Skip to content

P
pytensor
提交 bea31470 authored 作者: Frédéric Bastien's avatar Frédéric Bastien 提交者： GitHub
浏览文件
操作

Merge pull request #5782 from Amrithasuresh/master

Updated numpy as np #4218
上级 56da8ca8 65a48d98
隐藏空白字符变更
内嵌 并排
正在显示 包含 719 行增加 和 719 行删除
theano/sandbox/fourier.py
浏览文件 @ bea31470
...@@ -4,7 +4,7 @@ Provides Ops for FFT and DCT. ...@@ -4,7 +4,7 @@ Provides Ops for FFT and DCT.
""" """
from __future__ import absolute_import, print_function, division from __future__ import absolute_import, print_function, division
import numpy import numpy as np
import numpy.fft import numpy.fft
from six.moves import xrange from six.moves import xrange
...@@ -126,13 +126,13 @@ def dct_matrix(rows, cols, unitary=True): ...@@ -126,13 +126,13 @@ def dct_matrix(rows, cols, unitary=True):
This algorithm is adapted from Dan Ellis' Rastmat spec2cep.m, lines 15-20. This algorithm is adapted from Dan Ellis' Rastmat spec2cep.m, lines 15-20.
""" """
rval = numpy.zeros((rows, cols)) rval = np.zeros((rows, cols))
col_range = numpy.arange(cols) col_range = np.arange(cols)
scale = numpy.sqrt(2.0 / cols) scale = np.sqrt(2.0 / cols)
for i in xrange(rows): for i in xrange(rows):
rval[i] = numpy.cos( rval[i] = np.cos(
i * (col_range * 2 + 1) / (2.0 * cols) * numpy.pi) * scale i * (col_range * 2 + 1) / (2.0 * cols) * np.pi) * scale
if unitary: if unitary:
rval[0] *= numpy.sqrt(0.5) rval[0] *= np.sqrt(0.5)
return rval return rval
theano/sandbox/linalg/tests/test_linalg.py
浏览文件 @ bea31470
from __future__ import absolute_import, print_function, division from __future__ import absolute_import, print_function, division
import numpy import numpy as np
import numpy.linalg import numpy.linalg
import theano import theano
...@@ -39,9 +39,9 @@ def test_rop_lop(): ...@@ -39,9 +39,9 @@ def test_rop_lop():
non_sequences=[y, mx, mv]) non_sequences=[y, mx, mv])
scan_f = function([mx, mv], sy) scan_f = function([mx, mv], sy)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vx = numpy.asarray(rng.randn(4, 4), theano.config.floatX) vx = np.asarray(rng.randn(4, 4), theano.config.floatX)
vv = numpy.asarray(rng.randn(4, 4), theano.config.floatX) vv = np.asarray(rng.randn(4, 4), theano.config.floatX)
v1 = rop_f(vx, vv) v1 = rop_f(vx, vv)
v2 = scan_f(vx, vv) v2 = scan_f(vx, vv)
...@@ -61,7 +61,7 @@ def test_rop_lop(): ...@@ -61,7 +61,7 @@ def test_rop_lop():
'Op did not raised an error even though the function' 'Op did not raised an error even though the function'
' is not differentiable')) ' is not differentiable'))
vv = numpy.asarray(rng.uniform(size=(4,)), theano.config.floatX) vv = np.asarray(rng.uniform(size=(4,)), theano.config.floatX)
yv = tensor.Lop(y, mx, v) yv = tensor.Lop(y, mx, v)
lop_f = function([mx, v], yv) lop_f = function([mx, v], yv)
...@@ -75,21 +75,21 @@ def test_rop_lop(): ...@@ -75,21 +75,21 @@ def test_rop_lop():
def test_spectral_radius_bound(): def test_spectral_radius_bound():
tol = 10 ** (-6) tol = 10 ** (-6)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
x = theano.tensor.matrix() x = theano.tensor.matrix()
radius_bound = spectral_radius_bound(x, 5) radius_bound = spectral_radius_bound(x, 5)
f = theano.function([x], radius_bound) f = theano.function([x], radius_bound)
shp = (3, 4) shp = (3, 4)
m = rng.rand(*shp) m = rng.rand(*shp)
m = numpy.cov(m).astype(config.floatX) m = np.cov(m).astype(config.floatX)
radius_bound_theano = f(m) radius_bound_theano = f(m)
# test the approximation # test the approximation
mm = m mm = m
for i in range(5): for i in range(5):
mm = numpy.dot(mm, mm) mm = np.dot(mm, mm)
radius_bound_numpy = numpy.trace(mm) ** (2 ** (-5)) radius_bound_numpy = np.trace(mm) ** (2 ** (-5))
assert abs(radius_bound_numpy - radius_bound_theano) < tol assert abs(radius_bound_numpy - radius_bound_theano) < tol
# test the bound # test the bound
......
theano/sandbox/minimal.py
浏览文件 @ bea31470
...@@ -2,7 +2,7 @@ from __future__ import absolute_import, print_function, division ...@@ -2,7 +2,7 @@ from __future__ import absolute_import, print_function, division
import unittest import unittest
import numpy import numpy as np
from theano import gof, tensor, function from theano import gof, tensor, function
from theano.tests import unittest_tools as utt from theano.tests import unittest_tools as utt
...@@ -39,11 +39,11 @@ class Minimal(gof.Op): ...@@ -39,11 +39,11 @@ class Minimal(gof.Op):
# but do not modify any of the arguments [inplace]. # but do not modify any of the arguments [inplace].
print("perform got %i arguments" % len(inputs)) print("perform got %i arguments" % len(inputs))
print("Max of input[0] is ", numpy.max(inputs[0])) print("Max of input[0] is ", np.max(inputs[0]))
# return some computed value. # return some computed value.
# do not return something that is aliased to one of the inputs. # do not return something that is aliased to one of the inputs.
output[0] = numpy.asarray(0, dtype='int64') output[0] = np.asarray(0, dtype='int64')
minimal = Minimal() minimal = Minimal()
...@@ -55,7 +55,7 @@ minimal = Minimal() ...@@ -55,7 +55,7 @@ minimal = Minimal()
class T_minimal(unittest.TestCase): class T_minimal(unittest.TestCase):
def setUp(self): def setUp(self):
self.rng = numpy.random.RandomState(utt.fetch_seed(666)) self.rng = np.random.RandomState(utt.fetch_seed(666))
def test0(self): def test0(self):
A = tensor.matrix() A = tensor.matrix()
...@@ -66,6 +66,6 @@ class T_minimal(unittest.TestCase): ...@@ -66,6 +66,6 @@ class T_minimal(unittest.TestCase):
print('built') print('built')
Aval = self.rng.randn(5, 5) Aval = self.rng.randn(5, 5)
bval = numpy.arange(5, dtype=float) bval = np.arange(5, dtype=float)
f(Aval, bval) f(Aval, bval)
print('done') print('done')
theano/sandbox/multinomial.py
浏览文件 @ bea31470
from __future__ import absolute_import, print_function, division from __future__ import absolute_import, print_function, division
import numpy import numpy as np
import warnings import warnings
import theano import theano
...@@ -172,7 +172,7 @@ class MultinomialFromUniform(Op): ...@@ -172,7 +172,7 @@ class MultinomialFromUniform(Op):
raise ValueError("unis.shape[0] != pvals.shape[0] * n_samples", raise ValueError("unis.shape[0] != pvals.shape[0] * n_samples",
unis.shape[0], pvals.shape[0], n_samples) unis.shape[0], pvals.shape[0], n_samples)
if z[0] is None or z[0].shape != pvals.shape: if z[0] is None or z[0].shape != pvals.shape:
z[0] = numpy.zeros(pvals.shape, dtype=node.outputs[0].dtype) z[0] = np.zeros(pvals.shape, dtype=node.outputs[0].dtype)
else: else:
z[0].fill(0) z[0].fill(0)
...@@ -209,7 +209,7 @@ class MultinomialFromUniform(Op): ...@@ -209,7 +209,7 @@ class MultinomialFromUniform(Op):
# have the same answer as the c code as in the c code # have the same answer as the c code as in the c code
# the cumul is in double precission. # the cumul is in double precission.
cumsum = pvals[n].cumsum(dtype='float64') cumsum = pvals[n].cumsum(dtype='float64')
z[0][n, numpy.searchsorted(cumsum, unis_n)] += 1 z[0][n, np.searchsorted(cumsum, unis_n)] += 1
class ChoiceFromUniform(MultinomialFromUniform): class ChoiceFromUniform(MultinomialFromUniform):
...@@ -380,8 +380,8 @@ class ChoiceFromUniform(MultinomialFromUniform): ...@@ -380,8 +380,8 @@ class ChoiceFromUniform(MultinomialFromUniform):
else: else:
odtype = self.odtype odtype = self.odtype
if (z[0] is None or if (z[0] is None or
not numpy.all(z[0].shape == [pvals.shape[0], n_samples])): not np.all(z[0].shape == [pvals.shape[0], n_samples])):
z[0] = -1 * numpy.ones((pvals.shape[0], n_samples), dtype=odtype) z[0] = -1 * np.ones((pvals.shape[0], n_samples), dtype=odtype)
nb_multi = pvals.shape[0] nb_multi = pvals.shape[0]
nb_outcomes = pvals.shape[1] nb_outcomes = pvals.shape[1]
......
theano/sandbox/rng_mrg.py
浏览文件 @ bea31470
...@@ -8,7 +8,7 @@ http://www.iro.umontreal.ca/~simardr/ssj/indexe.html ...@@ -8,7 +8,7 @@ http://www.iro.umontreal.ca/~simardr/ssj/indexe.html
from __future__ import absolute_import, print_function, division from __future__ import absolute_import, print_function, division
import warnings import warnings
import numpy import numpy as np
from six import integer_types from six import integer_types
from six.moves import xrange from six.moves import xrange
...@@ -38,7 +38,7 @@ if theano.sandbox.cuda.cuda_available: ...@@ -38,7 +38,7 @@ if theano.sandbox.cuda.cuda_available:
def matVecModM(A, s, m): def matVecModM(A, s, m):
# TODO : need description for method, parameter and return # TODO : need description for method, parameter and return
assert A.dtype == 'int64' assert A.dtype == 'int64'
return numpy.int32(numpy.sum((A * s) % m, 1) % m) return np.int32(np.sum((A * s) % m, 1) % m)
def multMatVect(v, A, m1, B, m2): def multMatVect(v, A, m1, B, m2):
...@@ -97,7 +97,7 @@ class DotModulo(Op): ...@@ -97,7 +97,7 @@ class DotModulo(Op):
(out,) = outputs (out,) = outputs
o1 = matVecModM(A, s, m) o1 = matVecModM(A, s, m)
o2 = matVecModM(A2, s2, m2) o2 = matVecModM(A2, s2, m2)
out[0] = numpy.concatenate((o1, o2)) out[0] = np.concatenate((o1, o2))
def c_code_cache_version(self): def c_code_cache_version(self):
return (6,) return (6,)
...@@ -198,39 +198,39 @@ class DotModulo(Op): ...@@ -198,39 +198,39 @@ class DotModulo(Op):
# MRG31k3p # MRG31k3p
# generator constants : # generator constants :
M1 = numpy.asarray(numpy.int32(2147483647)) # 2^31 - 1 M1 = np.asarray(np.int32(2147483647)) # 2^31 - 1
M2 = numpy.asarray(numpy.int32(2147462579)) # 2^31 - 21069 M2 = np.asarray(np.int32(2147462579)) # 2^31 - 21069
MASK12 = numpy.int32(511) # 2^9 - 1 MASK12 = np.int32(511) # 2^9 - 1
MASK13 = numpy.int32(16777215) # 2^24 - 1 MASK13 = np.int32(16777215) # 2^24 - 1
MASK2 = numpy.int32(65535) # 2^16 - 1 MASK2 = np.int32(65535) # 2^16 - 1
MULT2 = numpy.int32(21069) MULT2 = np.int32(21069)
NORM = 4.656612873077392578125e-10 # 1./2^31 NORM = 4.656612873077392578125e-10 # 1./2^31
# A1p0 = numpy.asarray([[0, 4194304, 129], [1, 0, 0], [0, 1, 0]], # A1p0 = np.asarray([[0, 4194304, 129], [1, 0, 0], [0, 1, 0]],
# dtype='int64') # dtype='int64')
# A2p0 = numpy.asarray([[32768, 0, 32769], [1, 0, 0], [0, 1, 0]], # A2p0 = np.asarray([[32768, 0, 32769], [1, 0, 0], [0, 1, 0]],
# dtype='int64') # dtype='int64')
A1p72 = numpy.asarray([[1516919229, 758510237, 499121365], A1p72 = np.asarray([[1516919229, 758510237, 499121365],
[1884998244, 1516919229, 335398200], [1884998244, 1516919229, 335398200],
[601897748, 1884998244, 358115744]], [601897748, 1884998244, 358115744]],
dtype='int64') dtype='int64')
A2p72 = numpy.asarray([[1228857673, 1496414766, 954677935], A2p72 = np.asarray([[1228857673, 1496414766, 954677935],
[1133297478, 1407477216, 1496414766], [1133297478, 1407477216, 1496414766],
[2002613992, 1639496704, 1407477216]], [2002613992, 1639496704, 1407477216]],
dtype='int64') dtype='int64')
A1p134 = numpy.asarray( A1p134 = np.asarray(
[[1702500920, 1849582496, 1656874625], [[1702500920, 1849582496, 1656874625],
[828554832, 1702500920, 1512419905], [828554832, 1702500920, 1512419905],
[1143731069, 828554832, 102237247]], [1143731069, 828554832, 102237247]],
dtype='int64') dtype='int64')
A2p134 = numpy.asarray( A2p134 = np.asarray(
[[796789021, 1464208080, 607337906], [[796789021, 1464208080, 607337906],
[1241679051, 1431130166, 1464208080], [1241679051, 1431130166, 1464208080],
[1401213391, 1178684362, 1431130166]], [1401213391, 1178684362, 1431130166]],
dtype='int64') dtype='int64')
np_int32_vals = [numpy.int32(i) for i in (0, 7, 9, 15, 16, 22, 24)] np_int32_vals = [np.int32(i) for i in (0, 7, 9, 15, 16, 22, 24)]
def ff_2p134(rstate): def ff_2p134(rstate):
...@@ -246,14 +246,14 @@ def ff_2p72(rstate): ...@@ -246,14 +246,14 @@ def ff_2p72(rstate):
def mrg_next_value(rstate, new_rstate): def mrg_next_value(rstate, new_rstate):
# TODO : need description for method, parameter and return # TODO : need description for method, parameter and return
x11, x12, x13, x21, x22, x23 = rstate x11, x12, x13, x21, x22, x23 = rstate
assert type(x11) == numpy.int32 assert type(x11) == np.int32
i0, i7, i9, i15, i16, i22, i24 = np_int32_vals i0, i7, i9, i15, i16, i22, i24 = np_int32_vals
# first component # first component
y1 = (((x12 & MASK12) << i22) + (x12 >> i9) + y1 = (((x12 & MASK12) << i22) + (x12 >> i9) +
((x13 & MASK13) << i7) + (x13 >> i24)) ((x13 & MASK13) << i7) + (x13 >> i24))
assert type(y1) == numpy.int32 assert type(y1) == np.int32
if (y1 < 0 or y1 >= M1): # must also check overflow if (y1 < 0 or y1 >= M1): # must also check overflow
y1 -= M1 y1 -= M1
y1 += x13 y1 += x13
...@@ -266,11 +266,11 @@ def mrg_next_value(rstate, new_rstate): ...@@ -266,11 +266,11 @@ def mrg_next_value(rstate, new_rstate):
# second component # second component
y1 = ((x21 & MASK2) << i15) + (MULT2 * (x21 >> i16)) y1 = ((x21 & MASK2) << i15) + (MULT2 * (x21 >> i16))
assert type(y1) == numpy.int32 assert type(y1) == np.int32
if (y1 < 0 or y1 >= M2): if (y1 < 0 or y1 >= M2):
y1 -= M2 y1 -= M2
y2 = ((x23 & MASK2) << i15) + (MULT2 * (x23 >> i16)) y2 = ((x23 & MASK2) << i15) + (MULT2 * (x23 >> i16))
assert type(y2) == numpy.int32 assert type(y2) == np.int32
if (y2 < 0 or y2 >= M2): if (y2 < 0 or y2 >= M2):
y2 -= M2 y2 -= M2
y2 += x23 y2 += x23
...@@ -286,7 +286,7 @@ def mrg_next_value(rstate, new_rstate): ...@@ -286,7 +286,7 @@ def mrg_next_value(rstate, new_rstate):
# Must never return either 0 or M1+1 # Must never return either 0 or M1+1
new_rstate[...] = [x11, x12, x13, x21, x22, x23] new_rstate[...] = [x11, x12, x13, x21, x22, x23]
assert new_rstate.dtype == numpy.int32 assert new_rstate.dtype == np.int32
if (x11 <= x21): if (x11 <= x21):
return (x11 - x21 + M1) * NORM return (x11 - x21 + M1) * NORM
else: else:
...@@ -360,22 +360,22 @@ class mrg_uniform(mrg_uniform_base): ...@@ -360,22 +360,22 @@ class mrg_uniform(mrg_uniform_base):
# some rng_mrg tests) I also add this limit here. # some rng_mrg tests) I also add this limit here.
raise ValueError("rng_mrg does not support more then (2**31 -1) samples") raise ValueError("rng_mrg does not support more then (2**31 -1) samples")
rstate = numpy.asarray(rstate) # bring state from GPU if necessary rstate = np.asarray(rstate) # bring state from GPU if necessary
if not self.inplace: if not self.inplace:
rstate = rstate.copy() rstate = rstate.copy()
n_streams, _ = rstate.shape n_streams, _ = rstate.shape
rval = numpy.zeros(n_elements, dtype=self.output_type.dtype) rval = np.zeros(n_elements, dtype=self.output_type.dtype)
err_orig = numpy.seterr(over='ignore') err_orig = np.seterr(over='ignore')
try: try:
for i in xrange(n_elements): for i in xrange(n_elements):
sample = mrg_next_value(rstate[i % n_streams], sample = mrg_next_value(rstate[i % n_streams],
rstate[i % n_streams]) rstate[i % n_streams])
rval[i] = sample rval[i] = sample
finally: finally:
numpy.seterr(**err_orig) np.seterr(**err_orig)
# send to GPU if necessary # send to GPU if necessary
o_rstate[0] = node.outputs[0].type.filter(rstate) o_rstate[0] = node.outputs[0].type.filter(rstate)
...@@ -396,13 +396,13 @@ class mrg_uniform(mrg_uniform_base): ...@@ -396,13 +396,13 @@ class mrg_uniform(mrg_uniform_base):
'NPY_ARRAY_ENSURECOPY|NPY_ARRAY_C_CONTIGUOUS|' 'NPY_ARRAY_ENSURECOPY|NPY_ARRAY_C_CONTIGUOUS|'
'NPY_ARRAY_ALIGNED') 'NPY_ARRAY_ALIGNED')
ndim = self.output_type.ndim ndim = self.output_type.ndim
o_type_num = numpy.asarray(0, dtype=self.output_type.dtype).dtype.num o_type_num = np.asarray(0, dtype=self.output_type.dtype).dtype.num
fail = sub['fail'] fail = sub['fail']
if self.output_type.dtype == 'float32': if self.output_type.dtype == 'float32':
otype = 'float' otype = 'float'
NORM = '4.6566126e-10f' # numpy.float32(1.0/(2**31+65)) NORM = '4.6566126e-10f' # np.float32(1.0/(2**31+65))
# this was determined by finding the biggest number such that # this was determined by finding the biggest number such that
# numpy.float32(number * M1) < 1.0 # np.float32(number * M1) < 1.0
else: else:
otype = 'double' otype = 'double'
NORM = '4.656612873077392578125e-10' NORM = '4.656612873077392578125e-10'
...@@ -590,9 +590,9 @@ class GPU_mrg_uniform(mrg_uniform_base, GpuOp): ...@@ -590,9 +590,9 @@ class GPU_mrg_uniform(mrg_uniform_base, GpuOp):
def c_support_code_apply(self, node, nodename): def c_support_code_apply(self, node, nodename):
if self.output_type.dtype == 'float32': if self.output_type.dtype == 'float32':
otype = 'float' otype = 'float'
NORM = '4.6566126e-10f' # numpy.float32(1.0/(2**31+65)) NORM = '4.6566126e-10f' # np.float32(1.0/(2**31+65))
# this was determined by finding the biggest number such that # this was determined by finding the biggest number such that
# numpy.float32(number * M1) < 1.0 # np.float32(number * M1) < 1.0
else: else:
otype = 'double' otype = 'double'
NORM = '4.656612873077392578125e-10' NORM = '4.656612873077392578125e-10'
...@@ -686,7 +686,7 @@ class GPU_mrg_uniform(mrg_uniform_base, GpuOp): ...@@ -686,7 +686,7 @@ class GPU_mrg_uniform(mrg_uniform_base, GpuOp):
o_rstate, o_sample = out o_rstate, o_sample = out
inplace = int(self.inplace) inplace = int(self.inplace)
ndim = self.output_type.ndim ndim = self.output_type.ndim
o_type_num = numpy.asarray(0, dtype=self.output_type.dtype).dtype.num o_type_num = np.asarray(0, dtype=self.output_type.dtype).dtype.num
fail = sub['fail'] fail = sub['fail']
if self.output_type.dtype == 'float32': if self.output_type.dtype == 'float32':
...@@ -858,15 +858,15 @@ class GPUA_mrg_uniform(GpuKernelBase, mrg_uniform_base): ...@@ -858,15 +858,15 @@ class GPUA_mrg_uniform(GpuKernelBase, mrg_uniform_base):
otype = 'ga_half' otype = 'ga_half'
# limit the values of the state that we use. # limit the values of the state that we use.
mask = '& 0x7fff' mask = '& 0x7fff'
NORM = '3.0518e-05f' # numpy.float16(1.0/(2**15+8)) NORM = '3.0518e-05f' # np.float16(1.0/(2**15+8))
# this was determined by finding the biggest number such that # this was determined by finding the biggest number such that
# numpy.float16(number * (M1 & 0x7fff)) < 1.0 # np.float16(number * (M1 & 0x7fff)) < 1.0
elif self.output_type.dtype == 'float32': elif self.output_type.dtype == 'float32':
otype = 'float' otype = 'float'
mask = '' mask = ''
NORM = '4.6566126e-10f' # numpy.float32(1.0/(2**31+65)) NORM = '4.6566126e-10f' # np.float32(1.0/(2**31+65))
# this was determined by finding the biggest number such that # this was determined by finding the biggest number such that
# numpy.float32(number * M1) < 1.0 # np.float32(number * M1) < 1.0
elif self.output_type.dtype == 'float64': elif self.output_type.dtype == 'float64':
otype = 'double' otype = 'double'
mask = '' mask = ''
...@@ -969,7 +969,7 @@ class GPUA_mrg_uniform(GpuKernelBase, mrg_uniform_base): ...@@ -969,7 +969,7 @@ class GPUA_mrg_uniform(GpuKernelBase, mrg_uniform_base):
o_rstate, o_sample = out o_rstate, o_sample = out
inplace = int(self.inplace) inplace = int(self.inplace)
ndim = self.output_type.ndim ndim = self.output_type.ndim
o_type_num = numpy.asarray(0, dtype=self.output_type.dtype).dtype.num o_type_num = np.asarray(0, dtype=self.output_type.dtype).dtype.num
fail = sub['fail'] fail = sub['fail']
ctx = sub['params'] ctx = sub['params']
kname = self.gpu_kernels(node, nodename)[0].objvar kname = self.gpu_kernels(node, nodename)[0].objvar
...@@ -1176,7 +1176,7 @@ class MRG_RandomStreams(object): ...@@ -1176,7 +1176,7 @@ class MRG_RandomStreams(object):
raise ValueError('seed should not be 0', seed) raise ValueError('seed should not be 0', seed)
elif seed >= M2: elif seed >= M2:
raise ValueError('seed should be less than %i' % M2, seed) raise ValueError('seed should be less than %i' % M2, seed)
self.rstate = numpy.asarray([seed] * 6, dtype='int32') self.rstate = np.asarray([seed] * 6, dtype='int32')
elif len(seed) == 6: elif len(seed) == 6:
if seed[0] == 0 and seed[1] == 0 and seed[2] == 0: if seed[0] == 0 and seed[1] == 0 and seed[2] == 0:
raise ValueError( raise ValueError(
...@@ -1192,7 +1192,7 @@ class MRG_RandomStreams(object): ...@@ -1192,7 +1192,7 @@ class MRG_RandomStreams(object):
raise ValueError( raise ValueError(
'The last 3 values of seed should be less than %i' % M2, 'The last 3 values of seed should be less than %i' % M2,
seed) seed)
self.rstate = numpy.asarray(seed, dtype='int32') self.rstate = np.asarray(seed, dtype='int32')
else: else:
raise TypeError("seed should be 1 integer or 6 integers") raise TypeError("seed should be 1 integer or 6 integers")
...@@ -1234,7 +1234,7 @@ class MRG_RandomStreams(object): ...@@ -1234,7 +1234,7 @@ class MRG_RandomStreams(object):
""" """
# self.rstate = ff_2p134(self.rstate) # self.rstate = ff_2p134(self.rstate)
self.rstate = multMatVect(self.rstate, A1p134, M1, A2p134, M2) self.rstate = multMatVect(self.rstate, A1p134, M1, A2p134, M2)
assert self.rstate.dtype == numpy.int32 assert self.rstate.dtype == np.int32
@theano.configparser.change_flags(compute_test_value='off') @theano.configparser.change_flags(compute_test_value='off')
def get_substream_rstates(self, n_streams, dtype, inc_rstate=True): def get_substream_rstates(self, n_streams, dtype, inc_rstate=True):
...@@ -1247,7 +1247,7 @@ class MRG_RandomStreams(object): ...@@ -1247,7 +1247,7 @@ class MRG_RandomStreams(object):
assert isinstance(dtype, str) assert isinstance(dtype, str)
assert n_streams < 2**72 assert n_streams < 2**72
assert n_streams > 0 assert n_streams > 0
rval = numpy.zeros((n_streams, 6), dtype='int32') rval = np.zeros((n_streams, 6), dtype='int32')
rval[0] = self.rstate rval[0] = self.rstate
# If multMatVect.dot_modulo isn't compiled, compile it. # If multMatVect.dot_modulo isn't compiled, compile it.
...@@ -1276,7 +1276,7 @@ class MRG_RandomStreams(object): ...@@ -1276,7 +1276,7 @@ class MRG_RandomStreams(object):
# HACK - we use fact that int32 and float32 have same size to # HACK - we use fact that int32 and float32 have same size to
# sneak ints into the CudaNdarray type. # sneak ints into the CudaNdarray type.
# these *SHOULD NEVER BE USED AS FLOATS* # these *SHOULD NEVER BE USED AS FLOATS*
tmp_float_buf = numpy.frombuffer(rval.data, dtype='float32') tmp_float_buf = np.frombuffer(rval.data, dtype='float32')
assert tmp_float_buf.shape == rval.shape assert tmp_float_buf.shape == rval.shape
assert (tmp_float_buf.view('int32') == rval).all() assert (tmp_float_buf.view('int32') == rval).all()
rval = tmp_float_buf rval = tmp_float_buf
...@@ -1334,9 +1334,9 @@ class MRG_RandomStreams(object): ...@@ -1334,9 +1334,9 @@ class MRG_RandomStreams(object):
if isinstance(size, tuple): if isinstance(size, tuple):
msg = "size must be a tuple of int or a Theano variable" msg = "size must be a tuple of int or a Theano variable"
assert all([isinstance(i, (numpy.integer, integer_types, Variable)) assert all([isinstance(i, (np.integer, integer_types, Variable))
for i in size]), msg for i in size]), msg
if any([isinstance(i, (numpy.integer, integer_types)) and i <= 0 if any([isinstance(i, (np.integer, integer_types)) and i <= 0
for i in size]): for i in size]):
raise ValueError( raise ValueError(
"The specified size contains a dimension with value <= 0", "The specified size contains a dimension with value <= 0",
...@@ -1558,10 +1558,10 @@ class MRG_RandomStreams(object): ...@@ -1558,10 +1558,10 @@ class MRG_RandomStreams(object):
evened = False evened = False
constant = False constant = False
if (isinstance(size, tuple) and if (isinstance(size, tuple) and
all([isinstance(i, (numpy.integer, integer_types)) for i in size])): all([isinstance(i, (np.integer, integer_types)) for i in size])):
constant = True constant = True
# Force dtype because it defaults to float when size is empty # Force dtype because it defaults to float when size is empty
n_samples = numpy.prod(size, dtype='int64') n_samples = np.prod(size, dtype='int64')
if n_samples % 2 == 1: if n_samples % 2 == 1:
n_samples += 1 n_samples += 1
...@@ -1583,14 +1583,14 @@ class MRG_RandomStreams(object): ...@@ -1583,14 +1583,14 @@ class MRG_RandomStreams(object):
sqrt_ln_U1 = sqrt(-2.0 * log(U1)) sqrt_ln_U1 = sqrt(-2.0 * log(U1))
# TypeError: 'TensorVariable' object does not support item assignment # TypeError: 'TensorVariable' object does not support item assignment
# so this doesn't work... # so this doesn't work...
# normal_samples[:n_samples/2] = sqrt_ln_U1 * cos(2.0*numpy.pi*U2) # normal_samples[:n_samples/2] = sqrt_ln_U1 * cos(2.0*np.pi*U2)
# normal_samples[n_samples/2:] = sqrt_ln_U1 * sin(2.0*numpy.pi*U2) # normal_samples[n_samples/2:] = sqrt_ln_U1 * sin(2.0*np.pi*U2)
# so trying this instead # so trying this instead
first_half = sqrt_ln_U1 * cos( first_half = sqrt_ln_U1 * cos(
numpy.array(2.0 * numpy.pi, dtype=dtype) * U2) np.array(2.0 * np.pi, dtype=dtype) * U2)
second_half = sqrt_ln_U1 * sin( second_half = sqrt_ln_U1 * sin(
numpy.array(2.0 * numpy.pi, dtype=dtype) * U2) np.array(2.0 * np.pi, dtype=dtype) * U2)
normal_samples = join(0, first_half, second_half) normal_samples = join(0, first_half, second_half)
final_samples = None final_samples = None
......
theano/sandbox/tests/test_multinomial.py
浏览文件 @ bea31470
...@@ -4,7 +4,7 @@ import sys ...@@ -4,7 +4,7 @@ import sys
from six import reraise from six import reraise
from nose.plugins.skip import SkipTest from nose.plugins.skip import SkipTest
import numpy import numpy as np
import theano import theano
from theano import config, function, tensor from theano import config, function, tensor
...@@ -40,9 +40,9 @@ def test_n_samples_1(): ...@@ -40,9 +40,9 @@ def test_n_samples_1():
f = function([p, u, n], m, allow_input_downcast=True) f = function([p, u, n], m, allow_input_downcast=True)
numpy.random.seed(12345) np.random.seed(12345)
for i in [1, 5, 10, 100, 1000, 10000]: for i in [1, 5, 10, 100, 1000, 10000]:
uni = numpy.random.rand(2 * i).astype(config.floatX) uni = np.random.rand(2 * i).astype(config.floatX)
res = f([[1.0, 0.0], [0.0, 1.0]], uni, i) res = f([[1.0, 0.0], [0.0, 1.0]], uni, i)
utt.assert_allclose(res, [[i * 1.0, 0.0], [0.0, i * 1.0]]) utt.assert_allclose(res, [[i * 1.0, 0.0], [0.0, i * 1.0]])
...@@ -55,17 +55,17 @@ def test_n_samples_2(): ...@@ -55,17 +55,17 @@ def test_n_samples_2():
f = function([p, u, n], m, allow_input_downcast=True) f = function([p, u, n], m, allow_input_downcast=True)
numpy.random.seed(12345) np.random.seed(12345)
for i in [1, 5, 10, 100, 1000]: for i in [1, 5, 10, 100, 1000]:
uni = numpy.random.rand(i).astype(config.floatX) uni = np.random.rand(i).astype(config.floatX)
pvals = numpy.random.randint(1, 1000, (1, 1000)).astype(config.floatX) pvals = np.random.randint(1, 1000, (1, 1000)).astype(config.floatX)
pvals /= pvals.sum(1) pvals /= pvals.sum(1)
res = f(pvals, uni, i) res = f(pvals, uni, i)
assert res.sum() == i assert res.sum() == i
for i in [1, 5, 10, 100, 1000]: for i in [1, 5, 10, 100, 1000]:
uni = numpy.random.rand(i).astype(config.floatX) uni = np.random.rand(i).astype(config.floatX)
pvals = numpy.random.randint( pvals = np.random.randint(
1, 1000000, (1, 1000000)).astype(config.floatX) 1, 1000000, (1, 1000000)).astype(config.floatX)
pvals /= pvals.sum(1) pvals /= pvals.sum(1)
res = f(pvals, uni, i) res = f(pvals, uni, i)
...@@ -104,8 +104,8 @@ def test_n_samples_compatibility(): ...@@ -104,8 +104,8 @@ def test_n_samples_compatibility():
raise raise
f = theano.function([X], samples) f = theano.function([X], samples)
res = f(numpy.random.randn(20, 10)) res = f(np.random.randn(20, 10))
assert numpy.all(res.sum(axis=1) == 1) assert np.all(res.sum(axis=1) == 1)
def test_multinomial_0(): def test_multinomial_0():
...@@ -160,9 +160,9 @@ def test_multinomial_large(): ...@@ -160,9 +160,9 @@ def test_multinomial_large():
assert any([type(node.op) is multinomial.GpuMultinomialFromUniform assert any([type(node.op) is multinomial.GpuMultinomialFromUniform
for node in f.maker.fgraph.toposort()]) for node in f.maker.fgraph.toposort()])
pval = numpy.arange(10000 * 4, dtype='float32').reshape((10000, 4)) + 0.1 pval = np.arange(10000 * 4, dtype='float32').reshape((10000, 4)) + 0.1
pval = pval / pval.sum(axis=1)[:, None] pval = pval / pval.sum(axis=1)[:, None]
uval = numpy.ones_like(pval[:, 0]) * 0.5 uval = np.ones_like(pval[:, 0]) * 0.5
mval = f(pval, uval) mval = f(pval, uval)
assert mval.shape == pval.shape assert mval.shape == pval.shape
...@@ -175,7 +175,7 @@ def test_multinomial_large(): ...@@ -175,7 +175,7 @@ def test_multinomial_large():
else: else:
raise NotImplementedError(config.cast_policy) raise NotImplementedError(config.cast_policy)
utt.assert_allclose(mval.sum(axis=1), 2) utt.assert_allclose(mval.sum(axis=1), 2)
asdf = numpy.asarray([0, 0, 2, 0]) + 0 * pval asdf = np.asarray([0, 0, 2, 0]) + 0 * pval
utt.assert_allclose(mval, asdf) # broadcast over all rows utt.assert_allclose(mval, asdf) # broadcast over all rows
run_with_c(body) run_with_c(body)
if cuda.cuda_available: if cuda.cuda_available:
...@@ -216,9 +216,9 @@ def test_gpu_opt(): ...@@ -216,9 +216,9 @@ def test_gpu_opt():
f = function([p, u], m_gpu, allow_input_downcast=True, mode=get_mode(True)) f = function([p, u], m_gpu, allow_input_downcast=True, mode=get_mode(True))
assert any([type(node.op) is multinomial.GpuMultinomialFromUniform assert any([type(node.op) is multinomial.GpuMultinomialFromUniform
for node in f.maker.fgraph.toposort()]) for node in f.maker.fgraph.toposort()])
pval = numpy.arange(10000 * 4, dtype='float32').reshape((10000, 4)) + 0.1 pval = np.arange(10000 * 4, dtype='float32').reshape((10000, 4)) + 0.1
pval = pval / pval.sum(axis=1)[:, None] pval = pval / pval.sum(axis=1)[:, None]
uval = numpy.ones_like(pval[:, 0]) * 0.5 uval = np.ones_like(pval[:, 0]) * 0.5
f(pval, uval) f(pval, uval)
# Test with a row, it was failing in the past. # Test with a row, it was failing in the past.
...@@ -230,7 +230,7 @@ def test_gpu_opt(): ...@@ -230,7 +230,7 @@ def test_gpu_opt():
f = function([r, u], m_gpu, allow_input_downcast=True, mode=get_mode(True)) f = function([r, u], m_gpu, allow_input_downcast=True, mode=get_mode(True))
assert any([type(node.op) is multinomial.GpuMultinomialFromUniform assert any([type(node.op) is multinomial.GpuMultinomialFromUniform
for node in f.maker.fgraph.toposort()]) for node in f.maker.fgraph.toposort()])
pval = numpy.arange(1 * 4, dtype='float32').reshape((1, 4)) + 0.1 pval = np.arange(1 * 4, dtype='float32').reshape((1, 4)) + 0.1
pval = pval / pval.sum(axis=1)[:, None] pval = pval / pval.sum(axis=1)[:, None]
uval = numpy.ones_like(pval[:, 0]) * 0.5 uval = np.ones_like(pval[:, 0]) * 0.5
f(pval, uval) f(pval, uval)
theano/sandbox/tests/test_multinomial_wo_replacement.py
浏览文件 @ bea31470
from __future__ import absolute_import, print_function, division from __future__ import absolute_import, print_function, division
import numpy import numpy as np
import os import os
from theano import config, function, tensor from theano import config, function, tensor
from theano.compat import PY3 from theano.compat import PY3
...@@ -24,25 +24,25 @@ class test_OP(unittest.TestCase): ...@@ -24,25 +24,25 @@ class test_OP(unittest.TestCase):
n_elements = 1000 n_elements = 1000
all_indices = range(n_elements) all_indices = range(n_elements)
numpy.random.seed(12345) np.random.seed(12345)
expected = [ expected = [
numpy.asarray([[931, 318, 185, 209, 559]]), np.asarray([[931, 318, 185, 209, 559]]),
numpy.asarray([[477, 887, 2, 717, 333, 665, 159, 559, 348, 136]]), np.asarray([[477, 887, 2, 717, 333, 665, 159, 559, 348, 136]]),
numpy.asarray([[546, 28, 79, 665, 295, 779, 433, 531, 411, 716, 244, 234, 70, 88, 612, 639, 383, 335, np.asarray([[546, 28, 79, 665, 295, 779, 433, 531, 411, 716, 244, 234, 70, 88, 612, 639, 383, 335,
451, 100, 175, 492, 848, 771, 559, 214, 568, 596, 370, 486, 855, 925, 138, 300, 528, 507, 451, 100, 175, 492, 848, 771, 559, 214, 568, 596, 370, 486, 855, 925, 138, 300, 528, 507,
730, 199, 882, 357, 58, 195, 705, 900, 66, 468, 513, 410, 816, 672]])] 730, 199, 882, 357, 58, 195, 705, 900, 66, 468, 513, 410, 816, 672]])]
for i in [5, 10, 50, 100, 500, n_elements]: for i in [5, 10, 50, 100, 500, n_elements]:
uni = numpy.random.rand(i).astype(config.floatX) uni = np.random.rand(i).astype(config.floatX)
pvals = numpy.random.randint(1, 100, (1, n_elements)).astype(config.floatX) pvals = np.random.randint(1, 100, (1, n_elements)).astype(config.floatX)
pvals /= pvals.sum(1) pvals /= pvals.sum(1)
res = f(pvals, uni, i) res = f(pvals, uni, i)
for ii in range(len(expected)): for ii in range(len(expected)):
if expected[ii].shape == res.shape: if expected[ii].shape == res.shape:
assert (expected[ii] == res).all() assert (expected[ii] == res).all()
res = numpy.squeeze(res) res = np.squeeze(res)
assert len(res) == i assert len(res) == i
assert numpy.all(numpy.in1d(numpy.unique(res), all_indices)), res assert np.all(np.in1d(np.unique(res), all_indices)), res
def test_fail_select_alot(self): def test_fail_select_alot(self):
""" """
...@@ -58,9 +58,9 @@ class test_OP(unittest.TestCase): ...@@ -58,9 +58,9 @@ class test_OP(unittest.TestCase):
n_elements = 100 n_elements = 100
n_selected = 200 n_selected = 200
numpy.random.seed(12345) np.random.seed(12345)
uni = numpy.random.rand(n_selected).astype(config.floatX) uni = np.random.rand(n_selected).astype(config.floatX)
pvals = numpy.random.randint(1, 100, (1, n_elements)).astype(config.floatX) pvals = np.random.randint(1, 100, (1, n_elements)).astype(config.floatX)
pvals /= pvals.sum(1) pvals /= pvals.sum(1)
self.assertRaises(ValueError, f, pvals, uni, n_selected) self.assertRaises(ValueError, f, pvals, uni, n_selected)
...@@ -79,18 +79,18 @@ class test_OP(unittest.TestCase): ...@@ -79,18 +79,18 @@ class test_OP(unittest.TestCase):
n_elements = 100 n_elements = 100
n_selected = 10 n_selected = 10
mean_rtol = 0.0005 mean_rtol = 0.0005
numpy.random.seed(12345) np.random.seed(12345)
pvals = numpy.random.randint(1, 100, (1, n_elements)).astype(config.floatX) pvals = np.random.randint(1, 100, (1, n_elements)).astype(config.floatX)
pvals /= pvals.sum(1) pvals /= pvals.sum(1)
avg_pvals = numpy.zeros((n_elements,), dtype=config.floatX) avg_pvals = np.zeros((n_elements,), dtype=config.floatX)
for rep in range(10000): for rep in range(10000):
uni = numpy.random.rand(n_selected).astype(config.floatX) uni = np.random.rand(n_selected).astype(config.floatX)
res = f(pvals, uni, n_selected) res = f(pvals, uni, n_selected)
res = numpy.squeeze(res) res = np.squeeze(res)
avg_pvals[res] += 1 avg_pvals[res] += 1
avg_pvals /= avg_pvals.sum() avg_pvals /= avg_pvals.sum()
avg_diff = numpy.mean(abs(avg_pvals - pvals)) avg_diff = np.mean(abs(avg_pvals - pvals))
assert avg_diff < mean_rtol, avg_diff assert avg_diff < mean_rtol, avg_diff
...@@ -110,14 +110,14 @@ class test_function(unittest.TestCase): ...@@ -110,14 +110,14 @@ class test_function(unittest.TestCase):
n_elements = 1000 n_elements = 1000
all_indices = range(n_elements) all_indices = range(n_elements)
numpy.random.seed(12345) np.random.seed(12345)
for i in [5, 10, 50, 100, 500, n_elements]: for i in [5, 10, 50, 100, 500, n_elements]:
pvals = numpy.random.randint(1, 100, (1, n_elements)).astype(config.floatX) pvals = np.random.randint(1, 100, (1, n_elements)).astype(config.floatX)
pvals /= pvals.sum(1) pvals /= pvals.sum(1)
res = f(pvals, i) res = f(pvals, i)
res = numpy.squeeze(res) res = np.squeeze(res)
assert len(res) == i assert len(res) == i
assert numpy.all(numpy.in1d(numpy.unique(res), all_indices)), res assert np.all(np.in1d(np.unique(res), all_indices)), res
def test_fail_select_alot(self): def test_fail_select_alot(self):
""" """
...@@ -134,8 +134,8 @@ class test_function(unittest.TestCase): ...@@ -134,8 +134,8 @@ class test_function(unittest.TestCase):
n_elements = 100 n_elements = 100
n_selected = 200 n_selected = 200
numpy.random.seed(12345) np.random.seed(12345)
pvals = numpy.random.randint(1, 100, (1, n_elements)).astype(config.floatX) pvals = np.random.randint(1, 100, (1, n_elements)).astype(config.floatX)
pvals /= pvals.sum(1) pvals /= pvals.sum(1)
self.assertRaises(ValueError, f, pvals, n_selected) self.assertRaises(ValueError, f, pvals, n_selected)
...@@ -155,17 +155,17 @@ class test_function(unittest.TestCase): ...@@ -155,17 +155,17 @@ class test_function(unittest.TestCase):
n_elements = 100 n_elements = 100
n_selected = 10 n_selected = 10
mean_rtol = 0.0005 mean_rtol = 0.0005
numpy.random.seed(12345) np.random.seed(12345)
pvals = numpy.random.randint(1, 100, (1, n_elements)).astype(config.floatX) pvals = np.random.randint(1, 100, (1, n_elements)).astype(config.floatX)
pvals /= pvals.sum(1) pvals /= pvals.sum(1)
avg_pvals = numpy.zeros((n_elements,), dtype=config.floatX) avg_pvals = np.zeros((n_elements,), dtype=config.floatX)
for rep in range(10000): for rep in range(10000):
res = f(pvals, n_selected) res = f(pvals, n_selected)
res = numpy.squeeze(res) res = np.squeeze(res)
avg_pvals[res] += 1 avg_pvals[res] += 1
avg_pvals /= avg_pvals.sum() avg_pvals /= avg_pvals.sum()
avg_diff = numpy.mean(abs(avg_pvals - pvals)) avg_diff = np.mean(abs(avg_pvals - pvals))
assert avg_diff < mean_rtol assert avg_diff < mean_rtol
def test_unpickle_legacy_op(self): def test_unpickle_legacy_op(self):
......
theano/sandbox/tests/test_rng_mrg.py
浏览文件 @ bea31470
...@@ -8,7 +8,7 @@ import functools ...@@ -8,7 +8,7 @@ import functools
from nose.plugins.skip import SkipTest from nose.plugins.skip import SkipTest
from nose.tools import assert_raises from nose.tools import assert_raises
import numpy import numpy as np
from six.moves import xrange from six.moves import xrange
import theano import theano
...@@ -44,9 +44,9 @@ utt.seed_rng() ...@@ -44,9 +44,9 @@ utt.seed_rng()
# 12 streams # 12 streams
# 7 substreams for each stream # 7 substreams for each stream
# 5 samples drawn from each substream # 5 samples drawn from each substream
java_samples = numpy.loadtxt(os.path.join(os.path.split(theano.__file__)[0], java_samples = np.loadtxt(os.path.join(os.path.split(theano.__file__)[0],
'sandbox', 'sandbox',
'samples_MRG31k3p_12_7_5.txt')) 'samples_MRG31k3p_12_7_5.txt'))
def test_deterministic(): def test_deterministic():
...@@ -65,15 +65,15 @@ def test_deterministic(): ...@@ -65,15 +65,15 @@ def test_deterministic():
fsample1 = f() fsample1 = f()
fsample2 = f() fsample2 = f()
assert not numpy.allclose(fsample1, fsample2) assert not np.allclose(fsample1, fsample2)
R2 = MRG_RandomStreams(seed=seed, use_cuda=use_cuda) R2 = MRG_RandomStreams(seed=seed, use_cuda=use_cuda)
u2 = R2.uniform(size=sample_size) u2 = R2.uniform(size=sample_size)
g = theano.function([], u2) g = theano.function([], u2)
gsample1 = g() gsample1 = g()
gsample2 = g() gsample2 = g()
assert numpy.allclose(fsample1, gsample1) assert np.allclose(fsample1, gsample1)
assert numpy.allclose(fsample2, gsample2) assert np.allclose(fsample2, gsample2)
def test_consistency_randomstreams(): def test_consistency_randomstreams():
...@@ -102,12 +102,12 @@ def test_consistency_randomstreams(): ...@@ -102,12 +102,12 @@ def test_consistency_randomstreams():
for j in range(n_samples): for j in range(n_samples):
s = f() s = f()
stream_samples.append(s) stream_samples.append(s)
stream_samples = numpy.array(stream_samples) stream_samples = np.array(stream_samples)
stream_samples = stream_samples.T.flatten() stream_samples = stream_samples.T.flatten()
samples.append(stream_samples) samples.append(stream_samples)
samples = numpy.array(samples).flatten() samples = np.array(samples).flatten()
assert(numpy.allclose(samples, java_samples)) assert(np.allclose(samples, java_samples))
def test_get_substream_rstates(): def test_get_substream_rstates():
...@@ -117,7 +117,7 @@ def test_get_substream_rstates(): ...@@ -117,7 +117,7 @@ def test_get_substream_rstates():
n_streams = 100 n_streams = 100
dtype = 'float32' dtype = 'float32'
rng = MRG_RandomStreams(numpy.random.randint(2147462579)) rng = MRG_RandomStreams(np.random.randint(2147462579))
rng.get_substream_rstates(n_streams, dtype) rng.get_substream_rstates(n_streams, dtype)
...@@ -137,13 +137,13 @@ def test_consistency_cpu_serial(): ...@@ -137,13 +137,13 @@ def test_consistency_cpu_serial():
n_substreams = 7 n_substreams = 7
samples = [] samples = []
curr_rstate = numpy.array([seed] * 6, dtype='int32') curr_rstate = np.array([seed] * 6, dtype='int32')
for i in range(n_streams): for i in range(n_streams):
stream_rstate = curr_rstate.copy() stream_rstate = curr_rstate.copy()
for j in range(n_substreams): for j in range(n_substreams):
rstate = theano.shared(numpy.array([stream_rstate.copy()], rstate = theano.shared(np.array([stream_rstate.copy()],
dtype='int32')) dtype='int32'))
new_rstate, sample = rng_mrg.mrg_uniform.new(rstate, ndim=None, new_rstate, sample = rng_mrg.mrg_uniform.new(rstate, ndim=None,
dtype=config.floatX, dtype=config.floatX,
size=(1,)) size=(1,))
...@@ -164,8 +164,8 @@ def test_consistency_cpu_serial(): ...@@ -164,8 +164,8 @@ def test_consistency_cpu_serial():
# next stream # next stream
curr_rstate = rng_mrg.ff_2p134(curr_rstate) curr_rstate = rng_mrg.ff_2p134(curr_rstate)
samples = numpy.array(samples).flatten() samples = np.array(samples).flatten()
assert(numpy.allclose(samples, java_samples)) assert(np.allclose(samples, java_samples))
def test_consistency_cpu_parallel(): def test_consistency_cpu_parallel():
...@@ -180,14 +180,14 @@ def test_consistency_cpu_parallel(): ...@@ -180,14 +180,14 @@ def test_consistency_cpu_parallel():
n_substreams = 7 # 7 samples will be drawn in parallel n_substreams = 7 # 7 samples will be drawn in parallel
samples = [] samples = []
curr_rstate = numpy.array([seed] * 6, dtype='int32') curr_rstate = np.array([seed] * 6, dtype='int32')
for i in range(n_streams): for i in range(n_streams):
stream_samples = [] stream_samples = []
rstate = [curr_rstate.copy()] rstate = [curr_rstate.copy()]
for j in range(1, n_substreams): for j in range(1, n_substreams):
rstate.append(rng_mrg.ff_2p72(rstate[-1])) rstate.append(rng_mrg.ff_2p72(rstate[-1]))
rstate = numpy.asarray(rstate) rstate = np.asarray(rstate)
rstate = theano.shared(rstate) rstate = theano.shared(rstate)
new_rstate, sample = rng_mrg.mrg_uniform.new(rstate, ndim=None, new_rstate, sample = rng_mrg.mrg_uniform.new(rstate, ndim=None,
...@@ -205,13 +205,13 @@ def test_consistency_cpu_parallel(): ...@@ -205,13 +205,13 @@ def test_consistency_cpu_parallel():
s = f() s = f()
stream_samples.append(s) stream_samples.append(s)
samples.append(numpy.array(stream_samples).T.flatten()) samples.append(np.array(stream_samples).T.flatten())
# next stream # next stream
curr_rstate = rng_mrg.ff_2p134(curr_rstate) curr_rstate = rng_mrg.ff_2p134(curr_rstate)
samples = numpy.array(samples).flatten() samples = np.array(samples).flatten()
assert(numpy.allclose(samples, java_samples)) assert(np.allclose(samples, java_samples))
def test_consistency_GPU_serial(): def test_consistency_GPU_serial():
...@@ -233,16 +233,16 @@ def test_consistency_GPU_serial(): ...@@ -233,16 +233,16 @@ def test_consistency_GPU_serial():
n_substreams = 7 n_substreams = 7
samples = [] samples = []
curr_rstate = numpy.array([seed] * 6, dtype='int32') curr_rstate = np.array([seed] * 6, dtype='int32')
for i in range(n_streams): for i in range(n_streams):
stream_rstate = curr_rstate.copy() stream_rstate = curr_rstate.copy()
for j in range(n_substreams): for j in range(n_substreams):
substream_rstate = numpy.array(stream_rstate.copy(), dtype='int32') substream_rstate = np.array(stream_rstate.copy(), dtype='int32')
# HACK - we transfer these int32 to the GPU memory as float32 # HACK - we transfer these int32 to the GPU memory as float32
# (reinterpret_cast) # (reinterpret_cast)
tmp_float_buf = numpy.frombuffer(substream_rstate.data, tmp_float_buf = np.frombuffer(substream_rstate.data,
dtype='float32') dtype='float32')
# Transfer to device # Transfer to device
rstate = float32_shared_constructor(tmp_float_buf) rstate = float32_shared_constructor(tmp_float_buf)
...@@ -269,8 +269,8 @@ def test_consistency_GPU_serial(): ...@@ -269,8 +269,8 @@ def test_consistency_GPU_serial():
# next stream # next stream
curr_rstate = rng_mrg.ff_2p134(curr_rstate) curr_rstate = rng_mrg.ff_2p134(curr_rstate)
samples = numpy.array(samples).flatten() samples = np.array(samples).flatten()
assert(numpy.allclose(samples, java_samples)) assert(np.allclose(samples, java_samples))
def test_consistency_GPU_parallel(): def test_consistency_GPU_parallel():
...@@ -293,17 +293,17 @@ def test_consistency_GPU_parallel(): ...@@ -293,17 +293,17 @@ def test_consistency_GPU_parallel():
n_substreams = 7 # 7 samples will be drawn in parallel n_substreams = 7 # 7 samples will be drawn in parallel
samples = [] samples = []
curr_rstate = numpy.array([seed] * 6, dtype='int32') curr_rstate = np.array([seed] * 6, dtype='int32')
for i in range(n_streams): for i in range(n_streams):
stream_samples = [] stream_samples = []
rstate = [curr_rstate.copy()] rstate = [curr_rstate.copy()]
for j in range(1, n_substreams): for j in range(1, n_substreams):
rstate.append(rng_mrg.ff_2p72(rstate[-1])) rstate.append(rng_mrg.ff_2p72(rstate[-1]))
rstate = numpy.asarray(rstate).flatten() rstate = np.asarray(rstate).flatten()
# HACK - transfer these int32 to the GPU memory as float32 # HACK - transfer these int32 to the GPU memory as float32
# (reinterpret_cast) # (reinterpret_cast)
tmp_float_buf = numpy.frombuffer(rstate.data, dtype='float32') tmp_float_buf = np.frombuffer(rstate.data, dtype='float32')
# Transfer to device # Transfer to device
rstate = float32_shared_constructor(tmp_float_buf) rstate = float32_shared_constructor(tmp_float_buf)
...@@ -325,13 +325,13 @@ def test_consistency_GPU_parallel(): ...@@ -325,13 +325,13 @@ def test_consistency_GPU_parallel():
s = f() s = f()
stream_samples.append(s) stream_samples.append(s)
samples.append(numpy.array(stream_samples).T.flatten()) samples.append(np.array(stream_samples).T.flatten())
# next stream # next stream
curr_rstate = rng_mrg.ff_2p134(curr_rstate) curr_rstate = rng_mrg.ff_2p134(curr_rstate)
samples = numpy.array(samples).flatten() samples = np.array(samples).flatten()
assert(numpy.allclose(samples, java_samples)) assert(np.allclose(samples, java_samples))
def test_GPU_nstreams_limit(): def test_GPU_nstreams_limit():
...@@ -373,13 +373,13 @@ def test_consistency_GPUA_serial(): ...@@ -373,13 +373,13 @@ def test_consistency_GPUA_serial():
n_substreams = 7 n_substreams = 7
samples = [] samples = []
curr_rstate = numpy.array([seed] * 6, dtype='int32') curr_rstate = np.array([seed] * 6, dtype='int32')
for i in range(n_streams): for i in range(n_streams):
stream_rstate = curr_rstate.copy() stream_rstate = curr_rstate.copy()
for j in range(n_substreams): for j in range(n_substreams):
substream_rstate = numpy.array([stream_rstate.copy()], substream_rstate = np.array([stream_rstate.copy()],
dtype='int32') dtype='int32')
# Transfer to device # Transfer to device
rstate = gpuarray_shared_constructor(substream_rstate) rstate = gpuarray_shared_constructor(substream_rstate)
...@@ -407,8 +407,8 @@ def test_consistency_GPUA_serial(): ...@@ -407,8 +407,8 @@ def test_consistency_GPUA_serial():
# next stream # next stream
curr_rstate = rng_mrg.ff_2p134(curr_rstate) curr_rstate = rng_mrg.ff_2p134(curr_rstate)
samples = numpy.array(samples).flatten() samples = np.array(samples).flatten()
assert(numpy.allclose(samples, java_samples)) assert(np.allclose(samples, java_samples))
def test_consistency_GPUA_parallel(): def test_consistency_GPUA_parallel():
...@@ -424,14 +424,14 @@ def test_consistency_GPUA_parallel(): ...@@ -424,14 +424,14 @@ def test_consistency_GPUA_parallel():
n_substreams = 7 # 7 samples will be drawn in parallel n_substreams = 7 # 7 samples will be drawn in parallel
samples = [] samples = []
curr_rstate = numpy.array([seed] * 6, dtype='int32') curr_rstate = np.array([seed] * 6, dtype='int32')
for i in range(n_streams): for i in range(n_streams):
stream_samples = [] stream_samples = []
rstate = [curr_rstate.copy()] rstate = [curr_rstate.copy()]
for j in range(1, n_substreams): for j in range(1, n_substreams):
rstate.append(rng_mrg.ff_2p72(rstate[-1])) rstate.append(rng_mrg.ff_2p72(rstate[-1]))
rstate = numpy.asarray(rstate) rstate = np.asarray(rstate)
rstate = gpuarray_shared_constructor(rstate) rstate = gpuarray_shared_constructor(rstate)
new_rstate, sample = rng_mrg.GPUA_mrg_uniform.new(rstate, ndim=None, new_rstate, sample = rng_mrg.GPUA_mrg_uniform.new(rstate, ndim=None,
...@@ -452,13 +452,13 @@ def test_consistency_GPUA_parallel(): ...@@ -452,13 +452,13 @@ def test_consistency_GPUA_parallel():
s = f() s = f()
stream_samples.append(s) stream_samples.append(s)
samples.append(numpy.array(stream_samples).T.flatten()) samples.append(np.array(stream_samples).T.flatten())
# next stream # next stream
curr_rstate = rng_mrg.ff_2p134(curr_rstate) curr_rstate = rng_mrg.ff_2p134(curr_rstate)
samples = numpy.array(samples).flatten() samples = np.array(samples).flatten()
assert(numpy.allclose(samples, java_samples)) assert(np.allclose(samples, java_samples))
def test_GPUA_full_fill(): def test_GPUA_full_fill():
...@@ -496,16 +496,16 @@ def basictest(f, steps, sample_size, prefix="", allow_01=False, inputs=None, ...@@ -496,16 +496,16 @@ def basictest(f, steps, sample_size, prefix="", allow_01=False, inputs=None,
ival = f(*inputs) ival = f(*inputs)
assert ival.shape == sample_size assert ival.shape == sample_size
dt += time.time() - t0 dt += time.time() - t0
ival = numpy.asarray(ival) ival = np.asarray(ival)
if i == 0: if i == 0:
mean = numpy.array(ival, copy=True) mean = np.array(ival, copy=True)
avg_var = numpy.mean((ival - target_avg) ** 2) avg_var = np.mean((ival - target_avg) ** 2)
min_ = ival.min() min_ = ival.min()
max_ = ival.max() max_ = ival.max()
else: else:
alpha = 1.0 / (1 + i) alpha = 1.0 / (1 + i)
mean = alpha * ival + (1 - alpha) * mean mean = alpha * ival + (1 - alpha) * mean
avg_var = (alpha * numpy.mean((ival - target_avg) ** 2) + avg_var = (alpha * np.mean((ival - target_avg) ** 2) +
(1 - alpha) * avg_var) (1 - alpha) * avg_var)
min_ = min(min_, ival.min()) min_ = min(min_, ival.min())
max_ = max(max_, ival.max()) max_ = max(max_, ival.max())
...@@ -514,19 +514,19 @@ def basictest(f, steps, sample_size, prefix="", allow_01=False, inputs=None, ...@@ -514,19 +514,19 @@ def basictest(f, steps, sample_size, prefix="", allow_01=False, inputs=None,
assert max_ < 1 assert max_ < 1
if hasattr(target_avg, 'shape'): # looks if target_avg is an array if hasattr(target_avg, 'shape'): # looks if target_avg is an array
diff = numpy.mean(abs(mean - target_avg)) diff = np.mean(abs(mean - target_avg))
# print prefix, 'mean diff with mean', diff # print prefix, 'mean diff with mean', diff
assert numpy.all(diff < mean_rtol * (1 + abs(target_avg))), ( assert np.all(diff < mean_rtol * (1 + abs(target_avg))), (
'bad mean? %s %s' % (mean, target_avg)) 'bad mean? %s %s' % (mean, target_avg))
else: else:
# if target_avg is a scalar, then we can do the mean of # if target_avg is a scalar, then we can do the mean of
# `mean` to get something more precise # `mean` to get something more precise
mean = numpy.mean(mean) mean = np.mean(mean)
# print prefix, 'mean', mean # print prefix, 'mean', mean
assert abs(mean - target_avg) < mean_rtol * (1 + abs(target_avg)), ( assert abs(mean - target_avg) < mean_rtol * (1 + abs(target_avg)), (
'bad mean? %f %f' % (mean, target_avg)) 'bad mean? %f %f' % (mean, target_avg))
std = numpy.sqrt(avg_var) std = np.sqrt(avg_var)
# print prefix, 'var', avg_var # print prefix, 'var', avg_var
# print prefix, 'std', std # print prefix, 'std', std
if target_std is not None: if target_std is not None:
...@@ -556,9 +556,9 @@ def test_uniform(): ...@@ -556,9 +556,9 @@ def test_uniform():
for size, const_size, var_input, input in [ for size, const_size, var_input, input in [
(sample_size, sample_size, [], []), (sample_size, sample_size, [], []),
(x.shape, sample_size, [x], (x.shape, sample_size, [x],
[numpy.zeros(sample_size, dtype=config.floatX)]), [np.zeros(sample_size, dtype=config.floatX)]),
((x.shape[0], sample_size[1]), sample_size, [x], ((x.shape[0], sample_size[1]), sample_size, [x],
[numpy.zeros(sample_size, dtype=config.floatX)]), [np.zeros(sample_size, dtype=config.floatX)]),
# test empty size (scalar) # test empty size (scalar)
((), (), [], []), ((), (), [], []),
]: ]:
...@@ -586,7 +586,7 @@ def test_uniform(): ...@@ -586,7 +586,7 @@ def test_uniform():
# print cpu_out[-1, -10:] # print cpu_out[-1, -10:]
# Increase the number of steps if sizes implies only a few samples # Increase the number of steps if sizes implies only a few samples
if numpy.prod(const_size) < 10: if np.prod(const_size) < 10:
steps_ = steps * 100 steps_ = steps * 100
else: else:
steps_ = steps steps_ = steps
...@@ -607,15 +607,15 @@ def test_uniform(): ...@@ -607,15 +607,15 @@ def test_uniform():
theano.sandbox.rng_mrg.GPU_mrg_uniform) theano.sandbox.rng_mrg.GPU_mrg_uniform)
for node in f.maker.fgraph.toposort()]) for node in f.maker.fgraph.toposort()])
# theano.printing.debugprint(f) # theano.printing.debugprint(f)
gpu_out = numpy.asarray(f(*input)) gpu_out = np.asarray(f(*input))
# print 'GPU: random?[:10], random?[-10:]' # print 'GPU: random?[:10], random?[-10:]'
# print gpu_out[0, 0:10] # print gpu_out[0, 0:10]
# print gpu_out[-1, -10:] # print gpu_out[-1, -10:]
basictest(f, steps_, const_size, prefix='mrg gpu', inputs=input) basictest(f, steps_, const_size, prefix='mrg gpu', inputs=input)
numpy.testing.assert_array_almost_equal(cpu_out, gpu_out, np.testing.assert_array_almost_equal(cpu_out, gpu_out,
decimal=6) decimal=6)
# print '' # print ''
# print 'ON CPU w Numpy with size=(%s):' % str(size) # print 'ON CPU w Numpy with size=(%s):' % str(size)
...@@ -633,7 +633,7 @@ def test_broadcastable(): ...@@ -633,7 +633,7 @@ def test_broadcastable():
x = tensor.matrix() x = tensor.matrix()
size1 = (10, 1) size1 = (10, 1)
size2 = (x.shape[0], 1) size2 = (x.shape[0], 1)
pvals_1 = numpy.random.uniform(0, 1, size=size1) pvals_1 = np.random.uniform(0, 1, size=size1)
pvals_1 = pvals_1 / sum(pvals_1) pvals_1 = pvals_1 / sum(pvals_1)
pvals_2 = R.uniform(size=size2) pvals_2 = R.uniform(size=size2)
pvals_2 = pvals_2 / tensor.sum(pvals_2) pvals_2 = pvals_2 / tensor.sum(pvals_2)
...@@ -684,9 +684,9 @@ def test_binomial(): ...@@ -684,9 +684,9 @@ def test_binomial():
for size, const_size, var_input, input in [ for size, const_size, var_input, input in [
(sample_size, sample_size, [], []), (sample_size, sample_size, [], []),
(x.shape, sample_size, [x], (x.shape, sample_size, [x],
[numpy.zeros(sample_size, dtype=config.floatX)]), [np.zeros(sample_size, dtype=config.floatX)]),
((x.shape[0], sample_size[1]), sample_size, [x], ((x.shape[0], sample_size[1]), sample_size, [x],
[numpy.zeros(sample_size, dtype=config.floatX)]), [np.zeros(sample_size, dtype=config.floatX)]),
# test empty size (scalar) # test empty size (scalar)
((), (), [], []), ((), (), [], []),
]: ]:
...@@ -701,7 +701,7 @@ def t_binomial(mean, size, const_size, var_input, input, steps, rtol): ...@@ -701,7 +701,7 @@ def t_binomial(mean, size, const_size, var_input, input, steps, rtol):
out = f(*input) out = f(*input)
# Increase the number of steps if sizes implies only a few samples # Increase the number of steps if sizes implies only a few samples
if numpy.prod(const_size) < 10: if np.prod(const_size) < 10:
steps_ = steps * 100 steps_ = steps * 100
else: else:
steps_ = steps steps_ = steps
...@@ -717,13 +717,13 @@ def t_binomial(mean, size, const_size, var_input, input, steps, rtol): ...@@ -717,13 +717,13 @@ def t_binomial(mean, size, const_size, var_input, input, steps, rtol):
f = theano.function(var_input, theano.Out( f = theano.function(var_input, theano.Out(
theano.sandbox.cuda.basic_ops.gpu_from_host(u), theano.sandbox.cuda.basic_ops.gpu_from_host(u),
borrow=True), mode=mode_with_gpu) borrow=True), mode=mode_with_gpu)
gpu_out = numpy.asarray(f(*input)) gpu_out = np.asarray(f(*input))
basictest(f, steps_, const_size, prefix='mrg gpu', basictest(f, steps_, const_size, prefix='mrg gpu',
inputs=input, allow_01=True, inputs=input, allow_01=True,
target_avg=mean, mean_rtol=rtol) target_avg=mean, mean_rtol=rtol)
numpy.testing.assert_array_almost_equal(out, gpu_out, np.testing.assert_array_almost_equal(out, gpu_out,
decimal=6) decimal=6)
RR = theano.tensor.shared_randomstreams.RandomStreams(234) RR = theano.tensor.shared_randomstreams.RandomStreams(234)
...@@ -752,22 +752,22 @@ def test_normal0(): ...@@ -752,22 +752,22 @@ def test_normal0():
for size, const_size, var_input, input, avg, rtol, std_tol in [ for size, const_size, var_input, input, avg, rtol, std_tol in [
(sample_size, sample_size, [], [], -5., default_rtol, default_rtol), (sample_size, sample_size, [], [], -5., default_rtol, default_rtol),
(x.shape, sample_size, [x], (x.shape, sample_size, [x],
[numpy.zeros(sample_size, dtype=config.floatX)], [np.zeros(sample_size, dtype=config.floatX)],
-5., default_rtol, default_rtol), -5., default_rtol, default_rtol),
((x.shape[0], sample_size[1]), sample_size, [x], ((x.shape[0], sample_size[1]), sample_size, [x],
[numpy.zeros(sample_size, dtype=config.floatX)], [np.zeros(sample_size, dtype=config.floatX)],
-5., default_rtol, default_rtol), -5., default_rtol, default_rtol),
# test odd value # test odd value
(sample_size_odd, sample_size_odd, [], [], -5., (sample_size_odd, sample_size_odd, [], [], -5.,
default_rtol, default_rtol), default_rtol, default_rtol),
# test odd value # test odd value
(x.shape, sample_size_odd, [x], (x.shape, sample_size_odd, [x],
[numpy.zeros(sample_size_odd, dtype=config.floatX)], [np.zeros(sample_size_odd, dtype=config.floatX)],
-5., default_rtol, default_rtol), -5., default_rtol, default_rtol),
(sample_size, sample_size, [], [], (sample_size, sample_size, [], [],
numpy.arange(numpy.prod(sample_size), np.arange(np.prod(sample_size),
dtype='float32').reshape(sample_size), dtype='float32').reshape(sample_size),
10. * std / numpy.sqrt(steps), default_rtol), 10. * std / np.sqrt(steps), default_rtol),
# test empty size (scalar) # test empty size (scalar)
((), (), [], [], -5., default_rtol, 0.02), ((), (), [], [], -5., default_rtol, 0.02),
# test with few samples at the same time # test with few samples at the same time
...@@ -788,7 +788,7 @@ def test_normal0(): ...@@ -788,7 +788,7 @@ def test_normal0():
# print 'random?[:10]\n', out[0, 0:10] # print 'random?[:10]\n', out[0, 0:10]
# Increase the number of steps if size implies only a few samples # Increase the number of steps if size implies only a few samples
if numpy.prod(const_size) < 10: if np.prod(const_size) < 10:
steps_ = steps * 50 steps_ = steps * 50
else: else:
steps_ = steps steps_ = steps
...@@ -812,7 +812,7 @@ def test_normal0(): ...@@ -812,7 +812,7 @@ def test_normal0():
# theano.printing.debugprint(f) # theano.printing.debugprint(f)
sys.stdout.flush() sys.stdout.flush()
gpu_out = numpy.asarray(f(*input)) gpu_out = np.asarray(f(*input))
# print 'random?[:10]\n', gpu_out[0, 0:10] # print 'random?[:10]\n', gpu_out[0, 0:10]
# print '----' # print '----'
sys.stdout.flush() sys.stdout.flush()
...@@ -821,7 +821,7 @@ def test_normal0(): ...@@ -821,7 +821,7 @@ def test_normal0():
mean_rtol=rtol, std_tol=std_tol) mean_rtol=rtol, std_tol=std_tol)
# Need to allow some rounding error as their is float # Need to allow some rounding error as their is float
# computation that are done on the gpu vs cpu # computation that are done on the gpu vs cpu
assert numpy.allclose(out, gpu_out, rtol=5e-6, atol=5e-6) assert np.allclose(out, gpu_out, rtol=5e-6, atol=5e-6)
# print '' # print ''
# print 'ON CPU w NUMPY:' # print 'ON CPU w NUMPY:'
...@@ -838,26 +838,26 @@ def basic_multinomialtest(f, steps, sample_size, target_pvals, n_samples, ...@@ -838,26 +838,26 @@ def basic_multinomialtest(f, steps, sample_size, target_pvals, n_samples,
prefix="", mean_rtol=0.04): prefix="", mean_rtol=0.04):
dt = 0.0 dt = 0.0
avg_pvals = numpy.zeros(target_pvals.shape, dtype=config.floatX) avg_pvals = np.zeros(target_pvals.shape, dtype=config.floatX)
for i in xrange(steps): for i in xrange(steps):
t0 = time.time() t0 = time.time()
ival = f() ival = f()
assert ival.shape == sample_size assert ival.shape == sample_size
assert numpy.all(numpy.sum(ival, axis=1) == n_samples) assert np.all(np.sum(ival, axis=1) == n_samples)
dt += time.time() - t0 dt += time.time() - t0
avg_pvals += ival avg_pvals += ival
avg_pvals /= (steps * n_samples) avg_pvals /= (steps * n_samples)
assert numpy.mean(abs(avg_pvals - target_pvals)) < mean_rtol assert np.mean(abs(avg_pvals - target_pvals)) < mean_rtol
print('random?[:10]\n', numpy.asarray(f()[:10])) print('random?[:10]\n', np.asarray(f()[:10]))
print(prefix, 'mean', avg_pvals) print(prefix, 'mean', avg_pvals)
# < mean_rtol, 'bad mean? %s %s' % (str(avg_pvals), str(target_pvals)) # < mean_rtol, 'bad mean? %s %s' % (str(avg_pvals), str(target_pvals))
print(numpy.mean(abs(avg_pvals - target_pvals))) print(np.mean(abs(avg_pvals - target_pvals)))
print(prefix, 'time', dt) print(prefix, 'time', dt)
print(prefix, 'elements', steps * numpy.prod(target_pvals.shape)) print(prefix, 'elements', steps * np.prod(target_pvals.shape))
print(prefix, 'samples/sec', steps * numpy.prod(target_pvals.shape) / dt) print(prefix, 'samples/sec', steps * np.prod(target_pvals.shape) / dt)
def test_multinomial(): def test_multinomial():
...@@ -875,8 +875,8 @@ def test_multinomial(): ...@@ -875,8 +875,8 @@ def test_multinomial():
# print '' # print ''
# print 'ON CPU:' # print 'ON CPU:'
pvals = numpy.asarray(numpy.random.uniform(size=sample_size)) pvals = np.asarray(np.random.uniform(size=sample_size))
pvals = numpy.apply_along_axis(lambda row: row / numpy.sum(row), 1, pvals) pvals = np.apply_along_axis(lambda row: row / np.sum(row), 1, pvals)
R = MRG_RandomStreams(234, use_cuda=False) R = MRG_RandomStreams(234, use_cuda=False)
# Note: we specify `nstreams` to avoid a warning. # Note: we specify `nstreams` to avoid a warning.
m = R.multinomial(pvals=pvals, dtype=config.floatX, nstreams=30 * 256) m = R.multinomial(pvals=pvals, dtype=config.floatX, nstreams=30 * 256)
...@@ -892,7 +892,7 @@ def test_multinomial(): ...@@ -892,7 +892,7 @@ def test_multinomial():
# print '' # print ''
# print 'ON GPU:' # print 'ON GPU:'
R = MRG_RandomStreams(234, use_cuda=True) R = MRG_RandomStreams(234, use_cuda=True)
pvals = numpy.asarray(pvals, dtype='float32') pvals = np.asarray(pvals, dtype='float32')
# We give the number of streams to avoid a warning. # We give the number of streams to avoid a warning.
n = R.multinomial(pvals=pvals, dtype='float32', nstreams=30 * 256) n = R.multinomial(pvals=pvals, dtype='float32', nstreams=30 * 256)
# well, it's really that this test w GPU doesn't make sense otw # well, it's really that this test w GPU doesn't make sense otw
...@@ -907,7 +907,7 @@ def test_multinomial(): ...@@ -907,7 +907,7 @@ def test_multinomial():
sys.stdout.flush() sys.stdout.flush()
basic_multinomialtest(f, steps, sample_size, pvals, n_samples=1, basic_multinomialtest(f, steps, sample_size, pvals, n_samples=1,
prefix='gpu mrg ') prefix='gpu mrg ')
numpy.testing.assert_array_almost_equal(out, gpu_out, decimal=6) np.testing.assert_array_almost_equal(out, gpu_out, decimal=6)
def test_multinomial_n_samples(): def test_multinomial_n_samples():
...@@ -922,8 +922,8 @@ def test_multinomial_n_samples(): ...@@ -922,8 +922,8 @@ def test_multinomial_n_samples():
sample_size = (450, 6) sample_size = (450, 6)
mode_ = theano.compile.mode.get_mode(mode_) mode_ = theano.compile.mode.get_mode(mode_)
pvals = numpy.asarray(numpy.random.uniform(size=sample_size)) pvals = np.asarray(np.random.uniform(size=sample_size))
pvals = numpy.apply_along_axis(lambda row: row / numpy.sum(row), 1, pvals) pvals = np.apply_along_axis(lambda row: row / np.sum(row), 1, pvals)
R = MRG_RandomStreams(234, use_cuda=False) R = MRG_RandomStreams(234, use_cuda=False)
for n_samples, steps in zip([5, 10, 100, 1000], [20, 10, 1, 1]): for n_samples, steps in zip([5, 10, 100, 1000], [20, 10, 1, 1]):
...@@ -936,7 +936,7 @@ def test_multinomial_n_samples(): ...@@ -936,7 +936,7 @@ def test_multinomial_n_samples():
if mode != 'FAST_COMPILE' and cuda_available: if mode != 'FAST_COMPILE' and cuda_available:
R = MRG_RandomStreams(234, use_cuda=True) R = MRG_RandomStreams(234, use_cuda=True)
pvals = numpy.asarray(pvals, dtype='float32') pvals = np.asarray(pvals, dtype='float32')
n = R.multinomial(pvals=pvals, n=n_samples, n = R.multinomial(pvals=pvals, n=n_samples,
dtype='float32', nstreams=30 * 256) dtype='float32', nstreams=30 * 256)
assert n.dtype == 'float32' assert n.dtype == 'float32'
...@@ -999,14 +999,14 @@ def test_random_state_transfer(): ...@@ -999,14 +999,14 @@ def test_random_state_transfer():
for (su1, su2) in zip(g1.rng.state_updates, g2.rng.state_updates): for (su1, su2) in zip(g1.rng.state_updates, g2.rng.state_updates):
su2[0].set_value(su1[0].get_value()) su2[0].set_value(su1[0].get_value())
numpy.testing.assert_array_almost_equal(f1(), f2(), decimal=6) np.testing.assert_array_almost_equal(f1(), f2(), decimal=6)
def test_gradient_scan(): def test_gradient_scan():
# Test for a crash when using MRG inside scan and taking the gradient # Test for a crash when using MRG inside scan and taking the gradient
# See https://groups.google.com/d/msg/theano-dev/UbcYyU5m-M8/UO9UgXqnQP0J # See https://groups.google.com/d/msg/theano-dev/UbcYyU5m-M8/UO9UgXqnQP0J
theano_rng = MRG_RandomStreams(10) theano_rng = MRG_RandomStreams(10)
w = theano.shared(numpy.ones(1, dtype='float32')) w = theano.shared(np.ones(1, dtype='float32'))
def one_step(x): def one_step(x):
return x + theano_rng.uniform((1,), dtype='float32') * w return x + theano_rng.uniform((1,), dtype='float32') * w
...@@ -1015,7 +1015,7 @@ def test_gradient_scan(): ...@@ -1015,7 +1015,7 @@ def test_gradient_scan():
values, updates = theano.scan(one_step, outputs_info=x, n_steps=10) values, updates = theano.scan(one_step, outputs_info=x, n_steps=10)
gw = theano.grad(tensor.sum(values[-1]), w) gw = theano.grad(tensor.sum(values[-1]), w)
f = theano.function([x], gw) f = theano.function([x], gw)
f(numpy.arange(1, dtype='float32')) f(np.arange(1, dtype='float32'))
def test_multMatVect(): def test_multMatVect():
...@@ -1029,14 +1029,14 @@ def test_multMatVect(): ...@@ -1029,14 +1029,14 @@ def test_multMatVect():
g0 = rng_mrg.DotModulo()(A1, s1, m1, A2, s2, m2) g0 = rng_mrg.DotModulo()(A1, s1, m1, A2, s2, m2)
f0 = theano.function([A1, s1, m1, A2, s2, m2], g0) f0 = theano.function([A1, s1, m1, A2, s2, m2], g0)
i32max = numpy.iinfo(numpy.int32).max i32max = np.iinfo(np.int32).max
A1 = numpy.random.randint(0, i32max, (3, 3)).astype('int64') A1 = np.random.randint(0, i32max, (3, 3)).astype('int64')
s1 = numpy.random.randint(0, i32max, 3).astype('int32') s1 = np.random.randint(0, i32max, 3).astype('int32')
m1 = numpy.asarray(numpy.random.randint(i32max), dtype="int32") m1 = np.asarray(np.random.randint(i32max), dtype="int32")
A2 = numpy.random.randint(0, i32max, (3, 3)).astype('int64') A2 = np.random.randint(0, i32max, (3, 3)).astype('int64')
s2 = numpy.random.randint(0, i32max, 3).astype('int32') s2 = np.random.randint(0, i32max, 3).astype('int32')
m2 = numpy.asarray(numpy.random.randint(i32max), dtype="int32") m2 = np.asarray(np.random.randint(i32max), dtype="int32")
f0.input_storage[0].storage[0] = A1 f0.input_storage[0].storage[0] = A1
f0.input_storage[1].storage[0] = s1 f0.input_storage[1].storage[0] = s1
...@@ -1050,8 +1050,8 @@ def test_multMatVect(): ...@@ -1050,8 +1050,8 @@ def test_multMatVect():
f0.fn() f0.fn()
r_b = f0.output_storage[0].value r_b = f0.output_storage[0].value
assert numpy.allclose(r_a1, r_b[:3]) assert np.allclose(r_a1, r_b[:3])
assert numpy.allclose(r_a2, r_b[3:]) assert np.allclose(r_a2, r_b[3:])
def test_seed_fn(): def test_seed_fn():
...@@ -1079,13 +1079,13 @@ def test_seed_fn(): ...@@ -1079,13 +1079,13 @@ def test_seed_fn():
fn1_val0 = fn1() fn1_val0 = fn1()
fn1_val1 = fn1() fn1_val1 = fn1()
assert not numpy.allclose(fn1_val0, fn1_val1) assert not np.allclose(fn1_val0, fn1_val1)
fn2_val0 = fn2() fn2_val0 = fn2()
fn2_val1 = fn2() fn2_val1 = fn2()
assert not numpy.allclose(fn2_val0, fn2_val1) assert not np.allclose(fn2_val0, fn2_val1)
fn3_val0 = fn3([4]) fn3_val0 = fn3([4])
fn3_val1 = fn3([4]) fn3_val1 = fn3([4])
assert not numpy.allclose(fn3_val0, fn3_val1) assert not np.allclose(fn3_val0, fn3_val1)
assert fn1_val0.size == 4 assert fn1_val0.size == 4
assert fn2_val0.size == 9 assert fn2_val0.size == 9
...@@ -1097,12 +1097,12 @@ def test_seed_fn(): ...@@ -1097,12 +1097,12 @@ def test_seed_fn():
fn2_val3 = fn2() fn2_val3 = fn2()
fn3_val2 = fn3([4]) fn3_val2 = fn3([4])
fn3_val3 = fn3([4]) fn3_val3 = fn3([4])
assert numpy.allclose(fn1_val0, fn1_val2) == same assert np.allclose(fn1_val0, fn1_val2) == same
assert numpy.allclose(fn1_val1, fn1_val3) == same assert np.allclose(fn1_val1, fn1_val3) == same
assert numpy.allclose(fn2_val0, fn2_val2) == same assert np.allclose(fn2_val0, fn2_val2) == same
assert numpy.allclose(fn2_val1, fn2_val3) == same assert np.allclose(fn2_val1, fn2_val3) == same
assert numpy.allclose(fn3_val0, fn3_val2) == same assert np.allclose(fn3_val0, fn3_val2) == same
assert numpy.allclose(fn3_val1, fn3_val3) == same assert np.allclose(fn3_val1, fn3_val3) == same
def rng_mrg_overflow(sizes, fct, mode, should_raise_error): def rng_mrg_overflow(sizes, fct, mode, should_raise_error):
...@@ -1118,7 +1118,7 @@ def rng_mrg_overflow(sizes, fct, mode, should_raise_error): ...@@ -1118,7 +1118,7 @@ def rng_mrg_overflow(sizes, fct, mode, should_raise_error):
def test_overflow_cpu(): def test_overflow_cpu():
# run with THEANO_FLAGS=mode=FAST_RUN,device=cpu,floatX=float32 # run with THEANO_FLAGS=mode=FAST_RUN,device=cpu,floatX=float32
rng = MRG_RandomStreams(numpy.random.randint(1234)) rng = MRG_RandomStreams(np.random.randint(1234))
fct = rng.uniform fct = rng.uniform
# should raise error as the size overflows # should raise error as the size overflows
sizes = [(2**31, ), (2**32, ), (2**15, 2**16,), (2, 2**15, 2**15)] sizes = [(2**31, ), (2**32, ), (2**15, 2**16,), (2, 2**15, 2**15)]
...@@ -1127,8 +1127,8 @@ def test_overflow_cpu(): ...@@ -1127,8 +1127,8 @@ def test_overflow_cpu():
sizes = [(2**5, ), (2**5, 2**5), (2**5, 2**5, 2**5)] sizes = [(2**5, ), (2**5, 2**5), (2**5, 2**5, 2**5)]
rng_mrg_overflow(sizes, fct, config.mode, should_raise_error=False) rng_mrg_overflow(sizes, fct, config.mode, should_raise_error=False)
# should support int32 sizes # should support int32 sizes
sizes = [(numpy.int32(2**10), ), sizes = [(np.int32(2**10), ),
(numpy.int32(2), numpy.int32(2**10), numpy.int32(2**10))] (np.int32(2), np.int32(2**10), np.int32(2**10))]
rng_mrg_overflow(sizes, fct, config.mode, should_raise_error=False) rng_mrg_overflow(sizes, fct, config.mode, should_raise_error=False)
...@@ -1147,8 +1147,8 @@ def test_overflow_gpu_old_backend(): ...@@ -1147,8 +1147,8 @@ def test_overflow_gpu_old_backend():
sizes = [(2**5, ), (2**5, 2**5), (2**5, 2**5, 2**5)] sizes = [(2**5, ), (2**5, 2**5), (2**5, 2**5, 2**5)]
rng_mrg_overflow(sizes, fct, mode, should_raise_error=False) rng_mrg_overflow(sizes, fct, mode, should_raise_error=False)
# should support int32 sizes # should support int32 sizes
sizes = [(numpy.int32(2**10), ), sizes = [(np.int32(2**10), ),
(numpy.int32(2), numpy.int32(2**10), numpy.int32(2**10))] (np.int32(2), np.int32(2**10), np.int32(2**10))]
rng_mrg_overflow(sizes, fct, mode, should_raise_error=False) rng_mrg_overflow(sizes, fct, mode, should_raise_error=False)
...@@ -1159,11 +1159,11 @@ def test_overflow_gpu_new_backend(): ...@@ -1159,11 +1159,11 @@ def test_overflow_gpu_new_backend():
from theano.gpuarray.type import gpuarray_shared_constructor from theano.gpuarray.type import gpuarray_shared_constructor
seed = 12345 seed = 12345
n_substreams = 7 n_substreams = 7
curr_rstate = numpy.array([seed] * 6, dtype='int32') curr_rstate = np.array([seed] * 6, dtype='int32')
rstate = [curr_rstate.copy()] rstate = [curr_rstate.copy()]
for j in range(1, n_substreams): for j in range(1, n_substreams):
rstate.append(rng_mrg.ff_2p72(rstate[-1])) rstate.append(rng_mrg.ff_2p72(rstate[-1]))
rstate = numpy.asarray(rstate) rstate = np.asarray(rstate)
rstate = gpuarray_shared_constructor(rstate) rstate = gpuarray_shared_constructor(rstate)
fct = functools.partial(rng_mrg.GPUA_mrg_uniform.new, rstate, fct = functools.partial(rng_mrg.GPUA_mrg_uniform.new, rstate,
ndim=None, dtype='float32') ndim=None, dtype='float32')
...@@ -1174,8 +1174,8 @@ def test_overflow_gpu_new_backend(): ...@@ -1174,8 +1174,8 @@ def test_overflow_gpu_new_backend():
sizes = [(2**5, ), (2**5, 2**5), (2**5, 2**5, 2**5)] sizes = [(2**5, ), (2**5, 2**5), (2**5, 2**5, 2**5)]
rng_mrg_overflow(sizes, fct, mode, should_raise_error=False) rng_mrg_overflow(sizes, fct, mode, should_raise_error=False)
# should support int32 sizes # should support int32 sizes
sizes = [(numpy.int32(2**10), ), sizes = [(np.int32(2**10), ),
(numpy.int32(2), numpy.int32(2**10), numpy.int32(2**10))] (np.int32(2), np.int32(2**10), np.int32(2**10))]
rng_mrg_overflow(sizes, fct, mode, should_raise_error=False) rng_mrg_overflow(sizes, fct, mode, should_raise_error=False)
...@@ -1185,12 +1185,12 @@ def test_validate_input_types_gpuarray_backend(): ...@@ -1185,12 +1185,12 @@ def test_validate_input_types_gpuarray_backend():
from theano.configparser import change_flags from theano.configparser import change_flags
with change_flags(compute_test_value="raise"): with change_flags(compute_test_value="raise"):
rstate = numpy.zeros((7, 6), dtype="int32") rstate = np.zeros((7, 6), dtype="int32")
rstate = gpuarray_shared_constructor(rstate) rstate = gpuarray_shared_constructor(rstate)
mrg_uniform.new(rstate, ndim=None, dtype="float32", size=(3,)) mrg_uniform.new(rstate, ndim=None, dtype="float32", size=(3,))
if __name__ == "__main__": if __name__ == "__main__":
rng = MRG_RandomStreams(numpy.random.randint(2147462579)) rng = MRG_RandomStreams(np.random.randint(2147462579))
print(theano.__file__) print(theano.__file__)
pvals = theano.tensor.fmatrix() pvals = theano.tensor.fmatrix()
for i in range(10): for i in range(10):
......
theano/scan_module/scan.py
浏览文件 @ bea31470
...@@ -45,7 +45,7 @@ __contact__ = "Razvan Pascanu <r.pascanu@gmail>" ...@@ -45,7 +45,7 @@ __contact__ = "Razvan Pascanu <r.pascanu@gmail>"
import logging import logging
import numpy import numpy as np
import warnings import warnings
from collections import OrderedDict from collections import OrderedDict
...@@ -488,8 +488,8 @@ def scan(fn, ...@@ -488,8 +488,8 @@ def scan(fn,
# a sequence, though is highly unlikely in practice # a sequence, though is highly unlikely in practice
if 'taps' in seq: if 'taps' in seq:
# go through the indicated slice # go through the indicated slice
mintap = numpy.min(seq['taps']) mintap = np.min(seq['taps'])
maxtap = numpy.max(seq['taps']) maxtap = np.max(seq['taps'])
for k in seq['taps']: for k in seq['taps']:
# create one slice of the input # create one slice of the input
# Later on, if we decide not to use scan because we are # Later on, if we decide not to use scan because we are
...@@ -670,15 +670,15 @@ def scan(fn, ...@@ -670,15 +670,15 @@ def scan(fn,
elif init_out.get('taps', None): elif init_out.get('taps', None):
if numpy.any(numpy.array(init_out.get('taps', [])) > 0): if np.any(np.array(init_out.get('taps', [])) > 0):
# Make sure we do not have requests for future values of a # Make sure we do not have requests for future values of a
# sequence we can not provide such values # sequence we can not provide such values
raise ValueError('Can not use future taps of outputs', raise ValueError('Can not use future taps of outputs',
init_out) init_out)
# go through the taps # go through the taps
mintap = abs(numpy.min(init_out['taps'])) mintap = abs(np.min(init_out['taps']))
mit_sot_tap_array.append(init_out['taps']) mit_sot_tap_array.append(init_out['taps'])
idx_offset = abs(numpy.min(init_out['taps'])) idx_offset = abs(np.min(init_out['taps']))
# Sequence # Sequence
mit_sot_scan_inputs.append( mit_sot_scan_inputs.append(
scan_utils.expand_empty(init_out['initial'][:mintap], scan_utils.expand_empty(init_out['initial'][:mintap],
...@@ -725,9 +725,9 @@ def scan(fn, ...@@ -725,9 +725,9 @@ def scan(fn,
# a map); in that case we do not have to do anything .. # a map); in that case we do not have to do anything ..
# Re-order args # Re-order args
max_mit_sot = numpy.max([-1] + mit_sot_rightOrder) + 1 max_mit_sot = np.max([-1] + mit_sot_rightOrder) + 1
max_sit_sot = numpy.max([-1] + sit_sot_rightOrder) + 1 max_sit_sot = np.max([-1] + sit_sot_rightOrder) + 1
n_elems = numpy.max([max_mit_sot, max_sit_sot]) n_elems = np.max([max_mit_sot, max_sit_sot])
_ordered_args = [[] for x in xrange(n_elems)] _ordered_args = [[] for x in xrange(n_elems)]
offset = 0 offset = 0
for idx in xrange(n_mit_sot): for idx in xrange(n_mit_sot):
...@@ -1101,7 +1101,7 @@ def scan(fn, ...@@ -1101,7 +1101,7 @@ def scan(fn,
return out_ls return out_ls
offset = n_mit_mot offset = n_mit_mot
offsets = [abs(numpy.min(x)) for x in mit_sot_tap_array] offsets = [abs(np.min(x)) for x in mit_sot_tap_array]
mit_sot_outs = remove_dimensions( mit_sot_outs = remove_dimensions(
scan_outs[offset:offset + n_mit_sot], scan_outs[offset:offset + n_mit_sot],
mit_sot_return_steps, mit_sot_return_steps,
......
theano/scan_module/scan_op.py
浏览文件 @ bea31470
...@@ -54,7 +54,7 @@ import logging ...@@ -54,7 +54,7 @@ import logging
import time import time
from collections import OrderedDict from collections import OrderedDict
import numpy import numpy as np
from six import iteritems, integer_types, raise_from from six import iteritems, integer_types, raise_from
from six.moves import xrange from six.moves import xrange
...@@ -193,7 +193,7 @@ class Scan(PureOp): ...@@ -193,7 +193,7 @@ class Scan(PureOp):
self.info['name'] = self.name self.info['name'] = self.name
# Pre-computing some values to speed up perform # Pre-computing some values to speed up perform
self.mintaps = [numpy.min(x) for x in self.tap_array] self.mintaps = [np.min(x) for x in self.tap_array]
self.mintaps += [0 for x in xrange(self.n_nit_sot)] self.mintaps += [0 for x in xrange(self.n_nit_sot)]
self.seqs_arg_offset = 1 + self.n_seqs self.seqs_arg_offset = 1 + self.n_seqs
self.shared_arg_offset = (self.seqs_arg_offset + self.shared_arg_offset = (self.seqs_arg_offset +
...@@ -336,7 +336,7 @@ class Scan(PureOp): ...@@ -336,7 +336,7 @@ class Scan(PureOp):
the inner function) the inner function)
""" """
assert numpy.all(isinstance(i, gof.Variable) for i in inputs) assert np.all(isinstance(i, gof.Variable) for i in inputs)
# Check that the number of inputs to the Scan node corresponds to # Check that the number of inputs to the Scan node corresponds to
# the number of inputs of the inner function of scan # the number of inputs of the inner function of scan
n_outer_ins = len(inputs) - len(self.outer_nitsot(inputs)) - 1 n_outer_ins = len(inputs) - len(self.outer_nitsot(inputs)) - 1
...@@ -901,53 +901,53 @@ class Scan(PureOp): ...@@ -901,53 +901,53 @@ class Scan(PureOp):
try: try:
if impl == 'py': if impl == 'py':
raise theano.gof.cmodule.MissingGXX raise theano.gof.cmodule.MissingGXX
cython_mintaps = numpy.asarray(self.mintaps, dtype='int32') cython_mintaps = np.asarray(self.mintaps, dtype='int32')
cython_tap_array_len = \ cython_tap_array_len = \
numpy.asarray([len(x) for x in self.tap_array], np.asarray([len(x) for x in self.tap_array],
dtype='int32') dtype='int32')
if len(self.tap_array) == 0: if len(self.tap_array) == 0:
d1 = 0 d1 = 0
else: else:
d1 = numpy.max(cython_tap_array_len) d1 = np.max(cython_tap_array_len)
d0 = len(self.tap_array) d0 = len(self.tap_array)
cython_tap_array = numpy.zeros((d0, d1), dtype='int32') cython_tap_array = np.zeros((d0, d1), dtype='int32')
for _d0 in xrange(d0): for _d0 in xrange(d0):
for _d1 in xrange(cython_tap_array_len[_d0]): for _d1 in xrange(cython_tap_array_len[_d0]):
cython_tap_array[_d0, _d1] = self.tap_array[_d0][_d1] cython_tap_array[_d0, _d1] = self.tap_array[_d0][_d1]
cython_mit_mot_out_nslices = \ cython_mit_mot_out_nslices = \
numpy.asarray([len(x) for x in self.mit_mot_out_slices], np.asarray([len(x) for x in self.mit_mot_out_slices],
dtype='int32') dtype='int32')
if len(self.mit_mot_out_slices) == 0: if len(self.mit_mot_out_slices) == 0:
d1 = 0 d1 = 0
else: else:
d1 = numpy.max(cython_mit_mot_out_nslices) d1 = np.max(cython_mit_mot_out_nslices)
d0 = len(self.mit_mot_out_slices) d0 = len(self.mit_mot_out_slices)
cython_mit_mot_out_slices = numpy.zeros((d0, d1), cython_mit_mot_out_slices = np.zeros((d0, d1),
dtype='int32') dtype='int32')
for _d0 in xrange(d0): for _d0 in xrange(d0):
for _d1 in xrange(cython_mit_mot_out_nslices[_d0]): for _d1 in xrange(cython_mit_mot_out_nslices[_d0]):
cython_mit_mot_out_slices[_d0, _d1] = \ cython_mit_mot_out_slices[_d0, _d1] = \
self.mit_mot_out_slices[_d0][_d1] self.mit_mot_out_slices[_d0][_d1]
cython_vector_seqs = numpy.asarray(self.vector_seqs, cython_vector_seqs = np.asarray(self.vector_seqs,
dtype='int32')
cython_vector_outs = np.asarray(self.vector_outs,
dtype='int32')
cython_mitmots_preallocated = np.asarray(self.mitmots_preallocated,
dtype='int32')
cython_inps_is_tensor = np.asarray(self.inps_is_tensor,
dtype='int32') dtype='int32')
cython_vector_outs = numpy.asarray(self.vector_outs, cython_outs_is_tensor = np.asarray(self.outs_is_tensor,
dtype='int32') dtype='int32')
cython_mitmots_preallocated = numpy.asarray(self.mitmots_preallocated,
dtype='int32')
cython_inps_is_tensor = numpy.asarray(self.inps_is_tensor,
dtype='int32')
cython_outs_is_tensor = numpy.asarray(self.outs_is_tensor,
dtype='int32')
if hasattr(self, 'destroy_map'): if hasattr(self, 'destroy_map'):
cython_destroy_map = [x in self.destroy_map cython_destroy_map = [x in self.destroy_map
for x in xrange(len(node.outputs))] for x in xrange(len(node.outputs))]
else: else:
cython_destroy_map = [0 for x in xrange(len(node.outputs))] cython_destroy_map = [0 for x in xrange(len(node.outputs))]
cython_destroy_map = numpy.asarray(cython_destroy_map, cython_destroy_map = np.asarray(cython_destroy_map,
dtype='int32') dtype='int32')
from . import scan_perform_ext from . import scan_perform_ext
def p(node, args, outs): def p(node, args, outs):
...@@ -2200,9 +2200,9 @@ class Scan(PureOp): ...@@ -2200,9 +2200,9 @@ class Scan(PureOp):
# Seqs # Seqs
outer_inp_seqs = [x[::-1] for x in inputs[1:1 + self.n_seqs]] outer_inp_seqs = [x[::-1] for x in inputs[1:1 + self.n_seqs]]
for idx in xrange(self.n_mit_mot + self.n_mit_sot): for idx in xrange(self.n_mit_mot + self.n_mit_sot):
mintap = numpy.min(self.tap_array[idx]) mintap = np.min(self.tap_array[idx])
if idx < self.n_mit_mot: if idx < self.n_mit_mot:
outmaxtap = numpy.max(self.mitmot_out_taps()[idx]) outmaxtap = np.max(self.mitmot_out_taps()[idx])
else: else:
outmaxtap = 0 outmaxtap = 0
seq = outs[idx] seq = outs[idx]
...@@ -2226,7 +2226,7 @@ class Scan(PureOp): ...@@ -2226,7 +2226,7 @@ class Scan(PureOp):
# that. # that.
for taps, x in zip(self.mitsot_taps(), for taps, x in zip(self.mitsot_taps(),
self.outer_mitsot_outs(outs)): self.outer_mitsot_outs(outs)):
mintap = numpy.min(taps) mintap = np.min(taps)
if hasattr(x[::-1][:mintap], 'test_value'): if hasattr(x[::-1][:mintap], 'test_value'):
assert (x[::-1][:mintap].tag.test_value.shape[0] == assert (x[::-1][:mintap].tag.test_value.shape[0] ==
inputs[0].tag.test_value) inputs[0].tag.test_value)
...@@ -2238,7 +2238,7 @@ class Scan(PureOp): ...@@ -2238,7 +2238,7 @@ class Scan(PureOp):
if hasattr(x[::-1].tag, 'test_value'): if hasattr(x[::-1].tag, 'test_value'):
assert (x[::-1].tag.test_value.shape[0] == assert (x[::-1].tag.test_value.shape[0] ==
inputs[0].tag.test_value) inputs[0].tag.test_value)
outer_inp_seqs += [x[::-1][:numpy.min(taps)] outer_inp_seqs += [x[::-1][:np.min(taps)]
for taps, x in zip(self.mitsot_taps(), for taps, x in zip(self.mitsot_taps(),
self.outer_mitsot_outs(outs))] self.outer_mitsot_outs(outs))]
outer_inp_seqs += [x[::-1][:-1] for x in self.outer_sitsot_outs(outs)] outer_inp_seqs += [x[::-1][:-1] for x in self.outer_sitsot_outs(outs)]
...@@ -2726,8 +2726,8 @@ class Scan(PureOp): ...@@ -2726,8 +2726,8 @@ class Scan(PureOp):
b = e b = e
e = e + self.n_mit_mot e = e + self.n_mit_mot
ib = ie ib = ie
ie = ie + int(numpy.sum([len(x) for x in ie = ie + int(np.sum([len(x) for x in
self.tap_array[:self.n_mit_mot]])) self.tap_array[:self.n_mit_mot]]))
clean_eval_points = [] clean_eval_points = []
for inp, evp in zip(inputs[b:e], eval_points[b:e]): for inp, evp in zip(inputs[b:e], eval_points[b:e]):
if evp is not None: if evp is not None:
...@@ -2742,9 +2742,9 @@ class Scan(PureOp): ...@@ -2742,9 +2742,9 @@ class Scan(PureOp):
b = e b = e
e = e + self.n_mit_sot e = e + self.n_mit_sot
ib = ie ib = ie
ie = ie + int(numpy.sum([len(x) for x in ie = ie + int(np.sum([len(x) for x in
self.tap_array[self.n_mit_mot: self.tap_array[self.n_mit_mot:
self.n_mit_mot + self.n_mit_sot]])) self.n_mit_mot + self.n_mit_sot]]))
clean_eval_points = [] clean_eval_points = []
for inp, evp in zip(inputs[b:e], eval_points[b:e]): for inp, evp in zip(inputs[b:e], eval_points[b:e]):
if evp is not None: if evp is not None:
...@@ -2795,8 +2795,8 @@ class Scan(PureOp): ...@@ -2795,8 +2795,8 @@ class Scan(PureOp):
inner_other = self_inputs[ie:] + inner_eval_points[ib:] inner_other = self_inputs[ie:] + inner_eval_points[ib:]
# Outputs # Outputs
n_mit_mot_outs = int(numpy.sum([len(x) for x in n_mit_mot_outs = int(np.sum([len(x) for x in
self.mit_mot_out_slices])) self.mit_mot_out_slices]))
info['n_mit_mot_outs'] = n_mit_mot_outs * 2 info['n_mit_mot_outs'] = n_mit_mot_outs * 2
b = 0 b = 0
e = n_mit_mot_outs e = n_mit_mot_outs
......
theano/scan_module/scan_opt.py
浏览文件 @ bea31470
...@@ -54,7 +54,7 @@ import logging ...@@ -54,7 +54,7 @@ import logging
import copy import copy
from sys import maxsize from sys import maxsize
from collections import OrderedDict from collections import OrderedDict
import numpy import numpy as np
import theano import theano
from theano import tensor, scalar from theano import tensor, scalar
...@@ -636,7 +636,7 @@ class PushOutSeqScan(gof.Optimizer): ...@@ -636,7 +636,7 @@ class PushOutSeqScan(gof.Optimizer):
if out in op.inner_mitsot_outs(ls): if out in op.inner_mitsot_outs(ls):
odx = op.inner_mitsot_outs(ls).index(out) odx = op.inner_mitsot_outs(ls).index(out)
inp = op.outer_mitsot(node)[odx] inp = op.outer_mitsot(node)[odx]
st = abs(numpy.min(op.mitsot_taps())) st = abs(np.min(op.mitsot_taps()))
y = tensor.set_subtensor(inp[st:], _y) y = tensor.set_subtensor(inp[st:], _y)
elif out in op.inner_sitsot_outs(ls): elif out in op.inner_sitsot_outs(ls):
odx = op.inner_sitsot_outs(ls).index(out) odx = op.inner_sitsot_outs(ls).index(out)
...@@ -1373,7 +1373,7 @@ class ScanSaveMem(gof.Optimizer): ...@@ -1373,7 +1373,7 @@ class ScanSaveMem(gof.Optimizer):
# TODO: Simplify the number of steps needed. # TODO: Simplify the number of steps needed.
# FB: This need good testing, left to later. # FB: This need good testing, left to later.
# call get_scalar_constant_value()? it can # call get_scalar_constant_value()? it can
# return python/numpy scalar or numpy.ndarray # return python/numpy scalar or np.ndarray
# currently. # currently.
# pval = pre_greedy_local_optimizer(list_opt_slice, # pval = pre_greedy_local_optimizer(list_opt_slice,
# pval) # pval)
......
theano/scan_module/scan_perform_ext.py
浏览文件 @ bea31470
...@@ -12,7 +12,7 @@ import os ...@@ -12,7 +12,7 @@ import os
import sys import sys
import warnings import warnings
import numpy import numpy as np
import theano import theano
from theano import config from theano import config
...@@ -103,7 +103,7 @@ except ImportError: ...@@ -103,7 +103,7 @@ except ImportError:
# During scan cython development, it is helpful to keep the old interface, to don't manually edit the c file each time. # During scan cython development, it is helpful to keep the old interface, to don't manually edit the c file each time.
preargs.remove('-DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION') preargs.remove('-DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION')
else: else:
numpy_ver = [int(n) for n in numpy.__version__.split('.')[:2]] numpy_ver = [int(n) for n in np.__version__.split('.')[:2]]
# Add add some macro to lower the number of edit # Add add some macro to lower the number of edit
# needed to the c file. # needed to the c file.
if bool(numpy_ver >= [1, 7]): if bool(numpy_ver >= [1, 7]):
......
theano/scan_module/scan_utils.py
浏览文件 @ bea31470
...@@ -20,7 +20,7 @@ import logging ...@@ -20,7 +20,7 @@ import logging
import warnings import warnings
from collections import OrderedDict from collections import OrderedDict
import numpy import numpy as np
import theano import theano
from theano.compat import izip from theano.compat import izip
...@@ -589,8 +589,8 @@ def get_updates_and_outputs(ls): ...@@ -589,8 +589,8 @@ def get_updates_and_outputs(ls):
def isNaN_or_Inf_or_None(x): def isNaN_or_Inf_or_None(x):
isNone = x is None isNone = x is None
try: try:
isNaN = numpy.isnan(x) isNaN = np.isnan(x)
isInf = numpy.isinf(x) isInf = np.isinf(x)
isStr = isinstance(x, string_types) isStr = isinstance(x, string_types)
except Exception: except Exception:
isNaN = False isNaN = False
...@@ -599,8 +599,8 @@ def isNaN_or_Inf_or_None(x): ...@@ -599,8 +599,8 @@ def isNaN_or_Inf_or_None(x):
if not isNaN and not isInf: if not isNaN and not isInf:
try: try:
val = get_scalar_constant_value(x) val = get_scalar_constant_value(x)
isInf = numpy.isinf(val) isInf = np.isinf(val)
isNaN = numpy.isnan(val) isNaN = np.isnan(val)
except Exception: except Exception:
isNaN = False isNaN = False
isInf = False isInf = False
...@@ -959,7 +959,7 @@ def scan_can_remove_outs(op, out_idxs): ...@@ -959,7 +959,7 @@ def scan_can_remove_outs(op, out_idxs):
added = False added = False
for pos, idx in enumerate(out_idxs): for pos, idx in enumerate(out_idxs):
if (out_idxs_mask[pos] and if (out_idxs_mask[pos] and
numpy.any([x in required_inputs for x in out_ins[idx]])): np.any([x in required_inputs for x in out_ins[idx]])):
# This output is required .. # This output is required ..
out_idxs_mask[pos] = 0 out_idxs_mask[pos] = 0
required_inputs += gof.graph.inputs([op.outputs[idx]]) required_inputs += gof.graph.inputs([op.outputs[idx]])
......
theano/scan_module/tests/test_scan.py
浏览文件 @ bea31470
...@@ -11,7 +11,7 @@ from collections import OrderedDict ...@@ -11,7 +11,7 @@ from collections import OrderedDict
import six.moves.cPickle as pickle import six.moves.cPickle as pickle
from six.moves import xrange from six.moves import xrange
import numpy import numpy as np
from nose.plugins.skip import SkipTest from nose.plugins.skip import SkipTest
from nose.tools import assert_raises from nose.tools import assert_raises
from nose.tools import raises from nose.tools import raises
...@@ -89,7 +89,7 @@ class multiple_outputs_numeric_grad: ...@@ -89,7 +89,7 @@ class multiple_outputs_numeric_grad:
for i, p in enumerate(pt): for i, p in enumerate(pt):
if ndarray_mask[i]: if ndarray_mask[i]:
pt[i] = numpy.array(p) pt[i] = np.array(p)
_eps = type_eps[str(pt[i].dtype)] _eps = type_eps[str(pt[i].dtype)]
if _eps > dtype_eps: if _eps > dtype_eps:
dtype_eps = _eps dtype_eps = _eps
...@@ -116,12 +116,12 @@ class multiple_outputs_numeric_grad: ...@@ -116,12 +116,12 @@ class multiple_outputs_numeric_grad:
t[pos] += _eps t[pos] += _eps
t = t.reshape(pt[i].shape) t = t.reshape(pt[i].shape)
f_eps = f(*(pt[:i] + [t] + pt[i + 1:])) f_eps = f(*(pt[:i] + [t] + pt[i + 1:]))
_g.append(numpy.asarray((f_eps - f_x) / _eps)) _g.append(np.asarray((f_eps - f_x) / _eps))
gx.append(numpy.asarray(_g).reshape(pt[i].shape)) gx.append(np.asarray(_g).reshape(pt[i].shape))
else: else:
t = numpy.array(pt[i] + _eps) t = np.array(pt[i] + _eps)
f_eps = f(*(pt[:i] + [t] + pt[i + 1:])) f_eps = f(*(pt[:i] + [t] + pt[i + 1:]))
gx.append(numpy.asarray((f_eps - f_x) / _eps)) gx.append(np.asarray((f_eps - f_x) / _eps))
self.gx = gx self.gx = gx
@staticmethod @staticmethod
...@@ -137,8 +137,8 @@ class multiple_outputs_numeric_grad: ...@@ -137,8 +137,8 @@ class multiple_outputs_numeric_grad:
for i in xrange(len(_g_pt)): for i in xrange(len(_g_pt)):
if self.ndarray_mask[i]: if self.ndarray_mask[i]:
g_pt.append(_g_pt[i]) g_pt.append(_g_pt[i])
elif isinstance(_g_pt[i], numpy.ndarray): elif isinstance(_g_pt[i], np.ndarray):
assert numpy.all(_g_pt[i] == 0) assert np.all(_g_pt[i] == 0)
if len(g_pt) != len(self.gx): if len(g_pt) != len(self.gx):
raise ValueError('argument has wrong number of elements', raise ValueError('argument has wrong number of elements',
len(g_pt)) len(g_pt))
...@@ -149,12 +149,12 @@ class multiple_outputs_numeric_grad: ...@@ -149,12 +149,12 @@ class multiple_outputs_numeric_grad:
raise ValueError('argument element %i has wrong shape %s' % raise ValueError('argument element %i has wrong shape %s' %
(i, str((a.shape, b.shape)))) (i, str((a.shape, b.shape))))
vv = multiple_outputs_numeric_grad.abs_rel_err(a, b) vv = multiple_outputs_numeric_grad.abs_rel_err(a, b)
errs.append(numpy.max( errs.append(np.max(
multiple_outputs_numeric_grad.abs_rel_err(a, b))) multiple_outputs_numeric_grad.abs_rel_err(a, b)))
if numpy.all(numpy.isfinite(errs)): if np.all(np.isfinite(errs)):
return numpy.max(errs), numpy.argmax(errs) return np.max(errs), np.argmax(errs)
else: else:
return numpy.inf, 0 return np.inf, 0
# TODO: Test this function, and if it works, # TODO: Test this function, and if it works,
...@@ -262,11 +262,11 @@ class T_Scan(unittest.TestCase): ...@@ -262,11 +262,11 @@ class T_Scan(unittest.TestCase):
if tmpdir is not None: if tmpdir is not None:
shutil.rmtree(tmpdir) shutil.rmtree(tmpdir)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
state = rng.uniform() state = rng.uniform()
steps = 5 steps = 5
numpy_values = numpy.array([state * (2 ** (k + 1)) for k numpy_values = np.array([state * (2 ** (k + 1)) for k
in xrange(steps)]) in xrange(steps)])
theano_values = my_f(state, steps) theano_values = my_f(state, steps)
utt.assert_allclose(numpy_values, theano_values) utt.assert_allclose(numpy_values, theano_values)
...@@ -300,7 +300,7 @@ class T_Scan(unittest.TestCase): ...@@ -300,7 +300,7 @@ class T_Scan(unittest.TestCase):
assert all(i.value is None for i in scan_node.op.fn.input_storage) assert all(i.value is None for i in scan_node.op.fn.input_storage)
assert all(o.value is None for o in scan_node.op.fn.output_storage) assert all(o.value is None for o in scan_node.op.fn.output_storage)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
state = rng.uniform() state = rng.uniform()
steps = 5 steps = 5
...@@ -332,11 +332,11 @@ class T_Scan(unittest.TestCase): ...@@ -332,11 +332,11 @@ class T_Scan(unittest.TestCase):
updates=updates, updates=updates,
allow_input_downcast=True) allow_input_downcast=True)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
state = rng.uniform() state = rng.uniform()
steps = 5 steps = 5
numpy_values = numpy.array([state * (2 ** (k + 1)) for k numpy_values = np.array([state * (2 ** (k + 1)) for k
in xrange(steps)]) in xrange(steps)])
theano_values = my_f(state, steps) theano_values = my_f(state, steps)
utt.assert_allclose(numpy_values, theano_values[0]) utt.assert_allclose(numpy_values, theano_values[0])
...@@ -370,10 +370,10 @@ class T_Scan(unittest.TestCase): ...@@ -370,10 +370,10 @@ class T_Scan(unittest.TestCase):
# This assertation fails if savemem optimization failed on scan # This assertation fails if savemem optimization failed on scan
if theano.config.mode != "FAST_COMPILE": if theano.config.mode != "FAST_COMPILE":
assert nodes[0].op._scan_savemem_visited assert nodes[0].op._scan_savemem_visited
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
my_f(rng.uniform(size=(3,)), my_f(rng.uniform(size=(3,)),
4, 4,
numpy.int64([2, 2, 3])) np.int64([2, 2, 3]))
@attr('slow') @attr('slow')
def test_only_nonseq_inputs(self): def test_only_nonseq_inputs(self):
...@@ -388,9 +388,9 @@ class T_Scan(unittest.TestCase): ...@@ -388,9 +388,9 @@ class T_Scan(unittest.TestCase):
fun = theano.function([inp], [broadcasted_inp, gr]) fun = theano.function([inp], [broadcasted_inp, gr])
# Execute the Theano function and compare outputs to the expected outputs # Execute the Theano function and compare outputs to the expected outputs
inputs = numpy.array([[1, 2], [3, 4]], dtype=theano.config.floatX) inputs = np.array([[1, 2], [3, 4]], dtype=theano.config.floatX)
expected_out1 = numpy.repeat(inputs[None], n_steps, axis=0) expected_out1 = np.repeat(inputs[None], n_steps, axis=0)
expected_out2 = numpy.ones(inputs.shape, dtype="int8") * n_steps expected_out2 = np.ones(inputs.shape, dtype="int8") * n_steps
out1, out2 = fun(inputs) out1, out2 = fun(inputs)
utt.assert_allclose(out1, expected_out1) utt.assert_allclose(out1, expected_out1)
...@@ -420,14 +420,14 @@ class T_Scan(unittest.TestCase): ...@@ -420,14 +420,14 @@ class T_Scan(unittest.TestCase):
updates=updates, updates=updates,
allow_input_downcast=True) allow_input_downcast=True)
# get random initial values # get random initial values
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = rng.uniform(size=(4,), low=-5., high=5.) v_u = rng.uniform(size=(4,), low=-5., high=5.)
v_x0 = rng.uniform() v_x0 = rng.uniform()
W = rng.uniform() W = rng.uniform()
W_in = rng.uniform() W_in = rng.uniform()
# compute the output in numpy # compute the output in numpy
v_out = numpy.zeros((4,)) v_out = np.zeros((4,))
v_out[0] = v_u[0] * W_in + v_x0 * W v_out[0] = v_u[0] * W_in + v_x0 * W
for step in xrange(1, 4): for step in xrange(1, 4):
v_out[step] = v_u[step] * W_in + v_out[step - 1] * W v_out[step] = v_u[step] * W_in + v_out[step - 1] * W
...@@ -437,7 +437,7 @@ class T_Scan(unittest.TestCase): ...@@ -437,7 +437,7 @@ class T_Scan(unittest.TestCase):
# simple rnn, one input, one state, weights for each; input/state # simple rnn, one input, one state, weights for each; input/state
# are vectors, weights are scalars; using shared variables # are vectors, weights are scalars; using shared variables
def test_one_sequence_one_output_weights_shared(self): def test_one_sequence_one_output_weights_shared(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
u = theano.tensor.vector('u') u = theano.tensor.vector('u')
x0 = theano.tensor.scalar('x0') x0 = theano.tensor.scalar('x0')
W_in = theano.shared(asarrayX(rng.uniform()), name='w_in') W_in = theano.shared(asarrayX(rng.uniform()), name='w_in')
...@@ -462,19 +462,19 @@ class T_Scan(unittest.TestCase): ...@@ -462,19 +462,19 @@ class T_Scan(unittest.TestCase):
v_u = rng.uniform(size=(4,), low=-5., high=5.) v_u = rng.uniform(size=(4,), low=-5., high=5.)
v_x0 = rng.uniform() v_x0 = rng.uniform()
# compute the output i numpy # compute the output i numpy
v_out = numpy.zeros((4,)) v_out = np.zeros((4,))
v_out[0] = v_u[0] * W_in.get_value() + v_x0 * W.get_value() v_out[0] = v_u[0] * W_in.get_value() + v_x0 * W.get_value()
for step in xrange(1, 4): for step in xrange(1, 4):
v_out[step] = (v_u[step] * W_in.get_value() + v_out[step] = (v_u[step] * W_in.get_value() +
v_out[step - 1] * W.get_value()) v_out[step - 1] * W.get_value())
theano_values = f3(v_u, v_x0) theano_values = f3(v_u, v_x0)
assert numpy.allclose(theano_values, v_out) assert np.allclose(theano_values, v_out)
# some rnn with multiple outputs and multiple inputs; other # some rnn with multiple outputs and multiple inputs; other
# dimension instead of scalars/vectors # dimension instead of scalars/vectors
def test_multiple_inputs_multiple_outputs(self): def test_multiple_inputs_multiple_outputs(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in2 = asarrayX(rng.uniform(size=(2,), low=-5., high=5.)) vW_in2 = asarrayX(rng.uniform(size=(2,), low=-5., high=5.))
vW = asarrayX(rng.uniform(size=(2, 2), low=-5., high=5.)) vW = asarrayX(rng.uniform(size=(2, 2), low=-5., high=5.))
vWout = asarrayX(rng.uniform(size=(2,), low=-5., high=5.)) vWout = asarrayX(rng.uniform(size=(2,), low=-5., high=5.))
...@@ -511,15 +511,15 @@ class T_Scan(unittest.TestCase): ...@@ -511,15 +511,15 @@ class T_Scan(unittest.TestCase):
allow_input_downcast=True) allow_input_downcast=True)
# compute the values in numpy # compute the values in numpy
v_x = numpy.zeros((3, 2), dtype=theano.config.floatX) v_x = np.zeros((3, 2), dtype=theano.config.floatX)
v_y = numpy.zeros((3,), dtype=theano.config.floatX) v_y = np.zeros((3,), dtype=theano.config.floatX)
v_x[0] = (numpy.dot(v_u1[0], vW_in1) + v_u2[0] * vW_in2 + v_x[0] = (np.dot(v_u1[0], vW_in1) + v_u2[0] * vW_in2 +
numpy.dot(v_x0, vW)) np.dot(v_x0, vW))
v_y[0] = numpy.dot(v_x0, vWout) v_y[0] = np.dot(v_x0, vWout)
for i in xrange(1, 3): for i in xrange(1, 3):
v_x[i] = (numpy.dot(v_u1[i], vW_in1) + v_u2[i] * vW_in2 + v_x[i] = (np.dot(v_u1[i], vW_in1) + v_u2[i] * vW_in2 +
numpy.dot(v_x[i - 1], vW)) np.dot(v_x[i - 1], vW))
v_y[i] = numpy.dot(v_x[i - 1], vWout) v_y[i] = np.dot(v_x[i - 1], vWout)
(theano_x, theano_y) = f4(v_u1, v_u2, v_x0, v_y0, vW_in1) (theano_x, theano_y) = f4(v_u1, v_u2, v_x0, v_y0, vW_in1)
utt.assert_allclose(theano_x, v_x) utt.assert_allclose(theano_x, v_x)
...@@ -527,7 +527,7 @@ class T_Scan(unittest.TestCase): ...@@ -527,7 +527,7 @@ class T_Scan(unittest.TestCase):
def test_multiple_outs_taps(self): def test_multiple_outs_taps(self):
l = 5 l = 5
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in2 = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2)) vW_in2 = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2))
vW = asarrayX(rng.uniform(size=(2, 2), low=-.2, high=.2)) vW = asarrayX(rng.uniform(size=(2, 2), low=-.2, high=.2))
vWout = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2)) vWout = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2))
...@@ -579,40 +579,40 @@ class T_Scan(unittest.TestCase): ...@@ -579,40 +579,40 @@ class T_Scan(unittest.TestCase):
v_y0, v_y0,
vW_in1) vW_in1)
ny0 = numpy.zeros((5, 2)) ny0 = np.zeros((5, 2))
ny1 = numpy.zeros((5,)) ny1 = np.zeros((5,))
ny2 = numpy.zeros((5, 2)) ny2 = np.zeros((5, 2))
ny0[0] = numpy.dot(v_u1[0], vW_in1) + \ ny0[0] = np.dot(v_u1[0], vW_in1) + \
(v_u2[1] + v_u2[0] * v_u2[2]) * vW_in2 + numpy.dot(v_x0, vW) (v_u2[1] + v_u2[0] * v_u2[2]) * vW_in2 + np.dot(v_x0, vW)
ny1[0] = (v_y0[2] + v_y0[0]) * numpy.dot(v_x0, vWout) ny1[0] = (v_y0[2] + v_y0[0]) * np.dot(v_x0, vWout)
ny2[0] = numpy.dot(v_u1[0], vW_in1) ny2[0] = np.dot(v_u1[0], vW_in1)
ny0[1] = numpy.dot(v_u1[1], vW_in1) + \ ny0[1] = np.dot(v_u1[1], vW_in1) + \
(v_u2[2] + v_u2[1] * v_u2[3]) * vW_in2 + numpy.dot(ny0[0], vW) (v_u2[2] + v_u2[1] * v_u2[3]) * vW_in2 + np.dot(ny0[0], vW)
ny1[1] = (ny1[0] + v_y0[1]) * numpy.dot(ny0[0], vWout) ny1[1] = (ny1[0] + v_y0[1]) * np.dot(ny0[0], vWout)
ny2[1] = numpy.dot(v_u1[1], vW_in1) ny2[1] = np.dot(v_u1[1], vW_in1)
ny0[2] = numpy.dot(v_u1[2], vW_in1) + \ ny0[2] = np.dot(v_u1[2], vW_in1) + \
(v_u2[3] + v_u2[2] * v_u2[4]) * vW_in2 + \ (v_u2[3] + v_u2[2] * v_u2[4]) * vW_in2 + \
numpy.dot(ny0[1], vW) np.dot(ny0[1], vW)
ny1[2] = (ny1[1] + v_y0[2]) * numpy.dot(ny0[1], vWout) ny1[2] = (ny1[1] + v_y0[2]) * np.dot(ny0[1], vWout)
ny2[2] = numpy.dot(v_u1[2], vW_in1) ny2[2] = np.dot(v_u1[2], vW_in1)
ny0[3] = numpy.dot(v_u1[3], vW_in1) + \ ny0[3] = np.dot(v_u1[3], vW_in1) + \
(v_u2[4] + v_u2[3] * v_u2[5]) * vW_in2 + \ (v_u2[4] + v_u2[3] * v_u2[5]) * vW_in2 + \
numpy.dot(ny0[2], vW) np.dot(ny0[2], vW)
ny1[3] = (ny1[2] + ny1[0]) * numpy.dot(ny0[2], vWout) ny1[3] = (ny1[2] + ny1[0]) * np.dot(ny0[2], vWout)
ny2[3] = numpy.dot(v_u1[3], vW_in1) ny2[3] = np.dot(v_u1[3], vW_in1)
ny0[4] = numpy.dot(v_u1[4], vW_in1) + \ ny0[4] = np.dot(v_u1[4], vW_in1) + \
(v_u2[5] + v_u2[4] * v_u2[6]) * vW_in2 + \ (v_u2[5] + v_u2[4] * v_u2[6]) * vW_in2 + \
numpy.dot(ny0[3], vW) np.dot(ny0[3], vW)
ny1[4] = (ny1[3] + ny1[1]) * numpy.dot(ny0[3], vWout) ny1[4] = (ny1[3] + ny1[1]) * np.dot(ny0[3], vWout)
ny2[4] = numpy.dot(v_u1[4], vW_in1) ny2[4] = np.dot(v_u1[4], vW_in1)
def test_using_taps_sequence(self): def test_using_taps_sequence(self):
# this test refers to a bug reported by Nicolas # this test refers to a bug reported by Nicolas
...@@ -621,9 +621,9 @@ class T_Scan(unittest.TestCase): ...@@ -621,9 +621,9 @@ class T_Scan(unittest.TestCase):
y, updates = theano.scan(lambda x: [x], y, updates = theano.scan(lambda x: [x],
sequences=dict(input=x, taps=[-1]), sequences=dict(input=x, taps=[-1]),
outputs_info=[None]) outputs_info=[None])
inp = numpy.arange(5).astype('float64') inp = np.arange(5).astype('float64')
rval = theano.function([x], y, updates=updates)(inp) rval = theano.function([x], y, updates=updates)(inp)
assert numpy.all(rval == inp[:-1]) assert np.all(rval == inp[:-1])
def test_using_negative_taps_sequence(self): def test_using_negative_taps_sequence(self):
# This test refers to a bug reported on github on May 22 2015 by # This test refers to a bug reported on github on May 22 2015 by
...@@ -636,7 +636,7 @@ class T_Scan(unittest.TestCase): ...@@ -636,7 +636,7 @@ class T_Scan(unittest.TestCase):
f = theano.function([x], res, updates = upd) f = theano.function([x], res, updates = upd)
output = f([1, 2, 3, 4, 5]) output = f([1, 2, 3, 4, 5])
expected_output = numpy.array([1, 2, 3], dtype="float32") expected_output = np.array([1, 2, 3], dtype="float32")
utt.assert_allclose(output, expected_output) utt.assert_allclose(output, expected_output)
def test_connection_pattern(self): def test_connection_pattern(self):
...@@ -649,8 +649,8 @@ class T_Scan(unittest.TestCase): ...@@ -649,8 +649,8 @@ class T_Scan(unittest.TestCase):
def fn(a_m2, a_m1, b_m2, b_m1): def fn(a_m2, a_m1, b_m2, b_m1):
return a_m1, b_m1 return a_m1, b_m1
a0 = theano.shared(numpy.arange(2)) a0 = theano.shared(np.arange(2))
b0 = theano.shared(numpy.arange(2)) b0 = theano.shared(np.arange(2))
(a, b), _ = theano.scan(fn, (a, b), _ = theano.scan(fn,
outputs_info=[{'initial': a0, 'taps': [-2, -1]}, outputs_info=[{'initial': a0, 'taps': [-2, -1]},
...@@ -741,7 +741,7 @@ class T_Scan(unittest.TestCase): ...@@ -741,7 +741,7 @@ class T_Scan(unittest.TestCase):
# Call verify_grad to ensure the correctness of the second gradients # Call verify_grad to ensure the correctness of the second gradients
floatX = theano.config.floatX floatX = theano.config.floatX
inputs_test_values = [numpy.random.random((3)).astype(floatX)] inputs_test_values = [np.random.random((3)).astype(floatX)]
theano.tests.unittest_tools.verify_grad(get_sum_of_grad, theano.tests.unittest_tools.verify_grad(get_sum_of_grad,
inputs_test_values) inputs_test_values)
...@@ -768,8 +768,8 @@ class T_Scan(unittest.TestCase): ...@@ -768,8 +768,8 @@ class T_Scan(unittest.TestCase):
# Call verify_grad to ensure the correctness of the second gradients # Call verify_grad to ensure the correctness of the second gradients
floatX = theano.config.floatX floatX = theano.config.floatX
inputs_test_values = [numpy.random.random((2, 3)).astype(floatX), inputs_test_values = [np.random.random((2, 3)).astype(floatX),
numpy.random.random((3)).astype(floatX)] np.random.random((3)).astype(floatX)]
theano.tests.unittest_tools.verify_grad(get_sum_of_grad, theano.tests.unittest_tools.verify_grad(get_sum_of_grad,
inputs_test_values) inputs_test_values)
...@@ -781,7 +781,7 @@ class T_Scan(unittest.TestCase): ...@@ -781,7 +781,7 @@ class T_Scan(unittest.TestCase):
# forward pass # forward pass
W = theano.shared( W = theano.shared(
numpy.random.randn(2, 2).astype('float32'), np.random.randn(2, 2).astype('float32'),
name="W", borrow=True) name="W", borrow=True)
def forward_scanner(x_t): def forward_scanner(x_t):
...@@ -807,7 +807,7 @@ class T_Scan(unittest.TestCase): ...@@ -807,7 +807,7 @@ class T_Scan(unittest.TestCase):
# vectors, weights are scalars; using shared variables and past # vectors, weights are scalars; using shared variables and past
# taps (sequences and outputs) # taps (sequences and outputs)
def test_using_taps_input_output(self): def test_using_taps_input_output(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW = asarrayX(rng.uniform()) vW = asarrayX(rng.uniform())
vW_in = asarrayX(rng.uniform()) vW_in = asarrayX(rng.uniform())
vu = asarrayX(rng.uniform(size=(4,), low=-5., high=5.)) vu = asarrayX(rng.uniform(size=(4,), low=-5., high=5.))
...@@ -843,7 +843,7 @@ class T_Scan(unittest.TestCase): ...@@ -843,7 +843,7 @@ class T_Scan(unittest.TestCase):
# in scan) which might seem strange, but then again why not use # in scan) which might seem strange, but then again why not use
# v_0[t] instead of v_0[t-2] in a real application ?? # v_0[t] instead of v_0[t-2] in a real application ??
# also vx0[0] corresponds to vx0[-2], vx0[1] to vx0[-1] # also vx0[0] corresponds to vx0[-2], vx0[1] to vx0[-1]
numpy_out = numpy.zeros((2,)) numpy_out = np.zeros((2,))
numpy_out[0] = vu[0] * vW_in + vx0[1] * vW + vx0[0] numpy_out[0] = vu[0] * vW_in + vx0[1] * vW + vx0[0]
numpy_out[1] = vu[1] * vW_in + numpy_out[0] * vW + vx0[1] numpy_out[1] = vu[1] * vW_in + numpy_out[0] * vW + vx0[1]
utt.assert_allclose(numpy_out, theano_out) utt.assert_allclose(numpy_out, theano_out)
...@@ -852,7 +852,7 @@ class T_Scan(unittest.TestCase): ...@@ -852,7 +852,7 @@ class T_Scan(unittest.TestCase):
# vectors, weights are scalars; using shared variables and past # vectors, weights are scalars; using shared variables and past
# taps (sequences and outputs) and future taps for sequences # taps (sequences and outputs) and future taps for sequences
def test_past_future_taps_shared(self): def test_past_future_taps_shared(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW = asarrayX(rng.uniform()) vW = asarrayX(rng.uniform())
vW_in = asarrayX(rng.uniform()) vW_in = asarrayX(rng.uniform())
vu = asarrayX(rng.uniform(size=(6,), low=-5., high=5.)) vu = asarrayX(rng.uniform(size=(6,), low=-5., high=5.))
...@@ -880,7 +880,7 @@ class T_Scan(unittest.TestCase): ...@@ -880,7 +880,7 @@ class T_Scan(unittest.TestCase):
allow_input_downcast=True) allow_input_downcast=True)
theano_out = f8(vu, vx0) theano_out = f8(vu, vx0)
# compute output in numpy # compute output in numpy
numpy_out = numpy.zeros(2) numpy_out = np.zeros(2)
# think of vu[0] as vu[-2], vu[4] as vu[2] # think of vu[0] as vu[-2], vu[4] as vu[2]
# and vx0[0] as vx0[-2], vx0[1] as vx0[-1] # and vx0[0] as vx0[-2], vx0[1] as vx0[-1]
numpy_out[0] = (vu[0] + vu[4]) * vW_in + vx0[1] * vW + vx0[0] numpy_out[0] = (vu[0] + vu[4]) * vW_in + vx0[1] * vW + vx0[0]
...@@ -889,9 +889,9 @@ class T_Scan(unittest.TestCase): ...@@ -889,9 +889,9 @@ class T_Scan(unittest.TestCase):
# simple rnn ; compute inplace version 1 # simple rnn ; compute inplace version 1
def test_inplace1(self): def test_inplace1(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW = asarrayX(numpy.random.uniform()) vW = asarrayX(np.random.uniform())
vW_in = asarrayX(numpy.random.uniform()) vW_in = asarrayX(np.random.uniform())
vu0 = asarrayX(rng.uniform(size=(3,), low=-5., high=5.)) vu0 = asarrayX(rng.uniform(size=(3,), low=-5., high=5.))
vu1 = asarrayX(rng.uniform(size=(3,), low=-5., high=5.)) vu1 = asarrayX(rng.uniform(size=(3,), low=-5., high=5.))
vu2 = asarrayX(rng.uniform(size=(3,), low=-5., high=5.)) vu2 = asarrayX(rng.uniform(size=(3,), low=-5., high=5.))
...@@ -934,8 +934,8 @@ class T_Scan(unittest.TestCase): ...@@ -934,8 +934,8 @@ class T_Scan(unittest.TestCase):
assert 0 in scan_node[0].op.destroy_map.keys() assert 0 in scan_node[0].op.destroy_map.keys()
assert 1 in scan_node[0].op.destroy_map.keys() assert 1 in scan_node[0].op.destroy_map.keys()
# compute output in numpy # compute output in numpy
numpy_x0 = numpy.zeros((3,)) numpy_x0 = np.zeros((3,))
numpy_x1 = numpy.zeros((3,)) numpy_x1 = np.zeros((3,))
numpy_x0[0] = vu0[0] * vW_in + vx0 * vW + vu1[0] * vu2[0] numpy_x0[0] = vu0[0] * vW_in + vx0 * vW + vu1[0] * vu2[0]
numpy_x1[0] = vu0[0] * vW_in + vx1 * vW + vu1[0] + vu2[0] numpy_x1[0] = vu0[0] * vW_in + vx1 * vW + vu1[0] + vu2[0]
for i in xrange(1, 3): for i in xrange(1, 3):
...@@ -953,9 +953,9 @@ class T_Scan(unittest.TestCase): ...@@ -953,9 +953,9 @@ class T_Scan(unittest.TestCase):
# simple rnn ; compute inplace version 2 # simple rnn ; compute inplace version 2
def test_inplace2(self): def test_inplace2(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW = asarrayX(numpy.random.uniform()) vW = asarrayX(np.random.uniform())
vW_in = asarrayX(numpy.random.uniform()) vW_in = asarrayX(np.random.uniform())
vu0 = asarrayX(rng.uniform(size=(3,), low=-5., high=5.)) vu0 = asarrayX(rng.uniform(size=(3,), low=-5., high=5.))
vu1 = asarrayX(rng.uniform(size=(4,), low=-5., high=5.)) vu1 = asarrayX(rng.uniform(size=(4,), low=-5., high=5.))
vu2 = asarrayX(rng.uniform(size=(5,), low=-5., high=5.)) vu2 = asarrayX(rng.uniform(size=(5,), low=-5., high=5.))
...@@ -1006,8 +1006,8 @@ class T_Scan(unittest.TestCase): ...@@ -1006,8 +1006,8 @@ class T_Scan(unittest.TestCase):
assert 0 in scan_node[0].op.destroy_map.keys() assert 0 in scan_node[0].op.destroy_map.keys()
assert 1 in scan_node[0].op.destroy_map.keys() assert 1 in scan_node[0].op.destroy_map.keys()
# compute output in numpy # compute output in numpy
numpy_x0 = numpy.zeros((3,)) numpy_x0 = np.zeros((3,))
numpy_x1 = numpy.zeros((3,)) numpy_x1 = np.zeros((3,))
numpy_x0[0] = vu0[0] * vW_in + vx0 * vW + vu1[0] * vu1[1] numpy_x0[0] = vu0[0] * vW_in + vx0 * vW + vu1[0] * vu1[1]
numpy_x1[0] = vu0[0] * vW_in + vx1 * vW + vu2[0] + vu2[1] + vu2[2] numpy_x1[0] = vu0[0] * vW_in + vx1 * vW + vu2[0] + vu2[1] + vu2[2]
for i in xrange(1, 3): for i in xrange(1, 3):
...@@ -1024,7 +1024,7 @@ class T_Scan(unittest.TestCase): ...@@ -1024,7 +1024,7 @@ class T_Scan(unittest.TestCase):
utt.assert_allclose(theano_x1, numpy_x1) utt.assert_allclose(theano_x1, numpy_x1)
def test_inplace3(self): def test_inplace3(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vx0 = asarrayX(rng.uniform()) vx0 = asarrayX(rng.uniform())
vx1 = asarrayX(rng.uniform()) vx1 = asarrayX(rng.uniform())
...@@ -1035,7 +1035,7 @@ class T_Scan(unittest.TestCase): ...@@ -1035,7 +1035,7 @@ class T_Scan(unittest.TestCase):
[], [],
[x0, x1], [x0, x1],
n_steps=3) n_steps=3)
x0 = asarrayX(numpy.zeros((3,))) x0 = asarrayX(np.zeros((3,)))
x0[0] = vx0 x0[0] = vx0
x0 = theano.tensor.constant(x0) x0 = theano.tensor.constant(x0)
to_replace = outputs[0].owner.inputs[0].owner.inputs[1] to_replace = outputs[0].owner.inputs[0].owner.inputs[1]
...@@ -1053,7 +1053,7 @@ class T_Scan(unittest.TestCase): ...@@ -1053,7 +1053,7 @@ class T_Scan(unittest.TestCase):
# Shared variable with updates # Shared variable with updates
def test_shared_arguments_with_updates(self): def test_shared_arguments_with_updates(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW1 = asarrayX(rng.rand(2, 3)) vW1 = asarrayX(rng.rand(2, 3))
vW2 = asarrayX(rng.rand(3, 2)) vW2 = asarrayX(rng.rand(3, 2))
...@@ -1128,22 +1128,22 @@ class T_Scan(unittest.TestCase): ...@@ -1128,22 +1128,22 @@ class T_Scan(unittest.TestCase):
theano_y0, theano_y1, theano_y2 = allstuff theano_y0, theano_y1, theano_y2 = allstuff
# do things in numpy # do things in numpy
numpy_y0 = numpy.zeros((6, 2)) numpy_y0 = np.zeros((6, 2))
numpy_y1 = numpy.zeros((4, 2)) numpy_y1 = np.zeros((4, 2))
numpy_y2 = numpy.zeros((3, 3)) numpy_y2 = np.zeros((3, 3))
numpy_y0[:3] = vy0 numpy_y0[:3] = vy0
numpy_y1[0] = vy1 numpy_y1[0] = vy1
numpy_W1 = vW1.copy() numpy_W1 = vW1.copy()
numpy_W2 = vW2.copy() numpy_W2 = vW2.copy()
for idx in xrange(3): for idx in xrange(3):
numpy_y0[idx + 3] = numpy.dot(numpy.dot(vu1[idx, :], numpy_W1), numpy_y0[idx + 3] = np.dot(np.dot(vu1[idx, :], numpy_W1),
numpy_W2) + \ numpy_W2) + \
0.1 * numpy_y0[idx + 2] + \ 0.1 * numpy_y0[idx + 2] + \
0.33 * numpy_y0[idx + 1] + \ 0.33 * numpy_y0[idx + 1] + \
0.17 * numpy_y0[idx] 0.17 * numpy_y0[idx]
numpy_y1[idx + 1] = (numpy.dot(vu2[idx, :], numpy_W2) + numpy_y1[idx + 1] = (np.dot(vu2[idx, :], numpy_W2) +
numpy_y1[idx]) numpy_y1[idx])
numpy_y2[idx] = numpy.dot(vu1[idx, :], numpy_W1) numpy_y2[idx] = np.dot(vu1[idx, :], numpy_W1)
numpy_W1 = numpy_W1 + .1 numpy_W1 = numpy_W1 + .1
numpy_W2 = numpy_W2 + .05 numpy_W2 = numpy_W2 + .05
...@@ -1174,7 +1174,7 @@ class T_Scan(unittest.TestCase): ...@@ -1174,7 +1174,7 @@ class T_Scan(unittest.TestCase):
f = theano.function([c, x, y], [gX, gY], f = theano.function([c, x, y], [gX, gY],
allow_input_downcast=True) allow_input_downcast=True)
# Check for runtime errors # Check for runtime errors
f(numpy.int32(0), numpy.float32(1.), numpy.float32(.5)) f(np.int32(0), np.float32(1.), np.float32(.5))
def test_simple_shared_mrg_random(self): def test_simple_shared_mrg_random(self):
theano_rng = theano.sandbox.rng_mrg.MRG_RandomStreams(utt.fetch_seed()) theano_rng = theano.sandbox.rng_mrg.MRG_RandomStreams(utt.fetch_seed())
...@@ -1211,10 +1211,10 @@ class T_Scan(unittest.TestCase): ...@@ -1211,10 +1211,10 @@ class T_Scan(unittest.TestCase):
updates=updates, updates=updates,
allow_input_downcast=True) allow_input_downcast=True)
rng_seed = numpy.random.RandomState(utt.fetch_seed()).randint(2 ** 30) rng_seed = np.random.RandomState(utt.fetch_seed()).randint(2 ** 30)
rng = numpy.random.RandomState(int(rng_seed)) # int() is for 32bit rng = np.random.RandomState(int(rng_seed)) # int() is for 32bit
numpy_v = numpy.zeros((10, 2)) numpy_v = np.zeros((10, 2))
for i in xrange(10): for i in xrange(10):
numpy_v[i] = rng.uniform(-1, 1, size=(2,)) numpy_v[i] = rng.uniform(-1, 1, size=(2,))
...@@ -1224,12 +1224,12 @@ class T_Scan(unittest.TestCase): ...@@ -1224,12 +1224,12 @@ class T_Scan(unittest.TestCase):
utt.assert_allclose(theano_v, numpy_v[5:, :]) utt.assert_allclose(theano_v, numpy_v[5:, :])
def test_gibbs_chain(self): def test_gibbs_chain(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_W = numpy.array(rng.rand(20, 30) - .5, dtype='float32') v_W = np.array(rng.rand(20, 30) - .5, dtype='float32')
v_vsample = numpy.array(rng.binomial(1, .5, size=(3, 20),), v_vsample = np.array(rng.binomial(1, .5, size=(3, 20),),
dtype='float32') dtype='float32')
v_bvis = numpy.array(rng.rand(20) - .5, dtype='float32') v_bvis = np.array(rng.rand(20) - .5, dtype='float32')
v_bhid = numpy.array(rng.rand(30) - .5, dtype='float32') v_bhid = np.array(rng.rand(30) - .5, dtype='float32')
W = theano.shared(v_W, 'vW') W = theano.shared(v_W, 'vW')
bhid = theano.shared(v_bhid, 'vbhid') bhid = theano.shared(v_bhid, 'vbhid')
bvis = theano.shared(v_bvis, 'vbvis') bvis = theano.shared(v_bvis, 'vbvis')
...@@ -1261,24 +1261,24 @@ class T_Scan(unittest.TestCase): ...@@ -1261,24 +1261,24 @@ class T_Scan(unittest.TestCase):
updates=updates, updates=updates,
allow_input_downcast=True) allow_input_downcast=True)
_rng = numpy.random.RandomState(utt.fetch_seed()) _rng = np.random.RandomState(utt.fetch_seed())
rng_seed = _rng.randint(2 ** 30) rng_seed = _rng.randint(2 ** 30)
nrng1 = numpy.random.RandomState(int(rng_seed)) # int() is for 32bit nrng1 = np.random.RandomState(int(rng_seed)) # int() is for 32bit
rng_seed = _rng.randint(2 ** 30) rng_seed = _rng.randint(2 ** 30)
nrng2 = numpy.random.RandomState(int(rng_seed)) # int() is for 32bit nrng2 = np.random.RandomState(int(rng_seed)) # int() is for 32bit
def numpy_implementation(vsample): def numpy_implementation(vsample):
for idx in range(10): for idx in range(10):
hmean = 1. / (1. + numpy.exp(-(numpy.dot(vsample, v_W) +\ hmean = 1. / (1. + np.exp(-(np.dot(vsample, v_W) +\
v_bhid))) v_bhid)))
hsample = numpy.array(nrng1.binomial(1, hsample = np.array(nrng1.binomial(1,
hmean, hmean,
size=hmean.shape), size=hmean.shape),
dtype='float32') dtype='float32')
vmean = 1. / (1. + numpy.exp(-(numpy.dot(hsample, v_W.T) +\ vmean = 1. / (1. + np.exp(-(np.dot(hsample, v_W.T) +\
v_bvis))) v_bvis)))
vsample = numpy.array(nrng2.binomial(1, vsample = np.array(nrng2.binomial(1,
vmean, vmean,
size=vmean.shape), size=vmean.shape),
dtype='float32') dtype='float32')
...@@ -1290,7 +1290,7 @@ class T_Scan(unittest.TestCase): ...@@ -1290,7 +1290,7 @@ class T_Scan(unittest.TestCase):
utt.assert_allclose(t_result, n_result) utt.assert_allclose(t_result, n_result)
def test_only_shared_no_input_no_output(self): def test_only_shared_no_input_no_output(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_state = asarrayX(rng.uniform()) v_state = asarrayX(rng.uniform())
state = theano.shared(v_state, 'vstate') state = theano.shared(v_state, 'vstate')
...@@ -1331,7 +1331,7 @@ class T_Scan(unittest.TestCase): ...@@ -1331,7 +1331,7 @@ class T_Scan(unittest.TestCase):
outputs, outputs,
updates=updates, updates=updates,
allow_input_downcast=True) allow_input_downcast=True)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = rng.uniform(size=(5,), low=-5., high=5.) v_u = rng.uniform(size=(5,), low=-5., high=5.)
numpy_result = v_u + 3 numpy_result = v_u + 3
...@@ -1352,7 +1352,7 @@ class T_Scan(unittest.TestCase): ...@@ -1352,7 +1352,7 @@ class T_Scan(unittest.TestCase):
updates=abs_updates, updates=abs_updates,
allow_input_downcast=True) allow_input_downcast=True)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vals = rng.uniform(size=(10,), low=-5., high=5.) vals = rng.uniform(size=(10,), low=-5., high=5.)
abs_vals = abs(vals) abs_vals = abs(vals)
theano_vals = f(vals) theano_vals = f(vals)
...@@ -1380,14 +1380,14 @@ class T_Scan(unittest.TestCase): ...@@ -1380,14 +1380,14 @@ class T_Scan(unittest.TestCase):
updates=updates, updates=updates,
allow_input_downcast=True) allow_input_downcast=True)
# get random initial values # get random initial values
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = rng.uniform(size=(4,), low=-5., high=5.) v_u = rng.uniform(size=(4,), low=-5., high=5.)
v_x0 = rng.uniform() v_x0 = rng.uniform()
W = rng.uniform() W = rng.uniform()
W_in = rng.uniform() W_in = rng.uniform()
# compute the output in numpy # compute the output in numpy
v_out = numpy.zeros((4,)) v_out = np.zeros((4,))
v_out[0] = v_u[3] * W_in + v_x0 * W v_out[0] = v_u[3] * W_in + v_x0 * W
for step in xrange(1, 4): for step in xrange(1, 4):
v_out[step] = v_u[3 - step] * W_in + v_out[step - 1] * W v_out[step] = v_u[3 - step] * W_in + v_out[step - 1] * W
...@@ -1404,9 +1404,9 @@ class T_Scan(unittest.TestCase): ...@@ -1404,9 +1404,9 @@ class T_Scan(unittest.TestCase):
result, result,
updates=updates, updates=updates,
allow_input_downcast=True) allow_input_downcast=True)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_v = rng.uniform(size=(5,), low=-5., high=5.) v_v = rng.uniform(size=(5,), low=-5., high=5.)
assert abs(numpy.sum(v_v) - f(v_v, 0.)) < 1e-3 assert abs(np.sum(v_v) - f(v_v, 0.)) < 1e-3
def test_grad_one_output(self): def test_grad_one_output(self):
def f_rnn(u_t, x_tm1, W_in, W): def f_rnn(u_t, x_tm1, W_in, W):
...@@ -1440,12 +1440,12 @@ class T_Scan(unittest.TestCase): ...@@ -1440,12 +1440,12 @@ class T_Scan(unittest.TestCase):
allow_input_downcast=True) allow_input_downcast=True)
# get random initial values # get random initial values
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = numpy.array(rng.uniform(size=(10,), low=-.5, high=.5), v_u = np.array(rng.uniform(size=(10,), low=-.5, high=.5),
dtype=theano.config.floatX) dtype=theano.config.floatX)
v_x0 = numpy.array(rng.uniform(), dtype=theano.config.floatX) v_x0 = np.array(rng.uniform(), dtype=theano.config.floatX)
W = numpy.array(rng.uniform(), dtype=theano.config.floatX) W = np.array(rng.uniform(), dtype=theano.config.floatX)
W_in = numpy.array(rng.uniform(), dtype=theano.config.floatX) W_in = np.array(rng.uniform(), dtype=theano.config.floatX)
analytic_grad = grad_fn(v_u, v_x0, W_in, W) analytic_grad = grad_fn(v_u, v_x0, W_in, W)
num_grad = multiple_outputs_numeric_grad( num_grad = multiple_outputs_numeric_grad(
...@@ -1459,7 +1459,7 @@ class T_Scan(unittest.TestCase): ...@@ -1459,7 +1459,7 @@ class T_Scan(unittest.TestCase):
num_grad.gx[max_err_pos])) num_grad.gx[max_err_pos]))
def test_grad_multiple_outs(self): def test_grad_multiple_outs(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in2 = asarrayX(rng.uniform(size=(2,), low=-.1, high=.1)) vW_in2 = asarrayX(rng.uniform(size=(2,), low=-.1, high=.1))
vW = asarrayX(rng.uniform(size=(2, 2), low=-.1, high=.1)) vW = asarrayX(rng.uniform(size=(2, 2), low=-.1, high=.1))
vWout = asarrayX(rng.uniform(size=(2,), low=-.1, high=.1)) vWout = asarrayX(rng.uniform(size=(2,), low=-.1, high=.1))
...@@ -1524,7 +1524,7 @@ class T_Scan(unittest.TestCase): ...@@ -1524,7 +1524,7 @@ class T_Scan(unittest.TestCase):
@attr('slow') @attr('slow')
def test_grad_multiple_outs_taps(self): def test_grad_multiple_outs_taps(self):
l = 5 l = 5
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in2 = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2)) vW_in2 = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2))
vW = asarrayX(rng.uniform(size=(2, 2), low=-.2, high=.2)) vW = asarrayX(rng.uniform(size=(2, 2), low=-.2, high=.2))
vWout = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2)) vWout = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2))
...@@ -1618,7 +1618,7 @@ class T_Scan(unittest.TestCase): ...@@ -1618,7 +1618,7 @@ class T_Scan(unittest.TestCase):
@attr('slow') @attr('slow')
def test_grad_multiple_outs_taps_backwards(self): def test_grad_multiple_outs_taps_backwards(self):
l = 5 l = 5
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in2 = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2)) vW_in2 = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2))
vW = asarrayX(rng.uniform(size=(2, 2), low=-.2, high=.2)) vW = asarrayX(rng.uniform(size=(2, 2), low=-.2, high=.2))
vWout = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2)) vWout = asarrayX(rng.uniform(size=(2,), low=-.2, high=.2))
...@@ -1685,10 +1685,10 @@ class T_Scan(unittest.TestCase): ...@@ -1685,10 +1685,10 @@ class T_Scan(unittest.TestCase):
num_grad.gx[max_err_pos])) num_grad.gx[max_err_pos]))
def test_grad_multiple_outs_some_uncomputable(self): def test_grad_multiple_outs_some_uncomputable(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in = asarrayX(rng.uniform(size=(2, 2), low=-3., high=3.)) vW_in = asarrayX(rng.uniform(size=(2, 2), low=-3., high=3.))
v_u = asarrayX(rng.uniform(size=(5, 2), low=-3., high=3.)) v_u = asarrayX(rng.uniform(size=(5, 2), low=-3., high=3.))
v_u2 = numpy.array([1, 3, 4, 6, 8], dtype='int32') v_u2 = np.array([1, 3, 4, 6, 8], dtype='int32')
v_x0 = asarrayX(rng.uniform(size=(2,), low=-3., high=3.)) v_x0 = asarrayX(rng.uniform(size=(2,), low=-3., high=3.))
W_in = theano.tensor.matrix('win') W_in = theano.tensor.matrix('win')
...@@ -1730,9 +1730,9 @@ class T_Scan(unittest.TestCase): ...@@ -1730,9 +1730,9 @@ class T_Scan(unittest.TestCase):
def reset_rng_fn(fn, *args): def reset_rng_fn(fn, *args):
for idx, arg in enumerate(fn.maker.expanded_inputs): for idx, arg in enumerate(fn.maker.expanded_inputs):
if (arg.value and type(arg.value.data) == \ if (arg.value and type(arg.value.data) == \
type(numpy.random.RandomState(123))): type(np.random.RandomState(123))):
obj = fn.maker.expanded_inputs[idx].value obj = fn.maker.expanded_inputs[idx].value
obj.data = numpy.random.RandomState(123) obj.data = np.random.RandomState(123)
fn.maker.expanded_inputs[idx].value = obj fn.maker.expanded_inputs[idx].value = obj
return fn(*args) return fn(*args)
...@@ -1764,7 +1764,7 @@ class T_Scan(unittest.TestCase): ...@@ -1764,7 +1764,7 @@ class T_Scan(unittest.TestCase):
assert(result == expected_result) assert(result == expected_result)
def test_grad_multiple_outs_some_truncate(self): def test_grad_multiple_outs_some_truncate(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in = asarrayX(rng.uniform(size=(2, 2), low=-.1, high=.1)) vW_in = asarrayX(rng.uniform(size=(2, 2), low=-.1, high=.1))
v_u = asarrayX(rng.uniform(size=(5, 2), low=-.1, high=.1)) v_u = asarrayX(rng.uniform(size=(5, 2), low=-.1, high=.1))
v_x0 = asarrayX(rng.uniform(size=(2,), low=-.1, high=.1)) v_x0 = asarrayX(rng.uniform(size=(2,), low=-.1, high=.1))
...@@ -1807,9 +1807,9 @@ class T_Scan(unittest.TestCase): ...@@ -1807,9 +1807,9 @@ class T_Scan(unittest.TestCase):
def reset_rng_fn(fn, *args): def reset_rng_fn(fn, *args):
for idx, arg in enumerate(fn.maker.expanded_inputs): for idx, arg in enumerate(fn.maker.expanded_inputs):
if (arg.value and if (arg.value and
isinstance(arg.value.data, numpy.random.RandomState)): isinstance(arg.value.data, np.random.RandomState)):
obj = fn.maker.expanded_inputs[idx].value obj = fn.maker.expanded_inputs[idx].value
obj.data = numpy.random.RandomState(123) obj.data = np.random.RandomState(123)
fn.maker.expanded_inputs[idx].value = obj fn.maker.expanded_inputs[idx].value = obj
out = fn(*args) out = fn(*args)
return out return out
...@@ -1819,7 +1819,7 @@ class T_Scan(unittest.TestCase): ...@@ -1819,7 +1819,7 @@ class T_Scan(unittest.TestCase):
num_grad = multiple_outputs_numeric_grad( num_grad = multiple_outputs_numeric_grad(
reset_rng_cost_fn, [v_u, v_x0, vW_in]) reset_rng_cost_fn, [v_u, v_x0, vW_in])
analytic_grad = reset_rng_grad_fn(v_u, v_x0, vW_in) analytic_grad = reset_rng_grad_fn(v_u, v_x0, vW_in)
utt.assert_allclose(analytic_grad[0][:2], numpy.zeros((2, 2))) utt.assert_allclose(analytic_grad[0][:2], np.zeros((2, 2)))
def test_grad_multiple_outs_some_disconnected(self): def test_grad_multiple_outs_some_disconnected(self):
final_cost = self._grad_mout_helper(100, mode_nodebug) final_cost = self._grad_mout_helper(100, mode_nodebug)
...@@ -1833,7 +1833,7 @@ class T_Scan(unittest.TestCase): ...@@ -1833,7 +1833,7 @@ class T_Scan(unittest.TestCase):
def _grad_mout_helper(self, n_iters, mode): def _grad_mout_helper(self, n_iters, mode):
# Created on Tue Oct 07 13:28:51 2014 # Created on Tue Oct 07 13:28:51 2014
# @author: vaneetke # @author: vaneetke
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
n_hid = 3 n_hid = 3
n_in = 1 n_in = 1
n_out = 1 n_out = 1
...@@ -1897,10 +1897,10 @@ class T_Scan(unittest.TestCase): ...@@ -1897,10 +1897,10 @@ class T_Scan(unittest.TestCase):
mode=mode) mode=mode)
# artificial data # artificial data
x_v = numpy.arange(0., 10.49, 0.21, dtype=theano.config.floatX) x_v = np.arange(0., 10.49, 0.21, dtype=theano.config.floatX)
x_v = x_v.reshape(len(x_v), 1) x_v = x_v.reshape(len(x_v), 1)
s_v = numpy.sin(x_v) s_v = np.sin(x_v)
t_v = numpy.roll(s_v, -1)[:-1] t_v = np.roll(s_v, -1)[:-1]
s_v = s_v[:-1] s_v = s_v[:-1]
for i in xrange(n_iters): for i in xrange(n_iters):
cost = learn_rnn_fn(s_v, t_v) cost = learn_rnn_fn(s_v, t_v)
...@@ -1919,14 +1919,14 @@ class T_Scan(unittest.TestCase): ...@@ -1919,14 +1919,14 @@ class T_Scan(unittest.TestCase):
updates=updates, updates=updates,
allow_input_downcast=True) allow_input_downcast=True)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
nx = rng.uniform(size=(10, 10)) nx = rng.uniform(size=(10, 10))
ny1, nz1 = f(nx) ny1, nz1 = f(nx)
ny2, nz2 = f(nx) ny2, nz2 = f(nx)
utt.assert_allclose([ny1, ny1], nz1) utt.assert_allclose([ny1, ny1], nz1)
utt.assert_allclose([ny2, ny2], nz2) utt.assert_allclose([ny2, ny2], nz2)
assert not numpy.allclose(ny1, ny2) assert not np.allclose(ny1, ny2)
def test_grad_of_shared(self): def test_grad_of_shared(self):
x1 = theano.shared(3.) x1 = theano.shared(3.)
...@@ -1942,7 +1942,7 @@ class T_Scan(unittest.TestCase): ...@@ -1942,7 +1942,7 @@ class T_Scan(unittest.TestCase):
def test_computing_gradient(self): def test_computing_gradient(self):
x1 = theano.tensor.scalar('x1') x1 = theano.tensor.scalar('x1')
x2 = theano.shared(numpy.array([1, 2, 3, 4, 5]), name='x2') x2 = theano.shared(np.array([1, 2, 3, 4, 5]), name='x2')
K = x2 * x1 K = x2 * x1
out, updates = theano.scan(lambda i, v: theano.tensor.grad(K[i], v), out, updates = theano.scan(lambda i, v: theano.tensor.grad(K[i], v),
...@@ -1950,10 +1950,10 @@ class T_Scan(unittest.TestCase): ...@@ -1950,10 +1950,10 @@ class T_Scan(unittest.TestCase):
non_sequences=x1) non_sequences=x1)
f = theano.function([x1], out, allow_input_downcast=True) f = theano.function([x1], out, allow_input_downcast=True)
assert numpy.all(f(3.) != 0.) assert np.all(f(3.) != 0.)
def test_shared_updates(self): def test_shared_updates(self):
X = theano.shared(numpy.array(1)) X = theano.shared(np.array(1))
out, updates = theano.scan( out, updates = theano.scan(
lambda: OrderedDict([(X, (X + 1))]), lambda: OrderedDict([(X, (X + 1))]),
...@@ -1967,8 +1967,8 @@ class T_Scan(unittest.TestCase): ...@@ -1967,8 +1967,8 @@ class T_Scan(unittest.TestCase):
assert X.get_value() == 11 assert X.get_value() == 11
def test_memory_aliasing_updates(self): def test_memory_aliasing_updates(self):
x = theano.shared(numpy.array(1)) x = theano.shared(np.array(1))
y = theano.shared(numpy.array(1)) y = theano.shared(np.array(1))
out, updates = theano.scan( out, updates = theano.scan(
lambda: OrderedDict([(x, x + 1), (y, x)]), lambda: OrderedDict([(x, x + 1), (y, x)]),
...@@ -1979,7 +1979,7 @@ class T_Scan(unittest.TestCase): ...@@ -1979,7 +1979,7 @@ class T_Scan(unittest.TestCase):
f = theano.function([], [], updates=updates) f = theano.function([], [], updates=updates)
f() f()
assert not numpy.may_share_memory(x.container.storage[0], assert not np.may_share_memory(x.container.storage[0],
y.container.storage[0]) y.container.storage[0])
assert x.get_value() != y.get_value() assert x.get_value() != y.get_value()
...@@ -1998,7 +1998,7 @@ class T_Scan(unittest.TestCase): ...@@ -1998,7 +1998,7 @@ class T_Scan(unittest.TestCase):
""" """
a = theano.tensor.vector() a = theano.tensor.vector()
init_a = theano.tensor.vector() init_a = theano.tensor.vector()
b = theano.shared(numpy.random.rand(5, 4)) b = theano.shared(np.random.rand(5, 4))
def inner_func(a): def inner_func(a):
return a + 1, OrderedDict([(b, 2 * b)]) return a + 1, OrderedDict([(b, 2 * b)])
...@@ -2032,9 +2032,9 @@ class T_Scan(unittest.TestCase): ...@@ -2032,9 +2032,9 @@ class T_Scan(unittest.TestCase):
non_sequences=[gy, x]) non_sequences=[gy, x])
f = theano.function([x, A], hy, allow_input_downcast=True) f = theano.function([x, A], hy, allow_input_downcast=True)
vx = numpy.array([1., 1.], dtype=theano.config.floatX) vx = np.array([1., 1.], dtype=theano.config.floatX)
vA = numpy.array([[1., 1.], [1., 0.]], dtype=theano.config.floatX) vA = np.array([[1., 1.], [1., 0.]], dtype=theano.config.floatX)
vR = numpy.array([[3.6, 1.8], [1.8, 0.9]], dtype=theano.config.floatX) vR = np.array([[3.6, 1.8], [1.8, 0.9]], dtype=theano.config.floatX)
out = f(vx, vA) out = f(vx, vA)
utt.assert_allclose(out, vR) utt.assert_allclose(out, vR)
...@@ -2157,7 +2157,7 @@ class T_Scan(unittest.TestCase): ...@@ -2157,7 +2157,7 @@ class T_Scan(unittest.TestCase):
# some rnn with multiple outputs and multiple inputs; other # some rnn with multiple outputs and multiple inputs; other
# dimension instead of scalars/vectors # dimension instead of scalars/vectors
def test_reordering(self): def test_reordering(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in2 = asarrayX(rng.uniform(size=(2,), low=-5., high=5.)) vW_in2 = asarrayX(rng.uniform(size=(2,), low=-5., high=5.))
vW = asarrayX(rng.uniform(size=(2, 2), low=-5., high=5.)) vW = asarrayX(rng.uniform(size=(2, 2), low=-5., high=5.))
vWout = asarrayX(rng.uniform(size=(2,), low=-5., high=5.)) vWout = asarrayX(rng.uniform(size=(2,), low=-5., high=5.))
...@@ -2200,15 +2200,15 @@ class T_Scan(unittest.TestCase): ...@@ -2200,15 +2200,15 @@ class T_Scan(unittest.TestCase):
allow_input_downcast=True) allow_input_downcast=True)
# compute the values in numpy # compute the values in numpy
v_x = numpy.zeros((3, 2), dtype=theano.config.floatX) v_x = np.zeros((3, 2), dtype=theano.config.floatX)
v_y = numpy.zeros((3,), dtype=theano.config.floatX) v_y = np.zeros((3,), dtype=theano.config.floatX)
v_x[0] = numpy.dot(v_u1[0], vW_in1) + v_u2[0] * vW_in2 + \ v_x[0] = np.dot(v_u1[0], vW_in1) + v_u2[0] * vW_in2 + \
numpy.dot(v_x0, vW) np.dot(v_x0, vW)
v_y[0] = numpy.dot(v_x0, vWout) + v_y0[2] v_y[0] = np.dot(v_x0, vWout) + v_y0[2]
for i in xrange(1, 3): for i in xrange(1, 3):
v_x[i] = numpy.dot(v_u1[i], vW_in1) + v_u2[i] * vW_in2 + \ v_x[i] = np.dot(v_u1[i], vW_in1) + v_u2[i] * vW_in2 + \
numpy.dot(v_x[i - 1], vW) np.dot(v_x[i - 1], vW)
v_y[i] = numpy.dot(v_x[i - 1], vWout) + v_y[i - 1] v_y[i] = np.dot(v_x[i - 1], vWout) + v_y[i - 1]
(theano_dump1, theano_dump2, theano_x, theano_y) = f4(v_u1, (theano_dump1, theano_dump2, theano_x, theano_y) = f4(v_u1,
v_u2, v_u2,
...@@ -2247,7 +2247,7 @@ class T_Scan(unittest.TestCase): ...@@ -2247,7 +2247,7 @@ class T_Scan(unittest.TestCase):
allow_input_downcast=True) allow_input_downcast=True)
def test_save_mem(self): def test_save_mem(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in2 = asarrayX(rng.uniform(size=(2,), low=-5., high=5.)) vW_in2 = asarrayX(rng.uniform(size=(2,), low=-5., high=5.))
vW = asarrayX(rng.uniform(size=(2, 2), low=-5., high=5.)) vW = asarrayX(rng.uniform(size=(2, 2), low=-5., high=5.))
vWout = asarrayX(rng.uniform(size=(2,), low=-5., high=5.)) vWout = asarrayX(rng.uniform(size=(2,), low=-5., high=5.))
...@@ -2288,16 +2288,16 @@ class T_Scan(unittest.TestCase): ...@@ -2288,16 +2288,16 @@ class T_Scan(unittest.TestCase):
allow_input_downcast=True) allow_input_downcast=True)
# compute the values in numpy # compute the values in numpy
v_x = numpy.zeros((8, 2), dtype=theano.config.floatX) v_x = np.zeros((8, 2), dtype=theano.config.floatX)
v_y = numpy.zeros((8,), dtype=theano.config.floatX) v_y = np.zeros((8,), dtype=theano.config.floatX)
v_x[0] = numpy.dot(v_u1[0], vW_in1) + v_u2[0] * vW_in2 + \ v_x[0] = np.dot(v_u1[0], vW_in1) + v_u2[0] * vW_in2 + \
numpy.dot(v_x0, vW) np.dot(v_x0, vW)
v_y[0] = numpy.dot(v_x0, vWout) + v_y0[2] v_y[0] = np.dot(v_x0, vWout) + v_y0[2]
for i in xrange(1, 8): for i in xrange(1, 8):
v_x[i] = numpy.dot(v_u1[i], vW_in1) + v_u2[i] * vW_in2 + \ v_x[i] = np.dot(v_u1[i], vW_in1) + v_u2[i] * vW_in2 + \
numpy.dot(v_x[i - 1], vW) np.dot(v_x[i - 1], vW)
v_y[i] = numpy.dot(v_x[i - 1], vWout) + v_y[i - 1] v_y[i] = np.dot(v_x[i - 1], vWout) + v_y[i - 1]
(theano_dump, theano_x, theano_y) = f4(v_u1, v_u2, v_x0, v_y0, vW_in1) (theano_dump, theano_x, theano_y) = f4(v_u1, v_u2, v_x0, v_y0, vW_in1)
...@@ -2321,24 +2321,24 @@ class T_Scan(unittest.TestCase): ...@@ -2321,24 +2321,24 @@ class T_Scan(unittest.TestCase):
sh = expr.shape[0] sh = expr.shape[0]
v1 = theano.shared(numpy.ones(5, dtype=theano.config.floatX)) v1 = theano.shared(np.ones(5, dtype=theano.config.floatX))
v2 = theano.shared(numpy.ones((5, 5), dtype=theano.config.floatX)) v2 = theano.shared(np.ones((5, 5), dtype=theano.config.floatX))
shapef = theano.function([W], shapef = theano.function([W],
expr, expr,
givens=OrderedDict([(initial, v1), givens=OrderedDict([(initial, v1),
(inpt, v2)])) (inpt, v2)]))
# First execution to cache n_steps # First execution to cache n_steps
shapef(numpy.ones((5, 5), dtype=theano.config.floatX)) shapef(np.ones((5, 5), dtype=theano.config.floatX))
cost = expr.sum() cost = expr.sum()
d_cost_wrt_W = tensor.grad(cost, [W]) d_cost_wrt_W = tensor.grad(cost, [W])
f = theano.function( f = theano.function(
[W, inpt], d_cost_wrt_W, [W, inpt], d_cost_wrt_W,
givens=OrderedDict([(initial, theano.shared(numpy.zeros(5)))])) givens=OrderedDict([(initial, theano.shared(np.zeros(5)))]))
rval = numpy.asarray([[5187989] * 5] * 5, dtype=theano.config.floatX) rval = np.asarray([[5187989] * 5] * 5, dtype=theano.config.floatX)
arg1 = numpy.ones((5, 5), dtype=theano.config.floatX) arg1 = np.ones((5, 5), dtype=theano.config.floatX)
arg2 = numpy.ones((10, 5), dtype=theano.config.floatX) arg2 = np.ones((10, 5), dtype=theano.config.floatX)
utt.assert_allclose(f(arg1, arg2), rval) utt.assert_allclose(f(arg1, arg2), rval)
def test_save_mem_reduced_number_of_steps(self): def test_save_mem_reduced_number_of_steps(self):
...@@ -2372,7 +2372,7 @@ class T_Scan(unittest.TestCase): ...@@ -2372,7 +2372,7 @@ class T_Scan(unittest.TestCase):
updates=updates, updates=updates,
allow_input_downcast=True) allow_input_downcast=True)
# get random initial values # get random initial values
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = rng.uniform(size=(20,), low=-5., high=5.) v_u = rng.uniform(size=(20,), low=-5., high=5.)
# compute the output in numpy # compute the output in numpy
...@@ -2428,7 +2428,7 @@ class T_Scan(unittest.TestCase): ...@@ -2428,7 +2428,7 @@ class T_Scan(unittest.TestCase):
allow_input_downcast=True) allow_input_downcast=True)
# get random initial values # get random initial values
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = rng.uniform(size=(20,), low=-5., high=5.) v_u = rng.uniform(size=(20,), low=-5., high=5.)
# compute the output in numpy # compute the output in numpy
...@@ -2474,7 +2474,7 @@ class T_Scan(unittest.TestCase): ...@@ -2474,7 +2474,7 @@ class T_Scan(unittest.TestCase):
floatX = theano.config.floatX floatX = theano.config.floatX
init_value = 5.0 init_value = 5.0
seq_value = numpy.arange(4, dtype=floatX) seq_value = np.arange(4, dtype=floatX)
output1, output2 = fct(init_value, seq_value) output1, output2 = fct(init_value, seq_value)
expected_output1 = [init_value] expected_output1 = [init_value]
...@@ -2509,13 +2509,13 @@ class T_Scan(unittest.TestCase): ...@@ -2509,13 +2509,13 @@ class T_Scan(unittest.TestCase):
[out1_direct, out2_direct]) [out1_direct, out2_direct])
# Test that the function returns valid outputs # Test that the function returns valid outputs
x_val = numpy.arange(0, 4)[:, None] x_val = np.arange(0, 4)[:, None]
seq_val = numpy.arange(4, 8)[:, None] seq_val = np.arange(4, 8)[:, None]
out1, out2 = fct(x_val, seq_val) out1, out2 = fct(x_val, seq_val)
expected_out1 = numpy.zeros((5, 4, 1)) expected_out1 = np.zeros((5, 4, 1))
expected_out2 = numpy.zeros((5, 4, 1)) expected_out2 = np.zeros((5, 4, 1))
for i in range(4): for i in range(4):
expected_out2[i + 1] = expected_out2[i] + seq_val[i] expected_out2[i + 1] = expected_out2[i] + seq_val[i]
for i in range(5): for i in range(5):
...@@ -2565,7 +2565,7 @@ class T_Scan(unittest.TestCase): ...@@ -2565,7 +2565,7 @@ class T_Scan(unittest.TestCase):
diff = mitsot_m1 + seq1 diff = mitsot_m1 + seq1
next_mitsot_val = mitsot_m2 + diff next_mitsot_val = mitsot_m2 + diff
next_sitsot_val = sitsot_m1 - diff next_sitsot_val = sitsot_m1 - diff
nitsot_out = tensor.alloc(numpy.asarray(0., 'float32'), nitsot_out = tensor.alloc(np.asarray(0., 'float32'),
next_mitsot_val + next_mitsot_val +
next_sitsot_val) next_sitsot_val)
return next_sitsot_val, next_mitsot_val, nitsot_out return next_sitsot_val, next_mitsot_val, nitsot_out
...@@ -2584,7 +2584,7 @@ class T_Scan(unittest.TestCase): ...@@ -2584,7 +2584,7 @@ class T_Scan(unittest.TestCase):
assert(len(scan_nodes_from_fct(f)) == 1) assert(len(scan_nodes_from_fct(f)) == 1)
# This generate a scan crash during execution. # This generate a scan crash during execution.
# output_shape = f(numpy.arange(5), 5, [1, 2]) # output_shape = f(np.arange(5), 5, [1, 2])
# assert(all(output_shape == (5, 6))) # assert(all(output_shape == (5, 6)))
# The following test will fail in DebugMode if there are # The following test will fail in DebugMode if there are
...@@ -2608,7 +2608,7 @@ class T_Scan(unittest.TestCase): ...@@ -2608,7 +2608,7 @@ class T_Scan(unittest.TestCase):
go1 = theano.tensor.grad(o1.mean(), wrt=x) go1 = theano.tensor.grad(o1.mean(), wrt=x)
f = theano.function([x], go1, updates=updates, f = theano.function([x], go1, updates=updates,
allow_input_downcast=True, mode=mode_with_opt) allow_input_downcast=True, mode=mode_with_opt)
self.assertTrue(numpy.allclose(f([1, 2, 3]), 2. / 3)) self.assertTrue(np.allclose(f([1, 2, 3]), 2. / 3))
topo = f.maker.fgraph.toposort() topo = f.maker.fgraph.toposort()
# this new assert is here to test if scan_merging works .. # this new assert is here to test if scan_merging works ..
...@@ -2711,7 +2711,7 @@ class T_Scan(unittest.TestCase): ...@@ -2711,7 +2711,7 @@ class T_Scan(unittest.TestCase):
n.op, theano.scan_module.scan_op.Scan)] n.op, theano.scan_module.scan_op.Scan)]
self.assertTrue(len(scans) == 2) self.assertTrue(len(scans) == 2)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
x_val = rng.uniform(size=(4,)).astype(theano.config.floatX) x_val = rng.uniform(size=(4,)).astype(theano.config.floatX)
y_val = rng.uniform(size=(4,)).astype(theano.config.floatX) y_val = rng.uniform(size=(4,)).astype(theano.config.floatX)
# Run it so DebugMode can detect optimization problems. # Run it so DebugMode can detect optimization problems.
...@@ -2752,7 +2752,7 @@ class T_Scan(unittest.TestCase): ...@@ -2752,7 +2752,7 @@ class T_Scan(unittest.TestCase):
return M return M
# some initializations # some initializations
hypx = numpy.log(numpy.tile([1,1,1,1,1,1,0.01], (3,1))) hypx = np.log(np.tile([1,1,1,1,1,1,0.01], (3,1)))
# variables used in the following expressions # variables used in the following expressions
hyp = theano.shared(hypx) hyp = theano.shared(hypx)
...@@ -2763,10 +2763,10 @@ class T_Scan(unittest.TestCase): ...@@ -2763,10 +2763,10 @@ class T_Scan(unittest.TestCase):
M = init_predictive_output(inputs,targets,hyp,x_star,s_star) M = init_predictive_output(inputs,targets,hyp,x_star,s_star)
X = numpy.random.random((10,4)) X = np.random.random((10,4))
Y = numpy.random.random((10,3)) Y = np.random.random((10,3))
test_m = numpy.random.random((4,)) test_m = np.random.random((4,))
test_s = numpy.eye(4) test_s = np.eye(4)
# Compute expected outputs (jacobian of M wrt x_star) # Compute expected outputs (jacobian of M wrt x_star)
dfdm = theano.function([inputs,targets,x_star,s_star], dfdm = theano.function([inputs,targets,x_star,s_star],
...@@ -2851,9 +2851,9 @@ class T_Scan(unittest.TestCase): ...@@ -2851,9 +2851,9 @@ class T_Scan(unittest.TestCase):
x = theano.tensor.fmatrix('x') x = theano.tensor.fmatrix('x')
mem_val = numpy.zeros((2,), dtype='float32') mem_val = np.zeros((2,), dtype='float32')
memory = theano.shared(mem_val) memory = theano.shared(mem_val)
W = theano.shared(numpy.random.random((5, 2)).astype('float32')) W = theano.shared(np.random.random((5, 2)).astype('float32'))
def f(inp, mem): def f(inp, mem):
i = theano.tensor.join(0, inp, mem) i = theano.tensor.join(0, inp, mem)
...@@ -2867,7 +2867,7 @@ class T_Scan(unittest.TestCase): ...@@ -2867,7 +2867,7 @@ class T_Scan(unittest.TestCase):
f = theano.function([x], outs[0]) f = theano.function([x], outs[0])
f2 = theano.function([x], outs[1]) f2 = theano.function([x], outs[1])
x_val = numpy.random.random((4, 3)).astype('float32') x_val = np.random.random((4, 3)).astype('float32')
f_vals = f(x_val) f_vals = f(x_val)
memory.set_value(mem_val) memory.set_value(mem_val)
...@@ -2876,12 +2876,12 @@ class T_Scan(unittest.TestCase): ...@@ -2876,12 +2876,12 @@ class T_Scan(unittest.TestCase):
def test_reduce_memory_consumption(self): def test_reduce_memory_consumption(self):
x = theano.shared(numpy.asarray( x = theano.shared(np.asarray(
numpy.random.uniform(size=(10,)), dtype=theano.config.floatX)) np.random.uniform(size=(10,)), dtype=theano.config.floatX))
o, _ = theano.reduce(lambda v, acc: acc + v, o, _ = theano.reduce(lambda v, acc: acc + v,
x, x,
theano.tensor.constant( theano.tensor.constant(
numpy.asarray(0., np.asarray(0.,
dtype=theano.config.floatX))) dtype=theano.config.floatX)))
mode = theano.compile.mode.FAST_RUN mode = theano.compile.mode.FAST_RUN
mode = mode.excluding('inplace') mode = mode.excluding('inplace')
...@@ -2905,15 +2905,15 @@ class T_Scan(unittest.TestCase): ...@@ -2905,15 +2905,15 @@ class T_Scan(unittest.TestCase):
gx = theano.tensor.grad(o, x) gx = theano.tensor.grad(o, x)
f2 = theano.function([], gx) f2 = theano.function([], gx)
utt.assert_allclose(f2(), numpy.ones((10,))) utt.assert_allclose(f2(), np.ones((10,)))
def test_foldl_memory_consumption(self): def test_foldl_memory_consumption(self):
x = theano.shared(numpy.asarray( x = theano.shared(np.asarray(
numpy.random.uniform(size=(10,)), dtype=theano.config.floatX)) np.random.uniform(size=(10,)), dtype=theano.config.floatX))
o, _ = theano.foldl(lambda v, acc: acc + v, o, _ = theano.foldl(lambda v, acc: acc + v,
x, x,
theano.tensor.constant( theano.tensor.constant(
numpy.asarray(0., np.asarray(0.,
dtype=theano.config.floatX))) dtype=theano.config.floatX)))
mode = theano.compile.mode.FAST_RUN mode = theano.compile.mode.FAST_RUN
...@@ -2938,16 +2938,16 @@ class T_Scan(unittest.TestCase): ...@@ -2938,16 +2938,16 @@ class T_Scan(unittest.TestCase):
gx = theano.tensor.grad(o, x) gx = theano.tensor.grad(o, x)
f2 = theano.function([], gx) f2 = theano.function([], gx)
utt.assert_allclose(f2(), numpy.ones((10,))) utt.assert_allclose(f2(), np.ones((10,)))
def test_foldr_memory_consumption(self): def test_foldr_memory_consumption(self):
x = theano.shared(numpy.asarray( x = theano.shared(np.asarray(
numpy.random.uniform(size=(10,)), dtype=theano.config.floatX)) np.random.uniform(size=(10,)), dtype=theano.config.floatX))
o, _ = theano.foldr(lambda v, acc: acc + v, o, _ = theano.foldr(lambda v, acc: acc + v,
x, x,
theano.tensor.constant( theano.tensor.constant(
numpy.asarray(0., np.asarray(0.,
dtype=theano.config.floatX))) dtype=theano.config.floatX)))
mode = theano.compile.mode.FAST_RUN mode = theano.compile.mode.FAST_RUN
...@@ -2972,26 +2972,26 @@ class T_Scan(unittest.TestCase): ...@@ -2972,26 +2972,26 @@ class T_Scan(unittest.TestCase):
gx = theano.tensor.grad(o, x) gx = theano.tensor.grad(o, x)
f2 = theano.function([], gx) f2 = theano.function([], gx)
utt.assert_allclose(f2(), numpy.ones((10,))) utt.assert_allclose(f2(), np.ones((10,)))
@attr('slow') @attr('slow')
def test_rop2(self): def test_rop2(self):
seed = utt.fetch_seed() seed = utt.fetch_seed()
rng = numpy.random.RandomState(seed) rng = np.random.RandomState(seed)
floatX = theano.config.floatX floatX = theano.config.floatX
v_u = numpy.array(rng.uniform(size=(3, 5)) - .5, dtype=floatX) v_u = np.array(rng.uniform(size=(3, 5)) - .5, dtype=floatX)
v_W = numpy.array(rng.uniform(size=(5, 5)) - .5, dtype=floatX) v_W = np.array(rng.uniform(size=(5, 5)) - .5, dtype=floatX)
v_h0 = numpy.array(rng.uniform(size=(5,)) - .5, dtype=floatX) v_h0 = np.array(rng.uniform(size=(5,)) - .5, dtype=floatX)
v_eu = numpy.array(rng.uniform(size=(3, 5)) - .5, dtype=floatX) v_eu = np.array(rng.uniform(size=(3, 5)) - .5, dtype=floatX)
v_eW = numpy.array(rng.uniform(size=(5, 5)) - .5, dtype=floatX) v_eW = np.array(rng.uniform(size=(5, 5)) - .5, dtype=floatX)
v_eh0 = numpy.array(rng.uniform(size=(5,)) - .5, dtype=floatX) v_eh0 = np.array(rng.uniform(size=(5,)) - .5, dtype=floatX)
def rnn_fn(_u, _y, _W): def rnn_fn(_u, _y, _W):
srng = theano.tensor.shared_randomstreams.RandomStreams(seed) srng = theano.tensor.shared_randomstreams.RandomStreams(seed)
tmp_val = _u + _y + srng.uniform(size=v_h0.shape) *\ tmp_val = _u + _y + srng.uniform(size=v_h0.shape) *\
numpy.asarray(1e-6, dtype=floatX) np.asarray(1e-6, dtype=floatX)
sl_o = theano.tensor.tanh(theano.tensor.dot(_W, tmp_val)) sl_o = theano.tensor.tanh(theano.tensor.dot(_W, tmp_val))
return sl_o, tmp_val return sl_o, tmp_val
...@@ -3053,15 +3053,15 @@ class T_Scan(unittest.TestCase): ...@@ -3053,15 +3053,15 @@ class T_Scan(unittest.TestCase):
def test_rop(self): def test_rop(self):
seed = utt.fetch_seed() seed = utt.fetch_seed()
rng = numpy.random.RandomState(seed) rng = np.random.RandomState(seed)
floatX = theano.config.floatX floatX = theano.config.floatX
v_u = numpy.array(rng.uniform(size=(20, 5)), dtype=floatX) v_u = np.array(rng.uniform(size=(20, 5)), dtype=floatX)
v_W = numpy.array(rng.uniform(size=(5, 5)), dtype=floatX) v_W = np.array(rng.uniform(size=(5, 5)), dtype=floatX)
v_h0 = numpy.array(rng.uniform(size=(5,)), dtype=floatX) v_h0 = np.array(rng.uniform(size=(5,)), dtype=floatX)
v_eu = numpy.array(rng.uniform(size=(20, 5)), dtype=floatX) v_eu = np.array(rng.uniform(size=(20, 5)), dtype=floatX)
v_eW = numpy.array(rng.uniform(size=(5, 5)), dtype=floatX) v_eW = np.array(rng.uniform(size=(5, 5)), dtype=floatX)
v_eh0 = numpy.array(rng.uniform(size=(5,)), dtype=floatX) v_eh0 = np.array(rng.uniform(size=(5,)), dtype=floatX)
def rnn_fn(_u, _y, _W): def rnn_fn(_u, _y, _W):
sl_o = theano.tensor.tanh(theano.tensor.dot(_W, (_u + _y))) sl_o = theano.tensor.tanh(theano.tensor.dot(_W, (_u + _y)))
...@@ -3163,14 +3163,14 @@ class T_Scan(unittest.TestCase): ...@@ -3163,14 +3163,14 @@ class T_Scan(unittest.TestCase):
assert len(scan_nodes) == 0 assert len(scan_nodes) == 0
seed = utt.fetch_seed() seed = utt.fetch_seed()
rng = numpy.random.RandomState(seed) rng = np.random.RandomState(seed)
floatX = theano.config.floatX floatX = theano.config.floatX
v_h = numpy.array(rng.uniform(size=(2,)), dtype=floatX) v_h = np.array(rng.uniform(size=(2,)), dtype=floatX)
v_W1 = numpy.array(rng.uniform(size=(2, 2)), dtype=floatX) v_W1 = np.array(rng.uniform(size=(2, 2)), dtype=floatX)
v_W2 = numpy.array(rng.uniform(size=(2, 2)), dtype=floatX) v_W2 = np.array(rng.uniform(size=(2, 2)), dtype=floatX)
v_out = numpy.dot(v_h, v_W1 + v_W2) v_out = np.dot(v_h, v_W1 + v_W2)
sol = numpy.zeros((5, 2)) sol = np.zeros((5, 2))
# This line is here to make sol have the same shape as the output of # This line is here to make sol have the same shape as the output of
# theano. Note that what we ask theano to do is to repeat the 2 # theano. Note that what we ask theano to do is to repeat the 2
# elements vector v_out 5 times # elements vector v_out 5 times
...@@ -3206,9 +3206,9 @@ class T_Scan(unittest.TestCase): ...@@ -3206,9 +3206,9 @@ class T_Scan(unittest.TestCase):
f_ref = theano.function([W1, W2, step_indices], o, mode='FAST_COMPILE') f_ref = theano.function([W1, W2, step_indices], o, mode='FAST_COMPILE')
# Compare the results of the two implementations # Compare the results of the two implementations
input_values = [numpy.random.random((5, 5)).astype("float32"), input_values = [np.random.random((5, 5)).astype("float32"),
numpy.random.random((5, 5)).astype("float32"), np.random.random((5, 5)).astype("float32"),
numpy.arange(5).astype("float32")] np.arange(5).astype("float32")]
out = f(*input_values) out = f(*input_values)
out_ref = f_ref(*input_values) out_ref = f_ref(*input_values)
...@@ -3243,10 +3243,10 @@ class T_Scan(unittest.TestCase): ...@@ -3243,10 +3243,10 @@ class T_Scan(unittest.TestCase):
([i_t, i_tm1], _) = theano.scan( ([i_t, i_tm1], _) = theano.scan(
fn, sequences=[inp], fn, sequences=[inp],
outputs_info=[numpy.asarray([0.0, 0.0], theano.config.floatX), outputs_info=[np.asarray([0.0, 0.0], theano.config.floatX),
None]) None])
f = theano.function([inp], [i_t, i_tm1]) f = theano.function([inp], [i_t, i_tm1])
val = numpy.arange(10).reshape(5, 2).astype(theano.config.floatX) val = np.arange(10).reshape(5, 2).astype(theano.config.floatX)
ret = f(val) ret = f(val)
utt.assert_allclose(ret[0], val + 10) utt.assert_allclose(ret[0], val + 10)
utt.assert_allclose(ret[1], [[0., 0.], utt.assert_allclose(ret[1], [[0., 0.],
...@@ -3330,7 +3330,7 @@ class T_Scan(unittest.TestCase): ...@@ -3330,7 +3330,7 @@ class T_Scan(unittest.TestCase):
return x_t + 1, theano.scan_module.until(x_t > 3) return x_t + 1, theano.scan_module.until(x_t > 3)
o, _ = theano.scan(lambda_fn, x) o, _ = theano.scan(lambda_fn, x)
f = theano.function([x], o) f = theano.function([x], o)
vx = numpy.zeros((50,), dtype=theano.config.floatX) vx = np.zeros((50,), dtype=theano.config.floatX)
vx[23] = 4 vx[23] = 4
out = f(vx) out = f(vx)
assert len(out) == 24 assert len(out) == 24
...@@ -3344,11 +3344,11 @@ class T_Scan(unittest.TestCase): ...@@ -3344,11 +3344,11 @@ class T_Scan(unittest.TestCase):
o2, _ = theano.scan(lambda x_t: x_t + 2, x) o2, _ = theano.scan(lambda x_t: x_t + 2, x)
f = theano.function([x], [o, o2], mode=mode_with_opt) f = theano.function([x], [o, o2], mode=mode_with_opt)
vx = numpy.zeros((50,), dtype=theano.config.floatX) vx = np.zeros((50,), dtype=theano.config.floatX)
vx[23] = 4 vx[23] = 4
out, out2 = f(vx) out, out2 = f(vx)
assert len(out) == 24 assert len(out) == 24
assert numpy.all(out2 == vx + 2) assert np.all(out2 == vx + 2)
lssc = [x for x in f.maker.fgraph.toposort() lssc = [x for x in f.maker.fgraph.toposort()
if isinstance(x.op, theano.scan_module.scan_op.Scan)] if isinstance(x.op, theano.scan_module.scan_op.Scan)]
# One scan node gets optimnized out # One scan node gets optimnized out
...@@ -3402,7 +3402,7 @@ class T_Scan(unittest.TestCase): ...@@ -3402,7 +3402,7 @@ class T_Scan(unittest.TestCase):
polynomial3[-1], polynomial3[-1],
polynomial4[-1]]) polynomial4[-1]])
test_coeff = numpy.asarray([1, 0, 2], dtype=theano.config.floatX) test_coeff = np.asarray([1, 0, 2], dtype=theano.config.floatX)
# This will be tested by DEBUG_MODE # This will be tested by DEBUG_MODE
out = calculate_polynomial(test_coeff, 3) out = calculate_polynomial(test_coeff, 3)
assert out[0] == 19 assert out[0] == 19
...@@ -3480,7 +3480,7 @@ class T_Scan(unittest.TestCase): ...@@ -3480,7 +3480,7 @@ class T_Scan(unittest.TestCase):
x) x)
f = theano.function([x], [o, o2], mode=mode_with_opt) f = theano.function([x], [o, o2], mode=mode_with_opt)
vx = numpy.zeros((50,), dtype=theano.config.floatX) vx = np.zeros((50,), dtype=theano.config.floatX)
vx[23] = 4 vx[23] = 4
out, out2 = f(vx) out, out2 = f(vx)
assert len(out) == 24 assert len(out) == 24
...@@ -3497,7 +3497,7 @@ class T_Scan(unittest.TestCase): ...@@ -3497,7 +3497,7 @@ class T_Scan(unittest.TestCase):
o, _ = theano.scan(lambda_fn, x) o, _ = theano.scan(lambda_fn, x)
f = theano.function([x], o.shape[0], mode=mode_with_opt) f = theano.function([x], o.shape[0], mode=mode_with_opt)
vx = numpy.zeros((50,), dtype=theano.config.floatX) vx = np.zeros((50,), dtype=theano.config.floatX)
vx[23] = 4 vx[23] = 4
out = f(vx) out = f(vx)
assert out == 24 assert out == 24
...@@ -3516,7 +3516,7 @@ class T_Scan(unittest.TestCase): ...@@ -3516,7 +3516,7 @@ class T_Scan(unittest.TestCase):
[o1.shape[0], o2.shape[0]], [o1.shape[0], o2.shape[0]],
mode=mode_with_opt) mode=mode_with_opt)
vx = numpy.ones((10,), dtype=theano.config.floatX) vx = np.ones((10,), dtype=theano.config.floatX)
out1, out2 = f(vx) out1, out2 = f(vx)
assert out1 == 10 assert out1 == 10
assert out2 == 10 assert out2 == 10
...@@ -3535,7 +3535,7 @@ class T_Scan(unittest.TestCase): ...@@ -3535,7 +3535,7 @@ class T_Scan(unittest.TestCase):
[o1.shape[0], o2.shape[0]], [o1.shape[0], o2.shape[0]],
mode=mode_with_opt) mode=mode_with_opt)
vx = numpy.ones((30,), dtype=theano.config.floatX) vx = np.ones((30,), dtype=theano.config.floatX)
o1, o2 = f(vx) o1, o2 = f(vx)
assert o1 == 20 assert o1 == 20
assert o2 == 20 assert o2 == 20
...@@ -3635,13 +3635,13 @@ class T_Scan(unittest.TestCase): ...@@ -3635,13 +3635,13 @@ class T_Scan(unittest.TestCase):
# Run the function and validate the outputs # Run the function and validate the outputs
dtype = theano.config.floatX dtype = theano.config.floatX
seq_value = numpy.random.random((10, 3)).astype(dtype) seq_value = np.random.random((10, 3)).astype(dtype)
out_init_value = numpy.random.random((3, 3)).astype(dtype) out_init_value = np.random.random((3, 3)).astype(dtype)
non_seq_value = numpy.random.random((3)).astype(dtype) non_seq_value = np.random.random((3)).astype(dtype)
outputs = fct(seq_value, out_init_value, non_seq_value) outputs = fct(seq_value, out_init_value, non_seq_value)
expected_g_seq = numpy.array([[4, 4, 4], expected_g_seq = np.array([[4, 4, 4],
[3, 3, 3], [3, 3, 3],
[3, 3, 3], [3, 3, 3],
[3, 3, 3], [3, 3, 3],
...@@ -3652,7 +3652,7 @@ class T_Scan(unittest.TestCase): ...@@ -3652,7 +3652,7 @@ class T_Scan(unittest.TestCase):
[1, 1, 1], [1, 1, 1],
[1, 1, 1]]) [1, 1, 1]])
expected_g_out_init = expected_g_seq[:3] expected_g_out_init = expected_g_seq[:3]
expected_g_non_seq = numpy.array([22, 22, 22]) expected_g_non_seq = np.array([22, 22, 22])
utt.assert_allclose(outputs[0], expected_g_seq) utt.assert_allclose(outputs[0], expected_g_seq)
utt.assert_allclose(outputs[1], expected_g_out_init) utt.assert_allclose(outputs[1], expected_g_out_init)
...@@ -3729,7 +3729,7 @@ class T_Scan(unittest.TestCase): ...@@ -3729,7 +3729,7 @@ class T_Scan(unittest.TestCase):
assert tf([1.0, 2.0, -3.0, 4.0], 2.0) == 42 assert tf([1.0, 2.0, -3.0, 4.0], 2.0) == 42
def test_return_steps(self): def test_return_steps(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vW_in2 = asarrayX(rng.uniform(size=(2,), low=-5., high=5.)) vW_in2 = asarrayX(rng.uniform(size=(2,), low=-5., high=5.))
vW = asarrayX(rng.uniform(size=(2, 2), low=-5., high=5.)) vW = asarrayX(rng.uniform(size=(2, 2), low=-5., high=5.))
vWout = asarrayX(rng.uniform(size=(2,), low=-5., high=5.)) vWout = asarrayX(rng.uniform(size=(2,), low=-5., high=5.))
...@@ -3774,16 +3774,16 @@ class T_Scan(unittest.TestCase): ...@@ -3774,16 +3774,16 @@ class T_Scan(unittest.TestCase):
allow_input_downcast=True) allow_input_downcast=True)
# compute the values in numpy # compute the values in numpy
v_x = numpy.zeros((8, 2), dtype=theano.config.floatX) v_x = np.zeros((8, 2), dtype=theano.config.floatX)
v_y = numpy.zeros((8,), dtype=theano.config.floatX) v_y = np.zeros((8,), dtype=theano.config.floatX)
v_x[0] = numpy.dot(v_u1[0], vW_in1) + v_u2[0] * vW_in2 + \ v_x[0] = np.dot(v_u1[0], vW_in1) + v_u2[0] * vW_in2 + \
numpy.dot(v_x0, vW) np.dot(v_x0, vW)
v_y[0] = numpy.dot(v_x0, vWout) + v_y0[2] v_y[0] = np.dot(v_x0, vWout) + v_y0[2]
for i in xrange(1, 8): for i in xrange(1, 8):
v_x[i] = numpy.dot(v_u1[i], vW_in1) + v_u2[i] * vW_in2 + \ v_x[i] = np.dot(v_u1[i], vW_in1) + v_u2[i] * vW_in2 + \
numpy.dot(v_x[i - 1], vW) np.dot(v_x[i - 1], vW)
v_y[i] = numpy.dot(v_x[i - 1], vWout) + v_y[i - 1] v_y[i] = np.dot(v_x[i - 1], vWout) + v_y[i - 1]
(theano_dump, theano_x, theano_y) = f4(v_u1, v_u2, v_x0, v_y0, vW_in1) (theano_dump, theano_x, theano_y) = f4(v_u1, v_u2, v_x0, v_y0, vW_in1)
...@@ -3811,14 +3811,14 @@ class T_Scan(unittest.TestCase): ...@@ -3811,14 +3811,14 @@ class T_Scan(unittest.TestCase):
assert any([isinstance(node.op, tensor.blas.Dot22) assert any([isinstance(node.op, tensor.blas.Dot22)
for node in topo]) for node in topo])
vx = numpy.array([[1., 1.], [2., 2.]], dtype=theano.config.floatX) vx = np.array([[1., 1.], [2., 2.]], dtype=theano.config.floatX)
vA = numpy.array([[1., 1.], [1., 0.]], dtype=theano.config.floatX) vA = np.array([[1., 1.], [1., 0.]], dtype=theano.config.floatX)
vR = numpy.array([[[2, 1], [4, 2]], [[2, 1], [4, 2]]], vR = np.array([[[2, 1], [4, 2]], [[2, 1], [4, 2]]],
dtype=theano.config.floatX) dtype=theano.config.floatX)
utt.assert_allclose(f(vx, vA), vR) utt.assert_allclose(f(vx, vA), vR)
def test_savemem_opt(self): def test_savemem_opt(self):
y0 = theano.shared(numpy.ones((2, 10))) y0 = theano.shared(np.ones((2, 10)))
[y1, y2], updates = theano.scan(lambda y: [y, y], [y1, y2], updates = theano.scan(lambda y: [y, y],
outputs_info=[dict(initial=y0, outputs_info=[dict(initial=y0,
taps=[-2]), None], taps=[-2]), None],
...@@ -3860,9 +3860,9 @@ class T_Scan(unittest.TestCase): ...@@ -3860,9 +3860,9 @@ class T_Scan(unittest.TestCase):
f = theano.function(inputs=[x, w], outputs=get_outputs(x, w)) f = theano.function(inputs=[x, w], outputs=get_outputs(x, w))
# Test the function to ensure it returns valid results # Test the function to ensure it returns valid results
x_value = numpy.random.random((2, 2, 3)).astype(theano.config.floatX) x_value = np.random.random((2, 2, 3)).astype(theano.config.floatX)
w_value = numpy.random.random((3, 3)).astype(theano.config.floatX) w_value = np.random.random((3, 3)).astype(theano.config.floatX)
expected_output = numpy.tile(x_value[:, 0].sum(0), (3, 1)).transpose() expected_output = np.tile(x_value[:, 0].sum(0), (3, 1)).transpose()
output = f(x_value, w_value) output = f(x_value, w_value)
utt.assert_allclose(output, expected_output) utt.assert_allclose(output, expected_output)
...@@ -3891,17 +3891,17 @@ class T_Scan(unittest.TestCase): ...@@ -3891,17 +3891,17 @@ class T_Scan(unittest.TestCase):
gw, gx = tensor.grad(loss, [w, xinit]) gw, gx = tensor.grad(loss, [w, xinit])
grad_fn = theano.function([xinit, w], [gx, gw], grad_fn = theano.function([xinit, w], [gx, gw],
allow_input_downcast=True) allow_input_downcast=True)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
# If numbers are small, the gradients with respect to x are small # If numbers are small, the gradients with respect to x are small
# and the numeric differentiation becomes unstable. # and the numeric differentiation becomes unstable.
# To fix this issue I ensure we are sampling numbers larger in # To fix this issue I ensure we are sampling numbers larger in
# absolute value than 1. # absolute value than 1.
v_x = numpy.array(rng.uniform(size=(5, 2, 2), low=1., high=3.), v_x = np.array(rng.uniform(size=(5, 2, 2), low=1., high=3.),
dtype=theano.config.floatX) dtype=theano.config.floatX)
# Making some entries to be negative. # Making some entries to be negative.
pos = rng.uniform(size=(5, 2, 2), low=0., high=1) < .5 pos = rng.uniform(size=(5, 2, 2), low=0., high=1) < .5
v_x[pos] = -1 * v_x[pos] v_x[pos] = -1 * v_x[pos]
v_w = numpy.array(rng.uniform(size=(2, 2), low=1., high=3.), v_w = np.array(rng.uniform(size=(2, 2), low=1., high=3.),
dtype=theano.config.floatX) dtype=theano.config.floatX)
pos = rng.uniform(size=(2, 2), low=0., high=1.) < .5 pos = rng.uniform(size=(2, 2), low=0., high=1.) < .5
v_w[pos] = -1 * v_w[pos] v_w[pos] = -1 * v_w[pos]
...@@ -3916,11 +3916,11 @@ class T_Scan(unittest.TestCase): ...@@ -3916,11 +3916,11 @@ class T_Scan(unittest.TestCase):
num_grad.gx[max_err_pos])) num_grad.gx[max_err_pos]))
def test_grad_numeric_shared(self): def test_grad_numeric_shared(self):
shared_var = theano.shared(numpy.float32(1.)) shared_var = theano.shared(np.float32(1.))
def inner_fn(): def inner_fn():
return [], OrderedDict( return [], OrderedDict(
[(shared_var, shared_var + numpy.float32(1.))]) [(shared_var, shared_var + np.float32(1.))])
_, updates = theano.scan(inner_fn, _, updates = theano.scan(inner_fn,
n_steps=10, n_steps=10,
truncate_gradient=-1, truncate_gradient=-1,
...@@ -3940,7 +3940,7 @@ class T_Scan(unittest.TestCase): ...@@ -3940,7 +3940,7 @@ class T_Scan(unittest.TestCase):
n_pars = 1 * 3 + 3 * 3 n_pars = 1 * 3 + 3 * 3
# Allocate big parameter array. # Allocate big parameter array.
pars = theano.shared(numpy.empty(n_pars)) pars = theano.shared(np.empty(n_pars))
# Assign slices. # Assign slices.
W1 = pars[:3].reshape(W1shape) W1 = pars[:3].reshape(W1shape)
...@@ -3983,15 +3983,15 @@ class T_Scan(unittest.TestCase): ...@@ -3983,15 +3983,15 @@ class T_Scan(unittest.TestCase):
Hp = tensor.Rop(d_cost_wrt_pars, pars, p) Hp = tensor.Rop(d_cost_wrt_pars, pars, p)
def test_seq_tap_bug_jeremiah(self): def test_seq_tap_bug_jeremiah(self):
inp = numpy.arange(10).reshape(-1, 1).astype(theano.config.floatX) inp = np.arange(10).reshape(-1, 1).astype(theano.config.floatX)
exp_out = numpy.zeros((10, 1)).astype(theano.config.floatX) exp_out = np.zeros((10, 1)).astype(theano.config.floatX)
exp_out[4:] = inp[:-4] exp_out[4:] = inp[:-4]
def onestep(x, x_tm4): def onestep(x, x_tm4):
return x, x_tm4 return x, x_tm4
seq = tensor.matrix() seq = tensor.matrix()
initial_value = theano.shared(numpy.zeros((4, 1), initial_value = theano.shared(np.zeros((4, 1),
dtype=theano.config.floatX)) dtype=theano.config.floatX))
outputs_info = [OrderedDict( outputs_info = [OrderedDict(
[('initial', initial_value), ('taps', [-4])]), None] [('initial', initial_value), ('taps', [-4])]), None]
...@@ -4000,7 +4000,7 @@ class T_Scan(unittest.TestCase): ...@@ -4000,7 +4000,7 @@ class T_Scan(unittest.TestCase):
outputs_info=outputs_info) outputs_info=outputs_info)
f = theano.function([seq], results[1]) f = theano.function([seq], results[1])
assert numpy.all(exp_out == f(inp)) assert np.all(exp_out == f(inp))
def test_borrow_bug_jeremiah(self): def test_borrow_bug_jeremiah(self):
# This tests two things. The first is a bug occuring when scan wrongly # This tests two things. The first is a bug occuring when scan wrongly
...@@ -4008,29 +4008,29 @@ class T_Scan(unittest.TestCase): ...@@ -4008,29 +4008,29 @@ class T_Scan(unittest.TestCase):
# method will be able to remove the Scan node from the graph in this # method will be able to remove the Scan node from the graph in this
# case. # case.
inp = numpy.arange(10).reshape(-1, 1).astype(theano.config.floatX) inp = np.arange(10).reshape(-1, 1).astype(theano.config.floatX)
exp_out = numpy.zeros((10, 1)).astype(theano.config.floatX) exp_out = np.zeros((10, 1)).astype(theano.config.floatX)
exp_out[4:] = inp[:-4] exp_out[4:] = inp[:-4]
def onestep(x, x_tm4): def onestep(x, x_tm4):
return x, x_tm4 return x, x_tm4
seq = tensor.matrix() seq = tensor.matrix()
initial_value = theano.shared(numpy.zeros((4, 1), initial_value = theano.shared(np.zeros((4, 1),
dtype=theano.config.floatX)) dtype=theano.config.floatX))
outputs_info = [OrderedDict([('initial', initial_value), outputs_info = [OrderedDict([('initial', initial_value),
('taps', [-4])]), None] ('taps', [-4])]), None]
results, _ = theano.scan(fn=onestep, results, _ = theano.scan(fn=onestep,
sequences=seq, sequences=seq,
outputs_info=outputs_info) outputs_info=outputs_info)
sharedvar = theano.shared(numpy.zeros((1, 1), sharedvar = theano.shared(np.zeros((1, 1),
dtype=theano.config.floatX)) dtype=theano.config.floatX))
updates = OrderedDict([(sharedvar, results[0][-1:])]) updates = OrderedDict([(sharedvar, results[0][-1:])])
f = theano.function([seq], results[1], updates=updates) f = theano.function([seq], results[1], updates=updates)
# This fails if scan uses wrongly the borrow flag # This fails if scan uses wrongly the borrow flag
assert numpy.all(exp_out == f(inp)) assert np.all(exp_out == f(inp))
# This fails if Scan's infer_shape() is unable to remove the Scan # This fails if Scan's infer_shape() is unable to remove the Scan
# node from the graph. # node from the graph.
...@@ -4070,9 +4070,9 @@ class T_Scan(unittest.TestCase): ...@@ -4070,9 +4070,9 @@ class T_Scan(unittest.TestCase):
# Compare obtained outputs with expected outputs # Compare obtained outputs with expected outputs
floatX = theano.config.floatX floatX = theano.config.floatX
outputs = fct(numpy.arange(9, dtype=floatX).reshape(3,3)) outputs = fct(np.arange(9, dtype=floatX).reshape(3,3))
states = numpy.array([[0, 1, 2], states = np.array([[0, 1, 2],
[3, 4, 5], [3, 4, 5],
[6, 7, 8], [6, 7, 8],
[9, 12, 15], [9, 12, 15],
...@@ -4144,8 +4144,8 @@ class T_Scan(unittest.TestCase): ...@@ -4144,8 +4144,8 @@ class T_Scan(unittest.TestCase):
f = theano.function([v], gv) f = theano.function([v], gv)
# Ensure the output of the function is valid # Ensure the output of the function is valid
output = f(numpy.random.random(5)) output = f(np.random.random(5))
utt.assert_allclose(output, numpy.ones(5)) utt.assert_allclose(output, np.ones(5))
def test_dot_optimization(self): def test_dot_optimization(self):
A = tensor.matrix('A') A = tensor.matrix('A')
...@@ -4155,10 +4155,10 @@ class T_Scan(unittest.TestCase): ...@@ -4155,10 +4155,10 @@ class T_Scan(unittest.TestCase):
B.dimshuffle(0, 'x', 1)], B.dimshuffle(0, 'x', 1)],
outputs_info=[tensor.zeros_like(A)]) outputs_info=[tensor.zeros_like(A)])
f = theano.function([A, B], S.owner.inputs[0][-1]) f = theano.function([A, B], S.owner.inputs[0][-1])
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
vA = rng.uniform(size=(5, 5)).astype(theano.config.floatX) vA = rng.uniform(size=(5, 5)).astype(theano.config.floatX)
vB = rng.uniform(size=(5, 5)).astype(theano.config.floatX) vB = rng.uniform(size=(5, 5)).astype(theano.config.floatX)
utt.assert_allclose(f(vA, vB), numpy.dot(vA.T, vB)) utt.assert_allclose(f(vA, vB), np.dot(vA.T, vB))
def test_pregreedy_optimizer(self): def test_pregreedy_optimizer(self):
W = tensor.zeros((5, 4)) W = tensor.zeros((5, 4))
...@@ -4171,7 +4171,7 @@ class T_Scan(unittest.TestCase): ...@@ -4171,7 +4171,7 @@ class T_Scan(unittest.TestCase):
lambda x: tensor.dot(tensor.dot(x, W) + bh_t, W.T) + bv_t, lambda x: tensor.dot(tensor.dot(x, W) + bh_t, W.T) + bv_t,
outputs_info=v, outputs_info=v,
n_steps=2) n_steps=2)
theano.function([v], chain)(numpy.zeros((3, 5), theano.function([v], chain)(np.zeros((3, 5),
dtype=theano.config.floatX)) dtype=theano.config.floatX))
def test_savemem_does_not_duplicate_number_of_scan_nodes(self): def test_savemem_does_not_duplicate_number_of_scan_nodes(self):
...@@ -4210,7 +4210,7 @@ class T_Scan(unittest.TestCase): ...@@ -4210,7 +4210,7 @@ class T_Scan(unittest.TestCase):
updates=updates, updates=updates,
mode=theano.Mode(linker='py'), mode=theano.Mode(linker='py'),
allow_input_downcast=True) allow_input_downcast=True)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = asarrayX(rng.uniform(size=(5,))) v_u = asarrayX(rng.uniform(size=(5,)))
outs = f(v_u, [0, 0, 0], 0) outs = f(v_u, [0, 0, 0], 0)
utt.assert_allclose(outs[0], v_u + 1) utt.assert_allclose(outs[0], v_u + 1)
...@@ -4243,7 +4243,7 @@ class T_Scan(unittest.TestCase): ...@@ -4243,7 +4243,7 @@ class T_Scan(unittest.TestCase):
updates=updates, updates=updates,
mode=theano.Mode(linker='py'), mode=theano.Mode(linker='py'),
allow_input_downcast=True) allow_input_downcast=True)
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_w = asarrayX(rng.uniform()) v_w = asarrayX(rng.uniform())
outs = f(v_w, [0, 0, 0], 0) outs = f(v_w, [0, 0, 0], 0)
utt.assert_allclose(outs[0], v_w + 1) utt.assert_allclose(outs[0], v_w + 1)
...@@ -4252,7 +4252,7 @@ class T_Scan(unittest.TestCase): ...@@ -4252,7 +4252,7 @@ class T_Scan(unittest.TestCase):
utt.assert_allclose(sh.get_value(), v_w + 4) utt.assert_allclose(sh.get_value(), v_w + 4)
def test_grad_bug_disconnected_input(self): def test_grad_bug_disconnected_input(self):
W = theano.shared(numpy.zeros((3, 3)), name='W') W = theano.shared(np.zeros((3, 3)), name='W')
v = theano.tensor.ivector(name='v') v = theano.tensor.ivector(name='v')
y, _ = theano.scan(lambda i, W: W[i], sequences=v, outputs_info=None, non_sequences=W) y, _ = theano.scan(lambda i, W: W[i], sequences=v, outputs_info=None, non_sequences=W)
...@@ -4270,14 +4270,14 @@ class T_Scan(unittest.TestCase): ...@@ -4270,14 +4270,14 @@ class T_Scan(unittest.TestCase):
# theano.printing.debugprint(out) # theano.printing.debugprint(out)
return theano.function([], out)() return theano.function([], out)()
x = theano.shared(numpy.asarray(0., dtype=theano.config.floatX)) x = theano.shared(np.asarray(0., dtype=theano.config.floatX))
utt.assert_allclose(test(x, tensor.sum((x+1)**2), mention_y=False), utt.assert_allclose(test(x, tensor.sum((x+1)**2), mention_y=False),
1.21000003815) 1.21000003815)
utt.assert_allclose(test(x, tensor.sum((x+1)**2), mention_y=True), utt.assert_allclose(test(x, tensor.sum((x+1)**2), mention_y=True),
1.21000003815) 1.21000003815)
def test_grad_find_input(self): def test_grad_find_input(self):
w = theano.shared(numpy.array(0, dtype='float32'), name='w') w = theano.shared(np.array(0, dtype='float32'), name='w')
init = tensor.fscalar('init') init = tensor.fscalar('init')
out, _ = theano.scan( out, _ = theano.scan(
...@@ -4333,7 +4333,7 @@ class T_Scan(unittest.TestCase): ...@@ -4333,7 +4333,7 @@ class T_Scan(unittest.TestCase):
for out in [y1, y2, y3, y4, y5, y6]: for out in [y1, y2, y3, y4, y5, y6]:
# This used to raise an exception # This used to raise an exception
f = theano.function([W, v], out, mode=mode_with_opt) f = theano.function([W, v], out, mode=mode_with_opt)
f(numpy.zeros((3, 3), dtype=theano.config.floatX), [1, 2]) f(np.zeros((3, 3), dtype=theano.config.floatX), [1, 2])
scan_nodes = scan_nodes_from_fct(f) scan_nodes = scan_nodes_from_fct(f)
assert len(scan_nodes) == 1 assert len(scan_nodes) == 1
...@@ -4375,9 +4375,9 @@ class T_Scan(unittest.TestCase): ...@@ -4375,9 +4375,9 @@ class T_Scan(unittest.TestCase):
# This used to raise an exception # This used to raise an exception
f = theano.function([W, v, vv], out, on_unused_input='ignore', f = theano.function([W, v, vv], out, on_unused_input='ignore',
mode=mode_with_opt) mode=mode_with_opt)
f(numpy.zeros((3, 3), theano.config.floatX), f(np.zeros((3, 3), theano.config.floatX),
[1, 2], [1, 2],
numpy.zeros((3, 3), theano.config.floatX)) np.zeros((3, 3), theano.config.floatX))
scan_nodes = scan_nodes_from_fct(f) scan_nodes = scan_nodes_from_fct(f)
assert len(scan_nodes) == 1 assert len(scan_nodes) == 1
...@@ -4413,7 +4413,7 @@ class T_Scan(unittest.TestCase): ...@@ -4413,7 +4413,7 @@ class T_Scan(unittest.TestCase):
result_inner, _ = theano.scan( result_inner, _ = theano.scan(
fn=loss_inner, fn=loss_inner,
outputs_info=tensor.as_tensor_variable( outputs_info=tensor.as_tensor_variable(
numpy.asarray(0, dtype=numpy.float32)), np.asarray(0, dtype=np.float32)),
non_sequences=[W], non_sequences=[W],
n_steps=1, n_steps=1,
) )
...@@ -4422,7 +4422,7 @@ class T_Scan(unittest.TestCase): ...@@ -4422,7 +4422,7 @@ class T_Scan(unittest.TestCase):
result_outer, _ = theano.scan( result_outer, _ = theano.scan(
fn=loss_outer, fn=loss_outer,
outputs_info=tensor.as_tensor_variable( outputs_info=tensor.as_tensor_variable(
numpy.asarray(0, dtype=numpy.float32)), np.asarray(0, dtype=np.float32)),
non_sequences=[W], non_sequences=[W],
n_steps=n_steps, n_steps=n_steps,
return_list=True, return_list=True,
...@@ -4432,14 +4432,14 @@ class T_Scan(unittest.TestCase): ...@@ -4432,14 +4432,14 @@ class T_Scan(unittest.TestCase):
H = theano.gradient.hessian(cost, W) H = theano.gradient.hessian(cost, W)
print(".", file=sys.stderr) print(".", file=sys.stderr)
f = theano.function([W, n_steps], H) f = theano.function([W, n_steps], H)
f(numpy.ones((8,), dtype='float32'), 1) f(np.ones((8,), dtype='float32'), 1)
def test_strict_mode(self): def test_strict_mode(self):
n = 10 n = 10
w = numpy.array([[-1, 2], [3, -4]]).astype(theano.config.floatX) w = np.array([[-1, 2], [3, -4]]).astype(theano.config.floatX)
w_ = theano.shared(w) w_ = theano.shared(w)
x0 = numpy.array([1, 2]).astype(theano.config.floatX) x0 = np.array([1, 2]).astype(theano.config.floatX)
x0_ = tensor.vector(name='x0', dtype=theano.config.floatX) x0_ = tensor.vector(name='x0', dtype=theano.config.floatX)
def _scan_loose(x): def _scan_loose(x):
...@@ -4474,9 +4474,9 @@ class T_Scan(unittest.TestCase): ...@@ -4474,9 +4474,9 @@ class T_Scan(unittest.TestCase):
def test_strict_mode_ex(self): def test_strict_mode_ex(self):
n = 10 n = 10
w = numpy.array([[-1, 2], [3, -4]]).astype(theano.config.floatX) w = np.array([[-1, 2], [3, -4]]).astype(theano.config.floatX)
w_ = theano.shared(w) w_ = theano.shared(w)
x0 = numpy.array([1, 2]).astype(theano.config.floatX) x0 = np.array([1, 2]).astype(theano.config.floatX)
x0_ = tensor.vector(name='x0', dtype=theano.config.floatX) x0_ = tensor.vector(name='x0', dtype=theano.config.floatX)
def _scan_loose(x): def _scan_loose(x):
...@@ -4497,7 +4497,7 @@ class T_Scan(unittest.TestCase): ...@@ -4497,7 +4497,7 @@ class T_Scan(unittest.TestCase):
# Build a MonitorMode that counts how many values are greater than 10 # Build a MonitorMode that counts how many values are greater than 10
def detect_large_outputs(i, node, fn): def detect_large_outputs(i, node, fn):
for output in fn.outputs: for output in fn.outputs:
if isinstance(output[0], numpy.ndarray): if isinstance(output[0], np.ndarray):
detect_large_outputs.large_count += (output[0] > 10).sum() detect_large_outputs.large_count += (output[0] > 10).sum()
detect_large_outputs.large_count = 0 detect_large_outputs.large_count = 0
...@@ -4516,7 +4516,7 @@ class T_Scan(unittest.TestCase): ...@@ -4516,7 +4516,7 @@ class T_Scan(unittest.TestCase):
f = theano.function(inputs=[A, k], f = theano.function(inputs=[A, k],
outputs=final_result, outputs=final_result,
updates=updates) updates=updates)
f(numpy.asarray([2, 3, .1, 0, 1], dtype=theano.config.floatX), 4) f(np.asarray([2, 3, .1, 0, 1], dtype=theano.config.floatX), 4)
# There should be 3 outputs greater than 10: prior_result[0] at step 3, # There should be 3 outputs greater than 10: prior_result[0] at step 3,
# and prior_result[1] at steps 2 and 3. # and prior_result[1] at steps 2 and 3.
...@@ -4574,19 +4574,19 @@ class ScanGpuTests: ...@@ -4574,19 +4574,19 @@ class ScanGpuTests:
mode=self.mode_with_gpu) mode=self.mode_with_gpu)
# get random initial values # get random initial values
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = rng.uniform(size=(4,), low=-5., high=5.) v_u = rng.uniform(size=(4,), low=-5., high=5.)
v_x0 = rng.uniform() v_x0 = rng.uniform()
W = rng.uniform() W = rng.uniform()
W_in = rng.uniform() W_in = rng.uniform()
v_u = numpy.asarray(v_u, dtype='float32') v_u = np.asarray(v_u, dtype='float32')
v_x0 = numpy.asarray(v_x0, dtype='float32') v_x0 = np.asarray(v_x0, dtype='float32')
W = numpy.asarray(W, dtype='float32') W = np.asarray(W, dtype='float32')
W_in = numpy.asarray(W_in, dtype='float32') W_in = np.asarray(W_in, dtype='float32')
# compute the output in numpy # compute the output in numpy
v_out = numpy.zeros((4,)) v_out = np.zeros((4,))
v_out[0] = v_u[0] * W_in + v_x0 * W v_out[0] = v_u[0] * W_in + v_x0 * W
for step in xrange(1, 4): for step in xrange(1, 4):
v_out[step] = v_u[step] * W_in + v_out[step - 1] * W v_out[step] = v_u[step] * W_in + v_out[step - 1] * W
...@@ -4646,14 +4646,14 @@ class ScanGpuTests: ...@@ -4646,14 +4646,14 @@ class ScanGpuTests:
mode=self.mode_with_gpu) mode=self.mode_with_gpu)
# get random initial values # get random initial values
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = rng.uniform(size=(4,), low=-5., high=5.) v_u = rng.uniform(size=(4,), low=-5., high=5.)
v_x0 = rng.uniform() v_x0 = rng.uniform()
W = rng.uniform() W = rng.uniform()
W_in = rng.uniform() W_in = rng.uniform()
# compute the output in numpy # compute the output in numpy
v_out = numpy.zeros((4,)) v_out = np.zeros((4,))
v_out[0] = v_u[0] * W_in + v_x0 * W v_out[0] = v_u[0] * W_in + v_x0 * W
for step in xrange(1, 4): for step in xrange(1, 4):
v_out[step] = v_u[step] * W_in + v_out[step - 1] * W v_out[step] = v_u[step] * W_in + v_out[step - 1] * W
...@@ -4708,20 +4708,20 @@ class ScanGpuTests: ...@@ -4708,20 +4708,20 @@ class ScanGpuTests:
mode=self.mode_with_gpu) mode=self.mode_with_gpu)
# get random initial values # get random initial values
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_u = rng.uniform(size=(4,), low=-5., high=5.) v_u = rng.uniform(size=(4,), low=-5., high=5.)
v_x0 = rng.uniform() v_x0 = rng.uniform()
W = rng.uniform() W = rng.uniform()
W_in = rng.uniform() W_in = rng.uniform()
# compute the output in numpy # compute the output in numpy
v_out1 = numpy.zeros((4,)) v_out1 = np.zeros((4,))
v_out2 = numpy.zeros((4,), dtype='int64') v_out2 = np.zeros((4,), dtype='int64')
v_out1[0] = v_u[0] * W_in + v_x0 * W v_out1[0] = v_u[0] * W_in + v_x0 * W
v_out2[0] = v_u[0] + v_x0 v_out2[0] = v_u[0] + v_x0
for step in xrange(1, 4): for step in xrange(1, 4):
v_out1[step] = v_u[step] * W_in + v_out1[step - 1] * W v_out1[step] = v_u[step] * W_in + v_out1[step - 1] * W
v_out2[step] = numpy.int64(v_u[step] + v_out1[step - 1]) v_out2[step] = np.int64(v_u[step] + v_out1[step - 1])
theano_out1, theano_out2 = f2(v_u, v_x0, W_in, W) theano_out1, theano_out2 = f2(v_u, v_x0, W_in, W)
utt.assert_allclose(theano_out1, v_out1) utt.assert_allclose(theano_out1, v_out1)
...@@ -4735,8 +4735,8 @@ class ScanGpuTests: ...@@ -4735,8 +4735,8 @@ class ScanGpuTests:
assert self.is_scan_on_gpu(scan_node) assert self.is_scan_on_gpu(scan_node)
def test_gibbs_chain(self): def test_gibbs_chain(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = np.random.RandomState(utt.fetch_seed())
v_vsample = numpy.array(rng.binomial(1, .5, size=(3, 20),), v_vsample = np.array(rng.binomial(1, .5, size=(3, 20),),
dtype='float32') dtype='float32')
vsample = theano.shared(v_vsample) vsample = theano.shared(v_vsample)
trng = theano.sandbox.rng_mrg.MRG_RandomStreams( trng = theano.sandbox.rng_mrg.MRG_RandomStreams(
...@@ -4788,11 +4788,11 @@ class ScanGpuTests: ...@@ -4788,11 +4788,11 @@ class ScanGpuTests:
# Initialize the network parameters # Initialize the network parameters
floatX = theano.config.floatX floatX = theano.config.floatX
U = theano.shared(numpy.zeros((n_in, n_hid), dtype="float32"), U = theano.shared(np.zeros((n_in, n_hid), dtype="float32"),
name='W_xin_to_l1') name='W_xin_to_l1')
V = theano.shared(numpy.zeros((n_hid, n_hid), dtype="float32"), V = theano.shared(np.zeros((n_hid, n_hid), dtype="float32"),
name='W_l1_to_l1') name='W_l1_to_l1')
W = theano.shared(numpy.zeros((n_hid, n_out), dtype="float32"), W = theano.shared(np.zeros((n_hid, n_out), dtype="float32"),
name='W_l1_to_l2') name='W_l1_to_l2')
nparams = [U, V, W] nparams = [U, V, W]
...@@ -4802,7 +4802,7 @@ class ScanGpuTests: ...@@ -4802,7 +4802,7 @@ class ScanGpuTests:
def scan_l(baseline, last_step): def scan_l(baseline, last_step):
return baseline + tensor.dot(last_step, V) return baseline + tensor.dot(last_step, V)
zero_output = tensor.alloc(numpy.asarray(0., dtype="float32"), zero_output = tensor.alloc(np.asarray(0., dtype="float32"),
mb_size, n_hid) mb_size, n_hid)
l1_out, _ = theano.scan(scan_l, sequences=[l1_base], l1_out, _ = theano.scan(scan_l, sequences=[l1_base],
...@@ -4833,9 +4833,9 @@ class ScanGpuTests: ...@@ -4833,9 +4833,9 @@ class ScanGpuTests:
assert len(grad_scan_node.outputs) == 2, len(grad_scan_node.outputs) assert len(grad_scan_node.outputs) == 2, len(grad_scan_node.outputs)
# Call the theano function to ensure the absence of a memory error # Call the theano function to ensure the absence of a memory error
feval_backprop(numpy.zeros((mb_length, mb_size, n_in), feval_backprop(np.zeros((mb_length, mb_size, n_in),
dtype="float32"), dtype="float32"),
numpy.zeros((mb_length, mb_size, n_out), np.zeros((mb_length, mb_size, n_out),
dtype="float32")) dtype="float32"))
def test_memory_reuse_gpudimshuffle(self): def test_memory_reuse_gpudimshuffle(self):
...@@ -4864,11 +4864,11 @@ class ScanGpuTests: ...@@ -4864,11 +4864,11 @@ class ScanGpuTests:
fct = theano.function([input1, init], [out1, out2], fct = theano.function([input1, init], [out1, out2],
mode=self.mode_with_gpu) mode=self.mode_with_gpu)
output = fct(numpy.ones((2, 1, 1), dtype="float32"), output = fct(np.ones((2, 1, 1), dtype="float32"),
numpy.ones((1, 1, 1), dtype="float32")) np.ones((1, 1, 1), dtype="float32"))
expected_output = (numpy.array([2, 4], dtype="float32"), expected_output = (np.array([2, 4], dtype="float32"),
numpy.array([3, 7], dtype="float32")) np.array([3, 7], dtype="float32"))
utt.assert_allclose(output, expected_output) utt.assert_allclose(output, expected_output)
...@@ -4985,7 +4985,7 @@ def test_speed(): ...@@ -4985,7 +4985,7 @@ def test_speed():
if not theano.config.cxx: if not theano.config.cxx:
raise SkipTest("G++ not available, so we need to skip this test.") raise SkipTest("G++ not available, so we need to skip this test.")
r = numpy.arange(10000).astype(theano.config.floatX).reshape(1000, 10) r = np.arange(10000).astype(theano.config.floatX).reshape(1000, 10)
t0 = time.time() t0 = time.time()
for i in xrange(1, 1000): for i in xrange(1, 1000):
...@@ -4993,7 +4993,7 @@ def test_speed(): ...@@ -4993,7 +4993,7 @@ def test_speed():
t1 = time.time() t1 = time.time()
print('python', t1 - t0) print('python', t1 - t0)
r = numpy.arange(10000).astype(theano.config.floatX).reshape(1000, 10) r = np.arange(10000).astype(theano.config.floatX).reshape(1000, 10)
t0 = time.time() t0 = time.time()
r_i = iter(r[1:]) r_i = iter(r[1:])
r_ii = iter(r[:-1]) r_ii = iter(r[:-1])
...@@ -5015,7 +5015,7 @@ def test_speed(): ...@@ -5015,7 +5015,7 @@ def test_speed():
print('python with builtin iterator', t1 - t0) print('python with builtin iterator', t1 - t0)
if 1: if 1:
r = numpy.arange(10000).astype(theano.config.floatX).reshape(1000, 10) r = np.arange(10000).astype(theano.config.floatX).reshape(1000, 10)
s_r = tensor.matrix() s_r = tensor.matrix()
s_y, updates = theano.scan(fn=lambda ri, rii: ri + rii, s_y, updates = theano.scan(fn=lambda ri, rii: ri + rii,
sequences=[s_r[1:]], sequences=[s_r[1:]],
...@@ -5030,9 +5030,9 @@ def test_speed(): ...@@ -5030,9 +5030,9 @@ def test_speed():
print('theano (scan, cvm)', t3 - t2) print('theano (scan, cvm)', t3 - t2)
if 1: if 1:
r = numpy.arange(10000).astype(theano.config.floatX).reshape(-1, 10) r = np.arange(10000).astype(theano.config.floatX).reshape(-1, 10)
shared_r = theano.shared(r) shared_r = theano.shared(r)
s_i = theano.shared(numpy.array(1)) s_i = theano.shared(np.array(1))
s_rinc = tensor.inc_subtensor(shared_r[s_i], shared_r[s_i - 1], s_rinc = tensor.inc_subtensor(shared_r[s_i], shared_r[s_i - 1],
tolerate_inplace_aliasing=True) tolerate_inplace_aliasing=True)
# theano.printing.debugprint(s_rinc) # theano.printing.debugprint(s_rinc)
...@@ -5075,18 +5075,18 @@ def test_speed_rnn(): ...@@ -5075,18 +5075,18 @@ def test_speed_rnn():
L = 10000 L = 10000
N = 50 N = 50
numpy.random.seed(2523452) np.random.seed(2523452)
r = numpy.arange(L * N).astype(theano.config.floatX).reshape(L, N) r = np.arange(L * N).astype(theano.config.floatX).reshape(L, N)
w = numpy.random.randn(N, N).astype(theano.config.floatX) w = np.random.randn(N, N).astype(theano.config.floatX)
t0 = time.time() t0 = time.time()
for i in xrange(1, L): for i in xrange(1, L):
r[i] = numpy.tanh(numpy.dot(r[i - 1], w)) r[i] = np.tanh(np.dot(r[i - 1], w))
t1 = time.time() t1 = time.time()
print('python', t1 - t0) print('python', t1 - t0)
if 1: if 1:
r = numpy.arange(L * N).astype(theano.config.floatX).reshape(L, N) r = np.arange(L * N).astype(theano.config.floatX).reshape(L, N)
s_r = tensor.matrix() s_r = tensor.matrix()
s_y, updates = theano.scan( s_y, updates = theano.scan(
fn=lambda ri, rii: tensor.tanh(tensor.dot(rii, w)), fn=lambda ri, rii: tensor.tanh(tensor.dot(rii, w)),
...@@ -5102,7 +5102,7 @@ def test_speed_rnn(): ...@@ -5102,7 +5102,7 @@ def test_speed_rnn():
print('theano (scan, cvm)', t3 - t2) print('theano (scan, cvm)', t3 - t2)
if 1: if 1:
r = numpy.arange(L * N).astype(theano.config.floatX).reshape(L, N) r = np.arange(L * N).astype(theano.config.floatX).reshape(L, N)
s_w = theano.shared(w) s_w = theano.shared(w)
shared_r = theano.shared(r) shared_r = theano.shared(r)
s_i = theano.scalar.sharedvar.shared(1) s_i = theano.scalar.sharedvar.shared(1)
...@@ -5154,18 +5154,18 @@ def test_speed_batchrnn(): ...@@ -5154,18 +5154,18 @@ def test_speed_batchrnn():
B = 50 B = 50
N = 400 N = 400
numpy.random.seed(2523452) np.random.seed(2523452)
r = numpy.arange(B * L * N).astype(theano.config.floatX).reshape(L, B, N) r = np.arange(B * L * N).astype(theano.config.floatX).reshape(L, B, N)
w = numpy.random.randn(N, N).astype(theano.config.floatX) w = np.random.randn(N, N).astype(theano.config.floatX)
t0 = time.time() t0 = time.time()
for i in xrange(1, L): for i in xrange(1, L):
r[i] = numpy.tanh(numpy.dot(r[i - 1], w)) r[i] = np.tanh(np.dot(r[i - 1], w))
t1 = time.time() t1 = time.time()
print('python', t1 - t0) print('python', t1 - t0)
if 1: if 1:
r = numpy.arange(B * L * N).astype( r = np.arange(B * L * N).astype(
theano.config.floatX).reshape(L, B, N) theano.config.floatX).reshape(L, B, N)
s_w = theano.shared(w) s_w = theano.shared(w)
shared_r = theano.shared(r) shared_r = theano.shared(r)
...@@ -5328,9 +5328,9 @@ def test_compute_test_value(): ...@@ -5328,9 +5328,9 @@ def test_compute_test_value():
theano.config.compute_test_value = 'raise' theano.config.compute_test_value = 'raise'
try: try:
x = tensor.vector('x') x = tensor.vector('x')
xv = numpy.ones(3, dtype=theano.config.floatX) xv = np.ones(3, dtype=theano.config.floatX)
x.tag.test_value = xv x.tag.test_value = xv
y = theano.shared(numpy.arange(3, dtype=theano.config.floatX), y = theano.shared(np.arange(3, dtype=theano.config.floatX),
name='y') name='y')
z, updates = theano.scan( z, updates = theano.scan(
fn=lambda u, v: u + v, fn=lambda u, v: u + v,
...@@ -5351,10 +5351,10 @@ def test_compute_test_value_nonseq(): ...@@ -5351,10 +5351,10 @@ def test_compute_test_value_nonseq():
theano.config.compute_test_value = 'raise' theano.config.compute_test_value = 'raise'
try: try:
x = tensor.vector('x') x = tensor.vector('x')
xv = numpy.ones(3, dtype=theano.config.floatX) xv = np.ones(3, dtype=theano.config.floatX)
x.tag.test_value = xv x.tag.test_value = xv
y = theano.shared( y = theano.shared(
numpy.arange(9, dtype=theano.config.floatX).reshape(3, 3), np.arange(9, dtype=theano.config.floatX).reshape(3, 3),
name='y') name='y')
z, updates = theano.scan( z, updates = theano.scan(
fn=lambda u, v: u + v, fn=lambda u, v: u + v,
...@@ -5373,7 +5373,7 @@ def test_compute_test_value_nonseq(): ...@@ -5373,7 +5373,7 @@ def test_compute_test_value_nonseq():
def test_compute_test_value_grad(): def test_compute_test_value_grad():
# Test case originally reported by Bitton Tenessi # Test case originally reported by Bitton Tenessi
# https://groups.google.com/d/msg/theano-users/fAP3i2CbskQ/3OgBf4yjqiQJ # https://groups.google.com/d/msg/theano-users/fAP3i2CbskQ/3OgBf4yjqiQJ
WEIGHT = numpy.array([1, 2, 1, 3, 4, 1, 5, 6, 1, 7, 8, 1], WEIGHT = np.array([1, 2, 1, 3, 4, 1, 5, 6, 1, 7, 8, 1],
dtype='float32') dtype='float32')
old_compute_test_val = theano.config.compute_test_value old_compute_test_val = theano.config.compute_test_value
...@@ -5387,13 +5387,13 @@ def test_compute_test_value_grad(): ...@@ -5387,13 +5387,13 @@ def test_compute_test_value_grad():
W = W_flat.reshape((2, 2, 3)) W = W_flat.reshape((2, 2, 3))
outputs_mi = tensor.as_tensor_variable( outputs_mi = tensor.as_tensor_variable(
numpy.asarray(0, dtype='float32')) np.asarray(0, dtype='float32'))
outputs_mi.tag.test_value = numpy.asarray(0, dtype='float32') outputs_mi.tag.test_value = np.asarray(0, dtype='float32')
def loss_mi(mi, sum_mi, W): def loss_mi(mi, sum_mi, W):
outputs_ti = tensor.as_tensor_variable( outputs_ti = tensor.as_tensor_variable(
numpy.asarray(0, dtype='float32')) np.asarray(0, dtype='float32'))
outputs_ti.tag.test_value = numpy.asarray(0, dtype='float32') outputs_ti.tag.test_value = np.asarray(0, dtype='float32')
def loss_ti(ti, sum_ti, mi, W): def loss_ti(ti, sum_ti, mi, W):
return W.sum().sum().sum() + sum_ti return W.sum().sum().sum() + sum_ti
...@@ -5430,10 +5430,10 @@ def test_compute_test_value_grad_cast(): ...@@ -5430,10 +5430,10 @@ def test_compute_test_value_grad_cast():
theano.config.compute_test_value = 'raise' theano.config.compute_test_value = 'raise'
try: try:
h = tensor.matrix('h') h = tensor.matrix('h')
h.tag.test_value = numpy.array([[1, 2, 3, 4], [5, 6, 7, 8]], h.tag.test_value = np.array([[1, 2, 3, 4], [5, 6, 7, 8]],
dtype=floatX) dtype=floatX)
w = theano.shared(numpy.random.randn(4, 3).astype(floatX), name='w') w = theano.shared(np.random.randn(4, 3).astype(floatX), name='w')
outputs, _ = theano.scan(lambda i, h, w: (theano.dot(h[i], w), i), outputs, _ = theano.scan(lambda i, h, w: (theano.dot(h[i], w), i),
outputs_info=[None, 0], non_sequences=[h, w], outputs_info=[None, 0], non_sequences=[h, w],
...@@ -5473,10 +5473,10 @@ def test_outputs_taps_check(): ...@@ -5473,10 +5473,10 @@ def test_outputs_taps_check():
def test_default_value_broadcasted(): def test_default_value_broadcasted():
def floatx(X): def floatx(X):
return numpy.asarray(X, dtype=theano.config.floatX) return np.asarray(X, dtype=theano.config.floatX)
def init_weights(shape, name): def init_weights(shape, name):
return theano.shared(floatx(numpy.random.randn(*shape) * 0.1), name) return theano.shared(floatx(np.random.randn(*shape) * 0.1), name)
X = theano.tensor.matrix('X') X = theano.tensor.matrix('X')
in_size = 2 in_size = 2
...@@ -5494,14 +5494,14 @@ def test_default_value_broadcasted(): ...@@ -5494,14 +5494,14 @@ def test_default_value_broadcasted():
gW_x = theano.tensor.grad(cost, W_x) gW_x = theano.tensor.grad(cost, W_x)
updates = [(W_x, W_x - 0.1 * gW_x)] updates = [(W_x, W_x - 0.1 * gW_x)]
f = theano.function([X], outputs=cost, updates=updates) f = theano.function([X], outputs=cost, updates=updates)
f(numpy.random.rand(10, in_size).astype(X.dtype)) f(np.random.rand(10, in_size).astype(X.dtype))
class TestInconsistentBroadcast(unittest.TestCase): class TestInconsistentBroadcast(unittest.TestCase):
def test_raise_error(self): def test_raise_error(self):
x = tensor.tensor3() x = tensor.tensor3()
initial_x = tensor.constant(numpy.zeros((1, 10))) initial_x = tensor.constant(np.zeros((1, 10)))
y, updates = theano.scan(fn=lambda x, prev_x: x + prev_x, y, updates = theano.scan(fn=lambda x, prev_x: x + prev_x,
sequences=x, sequences=x,
outputs_info=[dict(initial=initial_x)]) outputs_info=[dict(initial=initial_x)])
......
theano/scan_module/tests/test_scan_checkpoints.py
浏览文件 @ bea31470
from __future__ import absolute_import, print_function, division from __future__ import absolute_import, print_function, division
import numpy import numpy as np
import unittest import unittest
import theano import theano
...@@ -39,14 +39,14 @@ class TestScanCheckpoint(unittest.TestCase): ...@@ -39,14 +39,14 @@ class TestScanCheckpoint(unittest.TestCase):
f = theano.function(inputs=[self.A, self.k], f = theano.function(inputs=[self.A, self.k],
outputs=[self.result, self.result_check]) outputs=[self.result, self.result_check])
out, out_check = f(range(10), 101) out, out_check = f(range(10), 101)
assert numpy.allclose(out, out_check) assert np.allclose(out, out_check)
def test_backward_pass(self): def test_backward_pass(self):
"""Test gradient computation of A**k.""" """Test gradient computation of A**k."""
f = theano.function(inputs=[self.A, self.k], f = theano.function(inputs=[self.A, self.k],
outputs=[self.grad_A, self.grad_A_check]) outputs=[self.grad_A, self.grad_A_check])
out, out_check = f(range(10), 101) out, out_check = f(range(10), 101)
assert numpy.allclose(out, out_check) assert np.allclose(out, out_check)
@unittest.skipUnless(PYGPU_AVAILABLE, 'Requires pygpu.') @unittest.skipUnless(PYGPU_AVAILABLE, 'Requires pygpu.')
def test_memory(self): def test_memory(self):
...@@ -59,7 +59,7 @@ class TestScanCheckpoint(unittest.TestCase): ...@@ -59,7 +59,7 @@ class TestScanCheckpoint(unittest.TestCase):
f_check = theano.function(inputs=[self.A, self.k], f_check = theano.function(inputs=[self.A, self.k],
outputs=self.grad_A_check, mode=mode_with_gpu) outputs=self.grad_A_check, mode=mode_with_gpu)
free_gmem = theano.gpuarray.type._context_reg[None].free_gmem free_gmem = theano.gpuarray.type._context_reg[None].free_gmem
data = numpy.ones(free_gmem // 3000, dtype=numpy.float32) data = np.ones(free_gmem // 3000, dtype=np.float32)
# Check that it works with the checkpoints # Check that it works with the checkpoints
f_check(data, 1000) f_check(data, 1000)
# Check that the basic scan fails in that case # Check that the basic scan fails in that case
......
theano/scan_module/tests/test_scan_opt.py
浏览文件 @ bea31470
from __future__ import absolute_import, print_function, division from __future__ import absolute_import, print_function, division
import numpy import numpy as np
import unittest import unittest
import theano import theano
...@@ -18,7 +18,7 @@ class TestGaussNewton(unittest.TestCase): ...@@ -18,7 +18,7 @@ class TestGaussNewton(unittest.TestCase):
This test case is based on code by Sigurd Spieckermann. This test case is based on code by Sigurd Spieckermann.
""" """
def setUp(self): def setUp(self):
self.rng = numpy.random.RandomState(utt.fetch_seed()) self.rng = np.random.RandomState(utt.fetch_seed())
def _run(self, num_features, num_timesteps, batch_size, mode): def _run(self, num_features, num_timesteps, batch_size, mode):
# determine shapes of inputs and targets depending on the batch size # determine shapes of inputs and targets depending on the batch size
...@@ -58,8 +58,8 @@ class TestGaussNewton(unittest.TestCase): ...@@ -58,8 +58,8 @@ class TestGaussNewton(unittest.TestCase):
W_hy = theano.shared( W_hy = theano.shared(
(0.01 * self.rng.uniform(size=(10, 1))).astype(config.floatX), (0.01 * self.rng.uniform(size=(10, 1))).astype(config.floatX),
borrow=True) borrow=True)
b_h = theano.shared(numpy.zeros(10).astype(config.floatX), borrow=True) b_h = theano.shared(np.zeros(10).astype(config.floatX), borrow=True)
b_y = theano.shared(numpy.zeros(1).astype(config.floatX), borrow=True) b_y = theano.shared(np.zeros(1).astype(config.floatX), borrow=True)
params = [W_xh, W_hh, W_hy, b_h, b_y] params = [W_xh, W_hh, W_hy, b_h, b_y]
...@@ -171,8 +171,8 @@ class TestPushOutScanOutputDot(object): ...@@ -171,8 +171,8 @@ class TestPushOutScanOutputDot(object):
# Ensure that the function compiled with the optimization produces # Ensure that the function compiled with the optimization produces
# the same results as the function compiled without # the same results as the function compiled without
v_value = numpy.random.random((4)).astype(config.floatX) v_value = np.random.random((4)).astype(config.floatX)
m_value = numpy.random.random((4, 5)).astype(config.floatX) m_value = np.random.random((4, 5)).astype(config.floatX)
output_opt = f_opt(v_value, m_value) output_opt = f_opt(v_value, m_value)
output_no_opt = f_no_opt(v_value, m_value) output_no_opt = f_no_opt(v_value, m_value)
...@@ -217,8 +217,8 @@ class TestPushOutScanOutputDot(object): ...@@ -217,8 +217,8 @@ class TestPushOutScanOutputDot(object):
# Ensure that the function compiled with the optimization produces # Ensure that the function compiled with the optimization produces
# the same results as the function compiled without # the same results as the function compiled without
a_value = numpy.random.random((3, 4)).astype(config.floatX) a_value = np.random.random((3, 4)).astype(config.floatX)
b_value = numpy.random.random((4, 5)).astype(config.floatX) b_value = np.random.random((4, 5)).astype(config.floatX)
output_opt = f_opt(a_value, b_value) output_opt = f_opt(a_value, b_value)
output_no_opt = f_no_opt(a_value, b_value) output_no_opt = f_no_opt(a_value, b_value)
...@@ -263,8 +263,8 @@ class TestPushOutScanOutputDot(object): ...@@ -263,8 +263,8 @@ class TestPushOutScanOutputDot(object):
# Ensure that the function compiled with the optimization produces # Ensure that the function compiled with the optimization produces
# the same results as the function compiled without # the same results as the function compiled without
a_value = numpy.random.random((3, 4)).astype(config.floatX) a_value = np.random.random((3, 4)).astype(config.floatX)
b_value = numpy.random.random((4, 5)).astype(config.floatX) b_value = np.random.random((4, 5)).astype(config.floatX)
output_opt = f_opt(a_value, b_value) output_opt = f_opt(a_value, b_value)
output_no_opt = f_no_opt(a_value, b_value) output_no_opt = f_no_opt(a_value, b_value)
...@@ -296,7 +296,7 @@ class TestPushOutSumOfDot(): ...@@ -296,7 +296,7 @@ class TestPushOutSumOfDot():
dim = 5 dim = 5
# Weight matrices # Weight matrices
U = theano.shared(numpy.random.normal(size=(dim, dim), U = theano.shared(np.random.normal(size=(dim, dim),
scale=0.0001).astype(config.floatX)) scale=0.0001).astype(config.floatX))
U.name = 'U' U.name = 'U'
V = theano.shared(U.get_value()) V = theano.shared(U.get_value())
...@@ -306,7 +306,7 @@ class TestPushOutSumOfDot(): ...@@ -306,7 +306,7 @@ class TestPushOutSumOfDot():
# Variables and their values # Variables and their values
x = T.tensor3('x') x = T.tensor3('x')
x_value = numpy.random.normal(size=(seq_len, batch_size, dim), x_value = np.random.normal(size=(seq_len, batch_size, dim),
scale=0.0001).astype(config.floatX) scale=0.0001).astype(config.floatX)
ri = T.tensor3('ri') ri = T.tensor3('ri')
...@@ -315,7 +315,7 @@ class TestPushOutSumOfDot(): ...@@ -315,7 +315,7 @@ class TestPushOutSumOfDot():
zi = T.tensor3('zi') zi = T.tensor3('zi')
zi_value = x_value zi_value = x_value
init = T.alloc(numpy.cast[config.floatX](0), batch_size, dim) init = T.alloc(np.cast[config.floatX](0), batch_size, dim)
def rnn_step1( def rnn_step1(
# sequences # sequences
x, ri, zi, x, ri, zi,
...@@ -375,8 +375,8 @@ class TestPushOutSumOfDot(): ...@@ -375,8 +375,8 @@ class TestPushOutSumOfDot():
input2 = T.tensor3() input2 = T.tensor3()
input3 = T.tensor3() input3 = T.tensor3()
W = theano.shared(numpy.random.normal(size=(4, 5))).astype(config.floatX) W = theano.shared(np.random.normal(size=(4, 5))).astype(config.floatX)
U = theano.shared(numpy.random.normal(size=(6, 7))).astype(config.floatX) U = theano.shared(np.random.normal(size=(6, 7))).astype(config.floatX)
def inner_fct(seq1, seq2, seq3, previous_output): def inner_fct(seq1, seq2, seq3, previous_output):
temp1 = T.dot(seq1, W) + seq3 temp1 = T.dot(seq1, W) + seq3
...@@ -384,7 +384,7 @@ class TestPushOutSumOfDot(): ...@@ -384,7 +384,7 @@ class TestPushOutSumOfDot():
dot_output = T.dot(temp1, temp2) dot_output = T.dot(temp1, temp2)
return previous_output + dot_output return previous_output + dot_output
init = T.as_tensor_variable(numpy.random.normal(size=(3, 7))) init = T.as_tensor_variable(np.random.normal(size=(3, 7)))
# Compile the function twice, once with the optimization and once # Compile the function twice, once with the optimization and once
# without # without
...@@ -410,9 +410,9 @@ class TestPushOutSumOfDot(): ...@@ -410,9 +410,9 @@ class TestPushOutSumOfDot():
# TODO # TODO
# Compare the outputs of the 2 functions # Compare the outputs of the 2 functions
input1_value = numpy.random.random((2, 3, 4)).astype(config.floatX) input1_value = np.random.random((2, 3, 4)).astype(config.floatX)
input2_value = numpy.random.random((2, 5, 6)).astype(config.floatX) input2_value = np.random.random((2, 5, 6)).astype(config.floatX)
input3_value = numpy.random.random((2, 3, 5)).astype(config.floatX) input3_value = np.random.random((2, 3, 5)).astype(config.floatX)
output_opt = f_opt(input1_value, input2_value, input3_value) output_opt = f_opt(input1_value, input2_value, input3_value)
output_no_opt = f_no_opt(input1_value, input2_value, input3_value) output_no_opt = f_no_opt(input1_value, input2_value, input3_value)
......
theano/scan_module/tests/test_scan_utils.py
浏览文件 @ bea31470
from __future__ import absolute_import, print_function, division from __future__ import absolute_import, print_function, division
import itertools import itertools
import unittest import unittest
import numpy import numpy as np
import theano import theano
from theano import tensor from theano import tensor
from theano.scan_module.scan_utils import equal_computations, map_variables from theano.scan_module.scan_utils import equal_computations, map_variables
...@@ -51,8 +51,8 @@ class TestMapVariables(unittest.TestCase): ...@@ -51,8 +51,8 @@ class TestMapVariables(unittest.TestCase):
s2, = map_variables(self.replacer, [s]) s2, = map_variables(self.replacer, [s])
f = theano.function([x, y, z], [s, s2]) f = theano.function([x, y, z], [s, s2])
rval = f(x=numpy.array([1, 2, 3], dtype=numpy.float32), y=1, z=2) rval = f(x=np.array([1, 2, 3], dtype=np.float32), y=1, z=2)
assert numpy.array_equal(rval, [[1, 2, 3], [2, 4, 6]]) assert np.array_equal(rval, [[1, 2, 3], [2, 4, 6]])
def test_scan(self): def test_scan(self):
x = tensor.vector('x') x = tensor.vector('x')
...@@ -64,7 +64,7 @@ class TestMapVariables(unittest.TestCase): ...@@ -64,7 +64,7 @@ class TestMapVariables(unittest.TestCase):
# should do this as well. # should do this as well.
outer = tensor.scalar("outer") outer = tensor.scalar("outer")
shared = theano.shared( shared = theano.shared(
numpy.array(1., dtype=theano.config.floatX), np.array(1., dtype=theano.config.floatX),
name="shared") name="shared")
constant = tensor.constant(1, name="constant") constant = tensor.constant(1, name="constant")
...@@ -77,7 +77,7 @@ class TestMapVariables(unittest.TestCase): ...@@ -77,7 +77,7 @@ class TestMapVariables(unittest.TestCase):
return r return r
s, _ = theano.scan(step, sequences=x, s, _ = theano.scan(step, sequences=x,
outputs_info=[numpy.array(0.)]) outputs_info=[np.array(0.)])
# ensure z is owned by the outer graph so map_variables() will need to # ensure z is owned by the outer graph so map_variables() will need to
# jump through additional hoops to placate FunctionGraph. # jump through additional hoops to placate FunctionGraph.
t = z * s t = z * s
...@@ -85,8 +85,8 @@ class TestMapVariables(unittest.TestCase): ...@@ -85,8 +85,8 @@ class TestMapVariables(unittest.TestCase):
t2 = z * s2 t2 = z * s2
f = theano.function([x, outer], [t, t2]) f = theano.function([x, outer], [t, t2])
rval = f(x=numpy.array([1, 2, 3], dtype=numpy.float32), outer=0.5) rval = f(x=np.array([1, 2, 3], dtype=np.float32), outer=0.5)
assert numpy.array_equal(rval, [[1, 3, 6], [-1, -3, -6]]) assert np.array_equal(rval, [[1, 3, 6], [-1, -3, -6]])
def test_scan_with_shared_update(self): def test_scan_with_shared_update(self):
x = tensor.vector('x') x = tensor.vector('x')
...@@ -104,7 +104,7 @@ class TestMapVariables(unittest.TestCase): ...@@ -104,7 +104,7 @@ class TestMapVariables(unittest.TestCase):
return r return r
s, _ = theano.scan(step, sequences=x, s, _ = theano.scan(step, sequences=x,
outputs_info=[numpy.array(0.)]) outputs_info=[np.array(0.)])
self.assertRaises(NotImplementedError, self.assertRaises(NotImplementedError,
map_variables, self.replacer, [s]) map_variables, self.replacer, [s])
...@@ -128,7 +128,7 @@ class TestMapVariables(unittest.TestCase): ...@@ -128,7 +128,7 @@ class TestMapVariables(unittest.TestCase):
return r + counter return r + counter
s, _ = theano.scan(step, sequences=x, s, _ = theano.scan(step, sequences=x,
outputs_info=[numpy.array(0.)]) outputs_info=[np.array(0.)])
self.assertRaises(NotImplementedError, self.assertRaises(NotImplementedError,
map_variables, self.replacer, [s]) map_variables, self.replacer, [s])
...@@ -137,7 +137,7 @@ class TestMapVariables(unittest.TestCase): ...@@ -137,7 +137,7 @@ class TestMapVariables(unittest.TestCase):
# inner graph. # inner graph.
outer = tensor.scalar("outer") outer = tensor.scalar("outer")
shared = theano.shared( shared = theano.shared(
numpy.array(1., dtype=theano.config.floatX), np.array(1., dtype=theano.config.floatX),
name="shared") name="shared")
constant = tensor.constant(1., name="constant") constant = tensor.constant(1., name="constant")
z = outer * (shared + constant) z = outer * (shared + constant)
......
theano/tensor/blas.py
浏览文件 @ bea31470
...@@ -130,10 +130,10 @@ import logging ...@@ -130,10 +130,10 @@ import logging
import os import os
import time import time
import numpy import numpy as np
import numpy.distutils import numpy.distutils
try: try:
import numpy.distutils.__config__ import numpy.distutils.__config__ # noqa
except ImportError: except ImportError:
pass pass
...@@ -166,10 +166,10 @@ try: ...@@ -166,10 +166,10 @@ try:
# `scipy.linalg.blas.fblas` with `scipy.linalg.blas`. # `scipy.linalg.blas.fblas` with `scipy.linalg.blas`.
# See http://github.com/scipy/scipy/pull/358 # See http://github.com/scipy/scipy/pull/358
fblas = scipy.linalg.blas fblas = scipy.linalg.blas
_blas_gemv_fns = {numpy.dtype('float32'): fblas.sgemv, _blas_gemv_fns = {np.dtype('float32'): fblas.sgemv,
numpy.dtype('float64'): fblas.dgemv, np.dtype('float64'): fblas.dgemv,
numpy.dtype('complex64'): fblas.cgemv, np.dtype('complex64'): fblas.cgemv,
numpy.dtype('complex128'): fblas.zgemv} np.dtype('complex128'): fblas.zgemv}
except ImportError as e: except ImportError as e:
have_fblas = False have_fblas = False
# This is used in Gemv and ScipyGer. We use CGemv and CGer # This is used in Gemv and ScipyGer. We use CGemv and CGer
...@@ -190,12 +190,12 @@ def check_init_y(): ...@@ -190,12 +190,12 @@ def check_init_y():
if not have_fblas: if not have_fblas:
check_init_y._result = False check_init_y._result = False
y = float('NaN') * numpy.ones((2,)) y = float('NaN') * np.ones((2,))
x = numpy.ones((2,)) x = np.ones((2,))
A = numpy.ones((2, 2)) A = np.ones((2, 2))
gemv = _blas_gemv_fns[y.dtype] gemv = _blas_gemv_fns[y.dtype]
gemv(1.0, A.T, x, 0.0, y, overwrite_y=True, trans=True) gemv(1.0, A.T, x, 0.0, y, overwrite_y=True, trans=True)
check_init_y._result = numpy.isnan(y).any() check_init_y._result = np.isnan(y).any()
return check_init_y._result return check_init_y._result
...@@ -269,7 +269,7 @@ class Gemv(Op): ...@@ -269,7 +269,7 @@ class Gemv(Op):
out_storage[0][0] = gemv(alpha, A.T, x, beta, y, out_storage[0][0] = gemv(alpha, A.T, x, beta, y,
overwrite_y=self.inplace, trans=True) overwrite_y=self.inplace, trans=True)
else: else:
out = numpy.dot(A, x) out = np.dot(A, x)
if alpha != 1: if alpha != 1:
out *= alpha out *= alpha
if beta != 0: if beta != 0:
...@@ -277,7 +277,7 @@ class Gemv(Op): ...@@ -277,7 +277,7 @@ class Gemv(Op):
out += beta * y out += beta * y
else: else:
out += y out += y
out_storage[0][0] = numpy.asarray(out, dtype=y.dtype) out_storage[0][0] = np.asarray(out, dtype=y.dtype)
def infer_shape(self, node, input_shapes): def infer_shape(self, node, input_shapes):
return [input_shapes[0]] return [input_shapes[0]]
...@@ -341,9 +341,9 @@ class Ger(Op): ...@@ -341,9 +341,9 @@ class Ger(Op):
else: else:
A = cA.copy() A = cA.copy()
if calpha != 1: if calpha != 1:
A += calpha * numpy.outer(cx, cy) A += calpha * np.outer(cx, cy)
else: else:
A += numpy.outer(cx, cy) A += np.outer(cx, cy)
cZ[0] = A cZ[0] = A
def infer_shape(self, node, input_shapes): def infer_shape(self, node, input_shapes):
...@@ -900,26 +900,26 @@ class Gemm(GemmRelated): ...@@ -900,26 +900,26 @@ class Gemm(GemmRelated):
if not self.inplace: if not self.inplace:
z = z.copy() # the original z will not be changed z = z.copy() # the original z will not be changed
if z.shape == (): if z.shape == ():
z.itemset(z * a + b * numpy.dot(x, y)) z.itemset(z * a + b * np.dot(x, y))
zout[0] = z zout[0] = z
else: else:
if b == 0.0: if b == 0.0:
if a == 1.0: if a == 1.0:
z[:] = numpy.dot(x, y) z[:] = np.dot(x, y)
elif a == -1.0: elif a == -1.0:
z[:] = -numpy.dot(x, y) z[:] = -np.dot(x, y)
else: else:
z[:] = a * numpy.dot(x, y) z[:] = a * np.dot(x, y)
elif b == 1.0: elif b == 1.0:
if a == 1.0: if a == 1.0:
z += numpy.dot(x, y) z += np.dot(x, y)
elif a == -1.0: elif a == -1.0:
z -= numpy.dot(x, y) z -= np.dot(x, y)
else: else:
z += a * numpy.dot(x, y) z += a * np.dot(x, y)
else: else:
z *= b z *= b
z += a * numpy.dot(x, y) z += a * np.dot(x, y)
zout[0] = z zout[0] = z
def infer_shape(self, node, input_shapes): def infer_shape(self, node, input_shapes):
...@@ -1066,7 +1066,7 @@ def _as_scalar(res, dtype=None): ...@@ -1066,7 +1066,7 @@ def _as_scalar(res, dtype=None):
"""Return None or a TensorVariable whose type is in T.float_scalar_types""" """Return None or a TensorVariable whose type is in T.float_scalar_types"""
if dtype is None: if dtype is None:
dtype = config.floatX dtype = config.floatX
if numpy.all(res.type.broadcastable): if np.all(res.type.broadcastable):
while res.owner and isinstance(res.owner.op, T.DimShuffle): while res.owner and isinstance(res.owner.op, T.DimShuffle):
res = res.owner.inputs[0] res = res.owner.inputs[0]
# may still have some number of True's # may still have some number of True's
...@@ -1216,7 +1216,7 @@ def _gemm_canonicalize(r, scale, rval, maxclients): ...@@ -1216,7 +1216,7 @@ def _gemm_canonicalize(r, scale, rval, maxclients):
vectors = [] vectors = []
matrices = [] matrices = []
for i in r.owner.inputs: for i in r.owner.inputs:
if numpy.all(i.type.broadcastable): if np.all(i.type.broadcastable):
while i.owner and isinstance(i.owner.op, T.DimShuffle): while i.owner and isinstance(i.owner.op, T.DimShuffle):
i = i.owner.inputs[0] i = i.owner.inputs[0]
if i.type.broadcastable: if i.type.broadcastable:
...@@ -1539,7 +1539,7 @@ class Dot22(GemmRelated): ...@@ -1539,7 +1539,7 @@ class Dot22(GemmRelated):
x, y = inp x, y = inp
z, = out z, = out
try: try:
z[0] = numpy.asarray(numpy.dot(x, y)) z[0] = np.asarray(np.dot(x, y))
except ValueError as e: except ValueError as e:
# The error raised by numpy has no shape information, we mean to # The error raised by numpy has no shape information, we mean to
# add that # add that
...@@ -1704,8 +1704,8 @@ def local_dot22_to_ger_or_gemv(node): ...@@ -1704,8 +1704,8 @@ def local_dot22_to_ger_or_gemv(node):
x, y = node.inputs x, y = node.inputs
xb = x.broadcastable xb = x.broadcastable
yb = y.broadcastable yb = y.broadcastable
one = T.as_tensor_variable(numpy.asarray(1, dtype=x.dtype)) one = T.as_tensor_variable(np.asarray(1, dtype=x.dtype))
zero = T.as_tensor_variable(numpy.asarray(0, dtype=x.dtype)) zero = T.as_tensor_variable(np.asarray(0, dtype=x.dtype))
if xb[1] and yb[0]: if xb[1] and yb[0]:
# x and y are both vectors so this might qualifies for a GER # x and y are both vectors so this might qualifies for a GER
xv = x.dimshuffle(0) xv = x.dimshuffle(0)
...@@ -1810,7 +1810,7 @@ class Dot22Scalar(GemmRelated): ...@@ -1810,7 +1810,7 @@ class Dot22Scalar(GemmRelated):
x, y, scalar = inp x, y, scalar = inp
z, = out z, = out
try: try:
z[0] = numpy.asarray(scalar * numpy.dot(x, y)) z[0] = np.asarray(scalar * np.dot(x, y))
except ValueError as e: except ValueError as e:
# The error raised by numpy has no shape information, we # The error raised by numpy has no shape information, we
# mean to add that # mean to add that
...@@ -2034,9 +2034,9 @@ class BatchedDot(Op): ...@@ -2034,9 +2034,9 @@ class BatchedDot(Op):
shape = self.infer_shape(node, [i.shape for i in inp])[0] shape = self.infer_shape(node, [i.shape for i in inp])[0]
dtype = node.outputs[0].dtype dtype = node.outputs[0].dtype
z0 = z[0] = numpy.empty(shape, dtype=dtype) z0 = z[0] = np.empty(shape, dtype=dtype)
for i in xrange(z0.shape[0]): for i in xrange(z0.shape[0]):
z0[i] = numpy.dot(x[i], y[i]) z0[i] = np.dot(x[i], y[i])
def c_support_code(self): def c_support_code(self):
batch_gemm_defn = """ batch_gemm_defn = """
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论