def _save_signal_data(self, db, dbhist, analysis, dbdet, iso, m, kind):
if not (len(m.xs) and len(m.ys)):
self.debug('no data for {} {}'.format(iso.name, kind))
return
self.debug('saving data {} {} xs={}'.format(iso.name, kind, len(m.xs)))
dbiso = db.add_isotope(analysis, iso.name, dbdet, kind=kind)
data = ''.join([struct.pack('>ff', x, y) for x, y in zip(m.xs, m.ys)])
db.add_signal(dbiso, data)
add_result = kind in ('baseline', 'signal')
if add_result:
fod = self._get_filter_outlier_dict(iso, kind)
m.set_filtering(fod)
if m.fit:
# add fit
db.add_fit(dbhist, dbiso,
fit=m.fit,
filter_outliers=fod.get('filter_outliers', False),
filter_outlier_iterations=fod.get('iterations', 1),
filter_outlier_std_devs=fod.get('std_devs', 2))
# add isotope result
# print 'a',m.value, m.error, type(m.error), type(nan)
v, e = float(m.value), float(m.error)
v = 0 if math.isnan(v) or math.isinf(v) else v
e = 0 if math.isnan(e) or math.isinf(e) else e
db.add_isotope_result(dbiso,
dbhist,
signal_=v, signal_err=e)
def adapt_canopsis_data_to_ember(data):
"""
Transform canopsis data to ember data (in changing ``id`` to ``cid``).
:param data: data to transform
"""
if isinstance(data, dict):
for key, item in data.iteritems():
if isinstance(item, float) and (isnan(item) or isinf(item)):
data[key] = None
else:
if isinstance(item, (tuple, frozenset)):
item = list(item)
data[key] = item
adapt_canopsis_data_to_ember(item)
elif isiterable(data, is_str=False):
for i in range(len(data)):
item = data[i]
if isinstance(item, float) and (isnan(item) or isinf(item)):
data[i] = None
else:
if isinstance(item, (tuple, frozenset)):
item = list(item)
data[i] = item
adapt_canopsis_data_to_ember(item)
def strictly_simpler(self, x, y):
if math.isnan(x):
return False
if math.isnan(y):
return True
if math.isinf(y) and not math.isinf(x):
return True
if math.isinf(x) and not math.isinf(y):
return False
if x < 0 and y >= 0:
return False
if y < 0 and x >= 0:
return True
if is_integral(x):
if not is_integral(y):
return True
return self.int_strategy.strictly_simpler(int(x), int(y))
if is_integral(y):
return False
if y > 0:
return 0 <= x < y
else:
# The y == 0 case is handled by is_integral(y)
assert y < 0
return x > y
def PureStrategyDominance(game, conditional=True, weak=False):
"""
pure-strategy dominance criterion for IEDS
conditional==0==False --> unconditional dominance
conditional==1==True ---> conditional dominance
conditional==2 ---------> extra-conservative conditional dominance
"""
undominated = {r:set(game.strategies[r]) for r in game.roles}
for r in game.roles:
for dominant, dominated in product(game.strategies[r], repeat=2):
if dominant == dominated or dominated not in undominated[r]:
continue
dominance_proved = False
for profile in game:
if dominated in profile[r]:
reg = regret(game, profile, r, dominated, dominant)
if reg > 0 and not isinf(reg):
dominance_proved = True
elif (reg < 0) or (reg == 0 and not weak) or \
(isinf(reg) and conditional):
dominance_proved = False
break
elif dominant in profile[r] and conditional > 1:
if profile.deviate(r, dominant, dominated) not in game:
dominance_proved = False
break
if dominance_proved:
undominated[r].remove(dominated)
return Subgame(game, undominated)
def test_pow(self):
import math
def pw(x, y):
return x ** y
def espeq(x, y):
return not abs(x-y) > 1e05
raises(ZeroDivisionError, pw, 0.0, -1)
assert pw(0, 0.5) == 0.0
assert espeq(pw(4.0, 0.5), 2.0)
assert pw(4.0, 0) == 1.0
assert pw(-4.0, 0) == 1.0
assert type(pw(-1.0, 0.5)) == complex
assert pw(-1.0, 2.0) == 1.0
assert pw(-1.0, 3.0) == -1.0
assert pw(-1.0, 1e200) == 1.0
if self.py26:
assert pw(0.0, float("-inf")) == float("inf")
assert math.isnan(pw(-3, float("nan")))
assert math.isnan(pw(-3., float("nan")))
assert pw(-1.0, -float('inf')) == 1.0
assert pw(-1.0, float('inf')) == 1.0
assert pw(float('inf'), 0) == 1.0
assert pw(float('nan'), 0) == 1.0
assert math.isinf(pw(-0.5, float('-inf')))
assert math.isinf(pw(+0.5, float('-inf')))
assert pw(-1.5, float('-inf')) == 0.0
assert pw(+1.5, float('-inf')) == 0.0
assert str(pw(float('-inf'), -0.5)) == '0.0'
assert str(pw(float('-inf'), -2.0)) == '0.0'
assert str(pw(float('-inf'), -1.0)) == '-0.0'
assert str(pw(float('-inf'), 1.0)) == '-inf'
assert str(pw(float('-inf'), 2.0)) == 'inf'
def py_quadInterpolate(C,X1,X2,X3,Y1,Y2,Y3): resL = quadInterpolate(-1*C,X1,X2,X3,Y1,Y2,Y3) resH = quadInterpolate(C,X1,X2,X3,Y1,Y2,Y3) if math.isnan(resL) or math.isinf(resL) or math.isnan(resH) or math.isinf(resL): return " - " if abs(resL - 1) < 0.00001 or abs(resL - 1) > 1: return " - " if abs(resH - 1) < 0.00001 or abs(resH - 1) > 1: return " - " return " %.3f/%.3f "%(resL,resH)
def xsString(xc, p, source):
if isinstance(source,bool):
return 'true' if source else 'false'
elif isinstance(source,float):
if isnan(source):
return "NaN"
elif isinf(source):
return "INF"
'''
numMagnitude = fabs(source)
if numMagnitude < 1000000 and numMagnitude > .000001:
# don't want floating notation which python does for more than 4 decimal places
s =
'''
s = str(source)
if s.endswith(".0"):
s = s[:-2]
return s
elif isinstance(source,Decimal):
if isnan(source):
return "NaN"
elif isinf(source):
return "INF"
return str(source)
elif isinstance(source,ModelValue.DateTime):
return ('{0:%Y-%m-%d}' if source.dateOnly else '{0:%Y-%m-%dT%H:%M:%S}').format(source)
return str(source)
def isclose(a, b, rel_tol = 1e-09, abs_tol = 0.0):
"""
returns True if a is close in value to b. False otherwise
:param a: one of the values to be tested
:param b: the other value to be tested
:param rel_tol=1e-9: The relative tolerance -- the amount of error
allowed, relative to the absolute value of the
larger input values.
:param abs_tol=0.0: The minimum absolute tolerance level -- useful
for comparisons to zero.
NOTES:
-inf, inf and NaN behave similarly to the IEEE 754 Standard. That
is, NaN is not close to anything, even itself. inf and -inf are
only close to themselves.
The function can be used with any type that supports comparison,
substratcion and multiplication, including Decimal, Fraction, and
Complex
Complex values are compared based on their absolute value.
See PEP-0485 for a detailed description
"""
if a == b:
return True
if rel_tol < 0.0 or abs_tol < 0.0:
raise ValueError('error tolerances must be non-negative')
if math.isinf(abs(a)) or math.isinf(abs(b)):
return False
diff = abs(b - a)
return diff <= abs(rel_tol * b) or diff <= abs(rel_tol * a) or diff <= abs_tol
def combine_rshift(range1, range2):
"""
Combiner for Right shift operation.
>>> import ast
>>> combine(Range(10, 100), Range(3, 8), ast.RShift())
Range(low=0, high=12)
>>> combine(Range(10, float("inf")), Range(3, 8),
... ast.RShift())
Range(low=0, high=inf)
>>> combine(Range(-float("inf"), 0), Range(3, 8),
... ast.RShift())
Range(low=-inf, high=0)
>>> combine(Range(-30, 10), Range(3, float('inf')),
... ast.RShift())
Range(low=-4, high=1)
"""
if range1.low <= 0:
if isinf(range1.low):
min_ = range1.low
else:
min_ = range1.low >> range2.low
elif isinf(range2.high):
min_ = 0
else:
min_ = range1.low >> range2.high
if isinf(range1.high):
max_ = range1.high
elif isinf(range2.low):
max_ = 0
else:
max_ = range1.high >> range2.low
return Range(min_, max_)
def _encode_numbers(self, obj):
"""Returns a JSON representation of a Python number (int, float or Decimal)"""
# strict checks first - for speed
if obj.__class__ is int:
if abs(obj) > JAVASCRIPT_MAXINT:
raise ValueError("Number out of range: {!r}".format(obj))
return str(obj)
if obj.__class__ is float:
if isnan(obj):
raise ValueError("NaN is not supported")
if isinf(obj):
raise ValueError("Infinity is not supported")
return repr(obj)
# more in-depth class analysis last
if isinstance(obj, int):
obj = int(obj)
if abs(obj) > JAVASCRIPT_MAXINT:
raise ValueError("Number out of range: {!r}".format(obj))
return str(obj)
if isinstance(obj, float):
if isnan(obj):
raise ValueError("NaN is not supported")
if isinf(obj):
raise ValueError("Infinity is not supported")
return repr(obj)
if isinstance(obj, Decimal):
return '"' + str(obj) + '"'
# for complex and other Numbers
return self._encode(self.default(obj))
def isclose(a, b, rel_tol=1e-09, abs_tol=0.0):
"""
Determine whether two floating point numbers are close in value.
rel_tol
maximum difference for being considered "close", relative to the
magnitude of the input values
abs_tol
maximum difference for being considered "close", regardless of the
magnitude of the input values
Return True if a is close in value to b, and False otherwise.
For the values to be considered close, the difference between them
must be smaller than at least one of the tolerances.
-inf, inf and NaN behave similarly to the IEEE 754 Standard. That
is, NaN is not close to anything, even itself. inf and -inf are
only close to themselves.
"""
if rel_tol < 0.0 or abs_tol < 0.0:
raise ValueError('error tolerances must be non-negative')
if a == b: # short-circuit exact equality
return True
if math.isinf(a) or math.isinf(b):
# This includes the case of two infinities of opposite sign, or
# one infinity and one finite number. Two infinities of opposite sign
# would otherwise have an infinite relative tolerance.
return False
diff = abs(b - a)
return (((diff <= abs(rel_tol * b)) and
(diff <= abs(rel_tol * a))) or
(diff <= abs_tol))
def __rshift__(range1, range2):
"""
Combiner for Right shift operation.
>>> Interval(10, 100) >> Interval(3, 8)
Interval(low=0, high=12)
>>> Interval(10, float("inf")) >> Interval(3, 8)
Interval(low=0, high=inf)
>>> Interval(-float("inf"), 0) >> Interval(3, 8)
Interval(low=-inf, high=0)
>>> Interval(-30, 10) >> Interval(3, float('inf'))
Interval(low=-4, high=1)
"""
if range1.low <= 0:
if isinf(range1.low):
min_ = range1.low
else:
min_ = range1.low >> range2.low
elif isinf(range2.high):
min_ = 0
else:
min_ = range1.low >> range2.high
if isinf(range1.high):
max_ = range1.high
elif isinf(range2.low):
max_ = 0
else:
max_ = range1.high >> range2.low
return Interval(min_, max_)
def toVector(longitude, latitude):
"""Converts a set of spherical coordinates to a 3-vector.
The conversion shall not be performed by any library, to ensure
that the test case does not duplicate the code being tested.
Parameters
----------
longitude : `Angle`
The longitude (right ascension, azimuth, etc.) of the
position.
latitude : `Angle`
The latitude (declination, elevation, etc.) of the
position.
Returns
-------
x, y, z : `number`
Components of the unit vector representation of
`(longitude, latitude)`
"""
alpha = longitude.asRadians()
delta = latitude.asRadians()
if math.isnan(alpha) or math.isinf(alpha) or math.isnan(delta) or math.isinf(delta):
return (nan, nan, nan)
x = math.cos(alpha)*math.cos(delta)
y = math.sin(alpha)*math.cos(delta)
z = math.sin(delta)
return (x, y, z)
def test_floats_in_tiny_interval_within_bounds(data, center):
assume(not (math.isinf(next_down(center)) or math.isinf(next_up(center))))
lo = Decimal.from_float(next_down(center)).next_plus()
hi = Decimal.from_float(next_up(center)).next_minus()
assert float(lo) < lo < center < hi < float(hi)
val = data.draw(st.floats(lo, hi))
assert lo < val < hi
def retreive_WMS_metadata(typename):
workspace, layername = decodeTypeName(typename)
# workspace is hard-coded in the importer
url = settings.OGC_SERVER['default']['LOCATION'] + workspace+"/"
url += layername + "/wms?request=GetCapabilities&version=1.1.1"
get_cap_data = CreateStoryLayerThumbnailTask.request_geoserver_with_credentials(
url)
wms = WebMapService(url, xml=get_cap_data)
# I found that some dataset advertise illegal bounds - fix them up
xmin = wms[layername].boundingBoxWGS84[0]
if math.isnan(xmin) or math.isinf(xmin) or xmin < -180:
xmin = -180
ymin = wms[layername].boundingBoxWGS84[1]
if math.isnan(ymin) or math.isinf(ymin) or ymin < -90:
ymin = -90
xmax = wms[layername].boundingBoxWGS84[2]
if math.isnan(xmax) or math.isinf(xmax) or xmax > 180:
xmax = 180
ymax = wms[layername].boundingBoxWGS84[3]
if math.isnan(ymax) or math.isinf(ymax) or ymax > 90:
ymax = 90
return [xmin, ymin, xmax, ymax], wms[layername].timepositions
def compute_logp(self, mu, r, u, c, v):
# This function computes the log posterior probability (with the weight
# parameter marginalized out).
if math.isinf(self.prior_param['obs_df']):
loglik = - np.sum((self.y_coo.data - mu) ** 2 * self.prior_param['weight'])/ 2
else:
loglik = - (self.prior_param['obs_df'] + 1) / 2 * np.sum(
np.log( 1 + (self.y_coo.data - mu) ** 2 * self.prior_param['weight'] / self.prior_param['obs_df'])
)
r_scaled = r / self.prior_param['row_bias_scale']
c_scaled = c / self.prior_param['col_bias_scale']
u_scaled = u / np.tile(self.prior_param['factor_scale'], (u.shape[0], 1))
v_scaled = v / np.tile(self.prior_param['factor_scale'], (v.shape[0], 1))
if math.isinf(self.prior_param['param_df']):
logp_prior = \
- np.sum(r_scaled ** 2) / 2 + \
- np.sum(u_scaled ** 2, (0, 1)) / 2 + \
- np.sum(c_scaled ** 2) / 2 + \
- np.sum(v_scaled ** 2, (0, 1)) / 2
else:
logp_prior = \
- (self.prior_param['param_df'] + 1) / 2 * \
np.sum(np.log(1 + r_scaled ** 2 / self.prior_param['param_df'])) + \
- (self.prior_param['param_df'] + 1) / 2 * \
np.sum(np.log(1 + u_scaled ** 2 / self.prior_param['param_df']), (0, 1)) + \
- (self.prior_param['param_df'] + 1) / 2 * \
np.sum(np.log(1 + c_scaled ** 2 / self.prior_param['param_df'])) + \
- (self.prior_param['param_df'] + 1) / 2 * \
np.sum(np.log(1 + v_scaled ** 2 / self.prior_param['param_df']), (0, 1))
return loglik + logp_prior
def isclose(a, b, rel_tol=1e-09, abs_tol=0.0):
'''
Python 2 implementation of Python 3.5 math.isclose()
https://hg.python.org/cpython/file/tip/Modules/mathmodule.c#l1993
'''
# sanity check on the inputs
if rel_tol < 0 or abs_tol < 0:
raise ValueError("tolerances must be non-negative")
# short circuit exact equality -- needed to catch two infinities of
# the same sign. And perhaps speeds things up a bit sometimes.
if a == b:
return True
# This catches the case of two infinities of opposite sign, or
# one infinity and one finite number. Two infinities of opposite
# sign would otherwise have an infinite relative tolerance.
# Two infinities of the same sign are caught by the equality check
# above.
if math.isinf(a) or math.isinf(b):
return False
# equality check above would return false for nan, but we want
# to return true
if math.isnan(a) and math.isnan(b):
return True
# now do the regular computation
# this is essentially the "weak" test from the Boost library
diff = math.fabs(b - a)
result = (((diff <= math.fabs(rel_tol * b)) or
(diff <= math.fabs(rel_tol * a))) or
(diff <= abs_tol))
return result
def exp(x):
z = _make_complex(x)
exp_special = [
[0+0j, None, complex(0, -0.0), 0+0j, None, 0+0j, 0+0j],
[nan+nanj, None, None, None, None, nan+nanj, nan+nanj],
[nan+nanj, None, 1-0j, 1+0j, None, nan+nanj, nan+nanj],
[nan+nanj, None, 1-0j, 1+0j, None, nan+nanj, nan+nanj],
[nan+nanj, None, None, None, None, nan+nanj, nan+nanj],
[inf+nanj, None, complex(float("inf"), -0.0), inf, None, inf+nanj, inf+nanj],
[nan+nanj, nan+nanj, complex(float("nan"), -0.0), nan, nan+nanj, nan+nanj, nan+nanj]
]
if not isfinite(z):
if math.isinf(z.real) and math.isfinite(z.imag) and z.imag != 0:
if z.real > 0:
ret = complex(math.copysign(inf, math.cos(z.imag)),
math.copysign(inf, math.sin(z.imag)))
else:
ret = complex(math.copysign(0, math.cos(z.imag)),
math.copysign(0, math.sin(z.imag)))
else:
ret = exp_special[_special_type(z.real)][_special_type(z.imag)]
if math.isinf(z.imag) and (math.isfinite(z.real) or
(math.isinf(z.real) and z.real > 0)):
raise ValueError
return ret
if z.real > _LOG_LARGE_DOUBLE:
ret = e * rect(math.exp(z.real - 1), z.imag)
else:
ret = rect(math.exp(z.real), z.imag)
if math.isinf(ret.real) or math.isinf(ret.imag):
raise OverflowError
return ret
def test_unusual_numbers(self):
j = '{ "aListOfDouble": ["inf", "-inf", "nan"]}'
stuff_read = readStuffFromJSON(j)
self.assertEqual(len(stuff_read.aListOfDouble), 3)
self.assertTrue(math.isinf(stuff_read.aListOfDouble[0]))
self.assertTrue(math.isinf(stuff_read.aListOfDouble[1]))
self.assertTrue(math.isnan(stuff_read.aListOfDouble[2]))
def cosh(x):
_cosh_special = [
[inf+nanj, None, inf, complex(float("inf"), -0.0), None, inf+nanj, inf+nanj],
[nan+nanj, None, None, None, None, nan+nanj, nan+nanj],
[nan, None, 1, complex(1, -0.0), None, nan, nan],
[nan, None, complex(1, -0.0), 1, None, nan, nan],
[nan+nanj, None, None, None, None, nan+nanj, nan+nanj],
[inf+nanj, None, complex(float("inf"), -0.0), inf, None, inf+nanj, inf+nanj],
[nan+nanj, nan+nanj, nan, nan, nan+nanj, nan+nanj, nan+nanj]
]
z = _make_complex(x)
if not isfinite(z):
if math.isinf(z.imag) and not math.isnan(z.real):
raise ValueError
if math.isinf(z.real) and math.isfinite(z.imag) and z.imag != 0:
if z.real > 0:
return complex(math.copysign(inf, math.cos(z.imag)),
math.copysign(inf, math.sin(z.imag)))
return complex(math.copysign(inf, math.cos(z.imag)),
-math.copysign(inf, math.sin(z.imag)))
return _cosh_special[_special_type(z.real)][_special_type(z.imag)]
if abs(z.real) > _LOG_LARGE_DOUBLE:
x_minus_one = z.real - math.copysign(1, z.real)
ret = complex(e * math.cos(z.imag) * math.cosh(x_minus_one),
e * math.sin(z.imag) * math.sinh(x_minus_one))
else:
ret = complex(math.cos(z.imag) * math.cosh(z.real),
math.sin(z.imag) * math.sinh(z.real))
if math.isinf(ret.real) or math.isinf(ret.imag):
raise OverflowError
return ret
def sinh(x):
_sinh_special = [
[inf+nanj, None, complex(-float("inf"), -0.0), -inf, None, inf+nanj, inf+nanj],
[nan+nanj, None, None, None, None, nan+nanj, nan+nanj],
[nanj, None, complex(-0.0, -0.0), complex(-0.0, 0.0), None, nanj, nanj],
[nanj, None, complex(0.0, -0.0), complex(0.0, 0.0), None, nanj, nanj],
[nan+nanj, None, None, None, None, nan+nanj, nan+nanj],
[inf+nanj, None, complex(float("inf"), -0.0), inf, None, inf+nanj, inf+nanj],
[nan+nanj, nan+nanj, complex(float("nan"), -0.0), nan, nan+nanj, nan+nanj, nan+nanj]
]
z = _make_complex(x)
if not isfinite(z):
if math.isinf(z.imag) and not math.isnan(z.real):
raise ValueError
if math.isinf(z.real) and math.isfinite(z.imag) and z.imag != 0:
if z.real > 0:
return complex(math.copysign(inf, math.cos(z.imag)),
math.copysign(inf, math.sin(z.imag)))
return complex(-math.copysign(inf, math.cos(z.imag)),
math.copysign(inf, math.sin(z.imag)))
return _sinh_special[_special_type(z.real)][_special_type(z.imag)]
if abs(z.real) > _LOG_LARGE_DOUBLE:
x_minus_one = z.real - math.copysign(1, z.real)
return complex(math.cos(z.imag) * math.sinh(x_minus_one) * e,
math.sin(z.imag) * math.cosh(x_minus_one) * e)
return complex(math.cos(z.imag) * math.sinh(z.real),
math.sin(z.imag) * math.cosh(z.real))
def weigh_objects(self, weighed_obj_list, weight_properties):
"""Override the weigh objects.
This override calls the parent to do the weigh objects and then
replaces any infinite weights with a value that is a multiple of the
delta between the min and max values.
NOTE(jecarey): the infinite weight value is only used when the
smallest value is being favored (negative multiplier). When the
largest weight value is being used a weight of -1 is used instead.
See _weigh_object method.
"""
tmp_weights = super(CapacityWeigher, self).weigh_objects(
weighed_obj_list, weight_properties)
if math.isinf(self.maxval):
# NOTE(jecarey): if all weights were infinite then parent
# method returns 0 for all of the weights. Thus self.minval
# cannot be infinite at this point
copy_weights = [w for w in tmp_weights if not math.isinf(w)]
self.maxval = max(copy_weights)
offset = (self.maxval - self.minval) * OFFSET_MULT
self.maxval += OFFSET_MIN if offset == 0.0 else offset
tmp_weights = [self.maxval if math.isinf(w) else w
for w in tmp_weights]
return tmp_weights
def parallel_line2d(m, b, d):
b_ = None
if math.isinf(m) or math.isinf(-m):
b_ = d
else:
b_ = b + d * math.sqrt(m * m + 1)
return m, b_
def getBoundingBox2D(self):
"""
Returns the bounding box of all active process elements in their 2D coordinate system.
@return :
@author
"""
#Helper variables
maxX = float("-inf")
maxY = float("-inf")
minX = float("inf")
minY = float("inf")
#Now iterate over all active process components
for active in (self.activeProcessComponents + self.messageExchanges):
if(hasattr(active, "hasAbstractVisualRepresentation")):
point = active.hasAbstractVisualRepresentation.getPoint2D()
#Max tests
if(maxX < point[0]):
maxX = point[0]
if(maxY < point[1]):
maxY = point[1]
#Min tests
if(minX > point[0]):
minX = point[0]
if(minY > point[1]):
minY = point[1]
#inf tests
if(math.isinf(maxX)):
maxX = 0
minX = 0
if(math.isinf(maxY)):
maxY = 0
minY = 0
return [[minX, minY], [maxX, maxY]]
def lgamma(x):
"""Compute the natural logarithm of the gamma function for x."""
if isnan(x):
return x
if isinf(x):
return INFINITY
if x == math.floor(x) and x <= 2.0:
if x <= 0.0:
raise ValueError("math range error")
return 0.0
absx = abs(x)
if absx < 1e-20:
return -math.log(absx)
if x > 0.0:
r = math.log(_lanczos_sum(x)) - _lanczos_g + (x - 0.5) * (math.log(x + _lanczos_g - 0.5) - 1)
else:
r = (
math.log(math.pi)
- math.log(abs(_sinpi(absx)))
- math.log(absx)
- (math.log(_lanczos_sum(absx)) - _lanczos_g + (absx - 0.5) * (math.log(absx + _lanczos_g - 0.5) - 1))
)
if isinf(r):
raise OverflowError("math domain error")
return r
def dataToScreen(self, values):
results = []
if not isinstance(values,list):
values = [values]
for value in values:
try:
value = float(value)
if math.isinf(value):
value = 'Missing'
elif math.isinf(value):
value = 'Allele Masked'
except (ValueError,TypeError):
if value == variant.MISSING or value == None:
value = 'Missing'
elif value == variant.ALLELE_MASKED:
value = 'Allele Masked'
if isinstance(value,str):
index = self.cats.get(value,len(self.cats)-1)
if index < 0:
results.append(self.labelTop)
elif index > self.latticeLength:
results.append(self.labelBottom)
else:
results.append(self.labelTop + (index+0.5)*self.cellSize)
else:
results.append(self.numericPixelLow + (value-self.numericDataLow)*self.dataToPixelRatio)
return results
def normalize(self):
"It recovers the values got bad during the GML write-read cycle."
if "normalized" in dir(self) and self.normalized:
return
if "type" in self.vs.attributes():
virtual = self.vs.select(type=1)
for vertex in virtual:
for attr in ("priority", "filesize",
"section", "summary", "version"):
vertex[attr] = None
if "architecture" in self.vs.attributes():
vertex["architecture"] = None
if "revision" in self.attributes():
revision = self["revision"]
if isinstance(revision, float) and not math.isnan(revision):
self["revision"] = int(revision)
del self.vs["id"]
integer_attributes = (
("type", self.vs),
("filesize", self.vs),
("type", self.es),
)
for attr, object_ in integer_attributes:
if attr in object_.attributes():
for item in object_:
value = item[attr]
if isinstance(value, float) and\
not math.isinf(value) and not math.isnan(value):
item[attr] = int(value)
elif value is not None and (math.isinf(value) or math.isnan(value)):
print("The value of the {0} attribute is {1} ({2})."
.format(attr, value, item["name"]))
self.normalized = True
def add_point(self, label, pos, setpos, dmin, dmax):
point = dict()
point['pos'] = int(pos)
label = label.upper()
# Calculate point calibration if required
if self.has_calibration():
# Set to current physical position if no value supplied as argument
if math.isinf(setpos):
point['set'] = self.motor.position
else:
point['set'] = float(setpos)
# If point exists, we use current min, max values
if label in self.keys() and math.isinf(dmin) and math.isinf(dmax):
p = self[label]
min_pos = point['set'] + p['set'] - p['min']
max_pos = point['set'] + p['set'] - p['max']
# else, new point has new calibration,
elif math.isinf(dmin) and math.isinf(dmax):
min_pos = point['set']
max_pos = point['set']
else:
min_pos = point['set'] + dmin
max_pos = point['set'] + dmax
point['min'] = min_pos
point['max'] = max_pos
self[label] = point
self._update()
def SetNuisanceBkgValue(self, nuisname, value, bkg_name, year=None):
if self.split_bkg_by_year and year is None:
raise Exception("ERROR: backgrounds are split by year, you must specify which year you're doing!")
if year not in self.years:
raise Exception("ERROR: year {0} not in list of years!".format(year))
if bkg_name not in self.bkg_names:
raise Exception("ERROR: bkg {0} not in list of backgrounds!".format(bkg_name))
fullname = self.GetFullNuisName(nuisname, year)
fullbkg = bkg_name + (str(year) if year is not None else "")
fullidx = self.split_bkg_names.index(fullbkg)
if type(value)==tuple:
if self.nuisances[fullname].type not in ["lnN","lnU"]:
raise Exception("ERROR: only lnN/lnU nuisances support 2-sided values!")
if value[0]>2.05 or value[1]>2.05 or value[0]<0.3 or value[1]<0.3:
print "WARNING: nuisance {0} has a large value {1} (year {2}). Card: {3}".format(fullname, value, year, self.name)
if isnan(value[0]) or isinf(value[0]) or isnan(value[1]) or isinf(value[1]):
raise Exception("ERROR: nuisance value is nan or inf for nuis {0}, background {1}, year {2}".format(nuisname, bkg_name, year))
elif type(value)==float:
if self.nuisances[fullname].type in ["lnN","lnU"] and (value > 2.05 or value < 0.3):
print "WARNING: nuisance {0} has a large value {1} (year {2}). Card: {3}".format(fullname, value, year, self.name)
if isnan(value) or isinf(value):
raise Exception("ERROR: nuisance value is nan or inf for nuis {0}, background {1}, year {2}".format(nuisname, bkg_name, year))
else:
raise Exception("ERROR: value must be a float or tuple of 2 float (upper,lower) (lnN/lnU only)")
self.nuisances[fullname].bkg_values[fullidx] = value
def sqrt(x):
sqrt_special = [
[inf-infj, 0-infj, 0-infj, infj, infj, inf+infj, nan+infj],
[inf-infj, None, None, None, None, inf+infj, nan+nanj],
[inf-infj, None, 0-0j, 0+0j, None, inf+infj, nan+nanj],
[inf-infj, None, 0-0j, 0+0j, None, inf+infj, nan+nanj],
[inf-infj, None, None, None, None, inf+infj, nan+nanj],
[inf-infj, complex(float("inf"), -0.0), complex(float("inf"), -0.0), inf, inf, inf+infj, inf+nanj],
[inf-infj, nan+nanj, nan+nanj, nan+nanj, nan+nanj, inf+infj, nan+nanj]
]
z = _make_complex(x)
if math.isinf(z.real) or math.isinf(z.imag):
return sqrt_special[_special_type(z.real)][_special_type(z.imag)]
abs_x, abs_y = abs(z.real), abs(z.imag)
if abs_x < _DBL_MIN and abs_y < _DBL_MIN:
if abs_x > 0 or abs_y > 0:
abs_x = math.ldexp(abs_x, _CM_SCALE_UP)
s = math.ldexp(math.sqrt(abs_x +
math.hypot(abs_x,
math.ldexp(abs_y,
_CM_SCALE_UP))),
_CM_SCALE_DOWN)
else:
return complex(0, z.imag)
else:
abs_x /= 8
s = 2 * math.sqrt(abs_x + math.hypot(abs_x, abs_y/8))
if z.real >= 0:
return complex(s, math.copysign(abs_y/(2*s), z.imag))
return complex(abs_y/(2*s), math.copysign(s, z.imag))
def is_valid_number(x):
return not (math.isnan(x) or math.isinf(x) or x > 1e4)
[kshell, clustering, degree, strength, betweenness, closeness])
start_node = np.random.randint(0, num_of_nodes, 50)
infected_list = []
for idx, node in enumerate(start_node):
infected = SI(network, sorted_flights, node, 0.5)
infected_list.append(infected)
infected_list_np = np.array(infected_list)
median_infection = []
for i in range(num_of_nodes):
node_simulation = infected_list_np[:, i]
infected_count = 0
for time in node_simulation:
if not isinf(time):
infected_count += 1
print(infected_count)
if infected_count >= 25:
median_infection.append(np.median(node_simulation))
else:
np.delete(centrality, i, axis=1)
y_label = 'Median infection times'
scatter_base_path = 'scatter'
titles = [
'k-shell', 'Clustering', 'Degree', 'Strength', 'Betweenness',
'Closeness'
]
for i in range(6):
def assert_stuff(m):
# This makes sure that automatically imported builtins go after docstrings.
assert m.__doc__ == u'This is a module docstring.'
assert m.mystring == "foofoofoo"
assert m.long_string == u"This is a very long string literal, which would surely exceed any limitations on how long a line or a string literal can be. The string literal alone exceeds 256 characters. It also has a character outside the Basic Multilingual Plane: 😂. Here's a double quote: \". Here are some escaped newlines:\n\n\nHere is a literal newline:\nCall me Ishmael. Some years ago—never mind how long precisely—having little or no money in my purse, and nothing particular to interest me on shore, I thought I would sail about a little and see the watery part of the world. It is a way I have of driving off the spleen and regulating the circulation. Whenever I find myself growing grim about the mouth; whenever it is a damp, drizzly November in my soul; whenever I find myself involuntarily pausing before coffin warehouses, and bringing up the rear of every funeral I meet; and especially whenever my hypos get such an upper hand of me, that it requires a strong moral principle to prevent me from deliberately stepping into the street, and methodically knocking people’s hats off—then, I account it high time to get to sea as soon as I can. This is my substitute for pistol and ball. With a philosophical flourish Cato throws himself upon his sword; I quietly take to the ship. There is nothing surprising in this. If they but knew it, almost all men in their degree, some time or other, cherish very nearly the same feelings towards the ocean with me."
assert getattr(
m, mangle(u"identifier-that-has☝️💯☝️-to-be-mangled")) == "ponies"
assert m.mynumber == 3
assert m.myhex == 0x123
assert m.mylong - 1234567890987654321234567890987654320 == 1
assert m.myfloat == 3.34e15
assert math.isnan(m.mynan)
assert math.isinf(m.pinf)
assert m.pinf > 0
assert math.isinf(m.ninf)
assert m.ninf < 0
assert math.isinf(m.mycomplex.real)
assert m.mycomplex.real < 0
assert m.mycomplex.imag == 5
assert math.isnan(m.mycomplex2.real)
assert math.isinf(m.mycomplex2.imag)
assert m.mycomplex2.imag < 0
assert m.num_expr == 9
assert m.mylist == [1, 2, 3]
assert m.mytuple == ("a", "b", "c")
assert m.myset == {4, 5, 6}
assert m.mydict == {7: 8, 9: 900, 10: 15}
assert m.emptylist == []
assert m.emptytuple == ()
assert m.emptyset == set()
assert m.emptydict == {}
assert m.mylistcomp == [1, 3, 5, 7, 9]
assert m.mysetcomp == {0, 2, 4}
assert m.mydictcomp == dict(a="A", b="B", d="D", e="E")
assert type(m.mygenexpr) is type((x for x in [1, 2, 3]))
assert list(itertools.islice(m.mygenexpr, 5)) == [1, 3, 1, 3, 1]
assert m.attr_ref is str.upper
assert m.subscript == "l"
assert m.myslice == "el"
assert m.call == 5
assert m.comparison is True
assert m.boolexpr is True
assert m.condexpr == "y"
assert type(m.mylambda) is type(lambda x: x + "z")
assert m.mylambda("a") == "az"
assert m.augassign == 25
assert m.delstatement == ["a", "c", "d", "e"]
assert m.math is math
assert m.sqrt is math.sqrt
assert m.sine is math.sin
import datetime
assert m.timedelta is datetime.timedelta
assert m.if_block == "cd"
assert m.while_block == "xxxxe"
assert m.cont_and_break == "xyzxyzxxyzxy"
assert m.for_block == "fufifo"
assert m.caught_assertion is True
assert m.ran_finally is True
assert m.myraise == "payload"
assert m.ran_try_else is True
assert type(m.fun) is type(lambda x: x)
assert m.fun.__doc__ == "function docstring"
assert m.funcall1 == [
1, 2, 3, 4, ("a", "b", "c"), [("k1", "v1"), ("k2", "v2")]
]
assert m.funcall2 == [7, 8, 9, 10, (11, ), [("x1", "y1"), ("x2", "y2")]]
assert m.myret == 1
assert m.myyield == ["a", "b", "c"]
assert m.mydecorated.newattr == "hello"
assert m.myglobal == 103
class C:
pass
assert type(m.C1) is type(C)
assert m.C2.__doc__ == "class docstring"
assert issubclass(m.C2, m.C1)
assert (m.C2.attr1, m.C2.attr2, m.C2.attr3) == (5, 6, 7)
assert m.closed == ["v2", "v1"]
def pack_half_1x16(f32, func_opts):
"""Component-wise function of packHalf2x16."""
assert (isinstance(f32, float32))
# The bit layout of a float16 is:
#
# sign: 15
# exponent: 10:14
# mantissa: 0:9
#
# The sign, exponent, and mantissa determine its value by:
#
# if e = 0 and m = 0, then zero: (-1)^s * 0
# if e = 0 and m != 0, then subnormal: (-1)^s * 2^(e - 14) * m / 2^10
# if 0 < e < 31, then normal: (-1)^s * 2^(e - 15) * (1 + m / 2^10)
# if e = 31 and m = 0, then inf: (-1)^s * inf
# if e = 31 and m != 0, then nan
#
# where 0 <= m < 2^10.
#
# Some key boundary values of float16 are:
#
# min_normal16 = 2^(1 - 15) * (1 + 0 / 2^10)
# max_normal16 = 2^(30 - 15) * (1 + 1023 / 2^10)
#
# The maximum float16 step value is:
#
# max_step16 = 2^5
#
# Observe that each of the above boundary values lies in the range of
# normal float32 values. If we represent each of the above boundary values
# in the form returned by frexpf() for normal float32 values, 2^E
# * F where 0.5 <= F < 1, then:
#
# min_normal16 = 2^(-13) * 0.5
# max_normal16 = 2^16 * 0.99951171875
# The resultant float16's sign, exponent, and mantissa bits.
s = 0
e = 0
m = 0
# Calculate sign bit.
# Use copysign() to handle the case where x is -0.0.
if copysign(1.0, f32) < 0.0:
s = 1
# To reduce the number of cases in the if-tree below, decompose `abs(f32)`
# rather than `f32`.
(F, E) = frexp(fabs(f32))
# The output of frexp falls into three classes:
# - If f32 is NaN, then F is NaN .
# - If f32 is ±inf, then F is ±inf .
# - If f32 is ±0.0, then F is ±0.0 .
# - Otherwise, f32 = 2^E * F where 0.5 <= F < 1.0 .
#
# Since we decomposed `abs(f32)`, we only need be concerned with the
# postive cases.
if isnan(F):
# The resultant float16 is NaN.
e = 31
m = 1
elif isinf(F):
# The resultant float16 is infinite.
e = 31
m = 0
elif F == 0:
# f32 is zero, therefore the resultant float16 is zero.
e = 0
m = 0
elif E < -13:
# f32 lies in the range (0.0, min_normal16). Round f32 to a nearby
# float16 value. The resultant float16 will be either zero, subnormal,
# or normal.
e = 0
m = int(func_opts.round(2**(E + 24) * F))
elif E <= 16:
# f32 lies in the range [min_normal16, max_normal16 + max_step16).
# Round f32 to a nearby float16 value. The resultant float16 will be
# either normal or infinite.
e = int(E + 14)
m = int(func_opts.round(2**11 * F - 2**10))
else:
# f32 lies in the range [max_normal16 + max_step16, inf), which is
# outside the range of finite float16 values. The resultant float16 is
# infinite.
e = 31
m = 0
if (m == 1024):
# f32 was rounded upwards into the range of the next exponent. This
# correctly handles the case where f32 should be rounded up to float16
# infinity.
e += 1
m = 0
assert (s == 0 or s == 1)
assert (0 <= e and e <= 31)
assert (0 <= m and m <= 1023)
return uint16((s << 15) | (e << 10) | m)
def paintEvent(self, event):
unused(event)
if self.isVisible():
# Read out most important values to limit hash table lookups
# Low-pass roll, pitch and yaw
self.rollLP = self.roll#rollLP * 0.2f + 0.8f * roll
self.pitchLP = self.pitch#pitchLP * 0.2f + 0.8f * pitch
self.yawLP = self.yaw if isinf(self.yaw) == False and isnan(self.yaw) == False else self.yawLP#yawLP * 0.2f + 0.8f * yaw
# Translate for yaw
maxYawTrans = 60.0
newYawDiff = self.yawDiff
if isinf(newYawDiff):
newYawDiff = self.yawDiff
if newYawDiff > M_PI:
newYawDiff = newYawDiff - M_PI
if newYawDiff < -M_PI:
newYawDiff = newYawDiff + M_PI
newYawDiff = self.yawDiff * 0.8 + newYawDiff * 0.2
self.yawDiff = newYawDiff
self.yawInt += newYawDiff
if self.yawInt > M_PI:
self.yawInt = M_PI
if self.yawInt < -M_PI:
self.yawInt = -M_PI
yawTrans = self.yawInt * maxYawTrans
self.yawInt *= 0.6
if (yawTrans < 5.0) and (yawTrans > -5.0):
yawTrans = 0
# Negate to correct direction
yawTrans = -yawTrans
yawTrans = 0
#qDebug() << "yaw translation" << yawTrans << "integral" << yawInt << "difference" << yawDiff << "yaw" << yaw
# And if either video or the data stream is enabled, draw the next frame.
if self.videoEnabled:
self.xImageFactor = self.width() / float(self.glImage.width())
self.yImageFactor = self.height() / float(self.glImage.height())
painter = QPainter()
painter.begin(self)
painter.setRenderHint(QPainter.Antialiasing, True)
painter.setRenderHint(QPainter.HighQualityAntialiasing, True)
pmap = QPixmap.fromImage(self.glImage).scaledToWidth(self.width())
painter.drawPixmap(0, (self.height() - pmap.height()) / 2, pmap)
# END OF OPENGL PAINTING
if self.HUDInstrumentsEnabled:
#glEnable(GL_MULTISAMPLE)
# QT PAINTING
#makeCurrent()
painter.translate((self.vwidth/2.0+self.xCenterOffset)*self.scalingFactor, (self.vheight/2.0+self.yCenterOffset)*self.scalingFactor)
# COORDINATE FRAME IS NOW (0,0) at CENTER OF WIDGET
# Draw all fixed indicators
# BATTERY
self.paintText(self.fuelStatus, self.fuelColor, 6.0, (-self.vwidth/2.0) + 10, -self.vheight/2.0 + 6, painter)
# Waypoint
self.paintText(self.waypointName, self.defaultColor, 6.0, (-self.vwidth/3.0) + 10, +self.vheight/3.0 + 15, painter)
linePen = QPen(Qt.SolidLine)
linePen.setWidth(self.refLineWidthToPen(1.0))
linePen.setColor(self.defaultColor)
painter.setBrush(Qt.NoBrush)
painter.setPen(linePen)
# YAW INDICATOR
#
# .
# . .
# .......
#
_yawIndicatorWidth = 12.0
_yawIndicatorY = self.vheight/2.0 - 15.0
yawIndicator = QPolygon(4)
yawIndicator.setPoint(0, QPoint(self.refToScreenX(0.0), self.refToScreenY(_yawIndicatorY)))
yawIndicator.setPoint(1, QPoint(self.refToScreenX(_yawIndicatorWidth/2.0), self.refToScreenY(_yawIndicatorY+_yawIndicatorWidth)))
yawIndicator.setPoint(2, QPoint(self.refToScreenX(-_yawIndicatorWidth/2.0), self.refToScreenY(_yawIndicatorY+_yawIndicatorWidth)))
yawIndicator.setPoint(3, QPoint(self.refToScreenX(0.0), self.refToScreenY(_yawIndicatorY)))
painter.drawPolyline(yawIndicator)
painter.setPen(linePen)
# CENTER
# HEADING INDICATOR
#
# __ __
# \/\/
#
_hIndicatorWidth = 20.0
_hIndicatorY = -25.0
_hIndicatorYLow = _hIndicatorY + _hIndicatorWidth / 6.0
_hIndicatorSegmentWidth = _hIndicatorWidth / 7.0
hIndicator = QPolygon(7)
hIndicator.setPoint(0, QPoint(self.refToScreenX(0.0-_hIndicatorWidth/2.0), self.refToScreenY(_hIndicatorY)))
hIndicator.setPoint(1, QPoint(self.refToScreenX(0.0-_hIndicatorWidth/2.0+_hIndicatorSegmentWidth*1.75), self.refToScreenY(_hIndicatorY)))
hIndicator.setPoint(2, QPoint(self.refToScreenX(0.0-_hIndicatorSegmentWidth*1.0), self.refToScreenY(_hIndicatorYLow)))
hIndicator.setPoint(3, QPoint(self.refToScreenX(0.0), self.refToScreenY(_hIndicatorY)))
hIndicator.setPoint(4, QPoint(self.refToScreenX(0.0+_hIndicatorSegmentWidth*1.0), self.refToScreenY(_hIndicatorYLow)))
hIndicator.setPoint(5, QPoint(self.refToScreenX(0.0+_hIndicatorWidth/2.0-_hIndicatorSegmentWidth*1.75), self.refToScreenY(_hIndicatorY)))
hIndicator.setPoint(6, QPoint(self.refToScreenX(0.0+_hIndicatorWidth/2.0), self.refToScreenY(_hIndicatorY)))
painter.drawPolyline(hIndicator)
# SETPOINT
_centerWidth = 8.0
painter.drawEllipse(
QPointF(self.refToScreenX(min(10.0, self.desiredRoll * 10.0)),
self.refToScreenY(min(10.0, self.desiredPitch * 10.0))),
self.refToScreenX(_centerWidth/2.0), self.refToScreenX(_centerWidth/2.0))
_centerCrossWidth = 20.0
# left
painter.drawLine(QPointF(self.refToScreenX(-_centerWidth / 2.0), self.refToScreenY(0.0)),
QPointF(self.refToScreenX(-_centerCrossWidth / 2.0), self.refToScreenY(0.0)))
# right
painter.drawLine(QPointF(self.refToScreenX(_centerWidth / 2.0), self.refToScreenY(0.0)),
QPointF(self.refToScreenX(_centerCrossWidth / 2.0), self.refToScreenY(0.0)))
# top
painter.drawLine(QPointF(self.refToScreenX(0.0), self.refToScreenY(-_centerWidth / 2.0)),
QPointF(self.refToScreenX(0.0), self.refToScreenY(-_centerCrossWidth / 2.0)))
# COMPASS
_compassY = -self.vheight/2.0 + 6.0
compassRect = QRectF(QPointF(self.refToScreenX(-12.0), self.refToScreenY(_compassY)),
QSizeF(self.refToScreenX(24.0), self.refToScreenY(12.0)))
painter.setBrush(Qt.NoBrush)
painter.setPen(linePen)
painter.drawRoundedRect(compassRect, 3, 3)
# YAW is in compass-human readable format, so 0 .. 360 deg.
_yawDeg = (self.yawLP / M_PI) * 180.0
if _yawDeg < 0:
_yawDeg += 360
if _yawDeg > 360:
_yawDeg -= 360
# final safeguard for really stupid systems
_yawAngle = '{:3d}'.format(int(_yawDeg) % 360)
self.paintText(_yawAngle, self.defaultColor, 8.5, -9.8, _compassY + 1.7, painter)
painter.setBrush(Qt.NoBrush)
painter.setPen(linePen)
# CHANGE RATE STRIPS
self.drawChangeRateStrip(-95.0, -60.0, 40.0, -10.0, 10.0, self.zSpeed, painter)
# CHANGE RATE STRIPS
self.drawChangeRateStrip(95.0, -60.0, 40.0, -10.0, 10.0, self.totalAcc, painter,True)
# GAUGES
# Left altitude gauge
_gaugeAltitude = self.alt if self.alt != 0 else -self.zPos
painter.setBrush(Qt.NoBrush)
painter.setPen(linePen)
self.drawChangeIndicatorGauge(-self.vGaugeSpacing, 35.0, 15.0, 10.0, _gaugeAltitude, self.defaultColor, painter, False)
self.paintText('alt m', self.defaultColor, 5.5, -73.0, 50, painter)
# Right speed gauge
self.drawChangeIndicatorGauge(self.vGaugeSpacing, 35.0, 15.0, 10.0, self.totalSpeed, self.defaultColor, painter, False)
self.paintText('v m/s', self.defaultColor, 5.5, 55.0, 50, painter)
# Waypoint name
if self.waypointName != '':
self.paintText(self.waypointName, self.defaultColor, 2.0, (-self.vwidth/3.0) + 10, +self.vheight/3.0 + 15, painter)
# MOVING PARTS
painter.translate(self.refToScreenX(yawTrans), 0)
attColor = painter.pen().color()
# Draw multi-component attitude
for key in self.attitudes:
att = self.attitudes[key]
attColor = attColor.darker(200)
painter.setPen(attColor)
# Rotate view and draw all roll-dependent indicators
painter.rotate((att.x()/M_PI)* -180.0)
painter.translate(0, (-att.y()/M_PI)* -180.0 * self.refToScreenY(1.8))
#qDebug() << "ROLL" << roll << "PITCH" << pitch << "YAW DIFF" << valuesDot.value("roll", 0.0)
# PITCH
self.paintPitchLines(att.y(), painter)
painter.translate(0, -(-att.y()/M_PI)* -180.0 * self.refToScreenY(1.8))
painter.rotate(-(att.x()/M_PI)* -180.0)
painter.end()
def updateComponentAttitude(self, uas, timestamp, component, roll, pitch, yaw):
unused(uas, timestamp)
if isnan(roll) == False and isinf(roll) == False \
and isnan(pitch) == False and isinf(pitch)== False \
and isnan(yaw) == False and isinf(yaw) == False:
self.attitudes[component] = QVector3D(roll, pitch*3.35, yaw) # Constant here is the 'focal length' of the projection onto the plane
def create_optical_system(self,
matdict=None,
options=None,
elementname="zmxelem"):
"""
Creates optical system from ZEMAX file with material
data and options.
"""
# It is intended that matdict and options should not
# be changed at a higher level from within this function.
if matdict is None:
matdict = {}
if options is None:
options = {}
self.info("Creating optical system from ZMX")
(name, _) = self.read_name_and_notes()
optical_system = OpticalSystem.p(name=name)
# extract surface blockstrings
self.info("Extract surface blockstrings")
surface_blockstrings = self.filter_block_strings("SURF")
# construct basis coordinate system
self.info("Construct basis coordinate system")
lc0 = optical_system.addLocalCoordinateSystem(
LocalCoordinates.p(name="object", decz=0.0),
refname=optical_system.rootcoordinatesystem.name)
elem = OpticalElement.p(lc0, name=elementname)
# construct materials
self.info("Construct materials")
if matdict != {}:
for (key, mat) in list(matdict.items()):
mat.lc = lc0
# different material coordinate systems are not supported
elem.addMaterial(key, mat)
else:
self.info("checking for external material" +
"objects in dict with the following identifiers")
found_necessary_glasses = False
for blk in surface_blockstrings:
surfres = self.read_surf_block(blk)
glass_dict = surfres.get("GLAS", None)
if glass_dict is not None:
self.debug(str(glass_dict))
material_name = glass_dict.get("name", None)
material_code = glass_dict.get("code", None)
self.debug("mat name \"%s\" mat code %s" %
(material_name, material_code))
if material_name != "MIRROR" and material_code != 1:
found_necessary_glasses = True
self.info(material_name)
self.info("Are there necessary glasses? " +
str(found_necessary_glasses))
if found_necessary_glasses:
self.error("found material names: exiting")
return (None, [("zmxelem", [])])
else:
self.info("found only mirrors or no material: continuing")
self.info("Reading field")
self.debug(self.read_field())
self.info("Reading surface blocks")
refname = lc0.name
# lastlc = lc0
lastmatname = None
lastsurfname = None
surfname = None
surflist_for_sequence = []
lastthickness = 0
thickness = 0
for blk in surface_blockstrings:
lastthickness = thickness
self.debug("----")
surfres = self.read_surf_block(blk)
self.debug("Found surface with contents (except GARR):")
self.debug([(k, v) for (k, v) in list(surfres.items())
if k != "GARR"])
surf_options_dict = {}
# comment = surfres.get("COMM", "")
lastsurfname = surfname
surfname = "surf" + str(surfres["SURF"])
thickness = surfres["DISZ"]
curv = surfres["CURV"]
conic = surfres.get("CONI", 0.0)
stop = surfres["STOP"]
surftype = surfres["TYPE"]
parms = surfres["PARM"]
xdat = surfres["XDAT"]
sqap = surfres.get("SQAP", None)
clap = surfres.get("CLAP", None)
obdc = surfres.get("OBDC", None)
if math.isinf(thickness):
self.info("infinite object distance!")
thickness = 0
if stop:
surf_options_dict["is_stop"] = True
localcoordinates = optical_system.addLocalCoordinateSystem(
LocalCoordinates.p(name=surfname, decz=lastthickness),
refname=refname)
read_glass = surfres.get("GLAS", None)
self.debug("MATERIAL: %s" % (str(read_glass), ))
matname = None
mat = None
if read_glass is not None:
if read_glass["name"] == "MIRROR":
matname = lastmatname
surf_options_dict["is_mirror"] = True
if matdict.get(read_glass["name"], None) is not None:
matname = read_glass["name"]
if read_glass["code"] == 1:
matname = "modelglass" + str(uuid.uuid4())
# TODO: use global known uuid function
nd_value = read_glass["nd"]
vd_value = read_glass["vd"]
if abs(vd_value) < numerical_tolerance:
mat = ConstantIndexGlass.p(localcoordinates, nd_value)
else:
mat = ModelGlass.p(localcoordinates)
mat.calcCoefficientsFrom_nd_vd(nd=nd_value,
vd=vd_value)
elem.addMaterial(matname, mat)
else:
if surftype == "COORDBRK":
matname = lastmatname
if obdc is None:
lcapdec = optical_system.addLocalCoordinateSystem(
LocalCoordinates.p(name=surfname + "_ap"),
refname=surfname)
else:
self.info("Aperture decenter %f %f" % tuple(obdc))
lcapdec = optical_system.addLocalCoordinateSystem(
LocalCoordinates.p(name=surfname + "_ap",
decx=obdc[0],
decy=obdc[1]),
refname=surfname)
if sqap is None and clap is None:
aper = None
elif sqap is not None and clap is None:
self.debug("Rectangular aperture %f x %f" % tuple(sqap))
aper = RectangularAperture.p(lcapdec,
width=sqap[0] * 2,
height=sqap[1] * 2)
elif clap is not None and sqap is None:
self.debug("Circular aperture min %f max %f" % tuple(clap))
aper = CircularAperture.p(lcapdec,
minradius=clap[0],
maxradius=clap[1])
else: # both are not None
aper = None
if surftype == "STANDARD":
self.debug("SURFACE: Standard surface found")
actsurf = Surface.p(localcoordinates,
name=surfname,
shape=Conic.p(localcoordinates,
curv=curv,
cc=conic),
aperture=aper)
elif surftype == "EVENASPH":
self.debug("SURFACE: Polynomial asphere surface found")
acoeffs = [parms.get(1 + i, 0.0) for i in range(8)]
self.debug(acoeffs)
actsurf = Surface.p(localcoordinates,
name=surfname,
shape=Asphere.p(localcoordinates,
curv=curv,
cc=conic,
coefficients=acoeffs),
aperture=aper)
elif surftype == "BICONICX":
self.debug("SURFACE: biconic surface found")
Rx = parms.get(1, 0.0)
if abs(Rx) < 1e-16:
curvx = 0.0
else:
curvx = 1 / Rx
ccx = parms.get(2, 0.0)
actsurf = Surface.p(localcoordinates,
name=surfname,
shape=Biconic.p(localcoordinates,
curvy=curv,
ccy=conic,
curvx=curvx,
ccx=ccx),
aperture=aper)
elif surftype == "FZERNSAG": # Zernike Fringe Sag
self.debug("SURFACE: Zernike standard surface found")
# ignore extrapolate flag
# curv, cc, parms, xdat
acoeffs = [parms.get(1 + i, 0.0) for i in range(8)]
extrapolate = parms.get(0, 1)
zdecx = parms.get(9, 0.)
zdecy = parms.get(10, 0.)
self.debug("extrapolate: %d zdecx: %f zdecy: %f" %
(extrapolate, zdecx, zdecy))
numterms = int(xdat[1]["value"])
normradius = xdat[2]["value"]
self.debug("numterms: %d normradius: %f" %
(numterms, normradius))
zcoeffs = [
xdat[i + 3].get("value", 0.) for i in range(numterms)
]
lcz = localcoordinates.addChild(
LocalCoordinates.p(name=surfname + "_zerndec",
decx=zdecx,
decy=zdecy))
actsurf =\
Surface.p(localcoordinates,
name=surfname,
shape=LinearCombination.p(
localcoordinates,
list_of_coefficients_and_shapes=[
(1.0,
Asphere.p(
localcoordinates,
curv=curv,
cc=conic,
name=surfname +
"_zernasph")),
(1.0,
ZernikeFringe.p(
lcz,
normradius=normradius,
coefficients=zcoeffs,
name=surfname+"_zernike"))]))
elif surftype == "GRID_SAG":
# TODO: conic + aspheric polynomials + zernike standard sag
self.debug("SURFACE: Grid Sag found")
(num_pts_x, num_pts_y, delta_x, delta_y) = surfres["GDAT"]
self.debug("nx %d ny %d dx %f dy %f" %
(num_pts_x, num_pts_y, delta_x, delta_y))
sagarray = np.array([
surfres["GARR"][key]
for key in sorted(surfres["GARR"].keys())
])
self.debug(sagarray)
xvector = np.linspace(-num_pts_x * delta_x * 0.5,
num_pts_x * delta_x * 0.5, num_pts_x)
yvector = np.linspace(-num_pts_y * delta_y * 0.5,
num_pts_y * delta_y * 0.5, num_pts_y)
zgrid = np.flipud(sagarray[:, 0].reshape(num_pts_x,
num_pts_y)).T
# first line
actsurf = Surface.p(localcoordinates,
name=surfname,
shape=GridSag.p(localcoordinates,
(xvector, yvector, zgrid)))
elif surftype == "COORDBRK":
# COORDBRK
# parm1: decx
# parm2: decy
# parm3: rx
# parm4: ry
# parm5: rz
# parm6: order 0: means 1. decx, 2. decy, 3. rx, 4. ry, 5. rz
# order 1: means 1. rz 2. ry 3. rz, 4. decy, 5. decx
# disz: thickness (where does this appear in the transform?
# step 6)
# all in all: 1st step: transform coordinate system by former
# disz, dec, tilt (or tilt, dec)
# 2nd step: update vertex
self.debug("SURFACE: Coordinate break found")
localcoordinates.decx.set_value(parms.get(1, 0.0))
localcoordinates.decy.set_value(parms.get(2, 0.0))
localcoordinates.tiltx.set_value(parms.get(3, 0.0) * degree)
localcoordinates.tilty.set_value(parms.get(4, 0.0) * degree)
localcoordinates.tiltz.set_value(parms.get(5, 0.0) * degree)
localcoordinates.tiltThenDecenter = bool(parms.get(6, 0))
localcoordinates.update()
actsurf = Surface.p(localcoordinates, name=surfname)
else:
actsurf = Surface.p(localcoordinates, name=surfname)
if lastsurfname is not None:
elem.addSurface(surfname, actsurf, (lastmatname, matname))
self.info("addsurf: %s at material boundary %s" %
(surfname, (lastmatname, matname)))
surflist_for_sequence.append((surfname, surf_options_dict))
lastmatname = matname
refname = localcoordinates.name
# lastlc = localcoordinates
optical_system.addElement(elementname, elem)
seq = [(elementname, surflist_for_sequence)]
self.info(optical_system.rootcoordinatesystem.pprint())
return (optical_system, seq)
def lgamma(x):
sgngam = 1
if math.isnan(x):
return x
if math.isinf(x) and x > 0:
return x
if math.isinf(x) and x < 0:
return -x
if x < -34.0:
q = -x
w = lgamma(q)
p = math.floor(q)
if p == q:
return float('inf')
i = p
if (i & 1) == 0:
sgngam = -1
else:
sgngam = 1
z = q - p
if z > 0.5:
p += 1.0
z = p - q
z = q * math.sin(math.pi * z)
if z == 0.0:
return float('inf')
z = LOGPI - math.log(z) - w;
return z
if x < 13.0:
z = 1.0
p = 0.0
u = x
while u >= 3.0:
p -= 1.0
u = x + p
z *= u
while u < 2.0:
if u == 0.0:
return float('inf')
z /= u
p += 1.0
u = x + p
if z < 0.0:
sgngam = -1
z = -z
else:
sgngam = 1
if u == 2.0:
return math.log(z)
p -= 2.0
x = x + p
p = x * polevl( x, B, 5 ) / polevl( x, C, 6)
return math.log(z) + p
if x > MAXLGM:
return sgngam * float('inf')
q = ( x - 0.5 ) * math.log(x) - x + LS2PI
if x > 1.0e8:
return q
p = 1.0/(x*x)
if x >= 1000.0:
q += (( 7.9365079365079365079365e-4 * p \
- 2.7777777777777777777778e-3) * p \
+ 0.0833333333333333333333) / x
else:
q += polevl( p, A, 4 ) / x
return q
psnrs.append(psnr_bicubic_up_scipy)
#psnrs.append(psnr_bicubic_up1)
#psnrs.append(psnr_bicubic_up_matlab)
xlabels = ['ORG', 'CNN_UP', 'BICUBIC_UP\n(pred_from_cnn_up)',
'BICUBIC_UP\n(pred_from_2d)',
'BICUBIC_UP\n(pred_from_label)',
'BICUBIC_UP\n(a=-0.5)', 'BICUBIC_UP\n(scipy)', 'BICUBIC_UP\n(matlab)']
# separate dir depending on PSNR
# psnr diff
#psnr_diff = psnr_bicubic_up_pred - psnr_bicubic_up
#psnr_diff = psnr_bicubic_up - psnr_bicubic_up_scipy
psnr_diff = psnr_pred_a_from_label - psnr_cnn_up
if math.isinf(psnr_diff):
psnr_diff = 99.0
psnr_diff_floor = math.floor(psnr_diff)
str_psnr_diff_floor = str(psnr_diff_floor)
output_path_psnr = os.path.join(output_path, 'group_psnr_' + str_psnr_diff_floor)
if not os.path.exists(output_path_psnr):
os.makedirs(output_path_psnr)
plot_diff(imgs, psnrs, xlabels, idx, each_frame_index, x, y, save_dir=output_path_psnr)
# to save the raw data
list_raw = [[each_yuv, each_frame_index, x, y,
best_a_from_cnn_up, psnr_cnn_up, psnr_bicubic_up_pred,
#psnr_bicubic_up_pred_2d,
def _preprocess(self, lines=None, build_layers=False, layer_callback=None):
"""Checks for imperial/relativeness settings and tool changes"""
if not lines:
lines = self.lines
imperial = self.imperial
relative = self.relative
relative_e = self.relative_e
current_tool = self.current_tool
current_x = self.current_x
current_y = self.current_y
current_z = self.current_z
offset_x = self.offset_x
offset_y = self.offset_y
offset_z = self.offset_z
# Extrusion computation
current_e = self.current_e
offset_e = self.offset_e
total_e = self.total_e
max_e = self.max_e
current_e_multi = self.current_e_multi[current_tool]
offset_e_multi = self.offset_e_multi[current_tool]
total_e_multi = self.total_e_multi[current_tool]
max_e_multi = self.max_e_multi[current_tool]
# Store this one out of the build_layers scope for efficiency
cur_layer_has_extrusion = False
# Initialize layers and other global computations
if build_layers:
# Bounding box computation
xmin = float("inf")
ymin = float("inf")
zmin = 0
xmax = float("-inf")
ymax = float("-inf")
zmax = float("-inf")
# Also compute extrusion-only values
xmin_e = float("inf")
ymin_e = float("inf")
xmax_e = float("-inf")
ymax_e = float("-inf")
# Duration estimation
# TODO:
# get device caps from firmware: max speed, acceleration/axis
# (including extruder)
# calculate the maximum move duration accounting for above ;)
lastx = lasty = lastz = laste = lastf = 0.0
lastdx = 0
lastdy = 0
x = y = e = f = 0.0
currenttravel = 0.0
moveduration = 0.0
totalduration = 0.0
acceleration = 2000.0 # mm/s^2
layerbeginduration = 0.0
# Initialize layers
all_layers = self.all_layers = []
all_zs = self.all_zs = set()
layer_idxs = self.layer_idxs = []
line_idxs = self.line_idxs = []
layer_id = 0
layer_line = 0
last_layer_z = None
prev_z = None
prev_base_z = (None, None)
cur_z = None
cur_lines = []
if self.line_class != Line:
get_line = lambda l: Line(l.raw)
else:
get_line = lambda l: l
for true_line in lines:
# # Parse line
# Use a heavy copy of the light line to preprocess
line = get_line(true_line)
split_raw = split(line)
if line.command:
# Update properties
if line.is_move:
line.relative = relative
line.relative_e = relative_e
line.current_tool = current_tool
elif line.command == "G20":
imperial = True
elif line.command == "G21":
imperial = False
elif line.command == "G90":
relative = False
relative_e = False
elif line.command == "G91":
relative = True
relative_e = True
elif line.command == "M82":
relative_e = False
elif line.command == "M83":
relative_e = True
elif line.command[0] == "T":
current_tool = int(line.command[1:])
while (current_tool + 1 > len(self.current_e_multi)):
self.current_e_multi += [0]
self.offset_e_multi += [0]
self.total_e_multi += [0]
self.max_e_multi += [0]
current_e_multi = self.current_e_multi[current_tool]
offset_e_multi = self.offset_e_multi[current_tool]
total_e_multi = self.total_e_multi[current_tool]
max_e_multi = self.max_e_multi[current_tool]
if line.command[0] == "G":
parse_coordinates(line, split_raw, imperial)
# Compute current position
if line.is_move:
x = line.x
y = line.y
z = line.z
if line.f is not None:
self.current_f = line.f
if line.relative:
x = current_x + (x or 0)
y = current_y + (y or 0)
z = current_z + (z or 0)
else:
if x is not None: x = x + offset_x
if y is not None: y = y + offset_y
if z is not None: z = z + offset_z
if x is not None: current_x = x
if y is not None: current_y = y
if z is not None: current_z = z
elif line.command == "G28":
home_all = not any([line.x, line.y, line.z])
if home_all or line.x is not None:
offset_x = 0
current_x = self.home_x
if home_all or line.y is not None:
offset_y = 0
current_y = self.home_y
if home_all or line.z is not None:
offset_z = 0
current_z = self.home_z
elif line.command == "G92":
if line.x is not None: offset_x = current_x - line.x
if line.y is not None: offset_y = current_y - line.y
if line.z is not None: offset_z = current_z - line.z
line.current_x = current_x
line.current_y = current_y
line.current_z = current_z
# # Process extrusion
if line.e is not None:
if line.is_move:
if line.relative_e:
line.extruding = line.e > 0
total_e += line.e
current_e += line.e
total_e_multi += line.e
current_e_multi += line.e
else:
new_e = line.e + offset_e
line.extruding = new_e > current_e
total_e += new_e - current_e
current_e = new_e
new_e_multi = line.e + offset_e_multi
total_e_multi += new_e_multi - current_e_multi
current_e_multi = new_e_multi
max_e = max(max_e, total_e)
max_e_multi = max(max_e_multi, total_e_multi)
cur_layer_has_extrusion |= line.extruding
elif line.command == "G92":
offset_e = current_e - line.e
offset_e_multi = current_e_multi - line.e
self.current_e_multi[current_tool] = current_e_multi
self.offset_e_multi[current_tool] = offset_e_multi
self.max_e_multi[current_tool] = max_e_multi
self.total_e_multi[current_tool] = total_e_multi
# # Create layers and perform global computations
if build_layers:
# Update bounding box
if line.is_move:
if line.extruding:
if line.current_x is not None:
xmin_e = min(xmin_e, line.current_x)
xmax_e = max(xmax_e, line.current_x)
if line.current_y is not None:
ymin_e = min(ymin_e, line.current_y)
ymax_e = max(ymax_e, line.current_y)
if max_e <= 0:
if line.current_x is not None:
xmin = min(xmin, line.current_x)
xmax = max(xmax, line.current_x)
if line.current_y is not None:
ymin = min(ymin, line.current_y)
ymax = max(ymax, line.current_y)
# Compute duration
if line.command == "G0" or line.command == "G1":
x = line.x if line.x is not None else lastx
y = line.y if line.y is not None else lasty
z = line.z if line.z is not None else lastz
e = line.e if line.e is not None else laste
# mm/s vs mm/m => divide by 60
f = line.f / 60.0 if line.f is not None else lastf
# given last feedrate and current feedrate calculate the
# distance needed to achieve current feedrate.
# if travel is longer than req'd distance, then subtract
# distance to achieve full speed, and add the time it took
# to get there.
# then calculate the time taken to complete the remaining
# distance
# FIXME: this code has been proven to be super wrong when 2
# subsquent moves are in opposite directions, as requested
# speed is constant but printer has to fully decellerate
# and reaccelerate
# The following code tries to fix it by forcing a full
# reacceleration if this move is in the opposite direction
# of the previous one
dx = x - lastx
dy = y - lasty
if dx * lastdx + dy * lastdy <= 0:
lastf = 0
currenttravel = math.hypot(dx, dy)
if currenttravel == 0:
if line.z is not None:
currenttravel = abs(
line.z) if line.relative else abs(line.z -
lastz)
elif line.e is not None:
currenttravel = abs(
line.e) if line.relative_e else abs(
line.e - laste)
# Feedrate hasn't changed, no acceleration/decceleration planned
if f == lastf:
moveduration = currenttravel / f if f != 0 else 0.
else:
# FIXME: review this better
# this looks wrong : there's little chance that the feedrate we'll decelerate to is the previous feedrate
# shouldn't we instead look at three consecutive moves ?
distance = 2 * abs(
((lastf + f) *
(f - lastf) * 0.5) / acceleration
) # multiply by 2 because we have to accelerate and decelerate
if distance <= currenttravel and lastf + f != 0 and f != 0:
moveduration = 2 * distance / (
lastf + f
) # This is distance / mean(lastf, f)
moveduration += (currenttravel - distance) / f
else:
moveduration = 2 * currenttravel / (
lastf + f
) # This is currenttravel / mean(lastf, f)
# FIXME: probably a little bit optimistic, but probably a much better estimate than the previous one:
# moveduration = math.sqrt(2 * distance / acceleration) # probably buggy : not taking actual travel into account
lastdx = dx
lastdy = dy
totalduration += moveduration
lastx = x
lasty = y
lastz = z
laste = e
lastf = f
elif line.command == "G4":
moveduration = P(line)
if moveduration:
moveduration /= 1000.0
totalduration += moveduration
# FIXME : looks like this needs to be tested with "lift Z on move"
if line.z is not None:
if line.command == "G92":
cur_z = line.z
elif line.is_move:
if line.relative and cur_z is not None:
cur_z += line.z
else:
cur_z = line.z
# FIXME: the logic behind this code seems to work, but it might be
# broken
if cur_z != prev_z:
if prev_z is not None and last_layer_z is not None:
offset = self.est_layer_height if self.est_layer_height else 0.01
if abs(prev_z - last_layer_z) < offset:
if self.est_layer_height is None:
zs = sorted([
l.z for l in all_layers
if l.z is not None
])
heights = [
round(zs[i + 1] - zs[i], 3)
for i in range(len(zs) - 1)
]
heights = [
height for height in heights if height
]
if len(heights) >= 2:
self.est_layer_height = heights[1]
elif heights:
self.est_layer_height = heights[0]
else:
self.est_layer_height = 0.1
base_z = round(
prev_z - (prev_z % self.est_layer_height),
2)
else:
base_z = round(prev_z, 2)
else:
base_z = prev_z
if base_z != prev_base_z:
new_layer = Layer(cur_lines, base_z)
new_layer.duration = totalduration - layerbeginduration
layerbeginduration = totalduration
all_layers.append(new_layer)
if cur_layer_has_extrusion and prev_z not in all_zs:
all_zs.add(prev_z)
cur_lines = []
cur_layer_has_extrusion = False
layer_id += 1
layer_line = 0
last_layer_z = base_z
if layer_callback is not None:
layer_callback(self, len(all_layers) - 1)
prev_base_z = base_z
if build_layers:
cur_lines.append(true_line)
layer_idxs.append(layer_id)
line_idxs.append(layer_line)
layer_line += 1
prev_z = cur_z
# ## Loop done
# Store current status
self.imperial = imperial
self.relative = relative
self.relative_e = relative_e
self.current_tool = current_tool
self.current_x = current_x
self.current_y = current_y
self.current_z = current_z
self.offset_x = offset_x
self.offset_y = offset_y
self.offset_z = offset_z
self.current_e = current_e
self.offset_e = offset_e
self.max_e = max_e
self.total_e = total_e
self.current_e_multi[current_tool] = current_e_multi
self.offset_e_multi[current_tool] = offset_e_multi
self.max_e_multi[current_tool] = max_e_multi
self.total_e_multi[current_tool] = total_e_multi
# Finalize layers
if build_layers:
if cur_lines:
new_layer = Layer(cur_lines, prev_z)
new_layer.duration = totalduration - layerbeginduration
layerbeginduration = totalduration
all_layers.append(new_layer)
if cur_layer_has_extrusion and prev_z not in all_zs:
all_zs.add(prev_z)
self.append_layer_id = len(all_layers)
self.append_layer = Layer([])
self.append_layer.duration = 0
all_layers.append(self.append_layer)
self.layer_idxs = array('I', layer_idxs)
self.line_idxs = array('I', line_idxs)
# Compute bounding box
all_zs = self.all_zs.union(set([zmin])).difference(set([None]))
zmin = min(all_zs)
zmax = max(all_zs)
self.filament_length = self.max_e
while len(self.filament_length_multi) < len(self.max_e_multi):
self.filament_length_multi += [0]
for i in enumerate(self.max_e_multi):
self.filament_length_multi[i[0]] = i[1]
if self.filament_length > 0:
self.xmin = xmin_e if not math.isinf(xmin_e) else 0
self.xmax = xmax_e if not math.isinf(xmax_e) else 0
self.ymin = ymin_e if not math.isinf(ymin_e) else 0
self.ymax = ymax_e if not math.isinf(ymax_e) else 0
else:
self.xmin = xmin if not math.isinf(xmin) else 0
self.xmax = xmax if not math.isinf(xmax) else 0
self.ymin = ymin if not math.isinf(ymin) else 0
self.ymax = ymax if not math.isinf(ymax) else 0
self.zmin = zmin if not math.isinf(zmin) else 0
self.zmax = zmax if not math.isinf(zmax) else 0
self.width = self.xmax - self.xmin
self.depth = self.ymax - self.ymin
self.height = self.zmax - self.zmin
# Finalize duration
totaltime = datetime.timedelta(seconds=int(totalduration))
self.duration = totaltime
def train_model(model):
args = model.args
unchanged, best_dev = 0, 200
unroll_size = args.unroll_size
batch_size = args.batch_size
criterion = nn.CrossEntropyLoss(size_average=False)
trainer = SGD(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
map_to_ids = model.embedding_layer.map_to_ids
train = read_corpus(args.train, shuffle=False)
train = create_batches(train, map_to_ids, args.batch_size, cuda=args.gpu)
dev = read_corpus(args.dev)
dev = create_batches(dev, map_to_ids, 1, cuda=args.gpu)
test = read_corpus(args.test)
test = create_batches(test, map_to_ids, 1, cuda=args.gpu)
for epoch in range(args.max_epoch):
start_time = time.time()
N = (len(train[0]) - 1) // unroll_size + 1
hidden = model.init_hidden(batch_size)
total_loss, cur_loss = 0.0, 0.0
for i in range(N):
model.train()
x = train[0][i * unroll_size:(i + 1) * unroll_size]
y = train[1][i * unroll_size:(i + 1) * unroll_size].view(-1)
x, y = Variable(x), Variable(y)
model.zero_grad()
output, hidden = model(x, hidden)
hidden = repackage_hidden(args, hidden)
assert x.size(1) == batch_size
loss = criterion(output, y) / x.size(1)
loss.backward()
if math.isnan(loss.data[0]) or math.isinf(loss.data[0]):
print("nan/inf loss encoutered in training.")
sys.exit(0)
return
total_loss += loss.data[0] / x.size(0)
cur_loss += loss.data[0] / x.size(0)
torch.nn.utils.clip_grad_norm(model.parameters(), args.clip_grad)
for p in model.parameters():
if not p.requires_grad:
continue
if p.grad is not None:
if args.weight_decay > 0:
p.data.mul_(1.0 - args.weight_decay)
p.data.add_(-args.lr, p.grad.data)
if (i + 1) % args.eval_ite == 0:
dev_ppl = eval_model(model, dev)
sys.stdout.write(
"| Epoch={} | ite={} | lr={:.4f} | train_ppl={:.2f} | dev_ppl={:.2f} |"
"\n".format(epoch, i + 1, trainer.defaults["lr"],
np.exp(cur_loss / args.eval_ite), dev_ppl))
model.print_pnorm()
sys.stdout.flush()
cur_loss = 0.0
if dev_ppl < best_dev:
unchanged = 0
best_dev = dev_ppl
test_ppl = eval_model(model, test)
sys.stdout.write(
"\t[eval] test_ppl={:.2f}\n".format(test_ppl))
sys.stdout.flush()
train_ppl = np.exp(total_loss / N)
dev_ppl = eval_model(model, dev)
sys.stdout.write("-" * 89 + "\n")
sys.stdout.write(
"| End of epoch {} | lr={:.4f} | train_ppl={:.2f} | dev_ppl={:.2f} |"
"[{:.2f}m] |\n".format(epoch, trainer.defaults["lr"], train_ppl,
dev_ppl, (time.time() - start_time) / 60.0))
sys.stdout.write("-" * 89 + "\n")
model.print_pnorm()
sys.stdout.flush()
if dev_ppl < best_dev:
unchanged = 0
best_dev = dev_ppl
start_time = time.time()
test_ppl = eval_model(model, test)
sys.stdout.write("\t[eval] test_ppl={:.2f}\t[{:.2f}m]\n".format(
test_ppl, (time.time() - start_time) / 60.0))
sys.stdout.flush()
else:
unchanged += 1
if args.lr_decay_epoch > 0 and epoch >= args.lr_decay_epoch:
args.lr *= args.lr_decay
if unchanged >= args.patience:
sys.stdout.write(
"Reached " + str(args.patience) +
" iterations without improving dev loss. Reducing learning rate."
)
args.lr /= 2
unchanged = 0
trainer.defaults["lr"] = args.lr
sys.stdout.write("\n")
return
def train_step(self, samples, dummy_batch=False, raise_oom=False):
"""Do forward, backward and parameter update."""
if self._dummy_batch is None:
self._dummy_batch = samples[0]
self._set_seed()
self.model.train()
self.criterion.train()
self.zero_grad()
if not dummy_batch:
self.meters['train_wall'].start()
# forward and backward pass
logging_outputs, sample_sizes, ooms = [], [], 0
for i, sample in enumerate(samples):
sample = self._prepare_sample(sample)
if sample is None:
# when sample is None, run forward/backward on a dummy batch
# and ignore the resulting gradients
sample = self._prepare_sample(self._dummy_batch)
ignore_grad = True
else:
ignore_grad = False
def maybe_no_sync():
"""
Whenever *samples* contains more than one mini-batch, we
want to accumulate gradients locally and only call
all-reduce in the last backwards pass.
"""
if (self.args.distributed_world_size > 1
and hasattr(self.model, 'no_sync')
and i < len(samples) - 1):
return self.model.no_sync()
else:
return contextlib.ExitStack() # dummy contextmanager
try:
with maybe_no_sync():
# forward and backward
loss, sample_size, logging_output = self.task.train_step(
sample, self.model, self.criterion, self.optimizer,
ignore_grad)
if not ignore_grad:
logging_outputs.append(logging_output)
sample_sizes.append(sample_size)
if self.fast_stat_sync:
self._all_reduce_list[0] += sample_size
self._all_reduce_list[1] += logging_output.get(
'nsentences', 0.0)
self._all_reduce_list[2] += logging_output.get(
'loss', 0.0)
self._all_reduce_list[3] += logging_output.get(
'nll_loss', 0.0)
self._all_reduce_list[4] += logging_output.get(
'ntokens', 0.0)
except RuntimeError as e:
if 'out of memory' in str(e):
msg = ('| WARNING: ran out of memory with exception: ' +
'{};'.format(e) + '\n Skipping batch')
# TODO: print should really go to logger, this print goes
# to stderr, which is buffered, which in many cases is not
# printed out if another exception happens.
# NB(jerry): added a flush to mitigate this
print(msg, file=sys.stderr)
if torch.cuda.is_available() and hasattr(
torch.cuda, "memory_summary"):
for device_idx in range(torch.cuda.device_count()):
print(torch.cuda.memory_summary(
device=torch.cuda.device(device_idx)),
file=sys.stderr)
sys.stderr.flush()
if raise_oom:
raise ValueError(msg)
ooms += 1
self.zero_grad()
else:
raise e
if self.fast_stat_sync:
self._all_reduce_list[5] += ooms
if ooms > 0 and self._oom_batch is not None:
self.handle_ooms(ooms)
if dummy_batch:
return None
# gather logging outputs from all replicas
if self.fast_stat_sync:
# rework all_gather_list
all_reduce_list_tensor = torch.cuda.DoubleTensor(
self._all_reduce_list)
if self._sync_stats():
torch.distributed.all_reduce(all_reduce_list_tensor)
# Normalize loss and nll_loss by "sample_size"
# and convert to log base 2
all_reduce_list_tensor[2:4].div_(
(all_reduce_list_tensor[0:1] *
torch.log(torch.cuda.DoubleTensor([2]))))
self._all_reduce_list = all_reduce_list_tensor.tolist()
logging_output = {}
[
sample_size,
logging_output['nsentences'],
logging_output['loss'],
logging_output['nll_loss'],
logging_output['ntokens'],
ooms,
] = self._all_reduce_list
elif self._sync_stats():
logging_outputs, sample_sizes, ooms, prev_norms = \
zip(*distributed_utils.all_gather_list(
[logging_outputs, sample_sizes, ooms, self._prev_grad_norm],
))
logging_outputs = list(chain.from_iterable(logging_outputs))
sample_sizes = list(chain.from_iterable(sample_sizes))
ooms = sum(ooms)
if not self.args.use_bmuf:
assert (
all(norm == prev_norms[0] for norm in prev_norms) or all(
math.isnan(norm) or math.isinf(norm)
for norm in prev_norms)
), 'Fatal error: gradients are inconsistent between workers'
self.meters['oom'].update(ooms, len(samples))
if ooms == self.args.distributed_world_size * len(samples):
print('| WARNING: OOM in all workers, skipping update')
self.zero_grad()
return None
if not self.fast_stat_sync:
# aggregate logging outputs and sample sizes
logging_output = self.task.aggregate_logging_outputs(
logging_outputs, self.get_criterion())
sample_size = self.task.grad_denom(sample_sizes,
self.get_criterion())
if not all(k in logging_output for k in ['ntokens', 'nsentences']):
raise Exception(
('Please update the {}.aggregate_logging_outputs() method to '
'return ntokens and nsentences').format(
self.task.__class__.__name__))
try:
# normalize grads by sample size
if sample_size > 0:
self.optimizer.multiply_grads(
self.args.distributed_world_size / float(sample_size))
# clip grads
grad_norm = self.optimizer.clip_grad_norm(self.args.clip_norm)
self._prev_grad_norm = grad_norm
# take an optimization step
self.optimizer.step()
self.set_num_updates(self.get_num_updates() + 1)
# task specific update per step
self.task.update_step(self._num_updates)
# update meters
ntokens = logging_output.get('ntokens', 0)
nsentences = logging_output.get('nsentences', 0)
self.meters['wps'].update(ntokens)
self.meters['ups'].update(1.)
self.meters['wpb'].update(ntokens)
self.meters['bsz'].update(nsentences)
self.meters['gnorm'].update(grad_norm)
self.meters['clip'].update(1. if grad_norm > self.args.clip_norm
and self.args.clip_norm > 0 else 0.)
self.meters['train_loss'].update(logging_output.get('loss', 0),
sample_size)
if 'train_acc' in self.meters:
self.meters['train_acc'].update(logging_output.get('acc', 0),
sample_size)
if 'nll_loss' in logging_output:
self.meters['train_nll_loss'].update(
logging_output.get('nll_loss', 0), ntokens)
# clear CUDA cache to reduce memory fragmentation
if (self.args.empty_cache_freq > 0 and
((self.get_num_updates() + self.args.empty_cache_freq - 1) %
self.args.empty_cache_freq) == 0
and torch.cuda.is_available() and not self.args.cpu):
torch.cuda.empty_cache()
except OverflowError as e:
print('| WARNING: overflow detected, ' + str(e))
self.zero_grad()
logging_output = None
if self.args.fp16:
self.meters['loss_scale'].reset()
self.meters['loss_scale'].update(self.optimizer.scaler.loss_scale)
self.clear_buffered_stats()
self.meters['train_wall'].stop()
return logging_output
print(1e100 == float('inf'))
# False
print(float('inf') == math.inf == np.inf)
# True
print(1e1000 == math.inf)
# True
print(1e100 == math.inf)
# False
print(float('inf') == float('inf') * 100)
# True
print(math.isinf(1e1000))
# True
print(math.isinf(1e100))
# False
print(math.isinf(-1e1000))
# True
a = np.array([1, np.inf, -np.inf])
print(a)
# [ 1. inf -inf]
print(np.isinf(a))
# [False True True]
best_node_mean='NaN',
best_node_sd='NaN',
linear_discr_mean='NaN',
linear_discr_sd='NaN',
naive_bayes_mean='NaN',
naive_bayes_sd='NaN'
WHERE id={};
'''.format(id)
conn.execute(update_statement)
continue
if 'nr_class' not in mf:
print('Calculation failed.')
continue
for key, value in mf.items():
if math.isinf(value):
if value > 0:
mf[key] = sys.maxsize
else:
mf[key] = -sys.maxsize
if np.isnan(value):
store = False
break
if not store:
print('Skipping {} due to missing value in {}'.format(id, key))
continue
update_statement = '''
UPDATE datasets SET
nr_inst={nr_inst},
def _get_factors(self, output_desired, t_upper):
"""
Apply the continued fractions to find r and the gcd to find the desired factors.
"""
x_value = int(output_desired, 2)
logger.info('In decimal, x_final value for this result is: %s.', x_value)
if x_value <= 0:
self._ret['results'][output_desired] = \
'x_value is <= 0, there are no continued fractions.'
return False
logger.debug('Running continued fractions for this case.')
# Calculate T and x/T
T = pow(2, t_upper)
x_over_T = x_value / T
# Cycle in which each iteration corresponds to putting one more term in the
# calculation of the Continued Fraction (CF) of x/T
# Initialize the first values according to CF rule
i = 0
b = array.array('i')
t = array.array('f')
b.append(math.floor(x_over_T))
t.append(x_over_T - b[i])
while i >= 0:
# From the 2nd iteration onwards, calculate the new terms of the CF based
# on the previous terms as the rule suggests
if i > 0:
b.append(math.floor(1 / t[i - 1]))
t.append((1 / t[i - 1]) - b[i])
# Calculate the CF using the known terms
aux = 0
j = i
while j > 0:
aux = 1 / (b[j] + aux)
j = j - 1
aux = aux + b[0]
# Get the denominator from the value obtained
frac = fractions.Fraction(aux).limit_denominator()
denominator = frac.denominator
logger.debug('Approximation number %s of continued fractions:', i + 1)
logger.debug("Numerator:%s \t\t Denominator: %s.", frac.numerator, frac.denominator)
# Increment i for next iteration
i = i + 1
if denominator % 2 == 1:
if i >= self._N:
self._ret['results'][output_desired] = \
'unable to find factors after too many attempts.'
return False
logger.debug('Odd denominator, will try next iteration of continued fractions.')
continue
# If denominator even, try to get factors of N
# Get the exponential a^(r/2)
exponential = 0
if denominator < 1000:
exponential = pow(self._a, denominator / 2)
# Check if the value is too big or not
if math.isinf(exponential) or exponential > 1000000000:
self._ret['results'][output_desired] = \
'denominator of continued fraction is too big.'
return False
# If the value is not to big (infinity),
# then get the right values and do the proper gcd()
putting_plus = int(exponential + 1)
putting_minus = int(exponential - 1)
one_factor = math.gcd(putting_plus, self._N)
other_factor = math.gcd(putting_minus, self._N)
# Check if the factors found are trivial factors or are the desired factors
if one_factor == 1 or one_factor == self._N or \
other_factor == 1 or other_factor == self._N:
logger.debug('Found just trivial factors, not good enough.')
# Check if the number has already been found,
# use i-1 because i was already incremented
if t[i - 1] == 0:
self._ret['results'][output_desired] = \
'the continued fractions found exactly x_final/(2^(2n)).'
return False
if i >= self._N:
self._ret['results'][output_desired] = \
'unable to find factors after too many attempts.'
return False
else:
logger.debug('The factors of %s are %s and %s.', self._N, one_factor, other_factor)
logger.debug('Found the desired factors.')
self._ret['results'][output_desired] = (one_factor, other_factor)
factors = sorted((one_factor, other_factor))
if factors not in self._ret['factors']:
self._ret['factors'].append(factors)
return True
def gauss_laplace_momentum(self, potential,potential_disc, kinetic,\
stpsze, x0, p0, logp, grad, aux, n_disc = 0,\
M = None):
'''Performs one numerical integration step of the DHMC mixed
integrator, Alogirthm 2 from Nishimura et al. 2017.
Parameters
----------
stpsze - A scalar representing step size.
x0 - A torch.Tensor \in mathbb{R}^{D x 1}. Represents the intial
starting value.
p0 - A torch.Tensor \in mathbb{R}^{D x 1}. Represents the intial,
sampled momentum.
logp - A function representing the log of the posterior.
grad - A torch.Tensor, representing the gradient of the potential
energy evaluated at the specified point.
** May not need this if we include the functionality in the
potential function. **
M - A torch.Tensor M \in mathb{R}^{D x D}. Represents the
diagonal of the mass matrix.
Outputs
--------
x - A torch.Tensor \mathbb{R}^{D x 1}. The proposed x
p - A torch.Tensor \mathbb{R}^{D x 1}. The proposed p
logp - A torch.Tensor \mathbb{R}^{1 x 1}. As previously defined
gradcont - A troch.Tensor \mathbb{R}^{D x 1}. A tensor of the
gradients of the continous parameters
aux - Output from the potential. ** may not be needed.
'''
if M is None:
M = torch.ones(x0.size()[0])
x = x0.clone()
p = p0.clone()
# performs first half step on continous parameters
p[:-n_disc] = p[:-n_disc] + 0.5 * stpsze * potential(p[:-n_disc],
grad=True)
if n_disc == 0:
x = x + stpsze * kinetic(p, M, mom='Gauss', grad=True)
else:
# Update continous parameters if any
if self.n_param != self.n_disc:
x[:-n_disc] = x[:-n_disc] + stpsze * 0.5 * kinetic(
p[:-n_disc], M[:-n_disc], mom='Gauss', grad=True)
logp, aux = potential(x, grad=False)
gradcont = potential(x, grad=True)
# Update discrete parameters
if math.isinf(logp[0]):
return x, p, grad, logp, n_feval, n_fupdate # not sure what to do here line 149
# creates a function to permute 'J' indices. I.e J is the
# the discontinous set of parameters. Line 3 of Algorithm 2
coord_order = x.size()[0] - n_disc + np.random.permutation(n_disc)
for index in coord_order:
# calls on Algorithm 1 from Nishimura et al.
x, p, logp, aux = self.coordwise_int(self, potential_disc,\
aux, index, x, p, M,\
stpsze, logp)
x[:-n_disc] = x[:-n_disc] + stpsze * 0.5 * kinetic(
p[:-n_disc], M[:-n_disc], mom='Gauss', grad=True)
if self.n_param != self.n_disc:
logp = potential(x, grad='False')
gradcont = potential(x, grad='True')
return x, p, gradcont, logp, aux
assert math.ceil(0.5) == 1
assert math.ceil(1.0) == 1
assert math.ceil(1.5) == 2
assert math.ceil(-0.5) == 0
assert math.ldexp(float("inf"), -10**20) == float("inf")
assert almost_equal(math.log1p(1 / math.e - 1), -1)
assert almost_equal(math.log1p(0), 0)
assert almost_equal(math.log1p(math.e - 1), 1)
assert almost_equal(math.log1p(1), math.log(2))
assert almost_equal(math.acosh(1), 0)
assert almost_equal(math.acosh(2), 1.3169578969248168)
assert math.isinf(math.asinh(float("inf")))
assert almost_equal(math.asinh(0), 0)
assert almost_equal(math.asinh(1), 0.88137358701954305)
assert almost_equal(math.asinh(-1), -0.88137358701954305)
assert almost_equal(math.atanh(0), 0)
assert almost_equal(math.atanh(0.5), 0.54930614433405489)
assert almost_equal(math.atanh(-0.5), -0.54930614433405489)
assert math.isnan(math.atanh(float("nan")))
assert math.trunc(1.9) == 1.0
dist_fun = euclidean_distances
else: # 1-D list of strings
from nltk.metrics.distance import edit_distance
# x = ['we', 'shelled', 'clams', 'for', 'the', 'chowder']
# y = ['class', 'too']
x = ['i', 'soon', 'found', 'myself', 'muttering', 'to', 'the', 'walls']
y = ['see', 'drown', 'himself']
# x = 'we talked about the situation'.split()
# y = 'we talked about the situation'.split()
dist_fun = edit_distance
dist, cost, acc, path = dtw(x, y, dist_fun, w=w, s=s)
# Vizualize
from matplotlib import pyplot as plt
plt.imshow(cost.T,
origin='lower',
cmap=plt.cm.Reds,
interpolation='nearest')
plt.plot(path[0], path[1], '-o') # relation
plt.xticks(range(len(x)), x)
plt.yticks(range(len(y)), y)
plt.xlabel('x')
plt.ylabel('y')
plt.axis('tight')
if isinf(w):
plt.title('Minimum distance: {}, slope weight: {}'.format(dist, s))
else:
plt.title(
'Minimum distance: {}, window widht: {}, slope weight: {}'.format(
dist, w, s))
plt.show()
def create_features(self, token: Token, invoicePage: InvoicePage):
################################### HELPER FUNCTIONS FOR TOKEN GENERATION #################################################
# calculates and returns min dist from token 1 to token 2 in pixels
def calc_min_dist(t1, t2):
# get bounding outer rectangle
outer_rect_left = min(t1.coordinates["x"], t2.coordinates["x"])
outer_rect_top = min(t1.coordinates["y"], t2.coordinates["y"])
outer_rect_bottom = max(
(t1.coordinates["y"] + t1.coordinates["height"]),
(t2.coordinates["y"] + t2.coordinates["height"]),
)
outer_rect_right = max(
(t1.coordinates["x"] + t1.coordinates["width"]),
(t2.coordinates["x"] + t2.coordinates["width"]),
)
outer_rect_width = outer_rect_right - outer_rect_left
outer_rect_heigth = outer_rect_bottom - outer_rect_top
inner_rect_width = max(
0,
outer_rect_width -
(t1.coordinates["width"] + t2.coordinates["width"]),
)
inner_rect_height = max(
0,
outer_rect_heigth -
(t1.coordinates["height"] + t2.coordinates["height"]),
)
pixel_dist = math.sqrt(inner_rect_width**2 + inner_rect_height**2)
return pixel_dist
# checks if two tokens are aligned vertically within a margin of error (checks midpoint, left boundary, right boundary)
def is_vert_aligned(t1, t2, moe):
"""Returns true if t2 is vertically aligned with t1 and is below t1"""
t2_below_t1 = t1.coordinates["y"] - t2.coordinates["y"] < 0
if not t2_below_t1:
return False
if abs(t1.coordinates["x"] - t2.coordinates["x"]) < moe:
return True
if (abs((t1.coordinates["x"] + t1.coordinates["width"]) -
(t2.coordinates["x"] + t2.coordinates["width"])) < moe):
return True
t1_midpt_x = t1.coordinates["x"] + (t1.coordinates["width"] / 2)
t2_midpt_x = t2.coordinates["x"] + (t2.coordinates["width"] / 2)
if abs(t1_midpt_x - t2_midpt_x) < moe:
return True
return False
# checks if two tokens are aligned horizontally within a margin of error (checks midpoint, top boundary, bottom boundary)
def is_hori_aligned(t1, t2, moe):
"""Returns true if t2 is horizontally aligned with t1 and is to the right of t1"""
t2_to_right_of_t1 = t1.coordinates["x"] - t2.coordinates["x"] < 0
if not t2_to_right_of_t1:
return False
if abs(t1.coordinates["y"] - t2.coordinates["y"]) < moe:
return True
if (abs((t1.coordinates["y"] + t1.coordinates["height"]) -
(t2.coordinates["y"] + t2.coordinates["height"])) < moe):
return True
t1_midpt_y = t1.coordinates["y"] + (t1.coordinates["height"] / 2)
t2_midpt_y = t2.coordinates["y"] + (t2.coordinates["height"] / 2)
if abs(t1_midpt_y - t2_midpt_y) < moe:
return True
return False
#############################################################################################################################
features = {}
# number of characters and words in token text string
features["char_count"] = len(token.text)
features["word_count"] = len(token.text.split(" "))
# height and width of token
features["height"] = token.coordinates["height"]
features["width"] = token.coordinates["width"]
# distance to edges of page (to nearest point on the box)
features[
"rel_dist_top"] = token.coordinates["y"] / invoicePage.size["y"]
features[
"rel_dist_left"] = token.coordinates["x"] / invoicePage.size["x"]
features["dist_bottom"] = invoicePage.size["y"] - (
token.coordinates["y"] + token.coordinates["height"])
features["dist_right"] = invoicePage.size["x"] - (
token.coordinates["x"] + token.coordinates["width"])
# distance to boundaries of outermost text on page (tokens nearest to edge of page)
min_x = math.inf
min_y = math.inf
max_y = 0
max_x = 0
for t in invoicePage.grouped_tokens:
if t.coordinates["x"] < min_x:
min_x = t.coordinates["x"]
if t.coordinates["y"] < min_y:
min_y = t.coordinates["y"]
if t.coordinates["x"] + t.coordinates["width"] > max_x:
max_x = t.coordinates["x"] + t.coordinates["width"]
if t.coordinates["y"] + t.coordinates["height"] > max_y:
max_y = t.coordinates["y"] + t.coordinates["height"]
features["dist_top_outer"] = token.coordinates["y"] - min_y
features["dist_left_outer"] = token.coordinates["x"] - min_x
features["dist_bottom_outer"] = max_y - (token.coordinates["y"] +
token.coordinates["height"])
features["dist_right_outer"] = max_x - (token.coordinates["x"] +
token.coordinates["width"])
# relative size of token box compared to page
features["rel_size_page_x"] = token.coordinates[
"width"] / invoicePage.size["x"]
features["rel_size_page_y"] = (token.coordinates["height"] /
invoicePage.size["y"])
# ave dist to neighbours (pixel and relative)
min_dist_neighbours = [
calc_min_dist(t, token) for t in invoicePage.grouped_tokens
]
features["average_dist_neighbours_pixel"] = sum(
min_dist_neighbours) / len(min_dist_neighbours)
invoice_diag = math.sqrt(invoicePage.size["y"]**2 +
invoicePage.size["x"]**2)
features["average_dist_neighbours_rel"] = (
features["average_dist_neighbours_pixel"] / invoice_diag)
N = 5 # N is arbitrary
min_dist_neighbours.sort()
N_nearest_neighbours = min_dist_neighbours[:N]
features["average_dist_N_nearest_neighbours_pixel"] = sum(
N_nearest_neighbours) / len(N_nearest_neighbours)
features["average_dist_N_nearest_neighbours_rel"] = (
features["average_dist_N_nearest_neighbours_pixel"] / invoice_diag)
# relative size compared to other tokens (percentile)
perc_w = 0
perc_h = 0
for t in invoicePage.grouped_tokens:
if t is not token:
if t.coordinates["width"] < token.coordinates["width"]:
perc_w += 1
if t.coordinates["height"] < token.coordinates["height"]:
perc_h += 1
features["percentile_width"] = perc_w / len(invoicePage.grouped_tokens)
features["percentile_height"] = perc_h / len(
invoicePage.grouped_tokens)
# boolean if token contains fields
features["contains_date"] = 1 if token.date_values else 0
features["contains_currency"] = 1 if token.currency else 0
features[
"contains_specific_currency"] = 1 if token.specific_currency else 0
features["contains_date_range"] = 1 if token.date_range else 0
features["contains_address"] = 1 if token.address else 0
features["contains_num_label"] = 1 if token.num_label else 0
features["contains_total_label"] = 1 if token.total_label else 0
features["contains_amount_label"] = 1 if token.amount_label else 0
features["contains_date_label"] = 1 if token.date_label else 0
features["contains_date_of_invoice_label"] = 1 if features[
"contains_date_label"] and len(
token.date_label.split(" ")
) > 1 else 0 # This is a more specific feature that the one above
features["contains_digit"] = 1 if token.contains_digit else 0
features["contains_company"] = 1 if token.company else 0
features["contains_tax_label"] = 1 if token.tax_label else 0
# boolean if aligned with selected tokens
moe = 10 # arbitary 10 pixle margin of error
features["vert_align_to_cell_w_date"] = 0
features["vert_align_to_cell_w_currency"] = 0
features["vert_align_to_cell_w_address"] = 0
features["vert_align_to_cell_w_datelabel"] = 0
features["vert_align_to_cell_w_dateofinvoicelabel"] = 0
features["vert_align_to_cell_w_numlabel"] = 0
features["vert_align_to_cell_w_totallabel"] = 0
features["vert_align_to_cell_w_amountlabel"] = 0
features["vert_align_to_cell_w_digit"] = 0
features["vert_align_to_cell_w_invoicenum_label"] = 0
features["vert_align_to_cell_w_accountnum_label"] = 0
features["vert_align_to_cell_w_ponum_label"] = 0
features["vert_align_to_cell_w_tax_label"] = 0
features["hori_align_to_cell_w_date"] = 0
features["hori_align_to_cell_w_currency"] = 0
features["hori_align_to_cell_w_address"] = 0
features["hori_align_to_cell_w_datelabel"] = 0
features["hori_align_to_cell_w_dateofinvoicelabel"] = 0
features["hori_align_to_cell_w_numlabel"] = 0
features["hori_align_to_cell_w_totallabel"] = 0
features["hori_align_to_cell_w_amountlabel"] = 0
features["hori_align_to_cell_w_digit"] = 0
features["hori_align_to_cell_w_invoicenum_label"] = 0
features["hori_align_to_cell_w_accountnum_label"] = 0
features["hori_align_to_cell_w_ponum_label"] = 0
features["hori_align_to_cell_w_tax_label"] = 0
for t in invoicePage.grouped_tokens:
if t is not token:
if is_vert_aligned(t, token, moe):
if t.date_values:
features["vert_align_to_cell_w_date"] = 1
if t.currency:
features["vert_align_to_cell_w_currency"] = 1
if t.address:
features["vert_align_to_cell_w_address"] = 1
if t.date_label:
features["vert_align_to_cell_w_datelabel"] = 1
if t.date_of_invoice_label:
features["vert_align_to_cell_w_dateofinvoicelabel"] = 1
if t.num_label:
features["vert_align_to_cell_w_numlabel"] = 1
if t.invoice_num_label:
features["vert_align_to_cell_w_invoicenum_label"] = 1
if t.acc_num_label:
features["vert_align_to_cell_w_accountnum_label"] = 1
if t.po_num_label:
features["vert_align_to_cell_w_ponum_label"] = 1
if t.total_label:
features["vert_align_to_cell_w_totallabel"] = 1
if t.amount_label:
features["vert_align_to_cell_w_accountnum_label"] = 1
if t.contains_digit:
features["vert_align_to_cell_w_digit"] = 1
if t.tax_label:
features["vert_align_to_cell_w_tax_label"] = 1
if is_hori_aligned(t, token, moe):
if t.date_values:
features["hori_align_to_cell_w_date"] = 1
if t.currency:
features["hori_align_to_cell_w_currency"] = 1
if t.address:
features["hori_align_to_cell_w_address"] = 1
if t.date_label:
features["hori_align_to_cell_w_datelabel"] = 1
if t.date_of_invoice_label:
features["hori_align_to_cell_w_dateofinvoicelabel"] = 1
if t.num_label:
features["hori_align_to_cell_w_numlabel"] = 1
if t.invoice_num_label:
features["hori_align_to_cell_w_invoicenum_label"] = 1
if t.acc_num_label:
features["hori_align_to_cell_w_accountnum_label"] = 1
if t.po_num_label:
features["hori_align_to_cell_w_ponum_label"] = 1
if t.total_label:
features["hori_align_to_cell_w_totallabel"] = 1
if t.amount_label:
features["hori_align_to_cell_w_accountnum_label"] = 1
if t.contains_digit:
features["hori_align_to_cell_w_digit"] = 1
if t.tax_label:
features["hori_align_to_cell_w_tax_label"] = 1
# dist to nearest cell with field (inf if no field in page)
features["dist_nearest_cell_w_date"] = math.inf
features["dist_nearest_cell_w_currency"] = math.inf
features["dist_nearest_cell_w_address"] = math.inf
features["dist_nearest_cell_w_datelabel"] = math.inf
features["dist_nearest_cell_w_invoicedatelabel"] = math.inf
features["dist_nearest_cell_w_numlabel"] = math.inf
features["dist_nearest_cell_w_invoicenumlabel"] = math.inf
features["dist_nearest_cell_w_accnumlabel"] = math.inf
features["dist_nearest_cell_w_ponumlabel"] = math.inf
features["dist_nearest_cell_w_totallabel"] = math.inf
features["dist_nearest_cell_w_amountlabel"] = math.inf
features["dist_nearest_cell_w_digit"] = math.inf
features["dist_nearest_cell_w_tax_label"] = math.inf
for t in invoicePage.grouped_tokens:
if t is not token:
dist = calc_min_dist(t, token)
if t.date_values and dist < features[
"dist_nearest_cell_w_date"]:
features["dist_nearest_cell_w_date"] = dist
if t.currency and dist < features[
"dist_nearest_cell_w_currency"]:
features["dist_nearest_cell_w_currency"] = dist
if t.address and dist < features["dist_nearest_cell_w_address"]:
features["dist_nearest_cell_w_address"] = dist
if t.date_label and dist < features[
"dist_nearest_cell_w_datelabel"]:
features["dist_nearest_cell_w_datelabel"] = dist
if t.date_of_invoice_label and dist < features[
"dist_nearest_cell_w_invoicedatelabel"]:
features["dist_nearest_cell_w_invoicedatelabel"] = dist
if t.num_label and dist < features[
"dist_nearest_cell_w_numlabel"]:
features["dist_nearest_cell_w_numlabel"] = dist
if t.invoice_num_label and dist < features[
"dist_nearest_cell_w_invoicenumlabel"]:
features["dist_nearest_cell_w_invoicenumlabel"] = dist
if t.acc_num_label and dist < features[
"dist_nearest_cell_w_accnumlabel"]:
features["dist_nearest_cell_w_accnumlabel"] = dist
if t.po_num_label and dist < features[
"dist_nearest_cell_w_ponumlabel"]:
features["dist_nearest_cell_w_ponumlabel"] = dist
if t.total_label and dist < features[
"dist_nearest_cell_w_totallabel"]:
features["dist_nearest_cell_w_totallabel"] = dist
if t.total_label and dist < features[
"dist_nearest_cell_w_amountlabel"]:
features["dist_nearest_cell_w_amountlabel"] = dist
if t.contains_digit and dist < features[
"dist_nearest_cell_w_digit"]:
features["dist_nearest_cell_w_digit"] = dist
if t.tax_label and dist < features[
"dist_nearest_cell_w_tax_label"]:
features["dist_nearest_cell_w_tax_label"] = dist
DEFAULT_DISTANCE = 0.75 # This is arbitrary
for feature in features:
if "dist_nearest" in feature and math.isinf(features[feature]):
features[feature] = invoice_diag * DEFAULT_DISTANCE
features["rel_dist_nearest_cell_w_date"] = features[
"dist_nearest_cell_w_date"] / invoice_diag
features["rel_dist_nearest_cell_w_currency"] = features[
"dist_nearest_cell_w_currency"] / invoice_diag
features["rel_dist_nearest_cell_w_address"] = features[
"dist_nearest_cell_w_address"] / invoice_diag
features["rel_dist_nearest_cell_w_datelabel"] = features[
"dist_nearest_cell_w_datelabel"] / invoice_diag
features["rel_dist_nearest_cell_w_invoicedatelabel"] = features[
"dist_nearest_cell_w_invoicedatelabel"] / invoice_diag
features["rel_dist_nearest_cell_w_numlabel"] = features[
"dist_nearest_cell_w_numlabel"] / invoice_diag
features["rel_dist_nearest_cell_w_invoicenumlabel"] = features[
"dist_nearest_cell_w_invoicenumlabel"] / invoice_diag
features["rel_dist_nearest_cell_w_accnumlabel"] = features[
"dist_nearest_cell_w_accnumlabel"] / invoice_diag
features["rel_dist_nearest_cell_w_ponumlabel"] = features[
"dist_nearest_cell_w_ponumlabel"] / invoice_diag
features["rel_dist_nearest_cell_w_totallabel"] = features[
"dist_nearest_cell_w_totallabel"] / invoice_diag
features["rel_dist_nearest_cell_w_amountlabel"] = features[
"dist_nearest_cell_w_amountlabel"] / invoice_diag
features["rel_dist_nearest_cell_w_digit"] = features[
"dist_nearest_cell_w_digit"] / invoice_diag
features["rel_dist_nearest_cell_w_tax_label"] = features[
"dist_nearest_cell_w_tax_label"] / invoice_diag
"""
features TODO:
-text in bold (contains, aligns, dist)
-text from an area that was shaded
-token was from a grid (can be close to grid lines or q far away in a well spaced out grid box) / near a summation line (above total payable amounts)
-percentage of black pixels in box?
"""
return features
def take(self, n=None, constructor=list):
"""
Returns a container with the n first elements from the Stream, or less if
there aren't enough. Use this without args if you need only one element
outside a list.
Parameters
----------
n :
Number of elements to be taken. Defaults to None.
Rounded when it's a float, and this can be ``inf`` for taking all.
constructor :
Container constructor function that can receie a generator as input.
Defaults to ``list``.
Returns
-------
The first ``n`` elements of the Stream sequence, created by the given
constructor unless ``n == None``, which means returns the next element
from the sequence outside any container.
If ``n`` is None, this can raise StopIteration due to lack of data in
the Stream. When ``n`` is a number, there's no such exception.
Examples
--------
>>> Stream(5).take(3) # Three elements
[5, 5, 5]
>>> Stream(1.2, 2, 3).take() # One element, outside a container
1.2
>>> Stream(1.2, 2, 3).take(1) # With n = 1 argument, it'll be in a list
[1.2]
>>> Stream(1.2, 2, 3).take(1, constructor=tuple) # Why not a tuple?
(1.2,)
>>> Stream([1, 2]).take(3) # More than the Stream size, n is integer
[1, 2]
>>> Stream([]).take() # More than the Stream size, n is None
Traceback (most recent call last):
...
StopIteration
Taking rounded float quantities and "up to infinity" elements
(don't try using ``inf`` with endless Stream instances):
>>> Stream([4, 3, 2, 3, 2]).take(3.4)
[4, 3, 2]
>>> Stream([4, 3, 2, 3, 2]).take(3.6)
[4, 3, 2, 3]
>>> Stream([4, 3, 2, 3, 2]).take(inf)
[4, 3, 2, 3, 2]
See Also
--------
Stream.peek :
Returns the n first elements from the Stream, without removing them.
Note
----
You should avoid using take() as if this would be an iterator. Streams
are iterables that can be easily part of a "for" loop, and their
iterators (the ones automatically used in for loops) are slightly faster.
Use iter() builtin if you need that, instead, or perhaps the blocks
method.
"""
if n is None:
return next(self._data)
if isinf(n) and n > 0:
return constructor(self._data)
if isinstance(n, float):
n = rint(n) if n > 0 else 0 # So this works with -inf and nan
return constructor(next(self._data) for _ in xrange(n))
def json_clean(obj):
"""Clean an object to ensure it's safe to encode in JSON.
Atomic, immutable objects are returned unmodified. Sets and tuples are
converted to lists, lists are copied and dicts are also copied.
Note: dicts whose keys could cause collisions upon encoding (such as a dict
with both the number 1 and the string '1' as keys) will cause a ValueError
to be raised.
Parameters
----------
obj : any python object
Returns
-------
out : object
A version of the input which will not cause an encoding error when
encoded as JSON. Note that this function does not *encode* its inputs,
it simply sanitizes it so that there will be no encoding errors later.
"""
# types that are 'atomic' and ok in json as-is.
atomic_ok = (unicode_type, type(None))
# containers that we need to convert into lists
container_to_list = (tuple, set, types.GeneratorType)
# Since bools are a subtype of Integrals, which are a subtype of Reals,
# we have to check them in that order.
if isinstance(obj, bool):
return obj
if isinstance(obj, numbers.Integral):
# cast int to int, in case subclasses override __str__ (e.g. boost enum, #4598)
return int(obj)
if isinstance(obj, numbers.Real):
# cast out-of-range floats to their reprs
if math.isnan(obj) or math.isinf(obj):
return repr(obj)
return float(obj)
if isinstance(obj, atomic_ok):
return obj
if isinstance(obj, bytes):
if py3compat.PY3:
# unanmbiguous binary data is base64-encoded
# (this probably should have happened upstream)
return b2a_base64(obj).decode('ascii')
else:
# Python 2 bytestr is ambiguous,
# needs special handling for possible binary bytestrings.
# imperfect workaround: if ascii, assume text.
# otherwise assume binary, base64-encode (py3 behavior).
try:
return obj.decode('ascii')
except UnicodeDecodeError:
return b2a_base64(obj).decode('ascii')
if isinstance(obj,
container_to_list) or (hasattr(obj, '__iter__')
and hasattr(obj, next_attr_name)):
obj = list(obj)
if isinstance(obj, list):
return [json_clean(x) for x in obj]
if isinstance(obj, dict):
# First, validate that the dict won't lose data in conversion due to
# key collisions after stringification. This can happen with keys like
# True and 'true' or 1 and '1', which collide in JSON.
nkeys = len(obj)
nkeys_collapsed = len(set(map(unicode_type, obj)))
if nkeys != nkeys_collapsed:
raise ValueError('dict cannot be safely converted to JSON: '
'key collision would lead to dropped values')
# If all OK, proceed by making the new dict that will be json-safe
out = {}
for k, v in iteritems(obj):
out[unicode_type(k)] = json_clean(v)
return out
if isinstance(obj, datetime):
return obj.strftime(ISO8601)
# we don't understand it, it's probably an unserializable object
raise ValueError("Can't clean for JSON: %r" % obj)
def _compute_and_update_PI_kernel(
i,
T_A,
T_B,
m,
QT_even,
QT_odd,
QT_first,
M_T,
Σ_T,
μ_Q,
σ_Q,
k,
ignore_trivial,
excl_zone,
profile,
indices,
compute_QT,
):
"""
A Numba CUDA kernel to update the matrix profile and matrix profile indices
Parameters
----------
i : int
sliding window `i`
T_A : ndarray
The time series or sequence for which to compute the dot product
T_B : ndarray
The time series or sequence that will be used to annotate T_A. For every
subsequence in T_A, its nearest neighbor in T_B will be recorded.
m : int
Window size
QT_even : ndarray
The input QT array (dot product between the query sequence,`Q`, and
time series, `T`) to use when `i` is even
QT_odd : ndarray
The input QT array (dot product between the query sequence,`Q`, and
time series, `T`) to use when `i` is odd
QT_first : ndarray
Dot product between the first query sequence,`Q`, and time series, `T`
M_T : ndarray
Sliding mean of time series, `T`
Σ_T : ndarray
Sliding standard deviation of time series, `T`
μ_Q : ndarray
Mean of the query sequence, `Q`
σ_Q : ndarray
Standard deviation of the query sequence, `Q`
k : int
The total number of sliding windows to iterate over
ignore_trivial : bool
Set to `True` if this is a self-join. Otherwise, for AB-join, set this to
`False`.
excl_zone : int
The half width for the exclusion zone relative to the current
sliding window
profile : ndarray
Matrix profile. The first column consists of the global matrix profile,
the second column consists of the left matrix profile, and the third
column consists of the right matrix profile.
indices : ndarray
The first column consists of the matrix profile indices, the second
column consists of the left matrix profile indices, and the third
column consists of the right matrix profile indices.
compute_QT : bool
A boolean flag for whether or not to compute QT
Returns
-------
None
Notes
-----
`DOI: 10.1109/ICDM.2016.0085 \
<https://www.cs.ucr.edu/~eamonn/STOMP_GPU_final_submission_camera_ready.pdf>`__
See Table II, Figure 5, and Figure 6
"""
start = cuda.grid(1)
stride = cuda.gridsize(1)
if i % 2 == 0:
QT_out = QT_even
QT_in = QT_odd
else:
QT_out = QT_odd
QT_in = QT_even
for j in range(start, QT_out.shape[0], stride):
zone_start = max(0, j - excl_zone)
zone_stop = min(k, j + excl_zone)
if compute_QT:
QT_out[j] = (QT_in[j - 1] - T_B[i - 1] * T_A[j - 1] +
T_B[i + m - 1] * T_A[j + m - 1])
QT_out[0] = QT_first[i]
if math.isinf(M_T[j]) or math.isinf(μ_Q[i]):
D = np.inf
else:
if (σ_Q[i] < config.STUMPY_STDDEV_THRESHOLD
or Σ_T[j] < config.STUMPY_STDDEV_THRESHOLD):
D = m
else:
denom = m * σ_Q[i] * Σ_T[j]
if math.fabs(
denom
) < config.STUMPY_DENOM_THRESHOLD: # pragma nocover
denom = config.STUMPY_DENOM_THRESHOLD
D = abs(2 * m * (1.0 -
(QT_out[j] - m * μ_Q[i] * M_T[j]) / denom))
if (σ_Q[i] < config.STUMPY_STDDEV_THRESHOLD
and Σ_T[j] < config.STUMPY_STDDEV_THRESHOLD
) or D < config.STUMPY_D_SQUARED_THRESHOLD:
D = 0
if ignore_trivial:
if i <= zone_stop and i >= zone_start:
D = np.inf
if D < profile[j, 1] and i < j:
profile[j, 1] = D
indices[j, 1] = i
if D < profile[j, 2] and i > j:
profile[j, 2] = D
indices[j, 2] = i
if D < profile[j, 0]:
profile[j, 0] = D
indices[j, 0] = i
def main(conf):
os.environ['CUDA_VISIBLE_DEVICES'] = ','.join(conf['main_args']['cuda'])
model_dir = conf['main_args']['model_dir']
exp_dir = conf['main_args']['exp_dir']
# Define Dataloader
assert len(conf['main_args']['train_metadata']) == len(
conf['main_args']['train_n_src'])
train_gens = []
for i in range(len(conf['main_args']['train_metadata'])):
train_set = LibriMix(csv_path=conf['main_args']['train_metadata'][i],
sample_rate=conf['data']['sample_rate'],
n_src=conf['main_args']['train_n_src'][i],
segment=conf['data']['segment'])
train_gen = DataLoader(
train_set,
shuffle=True,
batch_size=int(conf['training']['batch_size'] * 2 /
conf['main_args']['train_n_src'][i]),
num_workers=conf['training']['num_workers'],
drop_last=True)
train_gens.append(train_gen)
assert len(conf['main_args']['val_metadata']) == len(
conf['main_args']['val_n_src'])
val_gens = []
for i in range(len(conf['main_args']['val_metadata'])):
val_set = LibriMix(csv_path=conf['main_args']['val_metadata'][i],
sample_rate=conf['data']['sample_rate'],
n_src=conf['main_args']['val_n_src'][i],
segment=conf['data']['segment'])
val_gen = DataLoader(val_set,
shuffle=True,
batch_size=conf['training']['batch_size'],
num_workers=conf['training']['num_workers'],
drop_last=True)
print(val_gen)
val_gens.append(val_gen)
SPKID = LIBRISPEECH_SPKID[conf['data']['subset']]
# Loss functions
loss_fn = dict()
loss_fn['sisdr'] = PermInvariantSISDR(return_individual_results=True)
loss_fn['spk_circle'] = CircleLoss(m=0.25, gamma=15)
# Define model, optimizer + scheduler
if conf['main_args']['stage'] == 1:
model = CAE(conf['cae'])
model_path = os.path.join(exp_dir, 'cae', config_cae_path(conf['cae']))
elif conf['main_args']['stage'] == 2:
conf['tcn'].update({
'cae_path':
os.path.join(exp_dir, 'cae', config_cae_path(conf['cae']))
})
model = CAE_DANet(conf['tcn'])
model_path = os.path.join(exp_dir, model_dir)
if conf['loss_fn']['spk_ce'] > 0:
model.danet.add_softmax(output_size=len(SPKID), normalize=False)
loss_fn['spk_ce'] = model.danet.spk_softmax
else:
raise ValueError('Training stage should be either 1 or 2!')
model = torch.nn.DataParallel(model).cuda()
opt = torch.optim.Adam(model.module.parameters(), lr=conf['optim']['lr'])
# Validation metric
metric_name = 'SISDRi'
if metric_name == 'SISDRi':
SISDRi = PermInvariantSISDR(backward_loss=False,
improvement=True,
return_individual_results=True)
# Save config
# os.makedirs(exp_dir, exist_ok=True)
os.makedirs(model_path, exist_ok=True)
conf_path = os.path.join(model_path, 'conf.yml')
with open(conf_path, 'w') as outfile:
yaml.safe_dump(conf, outfile)
# Train model
tr_step = 0
val_step = 0
new_lr = conf['optim']['lr']
halving = False
best_val_loss = float("-inf")
val_no_impv = 0
for i in range(conf['training']['epochs']):
metric_dic = {
'train_{}'.format(metric_name): 0.,
'val_{}'.format(metric_name): 0.
}
print("Training stage {} || Epoch: {}/{}".format(
conf['main_args']['stage'], i + 1, conf['training']['epochs']))
model.train()
train_metric_mean = []
for data_set in zip(tqdm(train_gens[0], desc='Training'), train_gens[1]) if len(train_gens) == 2 \
else tqdm(train_gens[0], desc='Training'): # mini-batch
if not isinstance(data_set, tuple):
data_set = (data_set, )
for data in data_set:
opt.zero_grad()
m1wavs = data[0].unsqueeze(1).cuda()
clean_wavs = data[-1].cuda()
speaker_id = data[1]
if conf['main_args']['stage'] == 1:
recon_sources, enc_masks, enc_mixture = model.module(
m1wavs, clean_wavs)
if conf['main_args']['stage'] == 2:
estimated_masks, enc_masks, enc_mixture, Wx, phase = model(
m1wavs,
clean_wavs,
train=True,
n_sources=clean_wavs.shape[1])
V = estimated_masks[
1] # V (B, K, F*T), enc_masks (B, C, F*T)
A = estimated_masks[2] # (B, nspk, K)
estimated_masks = estimated_masks[
0] # estimated_masks (B, nspk, F*T)
recon_sources = model.module.get_rec_sources(
estimated_masks.view(m1wavs.shape[0],
estimated_masks.shape[1],
model.module.input_dim, -1),
enc_mixture,
phase=phase) # recovered waveform
l_dict = dict()
if conf['loss_fn']['sisdr'] > 0:
l_sisdr = loss_fn['sisdr'](recon_sources,
clean_wavs).mean()
l_dict.update(
{'sisdr': conf['loss_fn']['sisdr'] * l_sisdr})
if conf['loss_fn']['compact'] > 0:
enc_mixture = enc_mixture.view(enc_mixture.shape[0],
-1).unsqueeze(1)
w = -enc_mixture / torch.sum(
enc_mixture, dim=[1, 2], keepdim=True)
enc_masks[enc_masks <= 0.5] = 0
An = F.normalize(A.detach(), dim=2)
l_va = w * enc_masks * (torch.bmm(An, F.normalize(V,
dim=1)))
l_va = l_va.sum(dim=[1, 2]).mean()
l_dict.update(
{'compact': conf['loss_fn']['compact'] * l_va})
if conf['loss_fn']['spk_circle'] > 0:
L = torch.zeros(A.shape[0], A.shape[1]).cuda()
for j in range(A.shape[0]):
for k in range(A.shape[1]):
L[j][k] = SPKID.index(speaker_id[k][j])
inp_sp, inp_sn = convert_label_to_similarity(
A.view(-1, A.shape[2]), L.view(-1))
l_c = loss_fn['spk_circle'](inp_sp, inp_sn)
l_dict.update(
{'circle': conf['loss_fn']['spk_circle'] * l_c})
if conf['loss_fn']['spk_ce'] > 0:
label = torch.zeros(A.shape[0],
A.shape[1],
dtype=torch.int64).cuda()
for j in range(A.shape[0]):
for k in range(A.shape[1]):
label[j][k] = SPKID.index(speaker_id[k][j])
if conf['tcn']['sim'] == 'cos':
l_softmax = loss_fn['spk_ce'](F.normalize(A.view(
-1, A.shape[2]),
p=2,
dim=1),
label.view(-1))
else:
l_softmax = loss_fn['spk_ce'](A.view(-1, A.shape[2]),
label.view(-1))
l_dict.update(
{'spksoftmax': conf['loss_fn']['spk_ce'] * l_softmax})
# Loss back-propagation
l = torch.tensor(0.0).cuda()
for loss in l_dict.values():
if not math.isinf(loss) and not math.isnan(loss):
l = l + loss
if not conf['cae']['stft'] or conf['main_args']['stage'] == 2:
l.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 5)
opt.step()
train_metric = SISDRi(recon_sources,
clean_wavs,
initial_mixtures=m1wavs)
train_metric_mean += train_metric.tolist()
train_metric_mean = np.mean(train_metric_mean)
metric_dic['train_{}'.format(metric_name)] = train_metric_mean
tr_step += 1
if val_gens is not None:
model.eval()
with torch.no_grad():
val_metric_mean = []
for data_set in zip(tqdm(val_gens[0], desc='Validation'), val_gens[1]) if len(val_gens) == 2 \
else tqdm(val_gens[0], desc='Validation'): # mini-batch
if not isinstance(data_set, tuple):
data_set = (data_set, )
for data in data_set:
m1wavs = data[0].unsqueeze(1).cuda()
clean_wavs = data[-1].cuda()
if conf['main_args']['stage'] == 1:
recon_sources, _, _ = model.module(
m1wavs, clean_wavs)
if conf['main_args']['stage'] == 2:
estimated_masks, _, enc_mixture, _, phase = model(
m1wavs,
clean_wavs,
train=True,
n_sources=clean_wavs.shape[1])
V = estimated_masks[
1] # V (B, K, F*T), enc_masks (B, C, F*T)
A = estimated_masks[2] # (B, nspk, K)
estimated_masks = estimated_masks[
0] # estimated_masks (B, nspk, F*T)
recon_sources = model.module.get_rec_sources(
estimated_masks.view(m1wavs.shape[0],
estimated_masks.shape[1],
model.module.input_dim,
-1),
enc_mixture,
phase=phase) # recovered waveform
val_metric = SISDRi(recon_sources,
clean_wavs,
initial_mixtures=m1wavs)
val_metric_mean += val_metric.tolist()
val_metric_mean = np.mean(val_metric_mean)
metric_dic['val_{}'.format(metric_name)] = val_metric_mean
val_step += 1
# Adjust learning rate (halving)
if conf['training']['half_lr']:
val_loss = round(val_metric_mean, 2) # keep two decimal places
if val_loss <= best_val_loss:
val_no_impv += 1
if val_no_impv % 6 == 0:
halving = True
if val_no_impv >= 20 and conf['training']['early_stop']:
print("No imporvement for 20 epochs, early stopping.")
break
else:
best_val_loss = val_loss
val_no_impv = 0
if halving:
optim_state = opt.state_dict()
optim_state['param_groups'][0]['lr'] = \
optim_state['param_groups'][0]['lr'] / 2.0
opt.load_state_dict(optim_state)
print('Learning rate adjusted to: {lr:.6f}'.format(
lr=optim_state['param_groups'][0]['lr']))
halving = False
# val_no_impv = 0
CAE.save_if_best(save_dir=model_path,
model=model.module,
optimizer=opt,
epoch=tr_step,
tr_loss=train_metric_mean,
cv_loss=val_metric_mean,
cv_loss_name='SISDRi',
save_every=50)
pprint(metric_dic)
def on_frame(frame):
global points_agg, frame_count, stop, prev_angle
cv_img = bridge.imgmsg_to_cv2(frame, 'bgr8')
frame_count += 1
if frame_count > 5 or stop:
print(frame_count)
print('stopped')
publish_drive(0, 0)
#cv2.imshow('cv_img', cv_img)
#cv2.waitKey(1)
return
# publish_drive(15, -80)
#msg_speed = drive_speed()
#msg_speed.speed = 10
#pub_speed.publish(msg_speed)
#USE THIS
#publish_drive(15,0)
# msg_angle = drive_angle()
# msg_angle.angle = 0
# msg_angle.publish(msg_angle)
hsv_image = cv2.cvtColor(cv_img, cv2.COLOR_BGR2HSV)
gray_image = cv2.cvtColor(cv_img, cv2.COLOR_BGR2GRAY)
#ORIGINALLY 25-30
lower_yellow = np.array([30, 100, 100], dtype = 'uint8')
upper_yellow = np.array([35, 255, 255], dtype='uint8')
mask_yellow = cv2.inRange(hsv_image, lower_yellow, upper_yellow)
mask_white = cv2.inRange(gray_image, 200, 255)
mask_yw = cv2.bitwise_or(mask_white, mask_yellow)
mask_yw_image = cv2.bitwise_and(gray_image, mask_yellow)
#cv2.imshow('cv_img', mask_yw_image)
#cv2.waitKey(1)
#mask_yw_image = cv2.cvtColor(cv_img, cv2.COLOR_BGR2GRAY)
ros_image = bridge.cv2_to_imgmsg(mask_yw_image, 'mono8')
#camera1_pub.publish(ros_image)
# frame_count = 0
# return
kernel_size = (5,5)
gauss_gray = cv2.GaussianBlur(mask_yw_image,kernel_size, 0)
low_threshold = 50
high_threshold = 150
canny_edges = cv2.Canny(gauss_gray,low_threshold,high_threshold)
#cv2.imshow('cv_img', canny_edges)
#cv2.waitKey(1)
imshape = canny_edges.shape
lower_left = [0,imshape[0]]
lower_right = [imshape[1],imshape[0]]
top_left = [0,imshape[0]/2]
top_right = [imshape[1],imshape[0]/2]
vertices = [np.array([lower_left,top_left,top_right,lower_right],dtype=np.int32)]
roi_image = region_of_interest(canny_edges, vertices)
#rho and theta are the distance and angular resolution of the grid in Hough space
#same values as quiz
rho = 2
theta = np.pi/180
#threshold is minimum number of intersections in a grid for candidate line to go to output
threshold = 20
min_line_len = 50
max_line_gap = 200
lines = hough_lines(roi_image, rho, theta, threshold, min_line_len, max_line_gap)
line_image = lines_on_image(lines, roi_image)
#cv2.imshow('cv_img', line_image)
#cv2.waitKey(1)
points = list()
if lines == None:
print('no lines; skipping')
return
for line in lines:
for x1, y1, x2, y2 in line:
points.append([x1, y1])
points.append([x2, y2])
points_agg += points
if frame_count > 3:
line_image = points_on_image(points_agg, roi_image)
kmeans = KMeans(n_clusters=min(8, len(points_agg))).fit(points_agg)
line_image = points_on_image(kmeans.cluster_centers_.astype(int), line_image, size=15)
# print(kmeans.cluster_centers_.astype(int)[:,0])
result = weighted_img(line_image, cv_img, alpha=0.8, beta=1., l=0.)
bounding_path = nn_walk(kmeans.cluster_centers_.astype(int))
line_segments = [[[int(bounding_path[i][0]), int(bounding_path[i][1]), int(bounding_path[i+1][0]), int(bounding_path[i+1][1])]] for i in range(len(bounding_path)-1)]
draw_lines(result, line_segments)
# cv2.imshow('cv_img', result)
# cv2.waitKey(1)
# Pathfinder algorithm
LINE_Y = result.shape[0] - 150.0
# find candidate line segments that hit that y
candidates = []
for i in range(len(bounding_path) - 1):
if (bounding_path[i][1] <= LINE_Y and bounding_path[i + 1][1] >= LINE_Y) or (bounding_path[i][1] >= LINE_Y and bounding_path[i + 1][1] <= LINE_Y):
candidates.append([bounding_path[i], bounding_path[i + 1]])
cv2.line(result, (int(bounding_path[i][0]), int(bounding_path[i][1])), (int(bounding_path[i+1][0]), int(bounding_path[i+1][1])), [0, 0, 255], 5)
slopes = [get_slope(p[0][0], p[0][1], p[1][0], p[1][1]) for p in candidates]
candidates = [candidates[i] for i in range(len(candidates)) if (not math.isnan(slopes[i])) and slopes[i] != 0 and (not math.isinf(slopes[i]))]
slopes = [slopes[i] for i in range(len(slopes)) if (not math.isnan(slopes[i])) and slopes[i] != 0 and (not math.isinf(slopes[i]))]
x = list()
for i in range(len(candidates)):
x.append(int((LINE_Y - candidates[i][0][1] + slopes[i] * candidates[i][0][0]) / slopes[i]))
# for a in x:
# cv2.line(result, (a, 0), (a, 480), [0, 0, 255], 2)
#original values
#left_cap = -100
#right_cap = result.shape[1] + 100
left_cap = -1000
right_cap = result.shape[1] + 1000
for a in x:
if a < result.shape[1] / 2:
if a > left_cap:
left_cap = a
else:
if a < right_cap:
right_cap = a
#print("left" + str(left_cap))
#print("right" + str(right_cap))
# cv2.line(result, (left_cap, 0), (left_cap, 480), [0, 0, 255], 2)
# cv2.line(result, (right_cap, 0), (right_cap, 480), [0, 0, 255], 2)
midpoint = (left_cap + right_cap) / 2
# print (midpoint, int(LINE_Y))
tri_base = midpoint - 320
angle = 0
speed = 15
#how to make it slowdown..
if len(x) == 0:
angle = prev_angle
speed = 17
print("no lines")
elif len(x) == 1:
single_lane = -1 * np.polyfit(kmeans.cluster_centers_.astype(int)[:,0], kmeans.cluster_centers_.astype(int)[:,1], 1)
angle = np.degrees(np.arctan(single_lane[0]))
if angle > 0:
angle = 90 - angle
elif angle < 0:
angle = -90 - angle
speed = 17
print("one line - angle: " + str(angle))
elif tri_base != 0:
angle = np.degrees(np.arctan((480 - LINE_Y) / tri_base))
if angle > 0:
angle = 90 - angle
elif angle < 0:
angle = -90 - angle
print("two lines?")
if len(x) != 0:
prev_angle = angle
cv2.circle(result, (midpoint, int(LINE_Y)), 10, (0, 0, 255), -1)
#print(midpoint, int(LINE_Y))
#print(angle)
publish_drive(speed, angle)
#publish_drive(0,0)
#cv2.imshow('cv_img', result)
#cv2.waitKey(20)
ros_image = bridge.cv2_to_imgmsg(result, 'bgr8')
camera1_pub.publish(ros_image)
frame_count = 0
points_agg = list()
return
slopes = list()
for line in lines:
for x1, y1, x2, y2 in line:
slope = get_slope(x1, y1, x2, y2)
if not math.isinf(slope) and not math.isnan(slope):
slopes.append(abs(get_slope(x1, y1, x2, y2)))
# plt.hist(slopes)
# plt.show()
result = weighted_img(line_image, cv_img, alpha=0.8, beta=1., l=0.)
# cv2.arrowedLine(result, (0, 0), (100, 100), (0,0,255), 5)
ros_image = bridge.cv2_to_imgmsg(result, 'bgr8')
camera1_pub.publish(ros_image)
def main_compute(code_only=False):
def compile_callback(code):
return bytearray()
tvm.register_func('tvm_callback_cuda_compile',
compile_callback,
override=True)
default_tune_op = importlib.import_module('lang.generic')
import logging
import warnings
from tvm import autotvm
warnings.simplefilter("ignore")
logging.getLogger('autotvm').setLevel(logging.ERROR)
logging.getLogger('autotvm').addHandler(logging.StreamHandler(sys.stdout))
task = autotvm.task.create("template_op", args=(), target=tvm_target)
AntaresGlobal.default_tune_op = default_tune_op
AntaresGlobal.default_task = task
if verbose:
print(' >> Backend = %s, Python PID = %s, Task = %s;' %
(backend, os.getpid(), default_tune_op.__name__))
num_trials = int(os.environ['STEP']) if 'STEP' in os.environ else 0
config = os.environ.get('CONFIG', '').strip()
if config != '':
best_config = config
elif num_trials > 0:
dev_num = backend_config.get_execution_parallism()
if dev_num <= 0:
raise Exception("No valid device found for backend: %s." % backend)
batch_size = os.environ.get('BATCH', '')
batch_size = 16 if not batch_size else int(batch_size)
from concurrent.futures import ThreadPoolExecutor
worker_size = batch_size if batch_size < dev_num else dev_num
thread_pool = ThreadPoolExecutor(max_workers=worker_size)
tuner_type = os.environ.get('TUNER')
if not tuner_type:
explicit_ops = AntaresGlobal.attrs.explicit_ops
tuner_type = 'OpEvo'
print(' >> MAKE_PARA = %d/%d, EXEC_PARA = %d, TUNER = %s\n' %
(worker_size, batch_size, dev_num, tuner_type))
auto_commit = os.environ.get('COMMIT', '')
if auto_commit:
saved_code = codehub_db(os.environ['COMPUTE_V1'])
if saved_code is not None and auto_commit != 'force':
raise Exception(
"Saved code has existed in codehub. Please try COMMIT=force to override it."
)
os.environ.pop('COMMIT')
try:
if getattr(AntaresGlobal, 'mode', None) == 'antares':
task.search_space_v2 = backend_config.search_space(
AntaresGlobal.compute_graph)
else:
task.search_space_v2 = AntaresGlobal.attrs.auto_config.get_config_space(
)
task.n_parallel = batch_size
tuner = importlib.import_module('tuner.%s.main' % tuner_type)
tuner = tuner.MainTuner(task)
except:
raise Exception('>> Cannot import Antares Tuner: %s' % tuner_type)
if hasattr(tuner, 'cleanup'):
AntaresGlobal.cleanup_funcs.append(tuner.cleanup)
if tuner is not None:
AntaresGlobal.current_step = 0
AntaresGlobal.completed_trials = 0
AntaresGlobal.num_trials = num_trials
eval_client.init(backend_root=backend_root)
def measure_batch(inputs):
results, futures = [], []
target_sources, config_strs = [], []
for i in range(len(inputs)):
dir_sid = AntaresGlobal.current_step + i + 1
config_str = inputs[i].config if type(
inputs[i].config).__name__ == 'str' else 'null'
config_strs.append(config_str)
try:
target_source = get_target_source(
config_strs[i], dir_sid)
except:
# traceback.print_exc()
target_source = None
target_sources.append(target_source)
expected_timecost = tuner.task.best.timecost if not math.isinf(
tuner.task.best.timecost) else min(
30, float(os.environ.get('EXPECTED_TIMEOUT', 'inf')))
for i in range(len(inputs)):
dir_sid = AntaresGlobal.current_step + i + 1
futures.append(
thread_pool.submit(run_config_entity,
target_sources[i], config_strs[i],
dir_sid, expected_timecost,
i % dev_num))
best_slot = -1
for i in range(len(inputs)):
dir_sid = AntaresGlobal.current_step + i + 1
t = futures[i].result()
if t < tuner.task.best.timecost:
best_slot = dir_sid
tuner.task.best.timecost = t
tuner.task.best.config = inputs[i].config
tuner.task.best.occur = best_slot
results.append(
autotvm.measure.MeasureResult(costs=(t, ),
error_no=0,
all_cost=i,
timestamp=time.time()))
AntaresGlobal.current_step += len(results)
stage_logs = 'STEP[%d / %d] Current Best Config = %s, Perf = %g sec / op (%g Gflops), MemRatio = %g %%, Occur Step = %d;' % (
AntaresGlobal.current_step, num_trials,
tuner.task.best.config, tuner.task.best.timecost,
compute_gflops(tuner.task.flop, tuner.task.best.timecost),
compute_mem_ratio(
tuner.task.best.timecost), tuner.task.best.occur)
print('\n\033[93m%s\033[0m' %
('=' * min(120, len(stage_logs))))
print(stage_logs)
print('\033[93m%s\033[0m\n' %
('=' * min(120, len(stage_logs))))
if auto_commit and best_slot >= 0:
with open(local_get_dir_file('my_kernel.cc', best_slot),
'r') as fp:
device_source = fp.read()
with open(local_get_dir_file('result.txt', best_slot),
'r') as fp:
t = float(fp.read().split()[0])
kernel_path = codehub_db(
os.environ['COMPUTE_V1'],
source_code=device_source + code_suffix(
tpr=t, step_prod=best_slot, step_plan=num_trials))
print(' >> Update current code to codehub: %s' %
kernel_path)
return results
tuner.task.best = Mock()
tuner.task.best.timecost = float('inf')
tuner.task.best.config = None
tuner.task.best.occur = -1
tuner.measure_batch = measure_batch
tuner.measure_batch.n_parallel = batch_size
callbacks = []
history_log_for_transfer_learning = os.environ.get('RECORD', '')
if history_log_for_transfer_learning:
callbacks.append(
autotvm.callback.log_to_file(
history_log_for_transfer_learning))
# Enable Transfer Learning for Incremental Task
if os.path.exists(history_log_for_transfer_learning):
print(
' >> Loading incremental history from log file: %s ..'
% history_log_for_transfer_learning)
tuner.load_history(
autotvm.record.load_from_file(
history_log_for_transfer_learning))
tuner.tune(n_trial=num_trials,
callbacks=callbacks,
measure_option=None)
if math.isinf(tuner.task.best.timecost):
print(
f'[Error] No valid config found in the whole tuning. (Try other tuner types other than `TUNER={tuner_type}`?)'
)
cleanup_on_exit(0, None)
best_config = tuner.task.best.config
if auto_commit:
device_source = codehub_db(os.environ['COMPUTE_V1'])
codehub_db(os.environ['COMPUTE_V1'],
source_code=device_source +
'\n// Antares Tuning Completed in %d steps.' %
AntaresGlobal.current_step)
print(
"\n[Best Config] CONFIG='%s' ==> Performance is up to %f Gflops, occurred at step %d / %d; time per run = %g sec."
%
(best_config,
compute_gflops(tuner.task.flop, tuner.task.best.timecost),
tuner.task.best.occur, num_trials, tuner.task.best.timecost))
cleanup_on_exit(-1, None)
else:
raise Exception('Unrecognized tuner type: `%s`' % tuner_type)
exit(0)
else:
saved_code = codehub_db(os.environ['COMPUTE_V1'])
if saved_code is not None:
print(" >> Using Saved Code from Codehub:")
print(
"// ---------------------------------------------------------------------------"
)
print(saved_code)
print(
"// ---------------------------------------------------------------------------"
)
exit(0)
best_config = ''
assert isinstance(best_config, str)
best_config = best_config if best_config else 'null'
device_source, kernel_path = get_target_source(best_config)
if code_only:
return device_source
if verbose:
print()
print(
"// ---------------------------------------------------------------------------"
)
print(device_source)
print(
"// ---------------------------------------------------------------------------"
)
eval_client.init(backend_root=backend_root)
dev_id = int(os.environ.get('DEV_ID', '0'))
result = evaluate_perf(kernel_path, dev_id, device_source)
exit(0 if result is not None and len(result) > 1 else 1)
def margin_intersect_offset(gradient_left, gradient_right, base_width, margin):
r"""
mL <- gradient left
mR / <- gradient right
\ /\/ <- margin
x-/---- <- regular intersect
/ \ | <- intersect offset
/ x-\---- <- margin intersect
/ / \ \
/ / \ \
/_/_____\_\_____
o_/_______\_o___| <- margin
|<--------->| <- base width
OR
mL <- gradient left
mR / <- gradient right
| /
x---- <- regular intersect
/ | | <- intersect offset
/ x-|---- <- margin intersect
/ / | |
/ / | |
/_/_____|_|___
o_/_______|_o___| <- margin
|<--------->| <- base width
"""
logging.debug(f"Gradient Left / Right: "
f"{gradient_left: 7.3f} / {gradient_right: 7.3f}")
if isinf(gradient_left) and isinf(gradient_right):
return None
if np.isclose(gradient_left, gradient_right, atol=ATOL):
return None
regular_axis_intercept_left = 0
regular_axis_intercept_right = base_width if isinf(gradient_right) else \
- base_width * gradient_right
logging.debug(
f"Regular Axis Intercept Left / Right: "
f"{regular_axis_intercept_left: 7.3f} / {regular_axis_intercept_right:7.3f}"
)
regular_intersect_x, regular_intersect_y = intersect_lines(
gradient_left, regular_axis_intercept_left, gradient_right,
regular_axis_intercept_right)
if regular_intersect_x is None or regular_intersect_y is None:
return None
logging.debug(f"Regular Intersect X / Y: "
f"{regular_intersect_x: 7.3f} / {regular_intersect_y:7.3f}")
margin_axis_intercept_left = margin if isinf(gradient_left) else \
- abs(margin / gradient_cos(gradient_left))
margin_axis_intercept_right = base_width - margin if isinf(gradient_right) else \
regular_axis_intercept_right - abs(margin / gradient_cos(gradient_right))
logging.debug(
f"Margin Axis Intercept Left / Right: "
f"{margin_axis_intercept_left: 7.3f} / {margin_axis_intercept_right:7.3f}"
)
margin_intersect_x, margin_intersect_y = intersect_lines(
gradient_left, margin_axis_intercept_left, gradient_right,
margin_axis_intercept_right)
logging.debug(f"Margin Intersect X / Y: "
f"{margin_intersect_x: 7.3f} / {margin_intersect_y:7.3f}")
return regular_intersect_y - margin_intersect_y
def stringify_number(number):
bound_for_float = 7.0
if -bound_for_float < number < bound_for_float and number != 0.0:
return str(number)
return str(int(number))
pl.xlabel(x_axis)
pl.ylabel(y_axis)
if y_scale_log:
pl.yscale('log')
pl.legend(loc='best')
constraint_text_horizontal_position = 0.94 - len(plot_data) * 0.07
for key in constraints:
constraint_text = get_symbol[key]
if math.isinf(constraints[key][1]):
if math.isinf(constraints[key][0]):
continue
else:
constraint_text += "$\geq$" + stringify_number(
constraints[key][0]) + get_unit[key]
else:
if not math.isinf(constraints[key][0]):
constraint_text = stringify_number(
constraints[key]
[0]) + get_unit[key] + "$\leq$" + constraint_text
constraint_text += "$\leq$" + stringify_number(
constraints[key][1]) + get_unit[key]
pl.text(0.97,
constraint_text_horizontal_position,
constraint_text,
def measure_batch(inputs):
results, futures = [], []
target_sources, config_strs = [], []
for i in range(len(inputs)):
dir_sid = AntaresGlobal.current_step + i + 1
config_str = inputs[i].config if type(
inputs[i].config).__name__ == 'str' else 'null'
config_strs.append(config_str)
try:
target_source = get_target_source(
config_strs[i], dir_sid)
except:
# traceback.print_exc()
target_source = None
target_sources.append(target_source)
expected_timecost = tuner.task.best.timecost if not math.isinf(
tuner.task.best.timecost) else min(
30, float(os.environ.get('EXPECTED_TIMEOUT', 'inf')))
for i in range(len(inputs)):
dir_sid = AntaresGlobal.current_step + i + 1
futures.append(
thread_pool.submit(run_config_entity,
target_sources[i], config_strs[i],
dir_sid, expected_timecost,
i % dev_num))
best_slot = -1
for i in range(len(inputs)):
dir_sid = AntaresGlobal.current_step + i + 1
t = futures[i].result()
if t < tuner.task.best.timecost:
best_slot = dir_sid
tuner.task.best.timecost = t
tuner.task.best.config = inputs[i].config
tuner.task.best.occur = best_slot
results.append(
autotvm.measure.MeasureResult(costs=(t, ),
error_no=0,
all_cost=i,
timestamp=time.time()))
AntaresGlobal.current_step += len(results)
stage_logs = 'STEP[%d / %d] Current Best Config = %s, Perf = %g sec / op (%g Gflops), MemRatio = %g %%, Occur Step = %d;' % (
AntaresGlobal.current_step, num_trials,
tuner.task.best.config, tuner.task.best.timecost,
compute_gflops(tuner.task.flop, tuner.task.best.timecost),
compute_mem_ratio(
tuner.task.best.timecost), tuner.task.best.occur)
print('\n\033[93m%s\033[0m' %
('=' * min(120, len(stage_logs))))
print(stage_logs)
print('\033[93m%s\033[0m\n' %
('=' * min(120, len(stage_logs))))
if auto_commit and best_slot >= 0:
with open(local_get_dir_file('my_kernel.cc', best_slot),
'r') as fp:
device_source = fp.read()
with open(local_get_dir_file('result.txt', best_slot),
'r') as fp:
t = float(fp.read().split()[0])
kernel_path = codehub_db(
os.environ['COMPUTE_V1'],
source_code=device_source + code_suffix(
tpr=t, step_prod=best_slot, step_plan=num_trials))
print(' >> Update current code to codehub: %s' %
kernel_path)
return results
def assertNearlyEqual(self, x, y, sigfigs=3): magnitude = (abs(x) + abs(y)) / 2 if math.isinf(magnitude): magnitude = 1 self.assertAlmostEqual(x, y, delta=magnitude*(10**(-sigfigs)))
def default(obj, json_options=DEFAULT_JSON_OPTIONS):
# We preserve key order when rendering SON, DBRef, etc. as JSON by
# returning a SON for those types instead of a dict.
if isinstance(obj, ObjectId):
return {"$oid": str(obj)}
if isinstance(obj, DBRef):
return _json_convert(obj.as_doc(), json_options=json_options)
if isinstance(obj, datetime.datetime):
if (json_options.datetime_representation ==
DatetimeRepresentation.ISO8601):
if not obj.tzinfo:
obj = obj.replace(tzinfo=utc)
if obj >= EPOCH_AWARE:
off = obj.tzinfo.utcoffset(obj)
if (off.days, off.seconds, off.microseconds) == (0, 0, 0):
tz_string = 'Z'
else:
tz_string = obj.strftime('%z')
millis = int(obj.microsecond / 1000)
fracsecs = ".%03d" % (millis, ) if millis else ""
return {
"$date":
"%s%s%s" %
(obj.strftime("%Y-%m-%dT%H:%M:%S"), fracsecs, tz_string)
}
millis = bson._datetime_to_millis(obj)
if (json_options.datetime_representation ==
DatetimeRepresentation.LEGACY):
return {"$date": millis}
return {"$date": {"$numberLong": str(millis)}}
if json_options.strict_number_long and isinstance(obj, Int64):
return {"$numberLong": str(obj)}
if isinstance(obj, (RE_TYPE, Regex)):
flags = ""
if obj.flags & re.IGNORECASE:
flags += "i"
if obj.flags & re.LOCALE:
flags += "l"
if obj.flags & re.MULTILINE:
flags += "m"
if obj.flags & re.DOTALL:
flags += "s"
if obj.flags & re.UNICODE:
flags += "u"
if obj.flags & re.VERBOSE:
flags += "x"
if isinstance(obj.pattern, text_type):
pattern = obj.pattern
else:
pattern = obj.pattern.decode('utf-8')
if json_options.json_mode == JSONMode.LEGACY:
return SON([("$regex", pattern), ("$options", flags)])
return {
'$regularExpression': SON([("pattern", pattern),
("options", flags)])
}
if isinstance(obj, MinKey):
return {"$minKey": 1}
if isinstance(obj, MaxKey):
return {"$maxKey": 1}
if isinstance(obj, Timestamp):
return {"$timestamp": SON([("t", obj.time), ("i", obj.inc)])}
if isinstance(obj, Code):
if obj.scope is None:
return {'$code': str(obj)}
return SON([('$code', str(obj)),
('$scope', _json_convert(obj.scope, json_options))])
if isinstance(obj, Binary):
return _encode_binary(obj, obj.subtype, json_options)
if PY3 and isinstance(obj, bytes):
return _encode_binary(obj, 0, json_options)
if isinstance(obj, uuid.UUID):
if json_options.strict_uuid:
data = obj.bytes
subtype = OLD_UUID_SUBTYPE
if json_options.uuid_representation == CSHARP_LEGACY:
data = obj.bytes_le
elif json_options.uuid_representation == JAVA_LEGACY:
data = data[7::-1] + data[:7:-1]
elif json_options.uuid_representation == UUID_SUBTYPE:
subtype = UUID_SUBTYPE
return _encode_binary(data, subtype, json_options)
else:
return {"$uuid": obj.hex}
if isinstance(obj, Decimal128):
return {"$numberDecimal": str(obj)}
if isinstance(obj, bool):
return obj
if (json_options.json_mode == JSONMode.CANONICAL
and isinstance(obj, integer_types)):
if -2**31 <= obj < 2**31:
return {'$numberInt': text_type(obj)}
return {'$numberLong': text_type(obj)}
if json_options.json_mode != JSONMode.LEGACY and isinstance(obj, float):
if math.isnan(obj):
return {'$numberDouble': 'NaN'}
elif math.isinf(obj):
representation = 'Infinity' if obj > 0 else '-Infinity'
return {'$numberDouble': representation}
elif json_options.json_mode == JSONMode.CANONICAL:
# repr() will return the shortest string guaranteed to produce the
# original value, when float() is called on it. str produces a
# shorter string in Python 2.
return {'$numberDouble': text_type(repr(obj))}
raise TypeError("%r is not JSON serializable" % obj)