def test_validate(self):
try:
MyOp(Generic()(), MyType(1)()) # MyOp requires MyType instances
raise Exception("Expected an exception")
except Exception as e:
if str(e) != "Error 1":
raise
def test_invalid_modes(self):
# Modes 'r+', 'r', and 'w+' cannot work with Aesara, becausei
# the output array may be modified inplace, and that should not
# modify the original file.
path = Variable(Generic())
for mmap_mode in ("r+", "r", "w+", "toto"):
with pytest.raises(ValueError):
load(path, "int32", (False,), mmap_mode)
def test_basic(self):
path = Variable(Generic())
# Not specifying mmap_mode defaults to None, and the data is
# copied into main memory
x = load(path, "int32", (False,))
y = x * 2
fn = function([path], y)
assert (fn(self.filename) == (self.data * 2)).all()
def make_node(self):
return Apply(
self,
[],
[
Variable(Generic()),
tensor(self.dtype, shape=self.broadcastable),
],
)
def test1(self):
path = Variable(Generic())
# 'c' means "copy-on-write", which allow the array to be overwritten
# by an inplace Op in the graph, without modifying the underlying
# file.
x = load(path, "int32", (False,), "c")
# x ** 2 has been chosen because it will work inplace.
y = (x ** 2).sum()
fn = function([path], y)
# Call fn() twice, to check that inplace ops do not cause trouble
assert (fn(self.filename) == (self.data ** 2).sum()).all()
assert (fn(self.filename) == (self.data ** 2).sum()).all()
def test_hash_and_eq_params_type(self):
w1 = ParamsType(
a1=TensorType("int64", (False, False)),
a2=TensorType("int64", (False, True, False, False, True)),
a3=Generic(),
)
w2 = ParamsType(
a1=TensorType("int64", (False, False)),
a2=TensorType("int64", (False, True, False, False, True)),
a3=Generic(),
)
assert w1 == w2
assert not (w1 != w2)
assert hash(w1) == hash(w2)
assert w1.name == w2.name
# Changing attributes names only.
w2 = ParamsType(
a1=TensorType("int64", (False, False)),
other_name=TensorType(
"int64",
(False, True, False, False, True)), # a2 -> other_name
a3=Generic(),
)
assert w1 != w2
# Changing attributes types only.
w2 = ParamsType(
a1=TensorType("int64", (False, False)),
a2=Generic(), # changing class
a3=Generic(),
)
assert w1 != w2
# Changing attributes types characteristics only.
w2 = ParamsType(
a1=TensorType("int64", (False, True)), # changing broadcasting
a2=TensorType("int64", (False, True, False, False, True)),
a3=Generic(),
)
assert w1 != w2
def make_node(self, path):
if isinstance(path, str):
path = Constant(Generic(), path)
return Apply(self, [path],
[tensor(self.dtype, shape=self.broadcastable)])
def make_node(self, request, data):
return Apply(self, [request, data], [Variable(Generic())])
def make_node(self, data):
return Apply(self, [data], [Variable(Generic()), data.type()])
def test_memmap(self):
path = Variable(Generic())
x = load(path, "int32", (False,), mmap_mode="c")
fn = function([path], x)
assert type(fn(self.filename)) == np.core.memmap
class SearchsortedOp(COp):
"""Wrapper of numpy.searchsorted.
For full documentation, see :func:`searchsorted`.
See Also
--------
searchsorted : numpy-like function to use the SearchsortedOp
"""
params_type = Generic()
__props__ = ("side", )
check_input = False
def __init__(self, side="left"):
if side == "left" or side == "right":
self.side = side
else:
raise ValueError(
f"'{side}' is an invalid value for keyword 'side'")
def get_params(self, node):
return self.side
def make_node(self, x, v, sorter=None):
x = aet.as_tensor(x, ndim=1)
v = aet.as_tensor(v)
out_type = v.type.clone(dtype="int64")
if sorter is None:
return Apply(self, [x, v], [out_type()])
else:
sorter = aet.as_tensor(sorter, ndim=1)
if PYTHON_INT_BITWIDTH == 32 and sorter.dtype == "int64":
raise TypeError(
"numpy.searchsorted with Python 32bit do not support a"
" sorter of int64.")
if sorter.type not in int_vector_types:
raise TypeError("sorter must be an integer vector",
sorter.type)
return Apply(self, [x, v, sorter], [out_type()])
def infer_shape(self, fgraph, node, shapes):
return [shapes[1]]
def perform(self, node, inputs, output_storage, params):
x = inputs[0]
v = inputs[1]
if len(node.inputs) == 3:
sorter = inputs[2]
else:
sorter = None
z = output_storage[0]
z[0] = np.searchsorted(x, v, side=params,
sorter=sorter).astype(node.outputs[0].dtype)
def c_support_code_struct(self, node, name):
return f"""
int right_{name};
"""
def c_init_code_struct(self, node, name, sub):
side = sub["params"]
fail = sub["fail"]
return ("""
PyObject* tmp_%(name)s = PyUnicode_FromString("right");
if (tmp_%(name)s == NULL)
%(fail)s;
right_%(name)s = PyUnicode_Compare(%(side)s, tmp_%(name)s);
Py_DECREF(tmp_%(name)s);
""" % locals())
def c_code(self, node, name, inames, onames, sub):
sorter = None
if len(node.inputs) == 3:
x, v, sorter = inames
else:
x, v = inames
if not sorter:
sorter = "NULL"
(z, ) = onames
fail = sub["fail"]
return ("""
Py_XDECREF(%(z)s);
%(z)s = (PyArrayObject*) PyArray_SearchSorted(%(x)s, (PyObject*) %(v)s,
right_%(name)s ? NPY_SEARCHLEFT : NPY_SEARCHRIGHT, (PyObject*) %(sorter)s);
if (!%(z)s)
%(fail)s;
if (PyArray_TYPE(%(z)s) != NPY_INT64){
PyObject * tmp = PyArray_Cast(%(z)s, NPY_INT64);
Py_XDECREF(%(z)s);
%(z)s = (PyArrayObject*) tmp;
}
""" % locals())
def c_code_cache_version(self):
return (2, )
def grad(self, inputs, output_gradients):
num_ins = len(inputs)
if num_ins == 3:
x, v, sorter = inputs
else:
x, v = inputs
x_grad = _float_zeros_like(x)
v_grad = _float_zeros_like(v)
if num_ins == 3:
return [x_grad, v_grad, disconnected_type()]
else:
return [x_grad, v_grad]
def test_params_type_filtering(self):
shape_tensor5 = (1, 2, 2, 3, 2)
size_tensor5 = (shape_tensor5[0] * shape_tensor5[1] *
shape_tensor5[2] * shape_tensor5[3] * shape_tensor5[4])
random_tensor = np.random.normal(
size=size_tensor5).reshape(shape_tensor5)
w = ParamsType(
a1=TensorType("int32", (False, False)),
a2=TensorType("float64", (False, False, False, False, False)),
a3=Generic(),
)
# With a value that does not match the params type.
o = Params(
w,
a1=np.asarray([[1, 2, 3, 4, 5, 6], [7, 8, 9, 10, 11,
12]]).astype("int64"),
a2=random_tensor.astype("float32"),
a3=2000,
)
# should fail (o.a1 is not int32, o.a2 is not float64)
with pytest.raises(TypeError):
w.filter(o, True)
# should fail (o.a1 is not int32, o.a2 is not float64, and downcast is disallowed)
with pytest.raises(TypeError):
w.filter(o, False, False)
# Should pass.
w.filter(o, strict=False, allow_downcast=True)
# With a value that matches the params type.
o1 = Params(
w,
a1=np.asarray([[1, 2, 3, 4, 5, 6], [7, 8, 9, 10, 11,
12]]).astype("int32"),
a2=random_tensor.astype("float64"),
a3=2000,
)
# All should pass.
w.filter(o1, strict=True)
w.filter(o1, strict=False, allow_downcast=False)
w.filter(o1, strict=False, allow_downcast=True)
# Check values_eq and values_eq_approx.
o2 = Params(
w,
a1=np.asarray([[1, 2, 3, 4, 5, 6], [7, 8, 9, 10, 11,
12]]).astype("int32"),
a2=random_tensor.astype("float64"),
a3=2000,
)
assert w.values_eq(o1, o2)
assert w.values_eq_approx(o1, o2)
# Check value_eq_approx.
# NB: I don't know exactly which kind of differences is rejected by values_eq but accepted by values_eq_approx.
# So, I just play a little with float values.
o3 = Params(
w,
a1=np.asarray([[1, 2, 3, 4, 5, 6], [7, 8, 9, 10, 11,
12]]).astype("int32"),
a2=(random_tensor.astype("float32") * 10 / 2.2 * 2.19999999999 /
10).astype("float64"),
a3=2000.0 - 0.00000000000000001,
)
assert w.values_eq_approx(o1, o3)
def test_hash_and_eq_params(self):
wp1 = ParamsType(
a=Generic(),
array=TensorType("int64", (False, )),
floatting=Scalar("float64"),
npy_scalar=TensorType("float64", tuple()),
)
wp2 = ParamsType(
a=Generic(),
array=TensorType("int64", (False, )),
floatting=Scalar("float64"),
npy_scalar=TensorType("float64", tuple()),
)
w1 = Params(
wp1,
a=1,
array=np.asarray([1, 2, 4, 5, 7]),
floatting=-4.5,
npy_scalar=np.asarray(12),
)
w2 = Params(
wp2,
a=1,
array=np.asarray([1, 2, 4, 5, 7]),
floatting=-4.5,
npy_scalar=np.asarray(12),
)
assert w1 == w2
assert not (w1 != w2)
assert hash(w1) == hash(w2)
# Changing attributes names only (a -> other_name).
wp2_other = ParamsType(
other_name=Generic(),
array=TensorType("int64", (False, )),
floatting=Scalar("float64"),
npy_scalar=TensorType("float64", tuple()),
)
w2 = Params(
wp2_other,
other_name=1,
array=np.asarray([1, 2, 4, 5, 7]),
floatting=-4.5,
npy_scalar=np.asarray(12),
)
assert w1 != w2
# Changing attributes values only (now a=2).
w2 = Params(
wp2,
a=2,
array=np.asarray([1, 2, 4, 5, 7]),
floatting=-4.5,
npy_scalar=np.asarray(12),
)
assert w1 != w2
# Changing NumPy array values (5 -> -5).
w2 = Params(
wp2,
a=1,
array=np.asarray([1, 2, 4, -5, 7]),
floatting=-4.5,
npy_scalar=np.asarray(12),
)
assert w1 != w2
import numpy as np
import pytest
import aesara
from aesara import tensor as at
from aesara.graph.basic import Apply
from aesara.graph.op import COp, ExternalCOp
from aesara.graph.params_type import Params, ParamsType
from aesara.graph.type import EnumList, Generic
from aesara.scalar import Scalar
from aesara.tensor.type import TensorType, matrix
from tests import unittest_tools as utt
tensor_type_0d = TensorType("float64", tuple())
scalar_type = Scalar("float64")
generic_type = Generic()
# A test op to compute `y = a*x^2 + bx + c` for any tensor x, with a, b, c as op params.
class QuadraticOpFunc(COp):
__props__ = ("a", "b", "c")
params_type = ParamsType(a=tensor_type_0d, b=scalar_type, c=generic_type)
def __init__(self, a, b, c):
self.a = a
self.b = b
self.c = c
def make_node(self, x):
x = at.as_tensor_variable(x)
return Apply(self, [x], [x.type()])
class GpuMaxAndArgmax(COp):
"""
GPU version of MaxAndArgmax
"""
params_type = Generic()
__props__ = ("axis", )
argmax_dtype = "int64"
def __init__(self, axis):
assert isinstance(axis, (list, tuple))
self.axis = tuple(axis)
def get_params(self, node):
return self.axis
def make_node(self, X):
context_name = infer_context_name(X)
# We keep the original broadcastable flags for dimensions on which
# we do not perform the max / argmax.
all_axes = set(self.axis)
broadcastable = [
b for i, b in enumerate(X.type.broadcastable) if i not in all_axes
]
inputs = [as_gpuarray_variable(X, context_name)]
outputs = [
GpuArrayType(X.type.dtype,
broadcastable,
context_name=context_name)(),
GpuArrayType(self.argmax_dtype,
broadcastable,
context_name=context_name)(),
]
return Apply(self, inputs, outputs)
def c_headers(self, **kwargs):
return ["<numpy_compat.h>", "<gpuarray_helper.h>"]
def c_header_dirs(self, **kwargs):
return [pygpu.get_include(), gpuarray_helper_inc_dir()]
def c_code(self, node, name, input_names, output_names, sub):
# Recall: X = input_names[0]
# Recall: axes = sub['params']
# Recall: max, argmax = output_names
# Recall: fail = sub['fail']
max_typecode = pygpu.gpuarray.dtype_to_typecode(node.inputs[0].dtype)
argmax_typecode = pygpu.gpuarray.dtype_to_typecode(self.argmax_dtype)
ret = """
#if PY_MAJOR_VERSION >= 3
#ifndef PyInt_AS_LONG
#define PyInt_AS_LONG PyLong_AS_LONG
#endif
#endif
int err = 0;
unsigned %(name)s_redux_len = PyTuple_GET_SIZE(%(axes)s);
unsigned* %(name)s_axes_to_reduce = (unsigned*)malloc(%(name)s_redux_len * sizeof(unsigned));
for (unsigned i = 0; i < %(name)s_redux_len; ++i) {
PyObject* axis_object = PyTuple_GET_ITEM(%(axes)s, i);
%(name)s_axes_to_reduce[i] = (unsigned) PyInt_AS_LONG(axis_object);
}
size_t %(name)s_input_ndim = PyGpuArray_NDIM(%(X)s);
size_t %(name)s_output_ndim = %(name)s_input_ndim - %(name)s_redux_len;
size_t* %(name)s_output_dims = (size_t*)malloc(%(name)s_output_ndim * sizeof(size_t));
if (%(name)s_redux_len == 1) {
for (unsigned i = 0; i < %(name)s_axes_to_reduce[0]; ++i) {
%(name)s_output_dims[i] = PyGpuArray_DIM(%(X)s, i);
}
for (unsigned i = %(name)s_axes_to_reduce[0] + 1; i < %(name)s_input_ndim; ++i) {
%(name)s_output_dims[i-1] = PyGpuArray_DIM(%(X)s, i);
}
} else {
int64_t current_input_pos = -1;
int64_t current_output_pos = -1;
for (unsigned i = 0; i < %(name)s_redux_len; ++i) {
for (++current_input_pos; current_input_pos < %(name)s_axes_to_reduce[i]; ++current_input_pos) {
%(name)s_output_dims[++current_output_pos] = PyGpuArray_DIM(%(X)s, current_input_pos);
}
}
for (++current_input_pos; current_input_pos < %(name)s_input_ndim; ++current_input_pos) {
%(name)s_output_dims[++current_output_pos] = PyGpuArray_DIM(%(X)s, current_input_pos);
}
}
if (aesara_prep_output(&%(max)s, %(name)s_output_ndim, %(name)s_output_dims, %(max_typecode)s, GA_C_ORDER, %(X)s->context)) {
PyErr_SetString(PyExc_RuntimeError, "GpuMaxAndArgmax: unable to prepare max output.");
%(fail)s
}
if (aesara_prep_output(&%(argmax)s, %(name)s_output_ndim, %(name)s_output_dims, %(argmax_typecode)s, GA_C_ORDER, %(X)s->context)) {
PyErr_SetString(PyExc_RuntimeError, "GpuMaxAndArgmax: unable to prepare argmax output.");
%(fail)s
}
if (%(name)s_input_ndim == 0) {
/* GpuArray_maxandargmax can't handle a 0-d array
* because it expects that 1 <= redux_len <= input_ndim.
* As input_ndim == 0, then 1 <= redux_len <= 0 is false.
* To handle this case we copy input to max and we set argmax to 0.
*/
if (GA_NO_ERROR != GpuArray_setarray(&%(max)s->ga, &%(X)s->ga)) {
PyErr_SetString(PyExc_RuntimeError, "GpuMaxAndArgmax: unable to copy input to max when input is a scalar.");
%(fail)s
}
if (GA_NO_ERROR != GpuArray_memset(&%(argmax)s->ga, 0)) {
PyErr_SetString(PyExc_RuntimeError, "GpuMaxAndArgmax: unable to set argmax to 0 when input is a scalar.");
%(fail)s
}
} else if (GA_NO_ERROR != (err =
GpuArray_maxandargmax(&%(max)s->ga, &%(argmax)s->ga, &%(X)s->ga, %(name)s_redux_len, %(name)s_axes_to_reduce)
)) {
PyErr_Format(PyExc_RuntimeError,
"GpuMaxAndArgmax: unable to compute gpuarray maxandargmax: error %%d: %%s (%%s).",
err, gpuarray_error_str(err), GpuArray_error(&%(X)s->ga, err));
%(fail)s
}
"""
return ret % {
"X": input_names[0],
"axes": sub["params"],
"max": output_names[0],
"argmax": output_names[1],
"max_typecode": max_typecode,
"argmax_typecode": argmax_typecode,
"name": name,
"fail": sub["fail"],
}
def c_code_cleanup(self, node, name, inputs, outputs, sub):
return """
free(%(name)s_output_dims);
free(%(name)s_axes_to_reduce);
""" % {
"name": name,
}
def c_code_cache_version(self):
return (2, )
