def _infer_value_domain(cls, **kwargs):
# rely on the underlying distribution's logic to infer the event_shape given param domains
instance = cls.dist_class(**{k: _dummy_tensor(domain) for k, domain in kwargs.items()}, validate_args=False)
out_shape = instance.event_shape
if isinstance(instance.support, constraints._IntegerInterval):
out_dtype = int(instance.support.upper_bound + 1)
else:
out_dtype = 'real'
return Domain(dtype=out_dtype, shape=out_shape)
def test_unary(symbol, data):
dtype = 'real'
if symbol == '~':
data = bool(data)
dtype = 2
expected_data = unary_eval(symbol, data)
x = Number(data, dtype)
actual = unary_eval(symbol, x)
check_funsor(actual, {}, Domain((), dtype), expected_data)
def __init__(self, data, inputs=None, dtype="real"):
assert isinstance(data, np.ndarray) or np.isscalar(data)
assert isinstance(inputs, tuple)
assert all(isinstance(d.dtype, integer_types) for k, d in inputs)
inputs = OrderedDict(inputs)
output = Domain(data.shape[len(inputs):], dtype)
fresh = frozenset(inputs.keys())
bound = frozenset()
super(Array, self).__init__(inputs, output, fresh, bound)
self.data = data
def __init__(self, data, inputs=None, dtype="real"):
assert isinstance(data, torch.Tensor)
assert isinstance(inputs, tuple)
if not torch._C._get_tracing_state():
assert len(inputs) <= data.dim()
for (k, d), size in zip(inputs, data.shape):
assert d.dtype == size
inputs = OrderedDict(inputs)
output = Domain(data.shape[len(inputs):], dtype)
fresh = frozenset(inputs.keys())
bound = frozenset()
super(Tensor, self).__init__(inputs, output, fresh, bound)
self.data = data
def __init__(self, var, expr):
assert isinstance(var, Variable)
assert isinstance(var.dtype, int)
assert isinstance(expr, Funsor)
inputs = expr.inputs.copy()
inputs.pop(var.name, None)
shape = (var.dtype,) + expr.output.shape
output = Domain(shape, expr.dtype)
fresh = frozenset()
bound = frozenset({var.name})
super(Lambda, self).__init__(inputs, output, fresh, bound)
self.var = var
self.expr = expr
def __init__(self, data, dtype=None):
assert isinstance(data, numbers.Number)
if isinstance(dtype, int):
data = type(dtype)(data)
if dtype != 2: # booleans have bitwise interpretation
assert 0 <= data and data < dtype
else:
assert isinstance(dtype, str) and dtype == "real"
data = float(data)
inputs = OrderedDict()
output = Domain((), dtype)
super(Number, self).__init__(inputs, output)
self.data = data
def _infer_value_domain(cls, **kwargs):
# rely on the underlying distribution's logic to infer the event_shape given param domains
instance = cls.dist_class(**{
k: dummy_numeric_array(domain)
for k, domain in kwargs.items()
},
validate_args=False)
out_shape = instance.event_shape
if type(instance.support).__name__ == "_IntegerInterval":
out_dtype = int(instance.support.upper_bound + 1)
else:
out_dtype = 'real'
return Domain(dtype=out_dtype, shape=out_shape)
def test_binary(symbol, data1, data2):
dtype = 'real'
if symbol in BOOLEAN_OPS:
dtype = 2
data1 = bool(data1)
data2 = bool(data2)
try:
expected_data = binary_eval(symbol, data1, data2)
except ZeroDivisionError:
return
x1 = Number(data1, dtype)
x2 = Number(data2, dtype)
actual = binary_eval(symbol, x1, x2)
check_funsor(actual, {}, Domain((), dtype), expected_data)
def test_reduce_event(op, event_shape, dims):
sizes = {'a': 3, 'b': 4, 'c': 5}
batch_shape = tuple(sizes[d] for d in dims)
shape = batch_shape + event_shape
inputs = OrderedDict((d, bint(sizes[d])) for d in dims)
numeric_op = REDUCE_OP_TO_NUMERIC[op]
data = rand(shape) + 0.5
dtype = 'real'
if op in [ops.and_, ops.or_]:
data = ops.astype(data, 'uint8')
expected_data = numeric_op(data.reshape(batch_shape + (-1, )), -1)
x = Tensor(data, inputs, dtype=dtype)
op_name = numeric_op.__name__[1:] if op in [ops.min, ops.max
] else numeric_op.__name__
actual = getattr(x, op_name)()
check_funsor(actual, inputs, Domain((), dtype), expected_data)
def test_reduce_event(op, event_shape, dims):
sizes = {'a': 3, 'b': 4, 'c': 5}
batch_shape = tuple(sizes[d] for d in dims)
shape = batch_shape + event_shape
inputs = OrderedDict((d, bint(sizes[d])) for d in dims)
torch_op = REDUCE_OP_TO_TORCH[op]
data = torch.rand(shape) + 0.5
dtype = 'real'
if op in [ops.and_, ops.or_]:
data = data.byte()
expected_data = torch_op(data.reshape(batch_shape + (-1, )), -1)
if op in [ops.min, ops.max]:
expected_data = expected_data[0]
x = Tensor(data, inputs, dtype=dtype)
actual = getattr(x, torch_op.__name__)()
check_funsor(actual, inputs, Domain((), dtype), expected_data)
def test_binary(symbol, data1, data2):
dtype = 'real'
if symbol in BOOLEAN_OPS:
dtype = 2
data1 = bool(data1)
data2 = bool(data2)
try:
expected_data = binary_eval(symbol, data1, data2)
except ZeroDivisionError:
return
x1 = Number(data1, dtype)
x2 = Number(data2, dtype)
actual = binary_eval(symbol, x1, x2)
check_funsor(actual, {}, Domain((), dtype), expected_data)
with interpretation(normalize):
actual_reflect = binary_eval(symbol, x1, x2)
assert actual.output == actual_reflect.output
def torch_stack(parts, dim=0):
"""
Wrapper around :func:`torch.stack` to operate on real-valued Funsors.
Note this operates only on the ``output`` tensor. To stack funsors in a
new named dim, instead use :class:`~funsor.terms.Stack`.
"""
assert isinstance(dim, int)
assert isinstance(parts, tuple)
assert len(set(x.output for x in parts)) == 1
shape = parts[0].output.shape
if dim >= 0:
dim = dim - len(shape) - 1
assert dim < 0
split = dim + len(shape) + 1
shape = shape[:split] + (len(parts), ) + shape[split:]
output = Domain(shape, parts[0].dtype)
fn = functools.partial(_torch_stack, dim)
return Function(fn, output, parts)
def test_reduce_all(op):
x = Variable('x', bint(2))
y = Variable('y', bint(3))
z = Variable('z', bint(4))
f = x * y + z
dtype = f.dtype
check_funsor(f, {'x': bint(2), 'y': bint(3), 'z': bint(4)}, Domain((), dtype))
if op is ops.logaddexp:
pytest.skip()
with interpretation(sequential):
actual = f.reduce(op)
values = [f(x=i, y=j, z=k)
for i in x.output
for j in y.output
for k in z.output]
expected = reduce(op, values)
assert actual == expected
def test_binary_funsor_funsor(symbol, dims1, dims2):
sizes = {'a': 3, 'b': 4, 'c': 5}
shape1 = tuple(sizes[d] for d in dims1)
shape2 = tuple(sizes[d] for d in dims2)
inputs1 = OrderedDict((d, bint(sizes[d])) for d in dims1)
inputs2 = OrderedDict((d, bint(sizes[d])) for d in dims2)
data1 = rand(shape1) + 0.5
data2 = rand(shape2) + 0.5
dtype = 'real'
if symbol in BOOLEAN_OPS:
dtype = 2
data1 = ops.astype(data1, 'uint8')
data2 = ops.astype(data2, 'uint8')
x1 = Tensor(data1, inputs1, dtype)
x2 = Tensor(data2, inputs2, dtype)
inputs, aligned = align_tensors(x1, x2)
expected_data = binary_eval(symbol, aligned[0], aligned[1])
actual = binary_eval(symbol, x1, x2)
check_funsor(actual, inputs, Domain((), dtype), expected_data)
def test_reduce_subset(op, reduced_vars):
reduced_vars = frozenset(reduced_vars)
x = Variable('x', bint(2))
y = Variable('y', bint(3))
z = Variable('z', bint(4))
f = x * y + z
dtype = f.dtype
check_funsor(f, {'x': bint(2), 'y': bint(3), 'z': bint(4)}, Domain((), dtype))
if op is ops.logaddexp:
pytest.skip()
with interpretation(sequential):
actual = f.reduce(op, reduced_vars)
expected = f
for v in [x, y, z]:
if v.name in reduced_vars:
expected = reduce(op, [expected(**{v.name: i}) for i in v.output])
check_funsor(actual, expected.inputs, expected.output)
# TODO check data
if not reduced_vars:
assert actual is f
def tensor_to_funsor(tensor, event_inputs=(), event_output=0, dtype="real"):
"""
Convert a :class:`torch.Tensor` to a :class:`funsor.tensor.Tensor` .
Note this should not touch data, but may trigger a
:meth:`torch.Tensor.reshape` op.
:param torch.Tensor tensor: A PyTorch tensor.
:param tuple event_inputs: A tuple of names for rightmost tensor
dimensions. If ``tensor`` has these names, they will be converted to
``result.inputs``.
:param int event_output: The number of tensor dimensions assigned to
``result.output``. These must be on the right of any ``event_input``
dimensions.
:return: A funsor.
:rtype: funsor.tensor.Tensor
"""
assert isinstance(tensor, torch.Tensor)
assert isinstance(event_inputs, tuple)
assert isinstance(event_output, int) and event_output >= 0
inputs_shape = tensor.shape[:tensor.dim() - event_output]
output = Domain(dtype=dtype, shape=tensor.shape[tensor.dim() - event_output:])
dim_to_name = default_dim_to_name(inputs_shape, event_inputs)
return to_funsor(tensor, output, dim_to_name)