def test_seed_fn():
idx = ivector()
for new_seed, same in [(234, True), (None, True), (23, False)]:
random = MRG_RandomStream(234)
fn1 = function([], random.uniform((2, 2), dtype="float32"))
fn2 = function([], random.uniform((3, 3), nstreams=2, dtype="float32"))
fn3 = function([idx], random.uniform(idx, nstreams=3, ndim=1, dtype="float32"))
fn1_val0 = fn1()
fn1_val1 = fn1()
assert not np.allclose(fn1_val0, fn1_val1)
fn2_val0 = fn2()
fn2_val1 = fn2()
assert not np.allclose(fn2_val0, fn2_val1)
fn3_val0 = fn3([4])
fn3_val1 = fn3([4])
assert not np.allclose(fn3_val0, fn3_val1)
assert fn1_val0.size == 4
assert fn2_val0.size == 9
random.seed(new_seed)
fn1_val2 = fn1()
fn1_val3 = fn1()
fn2_val2 = fn2()
fn2_val3 = fn2()
fn3_val2 = fn3([4])
fn3_val3 = fn3([4])
assert np.allclose(fn1_val0, fn1_val2) == same
assert np.allclose(fn1_val1, fn1_val3) == same
assert np.allclose(fn2_val0, fn2_val2) == same
assert np.allclose(fn2_val1, fn2_val3) == same
assert np.allclose(fn3_val0, fn3_val2) == same
assert np.allclose(fn3_val1, fn3_val3) == same
def test_f16_nonzero(mode=None, op_to_check=rng_mrg.mrg_uniform):
srng = MRG_RandomStream(seed=utt.fetch_seed())
m = srng.uniform(size=(1000, 1000), dtype="float16")
assert m.dtype == "float16", m.type
f = function([], m, mode=mode)
assert any(isinstance(n.op, op_to_check) for n in f.maker.fgraph.apply_nodes)
m_val = f()
assert np.all((0 < m_val) & (m_val < 1))
def test_deterministic():
seed = utt.fetch_seed()
sample_size = (10, 20)
R = MRG_RandomStream(seed=seed)
u = R.uniform(size=sample_size)
f = function([], u)
fsample1 = f()
fsample2 = f()
assert not np.allclose(fsample1, fsample2)
R2 = MRG_RandomStream(seed=seed)
u2 = R2.uniform(size=sample_size)
g = function([], u2)
gsample1 = g()
gsample2 = g()
assert np.allclose(fsample1, gsample1)
assert np.allclose(fsample2, gsample2)
def test_bad_size(self):
R = MRG_RandomStream(234)
for size in [
(0, 100),
(-1, 100),
(1, 0),
]:
with pytest.raises(ValueError):
R.uniform(size)
with pytest.raises(ValueError):
R.binomial(size)
with pytest.raises(ValueError):
R.multinomial(size, 1, [])
with pytest.raises(ValueError):
R.normal(size)
with pytest.raises(ValueError):
R.truncated_normal(size)
def test_simple_shared_mrg_random():
aesara_rng = MRG_RandomStream(10)
values, updates = scan(
lambda: aesara_rng.uniform((2, ), -1, 1),
[],
[],
[],
n_steps=5,
truncate_gradient=-1,
go_backwards=False,
)
my_f = function([], values, updates=updates, allow_input_downcast=True)
# Just check for run-time errors
my_f()
my_f()
def test_target_parameter():
srng = MRG_RandomStream()
pvals = np.array([[0.98, 0.01, 0.01], [0.01, 0.49, 0.50]])
def basic_target_parameter_test(x):
f = function([], x)
assert isinstance(f(), np.ndarray)
basic_target_parameter_test(srng.uniform((3, 2), target="cpu"))
basic_target_parameter_test(srng.normal((3, 2), target="cpu"))
basic_target_parameter_test(srng.truncated_normal((3, 2), target="cpu"))
basic_target_parameter_test(srng.binomial((3, 2), target="cpu"))
basic_target_parameter_test(
srng.multinomial(pvals=pvals.astype("float32"), target="cpu"))
basic_target_parameter_test(
srng.choice(p=pvals.astype("float32"), replace=False, target="cpu"))
basic_target_parameter_test(
srng.multinomial_wo_replacement(pvals=pvals.astype("float32"),
target="cpu"))
def test_cpu_target_with_shared_variable():
srng = MRG_RandomStream()
s = np.random.rand(2, 3).astype("float32")
x = gpuarray_shared_constructor(s, name="x")
try:
# To have aesara.shared(x) try to move on the GPU
aesara.compile.shared_constructor(gpuarray_shared_constructor)
y = srng.uniform(x.shape, target="cpu")
y.name = "y"
z = (x * y).sum()
z.name = "z"
fz = aesara.function([], z, mode=mode)
nodes = fz.maker.fgraph.toposort()
assert not any(isinstance(node.op, GPUA_mrg_uniform) for node in nodes)
finally:
aesara.compile.shared_constructor(gpuarray_shared_constructor,
remove=True)
def test_GPUA_full_fill():
# Make sure the whole sample buffer is filled. Also make sure
# large samples are consistent with CPU results.
# This needs to be large to trigger the problem on GPU
size = (10, 1000)
R = MRG_RandomStream(234)
uni = R.uniform(size, nstreams=60 * 256)
f_cpu = aesara.function([], uni)
rstate_gpu = gpuarray_shared_constructor(
R.state_updates[-1][0].get_value())
new_rstate, sample = GPUA_mrg_uniform.new(rstate_gpu,
ndim=None,
dtype="float32",
size=size)
rstate_gpu.default_update = new_rstate
f_gpu = aesara.function([], sample, mode=mode)
utt.assert_allclose(f_cpu(), f_gpu())
def test_consistency_randomstreams():
# Verify that the random numbers generated by MRG_RandomStream
# are the same as the reference (Java) implementation by L'Ecuyer et al.
seed = 12345
n_samples = 5
n_streams = 12
n_substreams = 7
samples = []
rng = MRG_RandomStream(seed=seed)
for i in range(n_streams):
stream_samples = []
u = rng.uniform(size=(n_substreams, ), nstreams=n_substreams)
f = function([], u)
for j in range(n_samples):
s = f()
stream_samples.append(s)
stream_samples = np.array(stream_samples)
stream_samples = stream_samples.T.flatten()
samples.append(stream_samples)
samples = np.array(samples).flatten()
assert np.allclose(samples, java_samples)
def test_broadcastable():
R = MRG_RandomStream(234)
x = matrix()
size1 = (10, 1)
size2 = (x.shape[0], 1)
pvals_1 = np.random.uniform(0, 1, size=size1)
pvals_1 = pvals_1 / sum(pvals_1)
pvals_2 = R.uniform(size=size2)
pvals_2 = pvals_2 / at_sum(pvals_2)
for distribution in [
R.uniform,
R.normal,
R.truncated_normal,
R.binomial,
R.multinomial,
R.multinomial_wo_replacement,
]:
# multinomial or multinomial_wo_replacement does not support "size" argument,
# the sizes of them are implicitly defined with "pvals" argument.
if distribution in [R.multinomial, R.multinomial_wo_replacement]:
# check when all dimensions are constant
uu = distribution(pvals=pvals_1)
assert uu.broadcastable == (False, True)
# check when some dimensions are aesara variables
uu = distribution(pvals=pvals_2)
assert uu.broadcastable == (False, True)
else:
# check when all dimensions are constant
uu = distribution(size=size1)
assert uu.broadcastable == (False, True)
# check when some dimensions are aesara variables
uu = distribution(size=size2)
assert uu.broadcastable == (False, True)
class Graph:
def __init__(self, seed=123):
self.rng = MRG_RandomStream(seed)
self.y = self.rng.uniform(size=(1, ))
def test_uniform():
# TODO: test param low, high
# TODO: test size=None
# TODO: test ndim!=size.ndim
# TODO: test bad seed
# TODO: test size=Var, with shape that change from call to call
if (config.mode in ("DEBUG_MODE", "DebugMode", "FAST_COMPILE")
or config.mode == "Mode" and config.linker in ["py"]):
sample_size = (10, 100)
steps = 50
else:
sample_size = (500, 50)
steps = int(1e3)
x = matrix()
for size, const_size, var_input, input in [
(sample_size, sample_size, [], []),
(x.shape, sample_size, [x],
[np.zeros(sample_size, dtype=config.floatX)]),
(
(x.shape[0], sample_size[1]),
sample_size,
[x],
[np.zeros(sample_size, dtype=config.floatX)],
),
# test empty size (scalar)
((), (), [], []),
]:
# TEST CPU IMPLEMENTATION
# The python and C implementation are tested with DebugMode
x = matrix()
R = MRG_RandomStream(234)
# Note: we specify `nstreams` to avoid a warning.
# TODO Look for all occurrences of `guess_n_streams` and `30 * 256`
# for such situations: it would be better to instead filter the
# warning using the warning module.
u = R.uniform(size=size,
nstreams=rng_mrg.guess_n_streams(size, warn=False))
f = function(var_input, u)
assert any(
isinstance(node.op, mrg_uniform)
for node in f.maker.fgraph.toposort())
f(*input)
# Increase the number of steps if sizes implies only a few samples
if np.prod(const_size) < 10:
steps_ = steps * 100
else:
steps_ = steps
check_basics(f, steps_, const_size, prefix="mrg cpu", inputs=input)
RR = RandomStream(234)
uu = RR.uniform(size=size)
ff = function(var_input, uu)
# It's not our problem if numpy generates 0 or 1
check_basics(ff,
steps_,
const_size,
prefix="numpy",
allow_01=True,
inputs=input)
def test_undefined_grad():
srng = MRG_RandomStream(seed=1234)
# checking uniform distribution
low = scalar()
out = srng.uniform((), low=low)
with pytest.raises(NullTypeGradError):
grad(out, low)
high = scalar()
out = srng.uniform((), low=0, high=high)
with pytest.raises(NullTypeGradError):
grad(out, high)
out = srng.uniform((), low=low, high=high)
with pytest.raises(NullTypeGradError):
grad(out, (low, high))
# checking binomial distribution
prob = scalar()
out = srng.binomial((), p=prob)
with pytest.raises(NullTypeGradError):
grad(out, prob)
# checking multinomial distribution
prob1 = scalar()
prob2 = scalar()
p = [as_tensor_variable([prob1, 0.5, 0.25])]
out = srng.multinomial(size=None, pvals=p, n=4)[0]
with pytest.raises(NullTypeGradError):
grad(at_sum(out), prob1)
p = [as_tensor_variable([prob1, prob2])]
out = srng.multinomial(size=None, pvals=p, n=4)[0]
with pytest.raises(NullTypeGradError):
grad(at_sum(out), (prob1, prob2))
# checking choice
p = [as_tensor_variable([prob1, prob2, 0.1, 0.2])]
out = srng.choice(a=None, size=1, p=p, replace=False)[0]
with pytest.raises(NullTypeGradError):
grad(out[0], (prob1, prob2))
p = [as_tensor_variable([prob1, prob2])]
out = srng.choice(a=None, size=1, p=p, replace=False)[0]
with pytest.raises(NullTypeGradError):
grad(out[0], (prob1, prob2))
p = [as_tensor_variable([prob1, 0.2, 0.3])]
out = srng.choice(a=None, size=1, p=p, replace=False)[0]
with pytest.raises(NullTypeGradError):
grad(out[0], prob1)
# checking normal distribution
avg = scalar()
out = srng.normal((), avg=avg)
with pytest.raises(NullTypeGradError):
grad(out, avg)
std = scalar()
out = srng.normal((), avg=0, std=std)
with pytest.raises(NullTypeGradError):
grad(out, std)
out = srng.normal((), avg=avg, std=std)
with pytest.raises(NullTypeGradError):
grad(out, (avg, std))
# checking truncated normal distribution
avg = scalar()
out = srng.truncated_normal((), avg=avg)
with pytest.raises(NullTypeGradError):
grad(out, avg)
std = scalar()
out = srng.truncated_normal((), avg=0, std=std)
with pytest.raises(NullTypeGradError):
grad(out, std)
out = srng.truncated_normal((), avg=avg, std=std)
with pytest.raises(NullTypeGradError):
grad(out, (avg, std))
