Python numpy.finfo() Examples

def __init__(self, renders=True):
    # start the bullet physics server
    self._renders = renders
    if (renders):
	    p.connect(p.GUI)
    else:
    	p.connect(p.DIRECT)

    observation_high = np.array([
          np.finfo(np.float32).max,
          np.finfo(np.float32).max,
          np.finfo(np.float32).max,
          np.finfo(np.float32).max])
    action_high = np.array([0.1])

    self.action_space = spaces.Discrete(9)
    self.observation_space = spaces.Box(-observation_high, observation_high)

    self.theta_threshold_radians = 1
    self.x_threshold = 2.4
    self._seed()
#    self.reset()
    self.viewer = None
    self._configure() 

def batch_iou_pair(yx_min1, yx_max1, yx_min2, yx_max2, min=float(np.finfo(np.float32).eps)):
    """
    Pairwisely calculates the IoU of two lists (at the same size M) of bounding boxes for N independent batches.
    :author 申瑞珉 (Ruimin Shen)
    :param yx_min1: The top left coordinates (y, x) of the first lists (size [N, M, 2]) of bounding boxes.
    :param yx_max1: The bottom right coordinates (y, x) of the first lists (size [N, M, 2]) of bounding boxes.
    :param yx_min2: The top left coordinates (y, x) of the second lists (size [N, M, 2]) of bounding boxes.
    :param yx_max2: The bottom right coordinates (y, x) of the second lists (size [N, M, 2]) of bounding boxes.
    :return: The lists (size [N, M]) of the IoU.
    """
    yx_min = torch.max(yx_min1, yx_min2)
    yx_max = torch.min(yx_max1, yx_max2)
    size = torch.clamp(yx_max - yx_min, min=0)
    intersect_area = torch.prod(size, -1)
    area1 = torch.prod(yx_max1 - yx_min1, -1)
    area2 = torch.prod(yx_max2 - yx_min2, -1)
    union_area = torch.clamp(area1 + area2 - intersect_area, min=min)
    return intersect_area / union_area 

def test_rt_bias():
    # Check for bias in round trip
    # Parallel test to arraywriters
    rng = np.random.RandomState(20111214)
    mu, std, count = 100, 10, 100
    arr = rng.normal(mu, std, size=(count,))
    eps = np.finfo(np.float32).eps
    aff = np.eye(4)
    for in_dt in (np.float32, np.float64):
        arr_t = arr.astype(in_dt)
        for out_dt in IUINT_TYPES:
            img = Nifti1Image(arr_t, aff)
            img_back = round_trip(img)
            arr_back_sc = img_back.get_data()
            slope, inter = img_back.get_header().get_slope_inter()
            bias = np.mean(arr_t - arr_back_sc)
            # Get estimate for error
            max_miss = rt_err_estimate(arr_t, arr_back_sc.dtype, slope, inter)
            # Hokey use of max_miss as a std estimate
            bias_thresh = np.max([max_miss / np.sqrt(count), eps])
            assert_true(np.abs(bias) < bias_thresh) 

def test_complex128_pass(self):
        nulp = 5
        x = np.linspace(-20, 20, 50, dtype=np.float64)
        x = 10**x
        x = np.r_[-x, x]
        xi = x + x*1j

        eps = np.finfo(x.dtype).eps
        y = x + x*eps*nulp/2.
        assert_array_almost_equal_nulp(xi, x + y*1j, nulp)
        assert_array_almost_equal_nulp(xi, y + x*1j, nulp)
        # The test condition needs to be at least a factor of sqrt(2) smaller
        # because the real and imaginary parts both change
        y = x + x*eps*nulp/4.
        assert_array_almost_equal_nulp(xi, y + y*1j, nulp)

        epsneg = np.finfo(x.dtype).epsneg
        y = x - x*epsneg*nulp/2.
        assert_array_almost_equal_nulp(xi, x + y*1j, nulp)
        assert_array_almost_equal_nulp(xi, y + x*1j, nulp)
        y = x - x*epsneg*nulp/4.
        assert_array_almost_equal_nulp(xi, y + y*1j, nulp) 

def test_check_nmant_nexp():
    # Routine for checking number of sigificand digits and exponent
    for t in IEEE_floats:
        nmant = np.finfo(t).nmant
        maxexp = np.finfo(t).maxexp
        assert_true(_check_nmant(t, nmant))
        assert_false(_check_nmant(t, nmant - 1))
        assert_false(_check_nmant(t, nmant + 1))
        assert_true(_check_maxexp(t, maxexp))
        assert_false(_check_maxexp(t, maxexp - 1))
        assert_false(_check_maxexp(t, maxexp + 1))
    # Check against type_info
    for t in ok_floats():
        ti = type_info(t)
        if ti['nmant'] != 106: # This check does not work for PPC double pair
            assert_true(_check_nmant(t, ti['nmant']))
        assert_true(_check_maxexp(t, ti['maxexp'])) 

def test_complex64_pass(self):
        nulp = 5
        x = np.linspace(-20, 20, 50, dtype=np.float32)
        x = 10**x
        x = np.r_[-x, x]
        xi = x + x*1j

        eps = np.finfo(x.dtype).eps
        y = x + x*eps*nulp/2.
        assert_array_almost_equal_nulp(xi, x + y*1j, nulp)
        assert_array_almost_equal_nulp(xi, y + x*1j, nulp)
        y = x + x*eps*nulp/4.
        assert_array_almost_equal_nulp(xi, y + y*1j, nulp)

        epsneg = np.finfo(x.dtype).epsneg
        y = x - x*epsneg*nulp/2.
        assert_array_almost_equal_nulp(xi, x + y*1j, nulp)
        assert_array_almost_equal_nulp(xi, y + x*1j, nulp)
        y = x - x*epsneg*nulp/4.
        assert_array_almost_equal_nulp(xi, y + y*1j, nulp) 

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_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_float64_pass(self):
        # The number of units of least precision
        # In this case, use a few places above the lowest level (ie nulp=1)
        nulp = 5
        x = np.linspace(-20, 20, 50, dtype=np.float64)
        x = 10**x
        x = np.r_[-x, x]

        # Addition
        eps = np.finfo(x.dtype).eps
        y = x + x*eps*nulp/2.
        assert_array_almost_equal_nulp(x, y, nulp)

        # Subtraction
        epsneg = np.finfo(x.dtype).epsneg
        y = x - x*epsneg*nulp/2.
        assert_array_almost_equal_nulp(x, y, nulp) 

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 finish_episode(self, log_probas, saved_rewards):
        R = 0
        policy_loss = []
        rewards = []
        for r in saved_rewards:
            R = r + 0.8 * R
            rewards.insert(0, R)

        rewards = torch.Tensor(rewards)
        # update: we notice improved performance without reward normalization
        # rewards = (rewards - rewards.mean()) / (rewards.std() + np.finfo(np.float32).eps)

        for log_prob, reward in zip(log_probas, rewards):
            policy_loss.append((-log_prob * reward).unsqueeze(0))
        l = len(policy_loss)
        policy_loss = torch.cat(policy_loss).sum()
        return policy_loss / l 

def test_basic_property(self):
        # Check A = L L^H
        shapes = [(1, 1), (2, 2), (3, 3), (50, 50), (3, 10, 10)]
        dtypes = (np.float32, np.float64, np.complex64, np.complex128)

        for shape, dtype in itertools.product(shapes, dtypes):
            np.random.seed(1)
            a = np.random.randn(*shape)
            if np.issubdtype(dtype, np.complexfloating):
                a = a + 1j*np.random.randn(*shape)

            t = list(range(len(shape)))
            t[-2:] = -1, -2

            a = np.matmul(a.transpose(t).conj(), a)
            a = np.asarray(a, dtype=dtype)

            c = np.linalg.cholesky(a)

            b = np.matmul(c, c.transpose(t).conj())
            assert_allclose(b, a,
                            err_msg="{} {}\n{}\n{}".format(shape, dtype, a, c),
                            atol=500 * a.shape[0] * np.finfo(dtype).eps) 

def _mutual_information_varoquaux(x, y, bins=256, sigma=1, normalized=True):
    """Based on Gael Varoquaux's implementation: https://gist.github.com/GaelVaroquaux/ead9898bd3c973c40429."""
    jh = np.histogram2d(x, y, bins=bins)[0]

    # smooth the jh with a gaussian filter of given sigma
    scipy.ndimage.gaussian_filter(jh, sigma=sigma, mode="constant", output=jh)

    # compute marginal histograms
    jh = jh + np.finfo(float).eps
    sh = np.sum(jh)
    jh = jh / sh
    s1 = np.sum(jh, axis=0).reshape((-1, jh.shape[0]))
    s2 = np.sum(jh, axis=1).reshape((jh.shape[1], -1))

    if normalized:
        mi = ((np.sum(s1 * np.log(s1)) + np.sum(s2 * np.log(s2))) / np.sum(jh * np.log(jh))) - 1
    else:
        mi = np.sum(jh * np.log(jh)) - np.sum(s1 * np.log(s1)) - np.sum(s2 * np.log(s2))

    return mi 

def do_precision_upper_bound(self, float_small, float_large):
        eps = np.finfo(float_large).eps

        arr = np.array([1.0], float_small)
        range = np.array([0.0, 1.0 - eps], float_large)

        # test is looking for behavior when the bounds change between dtypes
        if range.astype(float_small)[-1] != 1:
            return

        # previously crashed
        count, x_loc = np.histogram(arr, bins=1, range=range)
        assert_equal(count, [1])

        # gh-10322 means that the type comes from arr - this may change
        assert_equal(x_loc.dtype, float_small) 

def do_precision_lower_bound(self, float_small, float_large):
        eps = np.finfo(float_large).eps

        arr = np.array([1.0], float_small)
        range = np.array([1.0 + eps, 2.0], float_large)

        # test is looking for behavior when the bounds change between dtypes
        if range.astype(float_small)[0] != 1:
            return

        # previously crashed
        count, x_loc = np.histogram(arr, bins=1, range=range)
        assert_equal(count, [1])

        # gh-10322 means that the type comes from arr - this may change
        assert_equal(x_loc.dtype, float_small) 

def merge(self, tiles: List[np.ndarray], dtype=np.float32):
        if len(tiles) != len(self.crops):
            raise ValueError

        channels = 1 if len(tiles[0].shape) == 2 else tiles[0].shape[2]
        target_shape = self.image_height + self.margin_bottom + self.margin_top, self.image_width + self.margin_right + self.margin_left, channels

        image = np.zeros(target_shape, dtype=np.float64)
        norm_mask = np.zeros(target_shape, dtype=np.float64)

        w = np.dstack([self.weight] * channels)

        for tile, (x, y, tile_width, tile_height) in zip(tiles, self.crops):
            # print(x, y, tile_width, tile_height, image.shape)
            image[y:y + tile_height, x:x + tile_width] += tile * w
            norm_mask[y:y + tile_height, x:x + tile_width] += w

        # print(norm_mask.min(), norm_mask.max())
        norm_mask = np.clip(norm_mask, a_min=np.finfo(norm_mask.dtype).eps, a_max=None)
        normalized = np.divide(image, norm_mask).astype(dtype)
        crop = self.crop_to_orignal_size(normalized)
        return crop 

def iou(yx_min1, yx_max1, yx_min2, yx_max2, min=None):
    """
    Calculates the IoU of two bounding boxes.
    :author 申瑞珉 (Ruimin Shen)
    :param yx_min1: The top left coordinates (y, x) of the first bounding boxe.
    :param yx_max1: The bottom right coordinates (y, x) of the first bounding boxe.
    :param yx_min2: The top left coordinates (y, x) of the second bounding boxe.
    :param yx_max2: The bottom right coordinates (y, x) of the second bounding boxe.
    :return: The IoU.
    """
    assert np.all(yx_min1 <= yx_max1)
    assert np.all(yx_min2 <= yx_max2)
    if min is None:
        min = np.finfo(yx_min1.dtype).eps
    yx_min = np.maximum(yx_min1, yx_min2)
    yx_max = np.minimum(yx_max1, yx_max2)
    intersect_area = np.multiply.reduce(np.maximum(0.0, yx_max - yx_min))
    area1 = np.multiply.reduce(yx_max1 - yx_min1)
    area2 = np.multiply.reduce(yx_max2 - yx_min2)
    assert np.all(intersect_area >= 0)
    assert np.all(intersect_area <= area1)
    assert np.all(intersect_area <= area2)
    union_area = np.maximum(area1 + area2 - intersect_area, min)
    return intersect_area / union_area 

def iou_matrix(yx_min1, yx_max1, yx_min2, yx_max2, min=None):
    """
    Calculates the IoU of two lists of bounding boxes.
    :author 申瑞珉 (Ruimin Shen)
    :param yx_min1: The top left coordinates (y, x) of the first list (size [N1, 2]) of bounding boxes.
    :param yx_max1: The bottom right coordinates (y, x) of the first list (size [N1, 2]) of bounding boxes.
    :param yx_min2: The top left coordinates (y, x) of the second list (size [N2, 2]) of bounding boxes.
    :param yx_max2: The bottom right coordinates (y, x) of the second list (size [N2, 2]) of bounding boxes.
    :return: The matrix (size [N1, N2]) of the IoU.
    """
    if min is None:
        min = np.finfo(yx_min1.dtype).eps
    assert np.all(yx_min1 <= yx_max1)
    assert np.all(yx_min2 <= yx_max2)
    intersect_area = intersection_area(yx_min1, yx_max1, yx_min2, yx_max2)
    area1 = np.expand_dims(np.multiply.reduce(yx_max1 - yx_min1, -1), 1)
    area2 = np.expand_dims(np.multiply.reduce(yx_max2 - yx_min2, -1), 0)
    assert np.all(intersect_area >= 0)
    assert np.all(intersect_area <= area1)
    assert np.all(intersect_area <= area2)
    union_area = np.maximum(area1 + area2 - intersect_area, min)
    return intersect_area / union_area 

def SpectralClustering(CKSym, n):
    # This is direct port of JHU vision lab code. Could probably use sklearn SpectralClustering.
    CKSym = CKSym.astype(float)
    N, _ = CKSym.shape
    MAXiter = 1000  # Maximum number of iterations for KMeans
    REPlic = 20  # Number of replications for KMeans

    DN = np.diag(np.divide(1, np.sqrt(np.sum(CKSym, axis=0) + np.finfo(float).eps)))
    LapN = identity(N).toarray().astype(float) - np.matmul(np.matmul(DN, CKSym), DN)
    _, _, vN = np.linalg.svd(LapN)
    vN = vN.T
    kerN = vN[:, N - n:N]
    normN = np.sqrt(np.sum(np.square(kerN), axis=1))
    kerNS = np.divide(kerN, normN.reshape(len(normN), 1) + np.finfo(float).eps)
    km = KMeans(n_clusters=n, n_init=REPlic, max_iter=MAXiter, n_jobs=-1).fit(kerNS)
    return km.labels_ 

def test_against_cmath(self):
        import cmath

        points = [-1-1j, -1+1j, +1-1j, +1+1j]
        name_map = {'arcsin': 'asin', 'arccos': 'acos', 'arctan': 'atan',
                    'arcsinh': 'asinh', 'arccosh': 'acosh', 'arctanh': 'atanh'}
        atol = 4*np.finfo(complex).eps
        for func in self.funcs:
            fname = func.__name__.split('.')[-1]
            cname = name_map.get(fname, fname)
            try:
                cfunc = getattr(cmath, cname)
            except AttributeError:
                continue
            for p in points:
                a = complex(func(np.complex_(p)))
                b = cfunc(p)
                assert_(abs(a - b) < atol, "%s %s: %s; cmath: %s" % (fname, p, a, b)) 

def get_fbank(voice, mfc_obj):
    # Extract log mel-spectrogra
    fbank = mfc_obj.sig2logspec(voice).astype('float32')

    # Mean and variance normalization of each mel-frequency 
    fbank = fbank - fbank.mean(axis=0)
    fbank = fbank / (fbank.std(axis=0)+np.finfo(np.float32).eps)

    # If the duration of a voice recording is less than 10 seconds (1000 frames),
    # repeat the recording until it is longer than 10 seconds and crop.
    full_frame_number = 1000
    init_frame_number = fbank.shape[0]
    while fbank.shape[0] < full_frame_number:
          fbank = np.append(fbank, fbank[0:init_frame_number], axis=0)
          fbank = fbank[0:full_frame_number,:]
    return fbank 

def test_singleton(self):
        ftype = finfo(double)
        ftype2 = finfo(double)
        assert_equal(id(ftype), id(ftype2)) 

def test_limits_arithmetic(self):
        tiny = np.finfo(float).tiny
        a = array([tiny, 1. / tiny, 0.])
        assert_equal(getmaskarray(a / 2), [0, 0, 0])
        assert_equal(getmaskarray(2 / a), [1, 0, 1]) 

def test_double(self):
        # Generate 1 + small deviation, check that adding eps gives a few UNL
        x = np.ones(10).astype(np.float64)
        x += 0.01 * np.random.randn(10).astype(np.float64)
        eps = np.finfo(np.float64).eps
        assert_array_max_ulp(x, x+eps, maxulp=200) 

def backup_best(self, cls_ap, path):
        try:
            with open(self.model_dir + '.pkl', 'rb') as f:
                best = np.mean(list(pickle.load(f).values()))
        except:
            best = np.finfo(np.float32).min
        metric = np.mean(list(cls_ap.values()))
        if metric > best:
            with open(self.model_dir + '.pkl', 'wb') as f:
                pickle.dump(cls_ap, f)
            shutil.copy(path, self.model_dir + '.pth')
            logging.info('best model (%f) saved into %s.*' % (metric, self.model_dir))
        else:
            logging.info('best metric %f >= %f' % (best, metric)) 

def test_float16_fail(self):
        nulp = 5
        x = np.linspace(-4, 4, 10, dtype=np.float16)
        x = 10**x
        x = np.r_[-x, x]

        eps = np.finfo(x.dtype).eps
        y = x + x*eps*nulp*2.
        assert_raises(AssertionError, assert_array_almost_equal_nulp,
                      x, y, nulp)

        epsneg = np.finfo(x.dtype).epsneg
        y = x - x*epsneg*nulp*2.
        assert_raises(AssertionError, assert_array_almost_equal_nulp,
                      x, y, nulp) 

def test_float32_fail(self):
        nulp = 5
        x = np.linspace(-20, 20, 50, dtype=np.float32)
        x = 10**x
        x = np.r_[-x, x]

        eps = np.finfo(x.dtype).eps
        y = x + x*eps*nulp*2.
        assert_raises(AssertionError, assert_array_almost_equal_nulp,
                      x, y, nulp)

        epsneg = np.finfo(x.dtype).epsneg
        y = x - x*epsneg*nulp*2.
        assert_raises(AssertionError, assert_array_almost_equal_nulp,
                      x, y, nulp) 

def test_float32_pass(self):
        nulp = 5
        x = np.linspace(-20, 20, 50, dtype=np.float32)
        x = 10**x
        x = np.r_[-x, x]

        eps = np.finfo(x.dtype).eps
        y = x + x*eps*nulp/2.
        assert_array_almost_equal_nulp(x, y, nulp)

        epsneg = np.finfo(x.dtype).epsneg
        y = x - x*epsneg*nulp/2.
        assert_array_almost_equal_nulp(x, y, nulp) 

def test_float64_fail(self):
        nulp = 5
        x = np.linspace(-20, 20, 50, dtype=np.float64)
        x = 10**x
        x = np.r_[-x, x]

        eps = np.finfo(x.dtype).eps
        y = x + x*eps*nulp*2.
        assert_raises(AssertionError, assert_array_almost_equal_nulp,
                      x, y, nulp)

        epsneg = np.finfo(x.dtype).epsneg
        y = x - x*epsneg*nulp*2.
        assert_raises(AssertionError, assert_array_almost_equal_nulp,
                      x, y, nulp) 

def test_uniform_range_bounds(self):
        fmin = np.finfo('float').min
        fmax = np.finfo('float').max

        func = np.random.uniform
        assert_raises(OverflowError, func, -np.inf, 0)
        assert_raises(OverflowError, func,  0,      np.inf)
        assert_raises(OverflowError, func,  fmin,   fmax)
        assert_raises(OverflowError, func, [-np.inf], [0])
        assert_raises(OverflowError, func, [0], [np.inf])

        # (fmax / 1e17) - fmin is within range, so this should not throw
        # account for i386 extended precision DBL_MAX / 1e17 + DBL_MAX >
        # DBL_MAX by increasing fmin a bit
        np.random.uniform(low=np.nextafter(fmin, 1), high=fmax / 1e17)