def prewittgrad(self):
""" prewitt gradient """
gradx = np.array([1.0, 1.0, 1.0]).reshape(3, 1) * np.array(
[1.0, 0.0, -1.0]) / 3.0
grady = np.array([1.0, 0.0, -1.0]).reshape(3, 1) * np.array(
[1.0, 1.0, 1.0]) / 3.0
return gradx, grady
def matchColorToImage(color, img):
"""
match the color to the image mode
"""
imgChan = img.shape[-1]
if isinstance(color, (float, int)):
if isFloat(img):
if isinstance(color, int):
color = color / 255.0
elif isinstance(color, int):
color = int(color * 255)
else:
colChan = len(color)
if imgChan > colChan:
msg = "Don't know how to convert color[%d] to image pixel[%d]" % (
len(color), img.shape[-1])
raise NotImplementedError(msg)
if imgChan < colChan:
color = color[0:imgChan]
imgFloat = isFloat(img)
colFloat = isFloat(color)
if imgFloat != colFloat:
if imgFloat:
color = np.array(color) / 255.0
else:
color = np.int(np.array(color) * 255)
return color
def has_ext():
try:
a = bh.array([[76,25,11], [27,89,51], [18, 60, 32]], dtype=bh.float32)
b = bh.array([[10], [7], [43]], dtype=bh.float32)
bh.lapack.gesv(a, b)
return True
except Exception as e:
print("\n\033[31m[ext] Cannot test LAPACK extension methods.\033[0m")
print(e)
return False
def has_ext():
try:
a = bh.array([[76,25,11], [27,89,51], [18, 60, 32]], dtype=bh.float32)
b = bh.array([[10], [7], [43]], dtype=bh.float32)
bh.lapack.gesv(a, b)
return True
except Exception as e:
print("\n\033[31m[ext] Cannot test LAPACK extension methods.\033[0m")
print(e)
return False
def has_ext():
try:
src = bh.array([[1, 1, 1], [1, 1, 1], [1, 1, 0]], dtype=bh.float32)
kernel = bh.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=bh.float32)
bh.opencv.erode(src, kernel)
return True
except Exception as e:
print("\n\033[31m[ext] Cannot test OpenCV extension methods.\033[0m")
print(e)
return False
def getChannel(self, channelId: str):
"""
get a channel in floating point form
"""
if channelId not in self._activeChannels:
if channelId in [
'r', 'g', 'b'
] or (channelId == 'a' and len(self._activeChannels) < 2):
if 'v' in self._activeChannels:
return self._activeChannels['v']
res = numpyArray(self.image, floatingPoint=True, loader=None)
if isinstance(res.size, int) or len(res.size) < 3: # L
self._activeChannels['v'] = res
return self._activeChannels['v']
if res.size[2] == 2: # LA
self._activeChannels['v'] = res[:, :, 0]
self._activeChannels['a'] = res[:, :, 1]
elif res.size[2] == 3: # RGB
self._activeChannels['r'] = res[:, :, 0]
self._activeChannels['g'] = res[:, :, 1]
self._activeChannels['b'] = res[:, :, 2]
elif res.size[2] == 3: # RGBA
self._activeChannels['r'] = res[:, :, 0]
self._activeChannels['g'] = res[:, :, 1]
self._activeChannels['b'] = res[:, :, 2]
self._activeChannels['a'] = res[:, :, 3]
elif channelId == 'a':
ret = np.array(self.image.size())
ret.fill(1.0)
self._activeChannels['a'] = ret
return ret
return self._activeChannels[channelId]
def random_galaxy(N):
""" Generate a galaxy of random bodies """
m = np.array((numpy.arange(0.0, 1.0, step=1.0 / N) + np.float64(10)) *
np.float64(m_sol / 10))
x = np.array((numpy.arange(0.0, 1.0, step=1.0 / N) - np.float64(0.5)) *
np.float64(r_ly / 100))
y = np.array((numpy.arange(0.0, 1.0, step=1.0 / N) - np.float64(0.5)) *
np.float64(r_ly / 100))
z = np.array((numpy.arange(0.0, 1.0, step=1.0 / N) - np.float64(0.5)) *
np.float64(r_ly / 100))
vx = np.zeros(N, dtype=np.float64)
vy = np.zeros(N, dtype=np.float64)
vz = np.zeros(N, dtype=np.float64)
assert len(m) == N
return m, x, y, z, vx, vy, vz
def prepare_inputs(*inputs, device):
out = [bh.array(k) for k in inputs]
for o in out:
# force allocation on target device
tmp = o * 1 # noqa: F841
bh.flush()
return out
def gaussianBlur(img, sizeX, sizeY=None, edge='clamp'):
"""
perform a gaussian blur on an image
:param img: can be a pil image, numpy array, etc
:param sizeX: the x radius of the blur
:param sizeY: the y radius of the blur (if None, use same as X)
:param edge: what to do for more pixels when we reach an edge
can be: "clamp","mirror","wrap", or a color
default is "clamp"
:returns: image in a numpy array
"""
if sizeY is None:
sizeY = sizeX
if edge == 'clamp':
mode = 'nearest'
cval = 0.0
elif edge == 'mirror':
mode = 'mirror'
cval = 0.0
elif edge == 'wrap':
mode = 'wrap'
cval = 0.0
else:
mode = 'constant'
cval = np.array(strToColor(edge))
img = numpyArray(img)
img = scipy.ndimage.gaussian_filter(img, (sizeY, sizeX),
mode=mode,
cval=cval)
return img
def dilate(a, b=None, c=None):
if b is None:
b = np.array([[0, 1, 0], [1, 1, 1], [0, 1, 0]], dtype=a.dtype)
if c is None:
c = np.empty_like(a)
ufuncs.extmethod("opencv_dilate", c, a, b)
return c
def my_kmax(R, k):
N = len(R)
f = np.ravel(R)
#print(f)
indices = np.array(heapq.nlargest(k, range(len(f)), f.__getitem__))
j = np.mod(indices, N)
i = np.floor(indices / N)
#print(indices)
return [i, j]
def generating_kernel(a):
"""
generate a 5x5 kernel
Comes from:
https://compvisionlab.wordpress.com/2013/05/13/image-blending-using-pyramid/
"""
w_1d = np.array([0.25 - a/2.0, 0.25, a, 0.25, 0.25 - a/2.0])
return np.outer(w_1d, w_1d)
def compressed_copy(ary, param):
a_min = ary.min()
a_range = ary.max() - ary.min() + 1
# Normalize `ary` into uint8
a = (ary - a_min) * 256 / a_range
assert (a.min() >= 0)
assert (a.max() < 256)
a = np.array(a, dtype=np.uint8)
# Copy `a` using `param`
a = _bh.mem_copy(a, param=param)
# un-normalize and convert back to the original dtype of `ary`
a = array_create.array(a, dtype=ary.dtype)
return (a * a_range + a_min) / 256.0
def directionToNormalColor(azimuth, elevation):
"""
given a direction, convert it into an RGB normal color
:param azimuth: compass direction
:param elevation: up/down direction
"""
azimuth = math.radians(float(azimuth))
elevation = math.radians(float(elevation))
return (
np.array([math.sin(azimuth),
math.cos(azimuth),
math.cos(elevation)]) + 1) / 2
def run(self):
for _ in range(100):
a, b, c, d = (np.random.randn(self.nz) for _ in range(4))
out_legacy = self.veros_legacy.fortran.solve_tridiag(a=a,
b=b,
c=c,
d=d,
n=self.nz)
if self.veros_new.backend_name == "bohrium":
a, b, c, d = (bh.array(v) for v in (a, b, c, d))
out_new = numerics.solve_tridiag(self.veros_new, a, b, c, d)
if not np.allclose(out_legacy, out_new):
return False
return True
def run(self):
a, b, c, d = (np.random.randn(self.nx, self.ny, self.nz) for _ in range(4))
out_legacy = np.zeros((self.nx, self.ny, self.nz))
for i in range(self.nx):
for j in range(self.ny):
out_legacy[i, j] = self.veros_legacy.call_fortran_routine(
'solve_tridiag', a=a[i, j], b=b[i, j], c=c[i, j], d=d[i, j], n=self.nz
)
if rs.backend == 'bohrium':
import bohrium as bh
a, b, c, d = (bh.array(v) for v in (a, b, c, d))
out_new = numerics.solve_tridiag(self.veros_new.state, a, b, c, d)
np.testing.assert_allclose(out_legacy, out_new)
def run(self):
a, b, c, d = (np.random.randn(self.nx, self.ny, self.nz)
for _ in range(4))
out_legacy = np.zeros((self.nx, self.ny, self.nz))
for i in range(self.nx):
for j in range(self.ny):
out_legacy[i, j] = self.veros_legacy.fortran.solve_tridiag(
a=a[i, j], b=b[i, j], c=c[i, j], d=d[i, j], n=self.nz)
if self.veros_new.backend_name == "bohrium":
a, b, c, d = (bh.array(v) for v in (a, b, c, d))
out_new = numerics.solve_tridiag(self.veros_new, a, b, c, d)
passed = np.allclose(out_legacy, out_new)
return passed
def rgb2hsvArray(rgb):
"""
Transform an rgb array to an hsv array
:param rgb: the input image. can be pil image or numpy array
:return hsv: the output array
This comes from scikit-image:
https://github.com/scikit-image/scikit-image/blob/master/skimage/color/colorconv.py
"""
rgb = numpyArray(rgb)
singleColor = len(rgb.shape) == 1
if singleColor: # convert to one-pixel image
rgb = np.array([[rgb]])
out = np.empty_like(rgb)
# -- V channel
out_v = rgb.max(-1)
# -- S channel
delta = rgb.ptp(-1)
# Ignore warning for zero divided by zero
old_settings = np.seterr(invalid='ignore')
out_s = delta / out_v
out_s[delta == 0.] = 0.
# -- H channel
# red is max
idx = (rgb[:, :, 0] == out_v)
out[idx, 0] = (rgb[idx, 1] - rgb[idx, 2]) / delta[idx]
# green is max
idx = (rgb[:, :, 1] == out_v)
out[idx, 0] = 2. + (rgb[idx, 2] - rgb[idx, 0]) / delta[idx]
# blue is max
idx = (rgb[:, :, 2] == out_v)
out[idx, 0] = 4. + (rgb[idx, 0] - rgb[idx, 1]) / delta[idx]
out_h = (out[:, :, 0] / 6.) % 1.
out_h[delta == 0.] = 0.
np.seterr(**old_settings)
# -- output
out[:, :, 0] = out_h
out[:, :, 1] = out_s
out[:, :, 2] = out_v
# remove NaN
out[np.isnan(out)] = 0
if singleColor:
out = out[0, 0]
return out
def load_data(self):
"""Load the npz archive specified by --inputfn or None is not set"""
if self.args.inputfn is None:
return None
else:
nobh_data = numpy.load(self.args.inputfn)
bhary_keys = nobh_data["_bhary_keys"].tolist()
ret = {}
for k in nobh_data.keys():
if k == "_bhary_keys":
continue
# Convert numpy arrays into bohrium arrays
if bh_is_loaded_as_np and k in bhary_keys:
a = nobh_data[k]
ret[k] = bohrium.array(a, bohrium=True)
else:
ret[k] = nobh_data[k]
return ret
def smoothNoise(size=(256,256),undersize=0.5,scaleMode=None,seed=None):
"""
create smooth noise by starting with undersized random noise
and then scaling it up
:param size: finished size
:param undersize: start undersized (percent) then scale up
:param scaleMode: PIL scale mode (default is BILINEAR)
:param seed: random seed to permit repeatability
"""
if seed is not None:
np.random.seed(seed)
if scaleMode is None:
scaleMode=Image.BILINEAR
noise=np.random.random((int(size[0]*undersize),int(size[0]*undersize)))
if undersize!=0:
size=(int(size[0]),int(size[1]))
noise=np.array(Image.fromarray(noise).resize(size,scaleMode))/255.0
return noise
def toWavelet(img,wavelet='haar',mode='symmetric',level=None):
"""
:param img: any supported image type to transform into wavelet space
:param wavelet: any common, named wavelet, including
'Haar' (default)
'Daubechies'
'Symlet'
'Coiflet'
'Biorthogonal'
'ReverseBiorthogonal'
'DiscreteMeyer'
'Gaussian'
'MexicanHat'
'Morlet'
'ComplexGaussian'
'Shannon'
'FrequencyBSpline'
'ComplexMorlet'
or a custom [ [lowpass_decomposition],
[highpass_decomposition],
[lowpass_reconstruction],
[highpass_reconstruction] ]
where each is a pair of floating point values
:param mode: str or 2-tuple of str, optional
Signal extension mode, see Modes (default: "symmetric"). This can also be a tuple containing a mode to apply along each axis in axes.
:param level: int, optional
Decomposition level (must be >= 0). If level is None (default) then it will be calculated using the dwt_max_level function.
See also:
https://pywavelets.readthedocs.io/en/latest/ref/index.html
"""
if mode is None:
mode='symmetric'
img=numpyArray(img)
colorMode=imageMode(img)
if len(colorMode)==1:
return pywt.wavedec2(img,_wavelet(wavelet),mode,level)
ret=[]
for ch in range(len(colorMode)):
ret.append(np.array(pywt.wavedec2(img[:,:,ch],_wavelet(wavelet),mode,level)))
ret=np.dstack(ret)
return ret
def getAlpha(image,alwaysCreate:bool=True):
"""
gets the alpha channel regardless of image type
:param image: the image whose mask to get
:param alwaysCreate: always returns a numpy array (otherwise, may return None)
:return: alpha channel as a PIL image, or numpy array,
or possibly None, depending on alwaysCreate
"""
ret=None
if image is None or not hasAlpha(image):
if alwaysCreate:
ret=np.array(image.size())
ret.fill(1.0)
elif isinstance(image,Image.Image):
ret=image.getalpha()
else:
ret=image[:,:,-1]
return ret
def get(self, type='smooth', mode='reflect'):
""" type can be 'smooth' or 'fine'
mode can be 'reflect','constant','nearest','mirror', 'wrap' for handling borders """
gradfn = {'smooth': self.prewittgrad, 'fine': self.basicgrad}[type]
gradx, grady = gradfn()
# x, y and z below are now the gradient matrices,
# each entry from x,y,z is a gradient vector at an image point
x = filters.convolve(self.img, gradx, mode=mode)
y = filters.convolve(self.img, grady, mode=mode)
# norm is the magnitude of the x,y,z vectors,
# each entry is the magnitude of the gradient at an image point and z*z = 1
norm = np.sqrt(x * x + y * y + 1)
# divide by the magnitude to normalise
# as well scale to an image: negative 0-127, positive 127-255
x, y = [a / norm * 127.0 + 128.0 for a in (x, y)]
z = np.ones(
self.shape) / norm # generate z, matrix of ones, then normalise
z = z * 255.0 # all positive
# x, -y gives blender form
# convert to int, transpose to rgb and return the normal map
return np.array([x, -y, z]).transpose(1, 2, 0).astype(np.uint8)
def voronoi(size=(256,256),npoints=30,mode='squared',invert=False,seed=None):
"""
:param mode: 'simple','squared','twoNearestDiff','twoNearestMult'
see also:
http://blackpawn.com/texts/cellular/default.html
TODO: this could be optomized usinf a cKDtree
https://stackoverflow.com/questions/10818546/finding-index-of-nearest-point-in-numpy-arrays-of-x-and-y-coordinates
"""
size=(int(size[0]),int(size[1]))
img=np.ndarray(size)
size=np.array(size)
if seed is not None:
np.random.seed(seed)
points=np.random.rand(int(npoints),2)*size # random points where voroni connections will appear
xy=xypoints(img).reshape((-1,2)) # flattened list of xy points for every pixel
dist=scipy.spatial.distance.cdist(xy,points) # for each pixel, determine distance to all voroni points
if mode=='simple':
img=np.min(dist,axis=1)
elif mode=='squared':
img=np.min(dist,axis=1)**2
elif mode=='twoNearestDiff':
#smallest=np.argsort(dist,axis=1)
smallest=np.argpartition(dist,2,axis=1)[:,0:2] # faster because it stops sorting after 2
img=dist[:,smallest[:,1]]-dist[:,smallest[:,0]]
elif mode=='twoNearestMult':
#smallest=np.argsort(dist,axis=1)
smallest=np.argpartition(dist,2,axis=1)[:,0:2] # faster because it stops sorting after 2
img=dist[:,smallest[:,1]]*dist[:,smallest[:,0]]
else:
raise NotImplementedError('mode="'+mode+'"')
# convert pixel list back into 2d array
img=normalize(img.reshape((size[0],size[1])))
if invert:
img=1.0-img
return img
def run(args):
print("*** Testing the equivalency of Bohrium-NumPy and NumPy ***")
test_suite_start_time = time.time()
for f in args.file:
if f.startswith("test_") and f.endswith(
"py") and f not in args.exclude:
# Remove ".py"
m = __import__(f[:-3])
# All test classes starts with "test_"
for cls in [o for o in dir(m) if o.startswith("test_") and \
(True if args.test and o in args.test or not args.test else False)]:
# Exclude specific test
if cls in args.exclude_test:
continue
cls_obj = getattr(m, cls)
cls_inst = cls_obj()
cls_inst.args = args
# Exclude benchmarks
import inspect
is_benchmark = BenchHelper.__name__ in [
c.__name__ for c in inspect.getmro(cls_obj)
]
if args.exclude_benchmarks and is_benchmark:
continue
test_okay = True
test_start_time = time.time()
# All test methods starts with "test_"
for mth in [o for o in dir(cls_obj) if o.startswith("test_")]:
name = "%s/%s/%s" % (f, cls[5:], mth[5:])
print("Testing " + _C.OKGREEN + str(name) + _C.ENDC,
end=" ")
sys.stdout.flush()
for (np_arys, cmd) in getattr(cls_inst, "init")():
# Exclude complex
if args.exclude_complex_dtype:
complex_nptypes = [
eval(dtype) for dtype in TYPES.COMPLEX
]
index = 0
non_complex = {}
for ary in np_arys.values():
if ary.dtype not in complex_nptypes:
non_complex[index] = ary
index += 1
np_arays = non_complex
# Get Bohrium arrays
bh_arys = []
for a in np_arys.values():
bh_arys.append(bh.array(a))
# Execute using NumPy
(res1, cmd1) = getattr(cls_inst, mth)(np_arys)
res1 = res1.copy()
# Execute using Bohrium
(res2, cmd2) = getattr(cls_inst, mth)(bh_arys)
cmd += cmd1
# Compare
try:
if not np.isscalar(res2) and bh.check(res2):
res2 = res2.copy2numpy()
except RuntimeError as error_msg:
test_okay = False
print()
print(" " + _C.OKBLUE + "[CMD] %s" % cmd +
_C.ENDC)
print(" " + _C.FAIL + str(error_msg) + _C.ENDC)
else:
rtol = cls_inst.config['maxerror']
atol = rtol * 0.1
if not np.allclose(
res1, res2, rtol=rtol, atol=atol):
test_okay = False
if 'warn_on_err' in cls_inst.config:
print()
print(_C.WARNING + " [Warning] %s" %
(name) + _C.ENDC)
print(_C.OKBLUE + " [CMD] %s" % cmd +
_C.ENDC)
print(_C.OKGREEN + " NumPy result: %s" %
(res1) + _C.ENDC)
print(_C.FAIL + " Bohrium result: %s" %
(res2) + _C.ENDC)
print(_C.WARNING + " " +
str(cls_inst.config['warn_on_err']) +
_C.ENDC)
print(_C.OKBLUE +
" Manual verification is needed." +
_C.ENDC)
else:
print()
print(_C.FAIL + " [Error] %s" % (name) +
_C.ENDC)
print(_C.OKBLUE + " [CMD] %s" % cmd +
_C.ENDC)
print(_C.OKGREEN + " NumPy result: %s" %
(res1) + _C.ENDC)
print(_C.FAIL + " Bohrium result: %s" %
(res2) + _C.ENDC)
sys.exit(1)
if test_okay:
print(
_C.OKBLUE +
"({:.2f}s)".format(time.time() - test_start_time) +
_C.ENDC, "✓")
print("*** Finished in: " + _C.OKBLUE +
"{:.2f}s".format(time.time() - test_suite_start_time) + _C.ENDC +
" ***")
def run(self, pseudo_input):
"""
Run the Benchmark script and return the result.
Benchmarks are assumed to be installed along with the Bohrium module.
"""
(target, dtype) = self.get_meta(pseudo_input)
# Setup output filename
outputfn = "/tmp/%s_%s_%s_output_%s.npz" % (self.script, dtype, target,
self.uuid)
# Execute the benchmark
import benchpress
benchmark_path = benchpress.suite_util.benchmark_path(
self.script, "python_numpy", ".py")
sys_exec = [sys.executable] if target.lower() == "none" else [
sys.executable, "-m", "bohrium"
]
# Setup command
cmd = sys_exec + [
benchmark_path, '--size=' + self.sizetxt, '--dtype=' + str(dtype),
'--target=' + target, '--outputfn=' + outputfn
]
# Setup the inputfn if one is needed/provided
if self.inputfn:
npt_path = os.path.dirname(sys.argv[0])
if not npt_path:
npt_path = "./"
inputfn = "%s/datasets/%s" % (npt_path, self.inputfn.format(dtype))
cmd.append('--inputfn')
cmd.append(inputfn)
if not os.path.exists(inputfn):
raise Exception('File does not exist: %s\n' % inputfn)
env = os.environ.copy()
env['BH_PROXY_PORT'] = "4201"
# SIP on macOS won't allow passing on DYLD_LIBRARY_PATH in env, so
# we attach it to the command instead.
if "DYLD_LIBRARY_PATH" in env:
cmd = ["DYLD_LIBRARY_PATH=" + env["DYLD_LIBRARY_PATH"]] + cmd
# Execute the benchmark
(out, err) = shell_cmd(cmd, verbose=self.args.verbose, env=env)
if 'elapsed-time' not in out:
raise Exception(
_C.FAIL +
"Cannot find elapsed time, output:\n%s\nWith error:\n%s\n" %
(out, err) + _C.ENDC)
if not os.path.exists(outputfn):
raise Exception(
_C.FAIL +
'Benchmark did not produce any output, expected: %s\n' %
outputfn + _C.ENDC)
# Load the result from disk
npzs = np.load(outputfn)
res = {}
for k in npzs:
res[k] = npzs[k]
# Delete npz
del npzs
# Delete the result from disk
if os.path.exists(outputfn):
os.remove(outputfn)
# Convert to whatever namespace it ought to be in
res['res'] = bh.array(res['res'], bohrium=target != "None")
return (res['res'], ' '.join(cmd))
def __threshold(name, a, thresh, maxval):
b = np.array([thresh, maxval], dtype=a.dtype)
c = np.empty_like(a)
ufuncs.extmethod(name, c, a, b)
return c
def run(self, pseudo_input):
"""
Run the Benchmark script and return the result.
Benchmarks are assumed to be installed along with the Bohrium module.
"""
(target, dtype) = self.get_meta(pseudo_input)
# Setup output filename
outputfn = "/tmp/%s_%s_%s_output_%s.npz" % (
self.script, dtype, target, self.uuid
)
benchmarks_dir, err = subprocess.Popen( # Execute the benchmark
['bp-info', '--benchmarks'],
stdout = subprocess.PIPE,
stderr = subprocess.PIPE,
).communicate()
sys_exec = [sys.executable] if target.lower() == "none" else [sys.executable, "-m", "bohrium"]
benchmark_path = os.sep.join([benchmarks_dir.strip(), self.script, "python_numpy", self.script+ ".py"])
# Setup command
cmd = sys_exec + [
benchmark_path,
'--size=' + self.sizetxt,
'--dtype=' + str(dtype),
'--target=' + target,
'--outputfn=' + outputfn
]
# Setup the inputfn if one is needed/provided
if self.inputfn:
npt_path = os.path.dirname(sys.argv[0])
if not npt_path:
npt_path = "./"
inputfn = "%s/datasets/%s" % (
npt_path,
self.inputfn.format(dtype)
)
cmd.append('--inputfn')
cmd.append(inputfn)
if not os.path.exists(inputfn):
raise Exception('File does not exist: %s\n' % inputfn)
env = os.environ.copy()
env['BH_PROXY_PORT'] = "4201"
out = shell_cmd(cmd, verbose=self.args.verbose, env=env) # Execute the benchmark
if 'elapsed-time' not in out:
raise Exception("Cannot find elapsed time, output:\n%s\n" % out)
if not os.path.exists(outputfn):
raise Exception('Benchmark did not produce any output, expected: %s\n' % outputfn)
npzs = np.load(outputfn) # Load the result from disk
res = {}
for k in npzs:
res[k] = npzs[k]
del npzs # Delete npz
if os.path.exists(outputfn): # Delete the result from disk
os.remove(outputfn)
# Convert to whatever namespace it ought to be in
res['res'] = bh.array(res['res'], bohrium=target!="None")
return (res['res'], ' '.join(cmd))
for (np_arys, cmd) in getattr(cls_inst,"init")():
if args.exclude_complex_dtype: # Exclude complex
complex_nptypes = [eval(dtype) for dtype in TYPES.COMPLEX]
index = 0
non_complex = {}
for ary in np_arys.values():
if ary.dtype not in complex_nptypes:
non_complex[index] = ary
index += 1
np_arays = non_complex
bh_arys = [] # Get Bohrium arrays
for a in np_arys.values():
bh_arys.append(bh.array(a))
# Execute using NumPy
(res1,cmd1) = getattr(cls_inst,mth)(np_arys)
res1 = res1.copy()
# Execute using Bohrium
(res2,cmd2) = getattr(cls_inst,mth)(bh_arys)
cmd += cmd1
try: # Compare
if not np.isscalar(res2) and bh.check(res2):
res2 = res2.copy2numpy()
except RuntimeError as error_msg:
test_okay = False
print()
print(" " + _C.OKBLUE + "[CMD] %s"%cmd + _C.ENDC)
print(" " + _C.FAIL + str(error_msg) + _C.ENDC)
else:
def colormap(img, colors=None):
"""
apply the colors to a grayscale image
if a color image is provided, convert it (thus, acts like a "colorize" function)
:param img: a grayscale image
:param colors: [(decimalPercent,color),(...)]
if no colors are given, then [(0.0,black),(1.0,white)]
if a single color and no percent is given, assume [(0.0,black),(0.5,theColor),(1.0,white)]
:return: the resulting image
"""
img = grayscale(img)
if not isFloat(img):
img = img / 255.0
if colors is None:
colors = [(0.0, (0.0, 0.0, 0.0)), (1.0, (1.0, 1.0, 1.0))]
elif not isinstance(colors[0], (tuple, list, np.ndarray)):
white = []
black = []
if isinstance(colors, str):
colors = strToColor(colors)
if isFloat(colors):
imax = 1.0
imin = 0.0
else:
imax = 255
imin = 0
for _ in range(len(colors)):
white.append(imax)
black.append(imin)
if len(colors) in [2, 4]: # keep same alpha value
black[-1] = colors[-1]
white[-1] = white[-1]
colors = [(0.0, black), (0.5, colors), (1.0, white)]
else:
colors.sort() # make sure we go from low to high
# make sure colors are in the shape we need
colors = [[
matchColorToImage(color[0], img),
np.array(strToColor(color[1]))
] for color in colors]
shape = (img.shape[0], img.shape[1], len(color[1]))
img2 = np.ndarray(shape)
img = img[..., None]
if True:
lastColor = None
for color in colors:
if lastColor is None:
img2 += color[1]
else:
percent = (img - lastColor[0]) * lastColor[0] / color[0]
img2 = np.where(
np.logical_and(img > lastColor[0], img <= color[0]),
(color[1] * percent) + (lastColor[1] * (1 - percent)),
img2)
lastColor = color
img2 = np.where(img > lastColor[0], lastColor[1], img2)
else:
def gradMap(c):
lastColor = None
for color in colors:
if c < color[0]:
if lastColor is None:
return color[1]
percent = (c - lastColor[0]) / color[0]
return (lastColor[1] * percent + color[1]) / (2 * percent)
lastColor = color
return lastColor[1]
img2 = perPixel(gradMap, img, clamp=False)
return img2
def run(self, pseudo_input):
"""
Run the Benchmark script and return the result.
Benchmarks are assumed to be installed along with the Bohrium module.
"""
(target, dtype) = self.get_meta(pseudo_input)
# Setup output filename
outputfn = "/tmp/%s_%s_%s_output_%s.npz" % (
self.script, dtype, target, self.uuid
)
benchmarks_dir, err = subprocess.Popen( # Execute the benchmark
['bp-info', '--benchmarks'],
stdout = subprocess.PIPE,
stderr = subprocess.PIPE,
).communicate()
sys_exec = [sys.executable, "-m", "bohrium"] if target else [sys.executable]
benchmark_path = os.sep.join([benchmarks_dir.strip(), self.script, "python_numpy", self.script+ ".py"])
# Setup command
cmd = sys_exec + [
benchmark_path,
'--size=' +self.sizetxt,
'--dtype=' +str(dtype),
'--target=' +target,
'--outputfn=' +outputfn
]
# Setup the inputfn if one is needed/provided
if self.inputfn:
npt_path = os.path.dirname(sys.argv[0])
if not npt_path:
npt_path = "./"
inputfn = "%s/datasets/%s" % (
npt_path,
self.inputfn.format(dtype)
)
cmd.append('--inputfn')
cmd.append(inputfn)
if not os.path.exists(inputfn):
raise Exception('File does not exist: %s' % inputfn)
p = subprocess.Popen( # Execute the benchmark
cmd,
stdout = subprocess.PIPE,
stderr = subprocess.PIPE,
)
out, err = p.communicate()
if 'elapsed-time' not in out:
raise Exception("Cannot find elapsed time got stdout(%s) stderr(%s)]" % (out, err))
if not os.path.exists(outputfn):
raise Exception('Benchmark did not produce any output, expected: %s' % outputfn)
#
# We silently accept these errors when output to stderr:
#
# * The Python object count
# * Unsupported fuse model
# * Unknown fuse model
#
if err and not re.match("\[[0-9]+ refs\]", err) \
and 'unsupported fuse model' not in err \
and 'unknown fuse model' not in err:
err_chunked = ", ".join(err.split("\n"))
print(_C.OKBLUE + "[CMD] %s" % " ".join(cmd) + _C.ENDC)
print(_C.WARNING + "[Warning] The CMD above wrote the following to stderr: [%s]" % err_chunked + _C.ENDC)
npzs = np.load(outputfn) # Load the result from disk
res = {}
for k in npzs:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
res[k] = npzs[k]
del npzs # Delete npz
if os.path.exists(outputfn): # Delete the result from disk
os.remove(outputfn)
# Convert to whatever namespace it ought to be in
res['res'] = bh.array(res['res'], bohrium=target!="None")
return (res['res'], ' '.join(cmd))
def tensordot(a, b, axes=2):
"""
Compute tensor dot product along specified axes for arrays >= 1-D.
Given two tensors (arrays of dimension greater than or equal to one),
`a` and `b`, and an array_like object containing two array_like
objects, ``(a_axes, b_axes)``, sum the products of `a`'s and `b`'s
elements (components) over the axes specified by ``a_axes`` and
``b_axes``. The third argument can be a single non-negative
integer_like scalar, ``N``; if it is such, then the last ``N``
dimensions of `a` and the first ``N`` dimensions of `b` are summed
over.
Parameters
----------
a, b : array_like, len(shape) >= 1
Tensors to "dot".
axes : variable type
* integer_like scalar
Number of axes to sum over (applies to both arrays); or
* (2,) array_like, both elements array_like of the same length
List of axes to be summed over, first sequence applying to `a`,
second to `b`.
See Also
--------
dot, einsum
Notes
-----
When there is more than one axis to sum over - and they are not the last
(first) axes of `a` (`b`) - the argument `axes` should consist of
two sequences of the same length, with the first axis to sum over given
first in both sequences, the second axis second, and so forth.
Examples
--------
A "traditional" example:
>>> a = np.arange(60.).reshape(3,4,5)
>>> b = np.arange(24.).reshape(4,3,2)
>>> c = np.tensordot(a,b, axes=([1,0],[0,1]))
>>> c.shape
(5, 2)
>>> c
array([[ 4400., 4730.],
[ 4532., 4874.],
[ 4664., 5018.],
[ 4796., 5162.],
[ 4928., 5306.]])
>>> # A slower but equivalent way of computing the same...
>>> d = np.zeros((5,2))
>>> for i in range(5):
... for j in range(2):
... for k in range(3):
... for n in range(4):
... d[i,j] += a[k,n,i] * b[n,k,j]
>>> c == d
array([[ True, True],
[ True, True],
[ True, True],
[ True, True],
[ True, True]], dtype=bool)
An extended example taking advantage of the overloading of + and \\*:
>>> a = np.array(range(1, 9))
>>> a.shape = (2, 2, 2)
>>> A = np.array(('a', 'b', 'c', 'd'), dtype=object)
>>> A.shape = (2, 2)
>>> a; A
array([[[1, 2],
[3, 4]],
[[5, 6],
[7, 8]]])
array([[a, b],
[c, d]], dtype=object)
>>> np.tensordot(a, A) # third argument default is 2
array([abbcccdddd, aaaaabbbbbbcccccccdddddddd], dtype=object)
>>> np.tensordot(a, A, 1)
array([[[acc, bdd],
[aaacccc, bbbdddd]],
[[aaaaacccccc, bbbbbdddddd],
[aaaaaaacccccccc, bbbbbbbdddddddd]]], dtype=object)
>>> np.tensordot(a, A, 0) # "Left for reader" (result too long to incl.)
array([[[[[a, b],
[c, d]],
...
>>> np.tensordot(a, A, (0, 1))
array([[[abbbbb, cddddd],
[aabbbbbb, ccdddddd]],
[[aaabbbbbbb, cccddddddd],
[aaaabbbbbbbb, ccccdddddddd]]], dtype=object)
>>> np.tensordot(a, A, (2, 1))
array([[[abb, cdd],
[aaabbbb, cccdddd]],
[[aaaaabbbbbb, cccccdddddd],
[aaaaaaabbbbbbbb, cccccccdddddddd]]], dtype=object)
>>> np.tensordot(a, A, ((0, 1), (0, 1)))
array([abbbcccccddddddd, aabbbbccccccdddddddd], dtype=object)
>>> np.tensordot(a, A, ((2, 1), (1, 0)))
array([acccbbdddd, aaaaacccccccbbbbbbdddddddd], dtype=object)
"""
try:
iter(axes)
except:
axes_a = list(range(-axes, 0))
axes_b = list(range(0, axes))
else:
axes_a, axes_b = axes
try:
na = len(axes_a)
axes_a = list(axes_a)
except TypeError:
axes_a = [axes_a]
na = 1
try:
nb = len(axes_b)
axes_b = list(axes_b)
except TypeError:
axes_b = [axes_b]
nb = 1
a, b = np.array(a), np.array(b)
as_ = a.shape
nda = len(a.shape)
bs = b.shape
ndb = len(b.shape)
equal = True
if (na != nb):
equal = False
else:
for k in range(na):
if as_[axes_a[k]] != bs[axes_b[k]]:
equal = False
break
if axes_a[k] < 0:
axes_a[k] += nda
if axes_b[k] < 0:
axes_b[k] += ndb
if not equal:
raise ValueError("shape-mismatch for sum")
# Move the axes to sum over to the end of "a"
# and to the front of "b"
notin = [k for k in range(nda) if k not in axes_a]
newaxes_a = notin + axes_a
N2 = 1
for axis in axes_a:
N2 *= as_[axis]
newshape_a = (-1, N2)
olda = [as_[axis] for axis in notin]
notin = [k for k in range(ndb) if k not in axes_b]
newaxes_b = axes_b + notin
N2 = 1
for axis in axes_b:
N2 *= bs[axis]
newshape_b = (N2, -1)
oldb = [bs[axis] for axis in notin]
at = a.transpose(newaxes_a).reshape(newshape_a)
bt = b.transpose(newaxes_b).reshape(newshape_b)
res = dot(at, bt)
return res.reshape(olda + oldb)
def init(self, grid, material, steps):
assert cfg['bh_stack'] in ['openmp', 'opencl', 'cuda']
os.environ["BH_STACK"] = self.cfg['bh_stack']
grid.T1 = bohrium.array(grid.T1)
grid.T2 = bohrium.array(grid.T2)
grid.T3 = bohrium.array(grid.T3)
grid.T4 = bohrium.array(grid.T4)
grid.T5 = bohrium.array(grid.T5)
grid.T6 = bohrium.array(grid.T6)
grid.ux_new = bohrium.array(grid.ux_new)
grid.uy_new = bohrium.array(grid.uy_new)
grid.uz_new = bohrium.array(grid.uz_new)
material.C = bohrium.array(material.C)
material.P = bohrium.array(material.P)
super().init(grid, material, steps)
def tensordot(a, b, axes=2):
"""
Compute tensor dot product along specified axes for arrays >= 1-D.
Given two tensors (arrays of dimension greater than or equal to one),
`a` and `b`, and an array_like object containing two array_like
objects, ``(a_axes, b_axes)``, sum the products of `a`'s and `b`'s
elements (components) over the axes specified by ``a_axes`` and
``b_axes``. The third argument can be a single non-negative
integer_like scalar, ``N``; if it is such, then the last ``N``
dimensions of `a` and the first ``N`` dimensions of `b` are summed
over.
Parameters
----------
a, b : array_like, len(shape) >= 1
Tensors to "dot".
axes : variable type
* integer_like scalar
Number of axes to sum over (applies to both arrays); or
* (2,) array_like, both elements array_like of the same length
List of axes to be summed over, first sequence applying to `a`,
second to `b`.
See Also
--------
dot, einsum
Notes
-----
When there is more than one axis to sum over - and they are not the last
(first) axes of `a` (`b`) - the argument `axes` should consist of
two sequences of the same length, with the first axis to sum over given
first in both sequences, the second axis second, and so forth.
Examples
--------
A "traditional" example:
>>> a = np.arange(60.).reshape(3,4,5)
>>> b = np.arange(24.).reshape(4,3,2)
>>> c = np.tensordot(a,b, axes=([1,0],[0,1]))
>>> c.shape
(5, 2)
>>> c
array([[ 4400., 4730.],
[ 4532., 4874.],
[ 4664., 5018.],
[ 4796., 5162.],
[ 4928., 5306.]])
>>> # A slower but equivalent way of computing the same...
>>> d = np.zeros((5,2))
>>> for i in range(5):
... for j in range(2):
... for k in range(3):
... for n in range(4):
... d[i,j] += a[k,n,i] * b[n,k,j]
>>> c == d
array([[ True, True],
[ True, True],
[ True, True],
[ True, True],
[ True, True]], dtype=bool)
An extended example taking advantage of the overloading of + and \\*:
>>> a = np.array(range(1, 9))
>>> a.shape = (2, 2, 2)
>>> A = np.array(('a', 'b', 'c', 'd'), dtype=object)
>>> A.shape = (2, 2)
>>> a; A
array([[[1, 2],
[3, 4]],
[[5, 6],
[7, 8]]])
array([[a, b],
[c, d]], dtype=object)
>>> np.tensordot(a, A) # third argument default is 2
array([abbcccdddd, aaaaabbbbbbcccccccdddddddd], dtype=object)
>>> np.tensordot(a, A, 1)
array([[[acc, bdd],
[aaacccc, bbbdddd]],
[[aaaaacccccc, bbbbbdddddd],
[aaaaaaacccccccc, bbbbbbbdddddddd]]], dtype=object)
>>> np.tensordot(a, A, 0) # "Left for reader" (result too long to incl.)
array([[[[[a, b],
[c, d]],
...
>>> np.tensordot(a, A, (0, 1))
array([[[abbbbb, cddddd],
[aabbbbbb, ccdddddd]],
[[aaabbbbbbb, cccddddddd],
[aaaabbbbbbbb, ccccdddddddd]]], dtype=object)
>>> np.tensordot(a, A, (2, 1))
array([[[abb, cdd],
[aaabbbb, cccdddd]],
[[aaaaabbbbbb, cccccdddddd],
[aaaaaaabbbbbbbb, cccccccdddddddd]]], dtype=object)
>>> np.tensordot(a, A, ((0, 1), (0, 1)))
array([abbbcccccddddddd, aabbbbccccccdddddddd], dtype=object)
>>> np.tensordot(a, A, ((2, 1), (1, 0)))
array([acccbbdddd, aaaaacccccccbbbbbbdddddddd], dtype=object)
"""
try:
iter(axes)
except:
axes_a = list(range(-axes, 0))
axes_b = list(range(0, axes))
else:
axes_a, axes_b = axes
try:
na = len(axes_a)
axes_a = list(axes_a)
except TypeError:
axes_a = [axes_a]
na = 1
try:
nb = len(axes_b)
axes_b = list(axes_b)
except TypeError:
axes_b = [axes_b]
nb = 1
a, b = np.array(a), np.array(b)
as_ = a.shape
nda = len(a.shape)
bs = b.shape
ndb = len(b.shape)
equal = True
if (na != nb):
equal = False
else:
for k in range(na):
if as_[axes_a[k]] != bs[axes_b[k]]:
equal = False
break
if axes_a[k] < 0:
axes_a[k] += nda
if axes_b[k] < 0:
axes_b[k] += ndb
if not equal:
raise ValueError("shape-mismatch for sum")
# Move the axes to sum over to the end of "a"
# and to the front of "b"
notin = [k for k in range(nda) if k not in axes_a]
newaxes_a = notin + axes_a
N2 = 1
for axis in axes_a:
N2 *= as_[axis]
newshape_a = (-1, N2)
olda = [as_[axis] for axis in notin]
notin = [k for k in range(ndb) if k not in axes_b]
newaxes_b = axes_b + notin
N2 = 1
for axis in axes_b:
N2 *= bs[axis]
newshape_b = (N2, -1)
oldb = [bs[axis] for axis in notin]
at = a.transpose(newaxes_a).reshape(newshape_a)
bt = b.transpose(newaxes_b).reshape(newshape_b)
res = dot(at, bt)
return res.reshape(olda + oldb)
def run(self, pseudo_input):
"""
Run the Benchmark script and return the result.
Benchmarks are assumed to be installed along with the Bohrium module.
"""
(target, dtype) = self.get_meta(pseudo_input)
# Setup output filename
outputfn = "/tmp/%s_%s_%s_output_%s.npz" % (
self.script, dtype, target, self.uuid
)
benchmarks_dir, err = subprocess.Popen( # Execute the benchmark
['bp-info', '--benchmarks'],
stdout = subprocess.PIPE,
stderr = subprocess.PIPE,
).communicate()
sys_exec = [sys.executable, "-m", "bohrium"] if target else [sys.executable]
benchmark_path = os.sep.join([benchmarks_dir.strip(), self.script, "python_numpy", self.script+ ".py"])
# Setup command
cmd = sys_exec + [
benchmark_path,
'--size=' +self.sizetxt,
'--dtype=' +str(dtype),
'--target=' +target,
'--outputfn=' +outputfn
]
# Setup the inputfn if one is needed/provided
if self.inputfn:
npt_path = os.path.dirname(sys.argv[0])
if not npt_path:
npt_path = "./"
inputfn = "%s/datasets/%s" % (
npt_path,
self.inputfn.format(dtype)
)
cmd.append('--inputfn')
cmd.append(inputfn)
if not os.path.exists(inputfn):
raise Exception('File does not exist: %s' % inputfn)
p = subprocess.Popen( # Execute the benchmark
cmd,
stdout = subprocess.PIPE,
stderr = subprocess.PIPE,
)
out, err = p.communicate()
if 'elapsed-time' not in out:
raise Exception("Benchmark error [stdout:%s,stderr:%s]" % (out, err))
if err and not re.match("\[[0-9]+ refs\]", err): #We accept the Python object count
raise Exception("Benchmark error[%s]" % err)
if not os.path.exists(outputfn):
raise Exception('Benchmark did not produce the output: %s' % outputfn)
npzs = np.load(outputfn) # Load the result from disk
res = {}
for k in npzs:
res[k] = npzs[k]
del npzs # Delete npz
if os.path.exists(outputfn): # Delete the result from disk
os.remove(outputfn)
# Convert to whatever namespace it ought to be in
res['res'] = bh.array(res['res'], bohrium=target!="None")
return (res['res'], ' '.join(cmd))
#!/usr/bin/dython2 import bohrium as bh a = bh.array(range(9 * 100000)) a = a + a b = a + 2 a = a + b print(a)
def connected_components(a, connectivity=8):
b = np.array([connectivity], dtype=a.dtype)
c = np.zeros_like(a)
ufuncs.extmethod("opencv_connected_components", c, a, b)
return c
