def test_default_updates_multiple(self):
x = shared(0)
y = shared(1)
x.default_update = x - 1
y.default_update = y + 1
f1 = pfunc([], [x, y])
f1()
assert x.get_value() == -1
assert y.get_value() == 2
f2 = pfunc([], [x, y], updates=[(x, (x - 2))], no_default_updates=[y])
f2()
assert x.get_value() == -3
assert y.get_value() == 2
f3 = pfunc([], [x, y], updates=[(x, (x - 2))], no_default_updates=True)
f3()
assert x.get_value() == -5
assert y.get_value() == 2
f4 = pfunc([], [x, y], updates=[(y, (y - 2))])
f4()
assert x.get_value() == -6
assert y.get_value() == 0
def test_ctors(self):
if PYTHON_INT_BITWIDTH == 32:
assert shared(7).type == iscalar, shared(7).type
else:
assert shared(7).type == lscalar, shared(7).type
assert shared(7.0).type == dscalar
assert shared(np.float32(7)).type == fscalar
# test tensor constructor
b = shared(np.zeros((5, 5), dtype="int32"))
assert b.type == TensorType("int32", broadcastable=[False, False])
b = shared(np.random.rand(4, 5))
assert b.type == TensorType("float64", broadcastable=[False, False])
b = shared(np.random.rand(5, 1, 2))
assert b.type == TensorType("float64",
broadcastable=[False, False, False])
assert shared([]).type == generic
def badfunc():
shared(7, bad_kw=False)
with pytest.raises(TypeError):
badfunc()
def test_doc(self):
# Ensure the code given in pfunc.txt works as expected
# Example #1.
a = lscalar()
b = shared(1)
f1 = pfunc([a], (a + b))
f2 = pfunc([In(a, value=44)], a + b, updates={b: b + 1})
assert b.get_value() == 1
assert f1(3) == 4
assert f2(3) == 4
assert b.get_value() == 2
assert f1(3) == 5
b.set_value(0)
assert f1(3) == 3
# Example #2.
a = lscalar()
b = shared(7)
f1 = pfunc([a], a + b)
f2 = pfunc([a], a * b)
assert f1(5) == 12
b.set_value(8)
assert f1(5) == 13
assert f2(4) == 32
def test_strict_generic(self):
# this should work, because
# generic can hold anything even when strict=True
u = shared("asdf", strict=False)
v = shared("asdf", strict=True)
u.set_value(88)
v.set_value(88)
def test_clone0(self):
x = shared(np.asarray([4, 4, 4]))
y = shared(np.asarray([4, 4, 4]))
z = shared(np.asarray([2, 2, 2]))
up = pfunc(
[], [], updates={x: (x * 5), y: ((x * 5) + y), z: (((x * 5) + y) ** z)}
)
up()
assert np.all(x.get_value() == 20)
assert np.all(y.get_value() == 24)
assert np.all(z.get_value() == (24 ** 2))
def test_scalar_floatX(self):
# the test should assure that floatX is not used in the shared
# constructor for scalars Shared values can change, and since we don't
# know the range they might take, we should keep the same
# bit width / precision as the original value used to create the
# shared variable.
# Since downcasting of a value now raises an Exception,
def f(var, val):
var.set_value(val)
b = shared(np.int64(7), allow_downcast=True)
assert b.type == aesara.tensor.lscalar
f(b, 8.23)
assert b.get_value() == 8
b = shared(np.int32(7), allow_downcast=True)
assert b.type == aesara.tensor.iscalar
f(b, 8.23)
assert b.get_value() == 8
b = shared(np.int16(7), allow_downcast=True)
assert b.type == aesara.tensor.wscalar
f(b, 8.23)
assert b.get_value() == 8
b = shared(np.int8(7), allow_downcast=True)
assert b.type == aesara.tensor.bscalar
f(b, 8.23)
assert b.get_value() == 8
b = shared(np.float64(7.234), allow_downcast=True)
assert b.type == aesara.tensor.dscalar
f(b, 8)
assert b.get_value() == 8
b = shared(np.float32(7.234), allow_downcast=True)
assert b.type == aesara.tensor.fscalar
f(b, 8)
assert b.get_value() == 8
b = shared(np.float(7.234), allow_downcast=True)
assert b.type == aesara.tensor.dscalar
f(b, 8)
assert b.get_value() == 8
b = shared(7.234, allow_downcast=True)
assert b.type == aesara.tensor.dscalar
f(b, 8)
assert b.get_value() == 8
b = shared(np.zeros((5, 5), dtype="float32"))
with pytest.raises(TypeError):
f(b, np.random.rand(5, 5))
def test_shared(self):
# CHECK: two functions (f1 and f2) can share w
w = shared(np.random.rand(2, 2), "w")
wval = w.get_value(borrow=False)
x = dmatrix()
out1 = w + x
out2 = w * x
f1 = pfunc([x], [out1])
f2 = pfunc([x], [out2])
xval = np.random.rand(2, 2)
assert np.all(f1(xval) == xval + wval)
assert np.all(f2(xval) == xval * wval)
# CHECK: updating a shared value
f3 = pfunc([x], out1, updates=[(w, (w - 1))])
# f3 changes the value of w
assert np.all(f3(xval) == xval + wval)
# this same value is read by f1
assert np.all(f1(xval) == xval + (wval - 1))
w.set_value(w.get_value(borrow=True) * 10, borrow=True)
# this same value is read by f1
assert np.all(f1(xval) == xval + w.get_value(borrow=True))
def gen(self, op, *args, **kwargs):
"""
Create a new random stream in this container.
Parameters
----------
op
A RandomFunction instance to
args
Interpreted by `op`.
kwargs
Interpreted by `op`.
Returns
-------
Tensor Variable
The symbolic random draw part of op()'s return value.
This function stores the updated RandomStateType Variable
for use at `build` time.
"""
seed = int(self.gen_seedgen.randint(2**30))
random_state_variable = shared(np.random.RandomState(seed))
# Add a reference to distinguish from other shared variables
random_state_variable.tag.is_rng = True
new_r, out = op(random_state_variable, *args, **kwargs)
out.rng = random_state_variable
out.update = (random_state_variable, new_r)
self.state_updates.append(out.update)
random_state_variable.default_update = new_r
return out
def test_shared_mutable(self):
bval = np.arange(5)
b = shared(bval)
b_out = b * 2
# shared vars copy args.
assert b.get_value(borrow=True) is not bval
# so we do this to get at the underlying data
bval = data_of(b)
# by default, shared are not mutable unless doing an explicit update
f = pfunc([], [b_out], mode="FAST_RUN")
assert (f() == np.arange(5) * 2).all()
assert np.all(b.get_value(borrow=True) == np.arange(5))
# using updates, b is now a mutable parameter
f = pfunc([], [b_out], updates=[(b, b_out)], mode="FAST_RUN")
assert (f() == (np.arange(5) * 2)).all()
# because of the update
assert (b.get_value(borrow=True) == (np.arange(5) * 2)).all()
assert (bval == (np.arange(5) * 2)).all() # because of mutable=True
# do not depend on updates being in-place though!
bval = np.arange(5)
b.set_value(bval, borrow=True)
bval = data_of(b)
f = pfunc([], [b_out], updates=[(b, (b_out + 3))], mode="FAST_RUN")
assert (f() == (np.arange(5) * 2)).all()
# because of the update
assert (b.get_value(borrow=True) == ((np.arange(5) * 2) + 3)).all()
# bval got modified to something...
assert not (bval == np.arange(5)).all()
# ... but not to b.value !
assert not (bval == b.get_value(borrow=True)).all()
def test_givens_replaces_shared_variable2(self):
a = shared(1.0, "a")
a.default_update = a + 3
c = a + 10
f = pfunc([], c, givens={a: (a + 10)})
assert f() == 21
assert f() == 34
def test_no_shared_as_input(self):
# Test that shared variables cannot be used as function inputs.
w_init = np.random.rand(2, 2)
w = shared(w_init.copy(), "w")
with pytest.raises(
TypeError, match=r"^Cannot use a shared variable \(w\) as explicit input"
):
pfunc([w], aet_sum(w * w))
def test_update_equiv(self):
# Like test_update_same, but the update expression is simplified until
# it is found to be equal to the original variable
a = shared(1.0, "a")
b = shared(np.ones((2, 3)), "b")
# See comment in test_update_same about why we try both
# shared variables.
f = function([], [], updates=[(a, a), (b, (2 * b - b))])
g = function([], [], updates=[(a, (a * 2 - a)), (b, b)])
f()
assert a.get_value(borrow=True).shape == (), a.get_value()
assert b.get_value(borrow=True).shape == (2, 3), b.get_value()
g()
assert a.get_value(borrow=True).shape == (), a.get_value()
assert b.get_value(borrow=True).shape == (2, 3), b.get_value()
def test_default_scalar_container(self):
# Similar in spirit to test_default_container, but updating a scalar
# variable. This is a sanity check for non mutable types.
x = shared(0.0, "x")
f = pfunc([], x)
assert f() == 0
x.set_value(x.get_value(borrow=True) + 1, borrow=True)
assert f() == 1
def test_default_updates_partial_graph(self):
a = shared(0)
a.default_update = a + 1 # Increment a each time it is used
b = 2 * a
# Use only the tip of the graph, a is not used
f = pfunc([b], b)
assert a.get_value() == 0
f(21)
assert a.get_value() == 0
def test_givens_replaces_shared_variable(self):
a = shared(1.0, "a")
a.default_update = a + 3.0
b = dscalar("b")
c = a + 10
f = pfunc([b], c, givens={a: b})
assert len(f.maker.fgraph.inputs) == 1
assert len(f.maker.fgraph.outputs) == 1
def test_tensor_floatX(self):
def f(var, val):
var.set_value(val)
b = shared(np.int64([7]), allow_downcast=True)
assert b.type == lvector
f(b, [8.23])
assert b.get_value() == 8
b = shared(np.int32([7]), allow_downcast=True)
assert b.type == ivector
f(b, [8.23])
assert b.get_value() == 8
b = shared(np.int16([7]), allow_downcast=True)
assert b.type == wvector
f(b, [8.23])
assert b.get_value() == 8
b = shared(np.int8([7]), allow_downcast=True)
assert b.type == bvector
f(b, [8.23])
assert b.get_value() == 8
b = shared(np.float64([7.234]), allow_downcast=True)
assert b.type == dvector
f(b, [8])
assert b.get_value() == 8
b = shared(np.float32([7.234]), allow_downcast=True)
assert b.type == fvector
f(b, [8])
assert b.get_value() == 8
# np.float([7.234]) don't work
# b = shared(np.float([7.234]))
# assert b.type == dvector
# f(b,[8])
# This generate a generic type. Should we cast? I don't think.
# b = shared([7.234])
# assert b.type == dvector
# f(b,[8])
b = shared(np.asarray([7.234], dtype=config.floatX),
allow_downcast=True)
assert b.dtype == config.floatX
f(b, [8])
assert b.get_value() == 8
b = shared(np.zeros((5, 5), dtype="float32"))
with pytest.raises(TypeError):
f(b, np.random.rand(5, 5))
def test_default_updates_expressions(self):
x = shared(0)
y = shared(1)
a = lscalar("a")
z = a * x
x.default_update = x + y
f1 = pfunc([a], z)
f1(12)
assert x.get_value() == 1
f2 = pfunc([a], z, no_default_updates=True)
assert f2(7) == 7
assert x.get_value() == 1
f3 = pfunc([a], z, no_default_updates=[x])
assert f3(9) == 9
assert x.get_value() == 1
def test_scalar_strict(self):
def f(var, val):
var.set_value(val)
b = shared(np.int64(7), strict=True)
assert b.type == aesara.tensor.lscalar
with pytest.raises(TypeError):
f(b, 8.23)
b = shared(np.int32(7), strict=True)
assert b.type == aesara.tensor.iscalar
with pytest.raises(TypeError):
f(b, 8.23)
b = shared(np.int16(7), strict=True)
assert b.type == aesara.tensor.wscalar
with pytest.raises(TypeError):
f(b, 8.23)
b = shared(np.int8(7), strict=True)
assert b.type == aesara.tensor.bscalar
with pytest.raises(TypeError):
f(b, 8.23)
b = shared(np.float64(7.234), strict=True)
assert b.type == aesara.tensor.dscalar
with pytest.raises(TypeError):
f(b, 8)
b = shared(np.float32(7.234), strict=True)
assert b.type == aesara.tensor.fscalar
with pytest.raises(TypeError):
f(b, 8)
b = shared(np.float(7.234), strict=True)
assert b.type == aesara.tensor.dscalar
with pytest.raises(TypeError):
f(b, 8)
b = shared(7.234, strict=True)
assert b.type == aesara.tensor.dscalar
with pytest.raises(TypeError):
f(b, 8)
b = shared(np.zeros((5, 5), dtype="float32"))
with pytest.raises(TypeError):
f(b, np.random.rand(5, 5))
def __init__(
self,
input=None,
target=None,
n_input=1,
n_hidden=1,
n_output=1,
lr=1e-3,
**kw,
):
super().__init__(**kw)
if input is None:
input = dvector("input")
if target is None:
target = dvector("target")
self.input = input
self.target = target
self.lr = shared(lr, "learning_rate")
self.w1 = shared(np.zeros((n_hidden, n_input)), "w1")
self.w2 = shared(np.zeros((n_output, n_hidden)), "w2")
# print self.lr.type
self.hidden = sigmoid(dot(self.w1, self.input))
self.output = dot(self.w2, self.hidden)
self.cost = aet_sum((self.output - self.target)**2)
self.sgd_updates = {
self.w1: self.w1 - self.lr * grad(self.cost, self.w1),
self.w2: self.w2 - self.lr * grad(self.cost, self.w2),
}
self.sgd_step = pfunc(
params=[self.input, self.target],
outputs=[self.output, self.cost],
updates=self.sgd_updates,
)
self.compute_output = pfunc([self.input], self.output)
self.output_from_hidden = pfunc([self.hidden], self.output)
def test_update(self):
# Test update mechanism in different settings.
# Simple value assignment.
x = shared(0)
assign = pfunc([], [], updates={x: 3})
assign()
assert x.get_value() == 3
# Basic increment function.
x.set_value(0)
inc = pfunc([], [], updates={x: x + 1})
inc()
assert x.get_value() == 1
# Increment by a constant value.
x.set_value(-1)
y = shared(2)
inc_by_y = pfunc([], [], updates={x: x + y})
inc_by_y()
assert x.get_value() == 1
def test_default_updates_input(self):
x = shared(0)
y = shared(1)
if PYTHON_INT_BITWIDTH == 32:
a = iscalar("a")
else:
a = lscalar("a")
x.default_update = y
y.default_update = y + a
f1 = pfunc([], x, no_default_updates=True)
f1()
assert x.get_value() == 0
assert y.get_value() == 1
f2 = pfunc([], x, no_default_updates=[x])
f2()
assert x.get_value() == 0
assert y.get_value() == 1
f3 = pfunc([], x, no_default_updates=[y])
f3()
assert x.get_value() == 1
assert y.get_value() == 1
f4 = pfunc([a], x)
f4(2)
assert x.get_value() == 1
assert y.get_value() == 3
f5 = pfunc([], x, updates={y: (y - 1)})
f5()
assert x.get_value() == 3
assert y.get_value() == 2
# a is needed as input if y.default_update is used
with pytest.raises(MissingInputError):
pfunc([], x)
def test_update_err_broadcast(self):
# Test that broadcastable dimensions raise error
data = np.random.rand(10, 10).astype("float32")
output_var = shared(name="output", value=data)
# the update_var has type matrix, and the update expression
# is a broadcasted scalar, and that should be allowed.
with pytest.raises(TypeError):
function(
inputs=[],
outputs=[],
updates={output_var: output_var.sum().dimshuffle("x", "x")},
)
def test_default_updates_chained(self):
x = shared(2)
y = shared(1)
z = shared(-1)
x.default_update = x - y
y.default_update = z
z.default_update = z - 1
f1 = pfunc([], [x])
f1()
assert x.get_value() == 1
assert y.get_value() == -1
assert z.get_value() == -2
f2 = pfunc([], [x, y])
f2()
assert x.get_value() == 2
assert y.get_value() == -2
assert z.get_value() == -3
f3 = pfunc([], [y])
f3()
assert x.get_value() == 2
assert y.get_value() == -3
assert z.get_value() == -4
f4 = pfunc([], [x, y], no_default_updates=[x])
f4()
assert x.get_value() == 2
assert y.get_value() == -4
assert z.get_value() == -5
f5 = pfunc([], [x, y, z], no_default_updates=[z])
f5()
assert x.get_value() == 6
assert y.get_value() == -5
assert z.get_value() == -5
def test_gradient_scan():
# 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
aesara_rng = MRG_RandomStream(10)
w = shared(np.ones(1, dtype="float32"))
def one_step(x):
return x + aesara_rng.uniform((1, ), dtype="float32") * w
x = vector(dtype="float32")
values, updates = scan(one_step, outputs_info=x, n_steps=10)
gw = grad(aet_sum(values[-1]), w)
f = function([x], gw)
f(np.arange(1, dtype="float32"))
def test_default_container(self):
# Ensure it is possible to (implicitly) use a shared variable in a
# function, as a 'state' that can be updated at will.
rng = np.random.RandomState(1827)
w_init = rng.rand(5)
w = shared(w_init.copy(), "w")
reg = aet_sum(w * w)
f = pfunc([], reg)
assert f() == np.sum(w_init * w_init)
# Change the value of w and ensure the output changes accordingly.
w.set_value(w.get_value(borrow=True) + 1.0, borrow=True)
assert f() == np.sum((w_init + 1) ** 2)
def test_ctors(self):
if aesara.configdefaults.python_int_bitwidth() == 32:
assert shared(7).type == aesara.tensor.iscalar, shared(7).type
else:
assert shared(7).type == aesara.tensor.lscalar, shared(7).type
assert shared(7.0).type == aesara.tensor.dscalar
assert shared(np.float32(7)).type == aesara.tensor.fscalar
# test tensor constructor
b = shared(np.zeros((5, 5), dtype="int32"))
assert b.type == TensorType("int32", broadcastable=[False, False])
b = shared(np.random.rand(4, 5))
assert b.type == TensorType("float64", broadcastable=[False, False])
b = shared(np.random.rand(5, 1, 2))
assert b.type == TensorType("float64", broadcastable=[False, False, False])
assert shared([]).type == generic
def badfunc():
shared(7, bad_kw=False)
with pytest.raises(TypeError):
badfunc()
def test_update_same(self):
# There was a bug in CVM, triggered when a shared variable
# was its own update expression.
a = shared(1.0, "a")
b = shared(np.ones((2, 3)), "b")
# The order of the variables is not determined, so we try
# both shared variables.
# TODO: explain the above comment. By "not determined" does
# this mean "not deterministic"?
# This test originally wrote the updates using dictionaries,
# and iterating over the dictionary was not deterministic.
# Is that all the comment above meant, or is the CVM intended
# to add extra non-determinism? Or is the CVM meant to
# deterministically but arbitrarily pick an order for the updates?
f = function([], [], updates=[(a, a), (b, (2 * b))])
g = function([], [], updates=[(a, (a * 2)), (b, b)])
f()
assert a.get_value(borrow=True).shape == (), a.get_value()
assert b.get_value(borrow=True).shape == (2, 3), b.get_value()
g()
assert a.get_value(borrow=True).shape == (), a.get_value()
assert b.get_value(borrow=True).shape == (2, 3), b.get_value()
def test_undefined_grad_opt():
# Make sure that undefined grad get removed in optimized graph.
random = MRG_RandomStream(np.random.randint(1, 2147462579))
pvals = shared(np.random.rand(10, 20).astype(config.floatX))
pvals = pvals / pvals.sum(axis=1)
pvals = zero_grad(pvals)
samples = random.multinomial(pvals=pvals, n=1)
samples = cast(samples, pvals.dtype)
samples = zero_grad(samples)
cost = tt_sum(samples + pvals)
grad_out = grad(cost, samples)
f = function([], grad_out)
assert not any(
[isinstance(node.op, UndefinedGrad) for node in f.maker.fgraph.apply_nodes]
)
def test_default_updates(self):
x = shared(0)
x.default_update = x + 1
f = pfunc([], [x])
f()
assert x.get_value() == 1
del x.default_update
f()
assert x.get_value() == 2
g = pfunc([], [x])
g()
assert x.get_value() == 2
def test_givens(self):
x = shared(0)
assign = pfunc([], x, givens={x: 3})
assert assign() == 3
assert x.get_value(borrow=True) == 0
y = ivector()
f = pfunc([y], (y * x), givens={x: 6})
assert np.all(f([1, 1, 1]) == [6, 6, 6])
assert x.get_value() == 0
z = ivector()
c = z * y
f = pfunc([y], (c + 7), givens={z: _asarray([4, 4, 4], dtype="int32")})
assert np.all(f([1, 1, 1]) == [11, 11, 11])
assert x.get_value() == 0
