling_type) 

def test_as_int_np_fix():
    # Test as_int works for integers.  We need as_int for integers because of a
    # numpy 1.4.1 bug such that int(np.uint32(2**32-1) == -1
    for t in np.sctypes['int'] + np.sctypes['uint']:
        info = np.iinfo(t)
        mn, mx = np.array([info.min, info.max], dtype=t)
        assert_equal((mn, mx), (as_int(mn), as_int(mx)))