import builtins
import warnings
from collections.abc import Sequence
from typing import TYPE_CHECKING, Optional
import numpy as np
from numpy.core.numeric import normalize_axis_tuple
from pytensor import config, printing
from pytensor import scalar as ps
from pytensor.gradient import DisconnectedType
from pytensor.graph.basic import Apply, Variable
from pytensor.graph.op import Op
from pytensor.graph.replace import _vectorize_node
from pytensor.link.c.op import COp
from pytensor.link.c.params_type import ParamsType
from pytensor.link.c.type import Generic
from pytensor.misc.safe_asarray import _asarray
from pytensor.printing import pprint
from pytensor.raise_op import Assert
from pytensor.scalar.basic import BinaryScalarOp
from pytensor.tensor import TensorLike
from pytensor.tensor.basic import (
alloc,
arange,
as_tensor_variable,
cast,
concatenate,
constant,
stack,
switch,
)
from pytensor.tensor.blockwise import Blockwise, vectorize_node_fallback
from pytensor.tensor.elemwise import (
CAReduce,
DimShuffle,
Elemwise,
get_normalized_batch_axes,
scalar_elemwise,
)
from pytensor.tensor.shape import shape, specify_broadcastable
from pytensor.tensor.type import (
DenseTensorType,
complex_dtypes,
continuous_dtypes,
discrete_dtypes,
int_dtypes,
integer_dtypes,
tensor,
uint_dtypes,
)
from pytensor.tensor.type_other import NoneConst
from pytensor.tensor.utils import as_list
from pytensor.tensor.variable import (
TensorConstant,
TensorVariable,
_tensor_py_operators,
)
if TYPE_CHECKING:
from numpy.typing import ArrayLike, DTypeLike
# We capture the builtins that we are going to replace to follow the numpy API
_abs = builtins.abs
if int(config.tensor__cmp_sloppy) > 1:
# This config variable is a quick-and-dirty way to get low-precision
# comparisons. For a more precise setting of these tolerances set
# them explicitly in your user code by assigning, for example,
# "pytensor.tensor.math.float32_atol = ..."
# When config.tensor__cmp_sloppy>1 we are even more sloppy. This is
# useful to test the GPU as they don't use extended precision and
# this cause some difference bigger then the normal sloppy.
float16_atol = 1e-2
float16_rtol = 5e-2
float32_atol = 5e-4
float32_rtol = 1e-3
float64_rtol = 1e-4
float64_atol = 1e-3
elif int(config.tensor__cmp_sloppy):
float16_atol = 5e-3
float16_rtol = 1e-2
float32_atol = 1e-4
float32_rtol = 1e-3
float64_rtol = 1e-4
float64_atol = 1e-3
else:
# If you change those value in test don't forget to put them back
# when the test end. Don't forget the case when the test fail.
float16_atol = 1e-3
float16_rtol = 1e-3
float32_atol = 1e-5
float32_rtol = 1e-5
# defaults in numpy.allclose
# Don't be more strict then numpy rtol
# It cause useless error.
float64_rtol = 1.0000000000000001e-05
float64_atol = 1e-8
def _get_atol_rtol(a, b):
tiny = ("float16",)
narrow = ("float32", "complex64")
if (str(a.dtype) in tiny) or (str(b.dtype) in tiny):
atol = float16_atol
rtol = float16_rtol
elif (str(a.dtype) in narrow) or (str(b.dtype) in narrow):
atol = float32_atol
rtol = float32_rtol
else:
atol = float64_atol
rtol = float64_rtol
return atol, rtol
def _allclose(a, b, rtol=None, atol=None):
a = np.asarray(a)
b = np.asarray(b)
atol_, rtol_ = _get_atol_rtol(a, b)
if rtol is not None:
rtol_ = rtol
if atol is not None:
atol_ = atol
return np.allclose(a, b, atol=atol_, rtol=rtol_)
class MaxAndArgmax(COp):
"""
Calculate the max and argmax over a given axis or over all axes.
"""
nin = 2 # tensor, axis
nout = 2 # max val, max idx
E_axis = "invalid axis"
params_type = Generic()
__props__ = ("axis",)
_f16_ok = True
def __init__(self, axis):
assert isinstance(axis, tuple | list)
self.axis = tuple(axis)
def get_params(self, node):
return self.axis
def make_node(self, x):
x = as_tensor_variable(x)
# Keep the original shapes for axes on which we do not perform the max/argmax.
all_axes = set(self.axis)
inputs = [x]
out_shape = tuple(s for i, s in enumerate(x.type.shape) if i not in all_axes)
outputs = [
tensor(dtype=x.type.dtype, shape=out_shape, name="max"),
tensor(dtype="int64", shape=out_shape, name="argmax"),
]
return Apply(self, inputs, outputs)
def perform(self, node, inp, outs):
x = inp[0]
axes = self.axis
max, max_idx = outs
if axes is None:
axes = tuple(range(x.ndim))
else:
axes = tuple(int(ax) for ax in axes)
max[0] = _asarray(np.max(x, axes), dtype=node.outputs[0].dtype)
# Numpy does not support multiple axes for argmax
# Work around
keep_axes = np.array([i for i in range(x.ndim) if i not in axes], dtype="int64")
# Not-reduced axes in front
transposed_x = np.transpose(x, np.concatenate((keep_axes, axes)))
kept_shape = transposed_x.shape[: len(keep_axes)]
reduced_shape = transposed_x.shape[len(keep_axes) :]
# Numpy.prod returns 1.0 when arg is empty, so we cast it to int64
# Otherwise reshape would complain citing float arg
new_shape = (*kept_shape, np.prod(reduced_shape, dtype="int64"))
reshaped_x = transposed_x.reshape(new_shape)
max_idx[0] = _asarray(np.argmax(reshaped_x, axis=-1), dtype="int64")
def c_code(self, node, name, inp, out, sub):
if len(self.axis) != 1 and len(self.axis) != node.inputs[0].ndim:
raise NotImplementedError(
"NumPy C-API can compute max and argmax only for 1 axis or for all axes."
)
x = inp[0]
axis = sub["params"]
max, argmax = out
fail = sub["fail"]
ret = """
#if PY_MAJOR_VERSION >= 3
#ifndef PyInt_AS_LONG
#define PyInt_AS_LONG PyLong_AS_LONG
#endif
#endif
int axis;
if (PyTuple_GET_SIZE(%(axis)s) == PyArray_NDIM(%(x)s)) {
axis = NPY_MAXDIMS;
} else if(PyTuple_GET_SIZE(%(axis)s) == 1) {
PyObject* axis_object = PyTuple_GET_ITEM(%(axis)s, 0);
axis = (int)PyInt_AS_LONG(axis_object);
if (axis > PyArray_NDIM(%(x)s)-1 || axis < -PyArray_NDIM(%(x)s)) {
PyErr_SetString(PyExc_ValueError,
"MaxAndArgmax: bad axis argument");
%(fail)s
}
} else {
PyErr_SetString(PyExc_NotImplementedError,
"MaxAndArgmax: NumPy C-API can compute max and argmax only for 1 axis or for all axes.");
%(fail)s
}
Py_CLEAR(%(max)s);
Py_CLEAR(%(argmax)s);//todo pass them as out parameter.
%(max)s = (PyArrayObject*)PyArray_Max(%(x)s, axis, NULL);
if (%(max)s == NULL) {
%(fail)s;
}
if (!PyArray_CheckExact(%(max)s)) {
%(max)s = (PyArrayObject*)PyArray_FromAny((PyObject*)%(max)s, NULL, 0, 0, NPY_ARRAY_ENSUREARRAY, NULL);
if(%(max)s == NULL){
%(fail)s;
}
}
%(argmax)s = (PyArrayObject*)PyArray_ArgMax(%(x)s, axis, NULL);
if (%(argmax)s == NULL) {
Py_CLEAR(%(max)s);
%(fail)s;
}
if (!PyArray_CheckExact(%(argmax)s)) {
%(argmax)s = (PyArrayObject*)PyArray_FromAny((PyObject*)%(argmax)s, NULL, 0, 0, NPY_ARRAY_ENSUREARRAY, NULL);
if(%(argmax)s == NULL){
%(fail)s;
}
}
if (PyArray_TYPE(%(argmax)s) != NPY_INT64) {
PyObject * tmp = PyArray_Cast(%(argmax)s, NPY_INT64);
if (NULL == tmp){
%(fail)s;
}
Py_DECREF(%(argmax)s);
%(argmax)s = (PyArrayObject*)tmp;
}
"""
return ret % locals()
def c_code_cache_version(self):
return (5,)
def infer_shape(self, fgraph, node, shapes):
ishape = shapes[0]
rval = tuple(
ishape[i]
for (i, b) in enumerate(node.inputs[0].type.broadcastable)
if i not in self.axis
)
return [rval, rval]
def R_op(self, inputs, eval_points):
if eval_points[0] is None:
return [None, None]
if len(self.axis) != 1:
raise ValueError("R_op supported for arg_max only for one axis!")
if self.axis[0] > 1:
raise ValueError("R_op supported for arg_max only when axis is 0 or 1")
if inputs[0].ndim != 2:
raise ValueError("R_op supported for arg_max only when input is a matrix")
max_vals, max_pos = self.make_node(*inputs).outputs
if self.axis[0] == 0:
return [eval_points[0][max_pos, arange(eval_points[0].shape[1])], None]
else:
return [eval_points[0][arange(eval_points[0].shape[0]), max_pos], None]
def grad(self, inp, grads):
# The strict sense mathematical gradient of the maximum function is
# not calculated here for it is not defined at every point where some
# coordinates are identical. However, since the latter set has null
# Lebesgue measure, the result may be interpreted as weak gradient.
# @note: This function should work correctly for L{vector}s.
# (x, y), (gz, gw)
# gz*dz/dx + gw*dw/dx, gz*dz/dy + gw*dw/dy
# gMax * dMax/dx + gArgMax * dArgMax/dx,
# gMax * dMax/daxis + gArgMax * dArgMax/daxis
# g_max has one less dimension than x, so you need to complete
# g_max to x's shape when axis=0 the broadcasting mechanism
# does it automatically
x = inp[0]
axis = as_tensor_variable(self.axis)
g_max, g_max_idx = grads
g_max_disconnected = isinstance(g_max.type, DisconnectedType)
g_max_idx_disconnected = isinstance(g_max_idx.type, DisconnectedType)
# if the op is totally disconnected, so are its inputs
if g_max_disconnected and g_max_idx_disconnected:
return [DisconnectedType()(), DisconnectedType()()]
# if the max is disconnected but the argmax is not,
# the gradient on its inputs is zero
if g_max_disconnected:
return [x.zeros_like()]
if NoneConst.equals(axis):
axis_ = list(range(x.ndim))
else:
axis_ = axis
xmax = max(x, axis_)
# Raise the g_max and xmax to the same number of dim as the input.
pattern = []
out_dim = 0
if NoneConst.equals(axis):
# We are taking the max/argmax over all dimensions.
axis = None
for i in range(x.ndim):
if axis is None or i in axis.data:
pattern.append("x")
else:
pattern.append(out_dim)
out_dim += 1
g_max_pad = DimShuffle(g_max.broadcastable, pattern)(g_max)
xmax_pad = DimShuffle(xmax.broadcastable, pattern)(xmax)
# Set the grad to the correct position.
g_x = eq(xmax_pad, x) * g_max_pad
return (g_x,)
class Argmax(COp):
"""
Calculate the argmax over a given axis or over all axes.
"""
nin = 2 # tensor, axis
nout = 1
E_axis = "invalid axis"
__props__ = ("axis",)
_f16_ok = True
params_type = ParamsType(c_axis=ps.int64)
def __init__(self, axis):
if axis is not None:
axis = tuple(axis)
self.axis = tuple(axis)
def get_params(self, node):
if self.axis is not None and len(self.axis) == 1:
c_axis = np.int64(self.axis[0])
else:
# The value here doesn't matter, it won't be used
c_axis = np.int64(-1)
return self.params_type.get_params(c_axis=c_axis)
def make_node(self, x, axis=None):
x = as_tensor_variable(x)
if self.axis is None:
all_axes = list(range(x.ndim))
else:
all_axes = self.axis
inputs = [x]
# We keep the original broadcastable flags for dimensions on which
# we do not perform the argmax.
out_shape = tuple(s for i, s in enumerate(x.type.shape) if i not in all_axes)
outputs = [tensor(dtype="int64", shape=out_shape, name="argmax")]
return Apply(self, inputs, outputs)
def prepare_node(self, node, storage_map, compute_map, impl):
if len(node.inputs) == 2:
raise ValueError(
"You are trying to compile a graph with an old Argmax node. Either reoptimize your graph or rebuild it to get the new node format."
)
def perform(self, node, inp, outs):
(x,) = inp
axes = self.axis
(max_idx,) = outs
if axes is None:
axes = tuple(range(x.ndim))
# Numpy does not support multiple axes for argmax
# Work around
keep_axes = np.array([i for i in range(x.ndim) if i not in axes], dtype="int64")
# Not-reduced axes in front
transposed_x = np.transpose(x, np.concatenate((keep_axes, axes)))
kept_shape = transposed_x.shape[: len(keep_axes)]
reduced_shape = transposed_x.shape[len(keep_axes) :]
new_shape = (*kept_shape, np.prod(reduced_shape))
reshaped_x = transposed_x.reshape(new_shape)
max_idx[0] = _asarray(np.argmax(reshaped_x, axis=-1), dtype="int64")
def c_code(self, node, name, inp, out, sub):
(x,) = inp
(argmax,) = out
fail = sub["fail"]
params = sub["params"]
if self.axis is None:
axis_code = "axis = NPY_MAXDIMS;"
else:
if len(self.axis) > 1:
raise NotImplementedError()
# params is only used here for now
axis_code = """
axis = {params}->c_axis;
if(axis > PyArray_NDIM({x})-1 || axis < -PyArray_NDIM({x})){{
PyErr_SetString(PyExc_ValueError,
"Argmax, bad axis argument");
{fail}
}}
""".format(**locals())
ret = """
int axis;
Py_CLEAR(%(argmax)s);//todo pass them as out parameter.
%(axis_code)s
%(argmax)s = (PyArrayObject*)PyArray_ArgMax(%(x)s, axis, NULL);
if(%(argmax)s == NULL){
%(fail)s;
}
if(!PyArray_CheckExact(%(argmax)s)){
%(argmax)s = (PyArrayObject*)PyArray_FromAny((PyObject*)%(argmax)s, NULL, 0, 0, NPY_ARRAY_ENSUREARRAY, NULL);
if(%(argmax)s == NULL){
%(fail)s;
}
}
if(PyArray_TYPE(%(argmax)s) != NPY_INT64){
PyObject * tmp = PyArray_Cast(%(argmax)s, NPY_INT64);
if (NULL == tmp){
%(fail)s;
}
Py_DECREF(%(argmax)s);
%(argmax)s = (PyArrayObject*)tmp;
}
"""
return ret % locals()
def c_code_cache_version(self):
return (1,)
def infer_shape(self, fgraph, node, shapes):
(ishape,) = shapes
if self.axis is None:
return [()]
rval = tuple(
[
ishape[i]
for (i, b) in enumerate(node.inputs[0].type.broadcastable)
if i not in self.axis
]
)
return [rval]
def grad(self, inp, grads):
(x,) = inp
return [x.zeros_like()]
@_vectorize_node.register(Argmax)
@_vectorize_node.register(MaxAndArgmax)
def vectorize_argmax_node(op, node, batch_x):
core_ndim = node.inputs[0].type.ndim
batch_ndim = batch_x.type.ndim - core_ndim
if not batch_ndim:
return node.op.make_node(batch_x)
batch_axes = get_normalized_batch_axes(op.axis, core_ndim, batch_ndim)
return type(op)(axis=batch_axes).make_node(batch_x)
def makeKeepDims(x, y, axis):
"""
Reintroduces in y with length one the axes of x which have been left out
in a prior reduction of x. With this option, the resulting tensor will
broadcast correctly against the original tensor x.
"""
x = as_tensor_variable(x)
y = as_tensor_variable(y)
if axis is None:
axis = list(range(x.type.ndim))
elif isinstance(axis, int | np.integer):
axis = [axis]
elif isinstance(axis, np.ndarray) and axis.ndim == 0:
axis = [int(axis)]
else:
axis = [int(a) for a in axis]
newaxis = []
for a in axis:
if not isinstance(a, int):
raise ValueError("keepdims option can be used only with constant axis")
if a < 0:
a += x.type.ndim
newaxis.append(a)
i = 0
new_dims = []
for j, _ in enumerate(x.type.broadcastable):
if j in newaxis:
new_dims.append("x")
else:
new_dims.append(i)
i += 1
return DimShuffle(y.type.broadcastable, new_dims)(y)
def check_and_normalize_axes(x, axis):
"""Check axes, normalize and convert them to a Python list of integers.
Parameters
----------
x: TensorVariable
axis: int, tuple or list of integers
Returns
-------
axis: list of integers
Return an empty list if argument is None.
"""
x = as_tensor_variable(x)
if axis is None:
axis = []
elif isinstance(axis, int | np.integer) or (
isinstance(axis, np.ndarray) and axis.ndim == 0
):
axis = [int(axis)]
elif isinstance(axis, tuple | list | np.ndarray):
axis = [int(i) for i in axis]
elif isinstance(axis, Variable):
if NoneConst.equals(axis):
axis = []
elif not isinstance(axis, TensorConstant):
raise TypeError(f"Computation needs a constant axis. Got {axis}")
else:
assert axis.dtype in integer_dtypes
if isinstance(axis.data, int | np.integer) or (
isinstance(axis.data, np.ndarray) and axis.data.ndim == 0
):
axis = [int(axis.data)]
elif isinstance(axis.data, list | np.ndarray):
axis = [int(i) for i in axis.data]
else:
raise TypeError(
f"Axis must be an integer, tuple, list of integers or a TensorVariable. Got {axis}"
)
if len(axis) > 0:
for i in range(len(axis)):
if axis[i] < 0:
axis[i] += x.type.ndim
if axis[i] < 0 or axis[i] >= x.type.ndim:
raise ValueError(
f"Computation needs a valid axis number for {int(x.type.ndim)}-D tensor. Got {int(axis[i])}"
)
axis = list(set(axis))
axis.sort()
return axis
def max_and_argmax(a, axis=None, keepdims=False):
"""
Returns maximum elements and their indices obtained by iterating over
given axis.
When axis is None (the default value), the max is performed
over the flattened tensor.
Parameters
----------
keepdims : bool
If this is set to True, the axes which are reduced are left in
the result as dimensions with size one. With this option, the result
will broadcast correctly against the original tensor.
"""
# Check axis and convert it to a Python list of integers.
# Axis will be used as an op param of MaxAndArgmax.
a = as_tensor_variable(a)
axis = check_and_normalize_axes(a, axis)
if len(axis) == 0:
axis = list(range(a.type.ndim))
out, argout = MaxAndArgmax(axis)(a)
if keepdims:
out = makeKeepDims(a, out, axis)
argout = makeKeepDims(a, argout, axis)
return [out, argout]
class FixedOpCAReduce(CAReduce):
def __str__(self):
return f"{type(self).__name__}{{{self._axis_str()}}}"
class NonZeroDimsCAReduce(FixedOpCAReduce):
def _c_all(self, node, name, inames, onames, sub):
decl, checks, alloc, loop, end = super()._c_all(node, name, inames, onames, sub)
# We add an additional check for zero-sized dimensions (This seems like
# something that could enabled in `elemwise_cgen.make_checks`.)
iname = inames[0]
axis = self.axis
if axis is None:
axis = list(range(len(node.inputs[0].type.broadcastable)))
pattern = [0] * len(node.inputs[0].broadcastable)
for i in axis:
pattern[i] = 1
pattern_ = str(pattern)[1:-1]
decl += f"""int tosum[]={{{pattern_}}};"""
alloc += f"""
for(int i=0;i<PyArray_NDIM({iname});i++){{
if(PyArray_DIMS({iname})[i]==0 && tosum[i]){{
PyErr_Format(PyExc_ValueError,
"Input of CAReduce{{{node.op.scalar_op}}} has zero-size on axis %%d",i);
{sub["fail"]};
}}
}}
"""
return decl, checks, alloc, loop, end
class Max(NonZeroDimsCAReduce):
nfunc_spec = ("max", 1, 1)
def __init__(self, axis):
super().__init__(ps.scalar_maximum, axis)
def clone(self, **kwargs):
axis = kwargs.get("axis", self.axis)
return type(self)(axis=axis)
class Min(NonZeroDimsCAReduce):
nfunc_spec = ("min", 1, 1)
def __init__(self, axis):
super().__init__(ps.scalar_minimum, axis)
def clone(self, **kwargs):
axis = kwargs.get("axis", self.axis)
return type(self)(axis=axis)
[docs]
def max(x, axis=None, keepdims=False):
"""
Returns maximum elements obtained by iterating over given axis.
When axis is None (the default value), the max is performed
over the flattened tensor.
Parameters
----------
keepdims: bool
If this is set to True, the axes which are reduced are left in
the result as dimensions with size one. With this option, the result
will broadcast correctly against the original tensor.
Notes
-----
We return an error as numpy when we reduce a dim with a shape of 0.
"""
# We have a choice of implementing this call with the
# CAReduce op or the MaxAndArgmax op.
# MaxAndArgmax supports grad and Rop, so we prefer to use that.
# CAReduce is faster, but optimizations will replace MaxAndArgmax[0]
# with CAReduce at compile time, so at this stage the important
# thing is supporting all user interface features, not speed.
# Some cases can be implemented only with CAReduce.
out = max_and_argmax(x, axis)[0]
if keepdims:
out = makeKeepDims(x, out, axis)
return out
def argmax(x, axis=None, keepdims=False):
"""
Returns indices of maximum elements obtained by iterating over given axis.
When axis is None (the default value), the argmax is performed
over the flattened tensor.
Parameters
----------
keepdims : bool
If this is set to True, the axes which are reduced are left in
the result as dimensions with size one. With this option, the result
will broadcast correctly against the original tensor.
"""
argout = max_and_argmax(x, axis)[1]
if keepdims:
argout = makeKeepDims(x, argout, axis)
return argout
[docs]
def min(x, axis=None, keepdims=False):
"""
Returns minimum elements obtained by iterating over given axis.
When axis is None (the default value), the min is performed
over the flattened tensor.
Parameters
----------
keepdims: bool
If this is set to True, the axes which are reduced are left in
the result as dimensions with size one. With this option, the result
will broadcast correctly against the original tensor.
"""
x = as_tensor_variable(x)
str_x_type = str(x.dtype)
if str_x_type.startswith("float") or str_x_type in int_dtypes:
return -max(-x, axis=axis, keepdims=keepdims)
elif str_x_type in uint_dtypes:
itype = np.iinfo(x.dtype)
max_val = np.array(itype.max, dtype=itype.dtype)
return max_val - max(max_val - x, axis=axis, keepdims=keepdims)
elif str_x_type == "bool":
return ~max(~x, axis=axis, keepdims=keepdims)
else:
# Be careful about unsigned integers, complex
raise NotImplementedError()
def argmin(x, axis=None, keepdims=False):
"""
Returns indices of minimum elements obtained by iterating over given axis.
When axis is None (the default value), the argmin is performed
over the flattened tensor.
Parameters
----------
keepdims: bool
If this is set to True, the axes which are reduced are left in
the result as dimensions with size one. With this option, the result
will broadcast correctly against the original tensor.
"""
x = as_tensor_variable(x)
str_x_type = str(x.dtype)
if str_x_type.startswith("float") or str_x_type in int_dtypes:
return argmax(-x, axis=axis, keepdims=keepdims)
elif str_x_type in uint_dtypes:
itype = np.iinfo(x.dtype)
return argmax(itype.max - x, axis=axis, keepdims=keepdims)
elif str_x_type == "bool":
return argmax(~x, axis=axis, keepdims=keepdims)
else:
# Be careful about unsigned integers, complex
raise NotImplementedError()
def smallest(*args):
"""
Return the [elementwise] smallest of a variable number of arguments.
Like python's min.
"""
if len(args) == 2:
a, b = args
return switch(a < b, a, b)
else:
return min(stack(args), axis=0)
def largest(*args):
"""
Return the [elementwise] largest of a variable number of arguments.
Like python's max.
"""
if len(args) == 2:
a, b = args
return switch(a > b, a, b)
else:
return max(stack(args), axis=0)
[docs]
@scalar_elemwise
def lt(a, b):
"""a < b"""
[docs]
@scalar_elemwise
def gt(a, b):
"""a > b"""
[docs]
@scalar_elemwise
def le(a, b):
"""a <= b"""
[docs]
@scalar_elemwise
def ge(a, b):
"""a >= b"""
[docs]
@scalar_elemwise
def eq(a, b):
"""a == b"""
[docs]
@scalar_elemwise
def neq(a, b):
"""a != b"""
@scalar_elemwise
def isnan(a):
"""isnan(a)"""
# Rename isnan to isnan_ to allow to bypass it when not needed.
# glibc 2.23 don't allow isnan on int, so we remove it from the graph.
isnan_ = isnan
def isnan(a):
"""isnan(a)"""
a = as_tensor_variable(a)
if a.dtype in discrete_dtypes:
return alloc(
np.asarray(False, dtype="bool"), *[a.shape[i] for i in range(a.ndim)]
)
return isnan_(a)
@scalar_elemwise
def isinf(a):
"""isinf(a)"""
# Rename isnan to isnan_ to allow to bypass it when not needed.
# glibc 2.23 don't allow isnan on int, so we remove it from the graph.
isinf_ = isinf
def isinf(a):
"""isinf(a)"""
a = as_tensor_variable(a)
if a.dtype in discrete_dtypes:
return alloc(
np.asarray(False, dtype="bool"), *[a.shape[i] for i in range(a.ndim)]
)
return isinf_(a)
def allclose(a, b, rtol=1.0e-5, atol=1.0e-8, equal_nan=False):
"""
Implement Numpy's ``allclose`` on tensors.
``absolute(a - b) <= (atol + rtol * absolute(b))``
Parameters
----------
a : tensor
Input to compare.
b : tensor
Input to compare.
rtol : float
The relative tolerance parameter.
atol : float
The absolute tolerance parameter.
equal_nan: bool
Whether to consider nan's in the same place to be close.
Returns
-------
bool
A boolean value (of type int8 returned by the tensor elementwise `all`
function) whether all elements in a and b are in the tolerance range
defined above.
Notes
-----
Not a symmetric equation. See Numpy's documentation.
"""
return all(isclose(a, b, rtol, atol, equal_nan))
def isclose(a, b, rtol=1.0e-5, atol=1.0e-8, equal_nan=False):
"""
Implements Numpy's ``isclose`` on tensors.
The tolerance values are positive, typically very small numbers. The
relative difference (`rtol` * abs(`b`)) and the absolute difference
`atol` are added together to compare against the absolute difference
between `a` and `b`.
``absolute(a - b) <= (atol + rtol * absolute(b))``
Parameters
----------
a : tensor
Input to compare.
b : tensor
Input to compare.
rtol : float
The relative tolerance parameter.
atol : float
The absolute tolerance parameter.
equal_nan : bool
Whether to consider nan's in the same place to be close
Returns
-------
int8
A boolean (int8) array where two arrays are element-wise equal
within a tolerance.
Notes
-----
Not a symmetric equation. See Numpy's documentation.
Examples
--------
>>> import pytensor
>>> import numpy as np
>>> a = _asarray([1e10, 1e-7], dtype="float64")
>>> b = _asarray([1.00001e10, 1e-8], dtype="float64")
>>> pytensor.tensor.isclose(a, b).eval()
array([1, 0], dtype=int8)
>>> a = _asarray([1e10, 1e-8], dtype="float64")
>>> b = _asarray([1.00001e10, 1e-9], dtype="float64")
>>> pytensor.tensor.isclose(a, b).eval()
array([1, 1], dtype=int8)
>>> a = _asarray([1e10, 1e-8], dtype="float64")
>>> b = _asarray([1.0001e10, 1e-9], dtype="float64")
>>> pytensor.tensor.isclose(a, b).eval()
array([0, 1], dtype=int8)
>>> a = _asarray([1.0, np.nan], dtype="float64")
>>> b = _asarray([1.0, np.nan], dtype="float64")
>>> pytensor.tensor.isclose(a, b).eval()
array([1, 0], dtype==int8)
>>> a = _asarray([1.0, np.nan], dtype="float64")
>>> b = _asarray([1.0, np.nan], dtype="float64")
>>> pytensor.tensor.isclose(a, b, equal_nan=True).eval()
array([1, 1], dtype==int8)
>>> a = _asarray([1.0, np.inf], dtype="float64")
>>> b = _asarray([1.0, -np.inf], dtype="float64")
>>> pytensor.tensor.isclose(a, b).eval()
array([1, 0], dtype==int8)
>>> a = _asarray([1.0, np.inf], dtype="float64")
>>> b = _asarray([1.0, np.inf], dtype="float64")
>>> pytensor.tensor.isclose(a, b).eval()
array([1, 1], dtype==int8)
"""
# close will be an int8 array of 1 where within tolerance
# and 0 where not within tolerance or there was a nan or inf value.
diff = _abs(a - b)
tolerance = atol + rtol * _abs(b)
close_prelim = le(diff, tolerance)
a_nan = isnan(a)
b_nan = isnan(b)
nans = bitwise_or(a_nan, b_nan)
a_inf = isinf(a)
b_inf = isinf(b)
infs = bitwise_or(a_inf, b_inf)
nans_or_infs = bitwise_or(nans, infs)
# close is now an array of 0's except where elements are not nan or inf
# and are within the tolerance.
close = bitwise_and(close_prelim, bitwise_not(nans_or_infs))
# deal with signed inf values. this will make an array inf_eq of 0's
# except where inf values have the same sign.
both_infs = bitwise_and(a_inf, b_inf)
inf_signs_eq = eq(a_inf * sign(a), b_inf * sign(b))
inf_eq = bitwise_and(both_infs, inf_signs_eq)
# now create the potential result combining close and inf_eq
close_with_infs = bitwise_or(close, inf_eq)
# deal with comparing nan's.
if equal_nan:
both_nans = bitwise_and(a_nan, b_nan)
return bitwise_or(close_with_infs, both_nans)
# otherwise nan's aren't considered close.
else:
return close_with_infs
##########################
# Bit-wise
##########################
[docs]
@scalar_elemwise
def and_(a, b):
"""bitwise a & b"""
[docs]
@scalar_elemwise
def or_(a, b):
"""bitwise a | b"""
@scalar_elemwise
def xor(a, b):
"""bitwise a ^ b"""
@scalar_elemwise
def invert(a):
"""bitwise ~a"""
##########################
# Math
##########################
[docs]
@scalar_elemwise
def abs(a):
"""|`a`|"""
pprint.assign(abs, printing.PatternPrinter(("|%(0)s|", -1000)))
[docs]
@scalar_elemwise
def exp(a):
"""e^`a`"""
@scalar_elemwise
def exp2(a):
"""2^`a`"""
@scalar_elemwise
def expm1(a):
"""e^`a` - 1"""
@scalar_elemwise
def neg(a):
"""-a"""
@scalar_elemwise
def reciprocal(a):
"""1.0/a"""
[docs]
@scalar_elemwise
def log(a):
"""base e logarithm of a"""
@scalar_elemwise
def log2(a):
"""base 2 logarithm of a"""
@scalar_elemwise
def log10(a):
"""base 10 logarithm of a"""
@scalar_elemwise
def log1p(a):
"""log(1+a)"""
@scalar_elemwise
def sign(a):
"""sign of a"""
[docs]
def sgn(a):
"""sign of a"""
warnings.warn(
"sgn is deprecated and will stop working in the future, use sign instead.",
FutureWarning,
)
return sign(a)
[docs]
@scalar_elemwise
def ceil(a):
"""ceiling of a"""
[docs]
@scalar_elemwise
def floor(a):
"""floor of a"""
@scalar_elemwise
def trunc(a):
"""trunc of a"""
def iround(a, mode=None):
"""cast(round(a,mode),'int64')"""
return cast(round(a, mode), "int64")
[docs]
def round(a, mode=None):
"""round_mode(a) with mode in [half_away_from_zero, half_to_even].
Default to half_to_even."""
if mode is None:
mode = "half_to_even"
if config.warn__round:
warnings.warn(
"pytensor.tensor.round() changed its default from"
" `half_away_from_zero` to `half_to_even` to have"
" the same default as NumPy. Use the PyTensor flag"
" `warn__round=False` to disable this warning."
)
if mode == "half_away_from_zero":
return round_half_away_from_zero(a)
elif mode == "half_to_even":
return round_half_to_even(a)
else:
raise Exception(f"round mode {mode} is not implemented.")
@scalar_elemwise
def round_half_to_even(a):
"""round_half_to_even(a)"""
@scalar_elemwise
def round_half_away_from_zero(a):
"""round_half_away_from_zero(a)"""
[docs]
@scalar_elemwise
def sqr(a):
"""square of a"""
def cov(m, y=None, rowvar=True, bias=False, ddof=None, fweights=None, aweights=None):
"""Calculate the covariance matrix.
Covariance indicates the level to which two variables vary together.
If we examine N-dimensional samples, :math:`m = [x_1, x_2, ... x_N]^T`,
then the covariance matrix element :math:`C_{ij}` is the covariance of
:math:`x_i` and :math:`x_j`. The element :math:`C_{ii}` is the variance
of :math:`x_i`. Code and docstring ported from numpy.
Parameters
==========
m : array_like
A 2-D array containing multiple variables and observations.
Each row of `m` represents a variable, and each column is
observations of all those variables.
y : array_like, optional
An additional set of variables and observations. `y` has the same form
as that of `m`.
rowvar : bool, optional
If `rowvar` is True (default), then each row represents a
variable, with observations in the columns. Otherwise, the relationship
is transposed: each column represents a variable, while the rows
contain observations.
bias : bool, optional
Default normalization (False) is by ``(N - 1)``, where ``N`` is the
number of observations given (unbiased estimate). If `bias` is True, then
normalization is by ``N``. These values can be overridden by using the
keyword ``ddof``.
ddof : int, optional
If not ``None`` the default value implied by `bias` is overridden.
The default value is ``None``.
Returns
=======
out : The covariance matrix of the variables.
"""
if fweights is not None:
raise NotImplementedError("fweights are not implemented")
if aweights is not None:
raise NotImplementedError("aweights are not implemented")
if not rowvar and m.shape[0] != 1:
m = m.T
if y is not None:
if not rowvar and y.shape[0] != 1:
y = y.T
m = concatenate((m, y), axis=0)
if ddof is None:
if not bias:
ddof = 1
else:
ddof = 0
# Determine the normalization
fact = m.shape[1] - ddof
m -= m.mean(axis=1, keepdims=1)
c = m.dot(m.T)
c *= constant(1) / fact
return c.squeeze()
[docs]
@scalar_elemwise
def sqrt(a):
"""square root of a"""
@scalar_elemwise
def deg2rad(a):
"""convert degree a to radian"""
@scalar_elemwise
def rad2deg(a):
"""convert radian a to degree"""
[docs]
@scalar_elemwise
def cos(a):
"""cosine of a"""
[docs]
@scalar_elemwise
def arccos(a):
"""arccosine of a"""
[docs]
@scalar_elemwise
def sin(a):
"""sine of a"""
[docs]
@scalar_elemwise
def arcsin(a):
"""arcsine of a"""
[docs]
@scalar_elemwise
def tan(a):
"""tangent of a"""
[docs]
@scalar_elemwise
def arctan(a):
"""arctangent of a"""
@scalar_elemwise
def arctan2(a, b):
"""arctangent of a / b"""
[docs]
@scalar_elemwise
def cosh(a):
"""hyperbolic cosine of a"""
[docs]
@scalar_elemwise
def arccosh(a):
"""hyperbolic arc cosine of a"""
[docs]
@scalar_elemwise
def sinh(a):
"""hyperbolic sine of a"""
[docs]
@scalar_elemwise
def arcsinh(a):
"""hyperbolic arc sine of a"""
[docs]
@scalar_elemwise
def tanh(a):
"""hyperbolic tangent of a"""
[docs]
@scalar_elemwise
def arctanh(a):
"""hyperbolic arc tangent of a"""
[docs]
@scalar_elemwise
def erf(a):
"""error function"""
[docs]
@scalar_elemwise
def erfc(a):
"""complementary error function"""
@scalar_elemwise
def erfcx(a):
"""scaled complementary error function"""
[docs]
@scalar_elemwise
def erfinv(a):
"""inverse error function"""
[docs]
@scalar_elemwise
def erfcinv(a):
"""inverse complementary error function"""
@scalar_elemwise
def owens_t(h, a):
"""owens t function"""
@scalar_elemwise
def gamma(a):
"""gamma function"""
@scalar_elemwise
def gammaln(a):
"""log gamma function"""
@scalar_elemwise
def psi(a):
"""derivative of log gamma function"""
digamma = psi
@scalar_elemwise
def tri_gamma(a):
"""second derivative of the log gamma function"""
@scalar_elemwise
def polygamma(n, x):
"""Polygamma function of order n evaluated at x"""
@scalar_elemwise
def chi2sf(x, k):
"""chi squared survival function"""
@scalar_elemwise
def gammainc(k, x):
"""Regularized lower gamma function"""
@scalar_elemwise
def gammaincc(k, x):
"""Regularized upper gamma function"""
@scalar_elemwise
def gammau(k, x):
"""Upper incomplete gamma function."""
@scalar_elemwise
def gammal(k, x):
"""Lower incomplete gamma function."""
@scalar_elemwise
def gammaincinv(k, x):
"""Inverse to the regularized lower incomplete gamma function"""
@scalar_elemwise
def gammainccinv(k, x):
"""Inverse of the regularized upper incomplete gamma function"""
@scalar_elemwise
def hyp2f1(a, b, c, z):
"""Gaussian hypergeometric function."""
@scalar_elemwise
def j0(x):
"""Bessel function of the first kind of order 0."""
@scalar_elemwise
def j1(x):
"""Bessel function of the first kind of order 1."""
@scalar_elemwise
def jv(v, x):
"""Bessel function of the first kind of order v (real)."""
@scalar_elemwise
def i0(x):
"""Modified Bessel function of the first kind of order 0."""
@scalar_elemwise
def i1(x):
"""Modified Bessel function of the first kind of order 1."""
@scalar_elemwise
def iv(v, x):
"""Modified Bessel function of the first kind of order v (real)."""
@scalar_elemwise
def ive(v, x):
"""Exponentially scaled modified Bessel function of the first kind of order v (real)."""
[docs]
@scalar_elemwise
def sigmoid(x):
"""Logistic sigmoid function (1 / (1 + exp(-x)), also known as expit or inverse logit"""
expit = sigmoid
@scalar_elemwise
def softplus(x):
"""Compute log(1 + exp(x)), also known as softplus or log1pexp"""
log1pexp = softplus
@scalar_elemwise
def log1mexp(x):
"""Compute log(1 - exp(x)), also known as log1mexp"""
@scalar_elemwise
def betainc(a, b, x):
"""Regularized incomplete beta function"""
@scalar_elemwise
def betaincinv(a, b, x):
"""Inverse of the regularized incomplete beta function"""
@scalar_elemwise
def real(z):
"""Return real component of complex-valued tensor `z`."""
_tensor_py_operators.real = property(real, doc=real.__doc__)
@scalar_elemwise
def imag(z):
"""Return imaginary component of complex-valued tensor `z`."""
_tensor_py_operators.imag = property(imag, doc=imag.__doc__)
@scalar_elemwise
def angle(z):
"""Return polar-coordinate angle of complex-valued tensor `z`"""
@scalar_elemwise # numpy.complex cannot build tensors
def complex(real, imag):
"""Return complex-valued tensor with `real` and `imag` components"""
@scalar_elemwise(symbolname="conj")
def _conj(z):
"""Return the complex conjugate of `z`."""
def conjugate(x):
_x = as_tensor_variable(x)
if _x.type.dtype not in complex_dtypes:
return _x
return _conj(_x)
conj = conjugate
@scalar_elemwise
def complex_from_polar(abs, angle):
"""Return complex-valued tensor from polar coordinate specification."""
class Mean(FixedOpCAReduce):
__props__ = ("axis",)
nfunc_spec = ("mean", 1, 1)
def __init__(self, axis=None):
super().__init__(ps.mean, axis)
assert self.axis is None or len(self.axis) == 1
def __str__(self):
if self.axis is not None:
return "Mean{%s}" % (", ".join(str(x) for x in self.axis))
else:
return "Mean"
def _output_dtype(self, idtype):
# we want to protect against overflow
return "float64"
def perform(self, node, inp, out):
(input,) = inp
(output,) = out
if self.axis is None:
axis = None
else:
axis = self.axis[0]
# numpy.asarray is needed as otherwise we can end up with a
# numpy scalar.
output[0] = np.asarray(np.mean(input, dtype="float64", axis=axis))
def c_code(self, node, name, inames, onames, sub):
ret = super().c_code(node, name, inames, onames, sub)
if self.axis is not None:
return ret
# TODO: c_code perform support only axis is None
return (
ret
+ f"""
*((double *)PyArray_DATA({onames[0]})) /= PyArray_SIZE({inames[0]});
"""
)
def clone(self, **kwargs):
axis = kwargs.get("axis", self.axis)
return type(self)(axis=axis)
# TODO: implement the grad. When done and tested, you can make this the default
# version.
# def grad(self, (x,), (gout,)):
# import pdb;pdb.set_trace()
# return grad(mean(x, self.axis, op=False),[x])
[docs]
def mean(input, axis=None, dtype=None, op=False, keepdims=False, acc_dtype=None):
"""
Computes the mean value along the given axis(es) of a tensor `input`.
Parameters
----------
axis : None or int or (list of int) (see `Sum`)
Compute the mean along this axis of the tensor.
None means all axes (like numpy).
dtype: None or string
Dtype to cast the result of the inner summation into.
For instance, by default, a sum of a float32 tensor will be
done in float64 (acc_dtype would be float64 by default),
but that result will be casted back in float32.
keepdims: bool
If this is set to True, the axes which are reduced are
left in the result as dimensions with size one. With this option,
the result will broadcast correctly against the original tensor.
acc_dtype: None or string
Dtype to use for the inner summation. This will not
necessarily be the dtype of the output (in particular
if it is a discrete (int/uint) dtype, the output will
be in a float type). If None, then we use the same rules as `sum()`.
"""
input = as_tensor_variable(input)
if op:
if dtype not in (None, "float64"):
raise NotImplementedError(
"The Mean op does not support the dtype argument, "
"and will always use float64. If you want to specify "
"the dtype, call tensor.mean(..., op=False).",
dtype,
)
if acc_dtype not in (None, "float64"):
raise NotImplementedError(
"The Mean op does not support the acc_dtype argument, "
"and will always use float64. If you want to specify "
"acc_dtype, call tensor.mean(..., op=False).",
dtype,
)
out = Mean(axis)(input)
if keepdims:
out = makeKeepDims(input, out, axis)
return out
if dtype is not None:
# The summation will be done with the specified dtype.
# sum() will complain if it is not suitable.
sum_dtype = dtype
else:
sum_dtype = None
# float16 overflows on the cast way too often
if input.dtype == "float16":
sum_dtype = "float32"
s = sum(input, axis=axis, dtype=sum_dtype, keepdims=keepdims, acc_dtype=acc_dtype)
shp = shape(input)
# Cast shp into a float type
# TODO Once we have a consistent casting policy, we could simply
# use true_div.
if s.dtype in ("float16", "float32", "complex64"):
shp = cast(shp, "float32")
else:
shp = cast(shp, "float64")
if axis is None:
axis = list(range(input.ndim))
elif isinstance(axis, int | np.integer):
axis = [axis]
elif isinstance(axis, np.ndarray) and axis.ndim == 0:
axis = [int(axis)]
else:
axis = [int(a) for a in axis]
# This sequential division will possibly be optimized by PyTensor:
for i in axis:
s = true_div(s, shp[i])
# This can happen when axis is an empty list/tuple
if s.dtype != shp.dtype and s.dtype in discrete_dtypes:
s = cast(s, shp.dtype)
if dtype == "float16" or (dtype is None and input.dtype == "float16"):
s = cast(s, "float16")
s.name = "mean"
return s
def var(input, axis=None, ddof=0, keepdims=False, corrected=False):
"""
Computes the variance along the given axis(es) of a tensor `input`.
Parameters
----------
axis: None or int or (list of int) (see `Sum`)
Compute the variance along this axis of the tensor.
None means all axes (like numpy).
ddof: Degrees of freedom; 0 would compute the ML estimate, 1 would compute
the unbiased estimate.
keepdims : bool
If this is set to True, the axes which are reduced are
left in the result as dimensions with size one. With this option,
the result will broadcast correctly against the original tensor.
corrected : bool
If this is set to True, the 'corrected_two_pass' algorithm is
used to compute the variance.
Refer : http://www.cs.yale.edu/publications/techreports/tr222.pdf
Notes
-----
Default uses the two-pass algorithm (reference below).
https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Two-pass_algorithm
Also supports 'corrected_two_pass' algorithm (using the 'corrected' flag)
which is numerically more stable. There exist other implementations that
offer better stability, but probably slower.
"""
if isinstance(ddof, (bool)):
raise ValueError(
"Parameter keepdims is now at index 3: (input, \
axis=None, ddof=0, keepdims=False, corrected=False)"
)
input_ndim = input.type.ndim
if axis is None:
axis = list(range(input_ndim))
elif isinstance(axis, int | np.integer):
axis = [axis]
elif isinstance(axis, np.ndarray) and axis.ndim == 0:
axis = [int(axis)]
else:
axis = [int(a) for a in axis]
# compute the axis-wise mean
mean_input = mean(input, axis, keepdims=True)
# center the input
centered_input = input - mean_input
# return the mean sqr
two = constant(2, dtype=centered_input.dtype)
if ddof == 0:
v = mean((centered_input**two), axis, keepdims=keepdims)
else:
shp = shape(input) - ddof
v = sum((centered_input**two), axis=axis, keepdims=keepdims)
for i in axis:
v = true_div(v, shp[i])
# use 'corrected_two_pass' algorithm
if corrected:
if ddof == 0:
error = mean(centered_input, axis, keepdims=keepdims) ** 2
else:
shp = shape(input) - ddof
shp_inp = shape(input)
error = sum(centered_input, axis=axis, keepdims=keepdims) ** 2
for i in axis:
error = true_div(error, shp[i] * shp_inp[i])
v = v - error
v.name = "var"
return v
def std(input, axis=None, ddof=0, keepdims=False, corrected=False):
"""
Computes the standard deviation along the given axis(es) of a tensor `input`.
Parameters
----------
axis: None or int or (list of int) (see `Sum`)
Compute the variance along this axis of the tensor.
None means all axes (like numpy).
ddof: Degrees of freedom; 0 would compute the ML estimate, 1 would compute
the unbiased estimate.
keepdims : bool
If this is set to True, the axes which are reduced are
left in the result as dimensions with size one. With this option,
the result will broadcast correctly against the original tensor.
corrected : bool
If this is set to True, the 'corrected_two_pass' algorithm is
used to compute the variance.
Refer : http://www.cs.yale.edu/publications/techreports/tr222.pdf
Notes
-----
It calls 'var()' and 'var()' uses the two-pass algorithm (reference below).
https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Two-pass_algorithm
Function 'var()' also supports 'corrected_two_pass' algorithm (using the
'corrected' flag) which is numerically more stable. There exist other
implementations that offer better stability, but probably slower.
"""
if isinstance(ddof, (bool)):
raise ValueError(
"Parameter keepdims is now at index 3: (input, \
axis=None, ddof=0, keepdims=False, corrected=False)"
)
ret = sqrt(
var(input=input, axis=axis, ddof=ddof, keepdims=keepdims, corrected=corrected)
)
ret.name = "std"
return ret
[docs]
@scalar_elemwise(symbolname="scalar_maximum")
def maximum(x, y):
"""elemwise maximum. See max for the maximum in one tensor"""
# see decorator for function body
[docs]
@scalar_elemwise(symbolname="scalar_minimum")
def minimum(x, y):
"""elemwise minimum. See min for the minimum in one tensor"""
# see decorator for function body
def divmod(x, y):
"""elementvise divmod, using floor_div and mod_check"""
return floor_div(x, y), mod_check(x, y)
@scalar_elemwise
def add(a, *other_terms):
"""elementwise addition"""
# see decorator for function body
@scalar_elemwise
def sub(a, b):
"""elementwise subtraction"""
# see decorator for function body
@scalar_elemwise
def mul(a, *other_terms):
"""elementwise multiplication"""
# see decorator for function body
@scalar_elemwise
def true_div(a, b):
"""elementwise [true] division (inverse of multiplication)"""
# see decorator for function body
@scalar_elemwise
def int_div(a, b):
"""elementwise [floor] division (inverse of multiplication)"""
# see decorator for function body
# floor_div and int_div are the same thing
floor_div = int_div
def ceil_intdiv(a, b):
"""Safely compute ``ceil(float_division(a, b))``.
Works for all dtypes, but mostly useful when `a` and `b` are ints.
"""
# If a and b are int with not many significant bits, we could
# cast them to float to avoid doing the modulo. We do not know if this
# is faster or not. But this is not safe for int64, because the cast will
# lose precision. For example:
# cast(cast(a, scalar.upcast(a.type.dtype, 'float32')) / b,
# ps.upcast(a.type.dtype, b.type.dtype))
# We cast for the case when a and b are uint*; otherwise, neq will
# force their upcast to int.
div = int_div(a, b)
ret = cast(neq(a % b, 0), div.dtype) + div
assert ret.dtype == ps.upcast(
div.owner.inputs[0].type.dtype, div.owner.inputs[1].type.dtype
)
return ret
def mod_check(x, y):
"""Make sure we do not try to use complex numbers."""
if (
as_tensor_variable(x).dtype in complex_dtypes
or as_tensor_variable(y).dtype in complex_dtypes
):
# Currently forbidden.
raise ps.Mod.complex_error
else:
return mod(x, y)
@scalar_elemwise
def mod(a, b):
"""elementwise modulo"""
# see decorator for function body
@scalar_elemwise
def pow(a, b):
"""elementwise power"""
# see decorator for function body
[docs]
@scalar_elemwise
def clip(x, min, max):
"""
Clip x to be between min and max.
Note that when `x` is equal to the boundaries, the output is considered
to be `x`, so at these points, the gradient of the cost wrt the output
will be propagated to `x`, not to `min` nor `max`. In other words,
on these points, the gradient wrt `x` will be equal to the gradient wrt
the output, and the gradient wrt `min` and `max` will be zero.
"""
# see decorator for function body
# for grep: clamp, bound
pprint.assign(add, printing.OperatorPrinter("+", -2, "either"))
pprint.assign(mul, printing.OperatorPrinter("*", -1, "either"))
pprint.assign(sub, printing.OperatorPrinter("-", -2, "left"))
pprint.assign(neg, printing.OperatorPrinter("-", 0, "either"))
pprint.assign(true_div, printing.OperatorPrinter("/", -1, "left"))
pprint.assign(int_div, printing.OperatorPrinter("//", -1, "left"))
pprint.assign(pow, printing.OperatorPrinter("**", 1, "right"))
class Dot(Op):
"""
Computes the dot product of two variables. For two matrices, this is
equivalent to matrix multiplication. For two vectors, this is the inner
product.
Notes
-----
Matrix-matrix products are sometimes optimized to Dot22 or Gemm ops
(see tensor.blas).
Vector-vector products are sometimes optimized to Ger or CGer (see
tensor.blas).
Matrix-vector products are sometimes optimized to Gemv, CGemv (see
tensor.blas).
"""
__props__ = ()
# the rationale for Dot22 is related to getting GEMM Ops into the
# graph. See Dot22 in tensor.blas for details.
def make_node(self, *inputs):
inputs = list(map(as_tensor_variable, inputs))
if len(inputs) != 2:
raise TypeError(f"Two arguments required, {len(inputs)} given ")
if inputs[0].ndim not in (1, 2):
raise TypeError(
"Input 0 (0-indexed) must have ndim of "
f"1 or 2, {int(inputs[0].ndim)} given. Consider calling "
"pytensor.tensor.dot instead."
)
if inputs[1].ndim not in (1, 2):
raise TypeError(
"Input 1 (0-indexed) must have ndim of "
f"1 or 2, {int(inputs[1].ndim)} given. Consider calling "
"pytensor.tensor.dot instead."
)
sx, sy = (input.type.shape for input in inputs)
if len(sy) == 2:
sz = sx[:-1] + sy[-1:]
elif len(sy) == 1:
sz = sx[:-1]
i_dtypes = [input.type.dtype for input in inputs]
outputs = [tensor(dtype=ps.upcast(*i_dtypes), shape=sz)]
return Apply(self, inputs, outputs)
def perform(self, node, inp, out):
x, y = inp
(z,) = out
# the asarray is here because dot between two vectors
# gives a numpy float object but we need to return a 0d
# ndarray
z[0] = np.asarray(np.dot(x, y))
def grad(self, inp, grads):
x, y = inp
(gz,) = grads
xdim, ydim, gdim = x.type.ndim, y.type.ndim, gz.type.ndim
# grad is scalar, so x is vector and y is vector
if gdim == 0:
xgrad = gz * y
ygrad = gz * x
# x is vector, y is matrix, grad is vector
elif xdim == 1 and ydim == 2:
xgrad = dot(gz, y.T)
ygrad = outer(x.T, gz)
# x is matrix, y is vector, grad is vector
elif xdim == 2 and ydim == 1:
xgrad = outer(gz, y.T)
ygrad = dot(x.T, gz)
# x is matrix, y is matrix, grad is matrix
elif xdim == ydim == 2:
xgrad = dot(gz, y.T)
ygrad = dot(x.T, gz)
# If x or y contain broadcastable dimensions but only one of
# them know that a matching dimensions is broadcastable, the
# above code don't always return the right broadcast pattern.
# This cause problem down the road. See gh-1461.
if xgrad.broadcastable != x.broadcastable:
xgrad = specify_broadcastable(
xgrad, *(ax for (ax, b) in enumerate(x.type.broadcastable) if b)
)
if ygrad.broadcastable != y.broadcastable:
ygrad = specify_broadcastable(
ygrad, *(ax for (ax, b) in enumerate(y.type.broadcastable) if b)
)
rval = xgrad, ygrad
for elem in rval:
assert elem.dtype.find("float") != -1
return rval
def R_op(self, inputs, eval_points):
# R_op for a \dot b evaluated at c for a and d for b is
# simply c \dot b + a \dot d
assert len(inputs) == 2
assert len(eval_points) == 2
if eval_points[0] is None and eval_points[1] is None:
return [None]
if eval_points[0]:
t1 = self(eval_points[0], inputs[1])
if eval_points[1]:
t2 = self(inputs[0], eval_points[1])
if eval_points[0] and eval_points[1]:
return [t1 + t2]
elif eval_points[0]:
return [t1]
else:
return [t2]
def infer_shape(self, fgraph, node, shapes):
xshp, yshp = shapes
x, y = node.inputs
# vector / vector
if x.ndim == 1 and y.ndim == 1:
return [()]
# matrix / vector
if x.ndim == 2 and y.ndim == 1:
return [xshp[:-1]]
# vector / matrix
if x.ndim == 1 and y.ndim == 2:
return [yshp[-1:]]
# matrix / matrix
if x.ndim == 2 and y.ndim == 2:
return [xshp[:-1] + yshp[-1:]]
raise NotImplementedError()
def __str__(self):
return "dot"
_dot = Dot()
pprint.assign(
_dot, printing.OperatorPrinter(printing.special["middle_dot"], -1, "left")
)
[docs]
def dot(l, r):
"""Return a symbolic dot product.
This is designed to work with both sparse and dense tensors types.
"""
if not isinstance(l, Variable):
l = as_tensor_variable(l)
if not isinstance(r, Variable):
r = as_tensor_variable(r)
try:
res = l.__dot__(r)
if res is NotImplemented:
raise NotImplementedError
except (NotImplementedError, AttributeError, TypeError):
res = r.__rdot__(l)
if res is NotImplemented:
raise NotImplementedError()
return res
def dense_dot(a, b):
"""
Computes the dot product of two variables.
For two matrices, this is equivalent to matrix multiplication.
For two vectors, this is the inner product.
When one variable is a scalar, this is like elementwise multiplication.
For N dimensions, this is a sum product over the last axis
of the first array and the second-to-last axis of the second array:
dot(a, b)[i,j,k,m] = sum(a[i,j,:] * b[k,:,m])
Note that this dot function does one of three things, in the following
sequence:
1. If either a or b is scalar, it returns the elementwise product
without calling the PyTensor Dot op.
2. If either a or b has more than 2 dimensions, it calls PyTensor's
tensordot function with appropriate axes. The tensordot function
expresses high-dimensional dot products in terms of 2D matrix
multiplications, so it may be possible to further optimize for
performance.
3. If both a and b have either 1 or 2 dimensions, it calls PyTensor's
Dot op on a and b.
Notes
-----
Matrix-matrix products are sometimes optimized to Dot22 or Gemm ops
(see tensor.blas).
Vector-vector products are sometimes optimized to Ger or CGer (see
tensor.blas).
Matrix-vector products are sometimes optimized to Gemv, CGemv (see
tensor.blas).
"""
a, b = as_tensor_variable(a), as_tensor_variable(b)
if not isinstance(a.type, DenseTensorType) or not isinstance(
b.type, DenseTensorType
):
raise TypeError("The dense dot product is only supported for dense types")
if a.ndim == 0 or b.ndim == 0:
return a * b
elif a.ndim > 2 or b.ndim > 2:
return tensordot(a, b, [[a.ndim - 1], [np.maximum(0, b.ndim - 2)]])
else:
return _dot(a, b)
def _tensordot_as_dot(a, b, axes, dot, batched):
"""
Reduces a tensor dot product to a matrix or vector dot product. Based
on code from Tijmen Tieleman's gnumpy
(http://www.cs.toronto.edu/~tijmen/gnumpy.html).
Please see the documentation of tensordot for the meaning of the a, b
and axes arguments.
:param dot: a function that accepts two symbolic variables and computes
the appropriate dot product (e.g. dot, batched_dot)
:type dot: function
:param batched: whether to treat the first axis of a and b as a batch
axis. If so, this axis will be preserved in the output,
allowing this function to be used also for batched
tensor dot products.
:type batched: boolean
:returns: a tensor with shape equal to the concatenation of a's shape
(less any dimensions that were summed over) and b's shape
(less the first dimension and any dimensions that were summed
over).
:rtype: symbolic tensor
"""
a, b = as_tensor_variable(a), as_tensor_variable(b)
if not np.isscalar(axes) and len(axes) != 2:
raise ValueError(
"Axes should be an integer or a "
"list/tuple of len 2 ({axes} was provided)"
)
# if 'axes' is a number of axes to multiply and sum over (trailing axes
# of a, leading axes of b), we can just reshape and use dot.
elif np.isscalar(axes):
axes = int(axes)
for operand_name, operand in (("a", a), ("b", b)):
if axes > operand.ndim:
raise ValueError(
f"axes can not be larger than the dimension of {operand_name} "
f"({operand_name}.ndim={operand.ndim}, axes={axes})"
)
if batched and axes == operand.ndim:
raise ValueError(
"axes to sum over must not include the batch axis "
f"of {operand_name} ({operand_name}.ndim={operand.ndim}, axes={axes})"
)
batch_axes = 1 if batched else 0
a_outaxes = slice(0, a.ndim - axes)
b_outaxes = slice(batch_axes + axes, b.ndim)
outshape = concatenate([a.shape[a_outaxes], b.shape[b_outaxes]])
outbcast = a.broadcastable[a_outaxes] + b.broadcastable[b_outaxes]
outndim = len(outbcast)
a_shape = [1] * 2
b_shape = [1] * 2
# compute total size of summed axes
for i in range(0, axes):
a_shape[1] *= a.shape[-(i + 1)]
b_shape[0] *= b.shape[batch_axes + i]
# compute total size of other axes
for i in range(0, a.ndim - axes - batch_axes):
a_shape[0] *= a.shape[batch_axes + i]
for i in range(0, b.ndim - axes - batch_axes):
b_shape[1] *= b.shape[-(i + 1)]
if batched:
a_shape.insert(0, a.shape[0])
b_shape.insert(0, b.shape[0])
a_reshaped = a.reshape(a_shape)
b_reshaped = b.reshape(b_shape)
out_reshaped = dot(a_reshaped, b_reshaped)
out = out_reshaped.reshape(outshape, ndim=outndim)
# Make sure the broadcastable pattern of the result is correct,
# since some shape information can be lost in the reshapes.
if out.type.broadcastable != outbcast:
out = specify_broadcastable(
out, *(ax for (ax, b) in enumerate(outbcast) if b)
)
return out
# if 'axes' is a list, transpose a and b such that the summed axes of a
# are last and the summed axes of b are first.
else:
axes = [as_list(axes_) for axes_ in axes]
if len(axes[0]) != len(axes[1]):
raise ValueError("Axes elements must have the same length.")
for i, (operand_name, operand) in enumerate((("a", a), ("b", b))):
if len(axes[i]) > operand.ndim:
raise ValueError(
f"axes[{i}] should be array_like with length less than "
f"the dimensions of {operand_name} ({operand_name}.ndim={operand.ndim}, len(axes[0])={len(axes[i])})."
)
if len(axes[i]) > 0 and np.max(axes[i]) >= operand.ndim:
raise ValueError(
f"axes[{i}] contains dimensions greater than or equal "
f"to {operand_name}.ndim ({operand_name}.ndim={operand.ndim}, max(axes[0])={np.max(np.array(axes[i]))})."
)
if batched and 0 in axes[i]:
raise ValueError(
"axes to sum over must not contain the batch axis "
f"(axes[{i}]={axes[i]})"
)
batch_axes = [0] if batched else []
other_axes = [
[x for x in range(operand.ndim) if x not in axes[i] and x not in batch_axes]
for i, operand in enumerate((a, b))
]
a_shuffled = a.dimshuffle(batch_axes + other_axes[0] + axes[0])
b_shuffled = b.dimshuffle(batch_axes + axes[1] + other_axes[1])
# now that a and b are in the right order, recur with integer axes
return _tensordot_as_dot(
a_shuffled, b_shuffled, len(axes[0]), dot=dot, batched=batched
)
def tensordot(
a: TensorLike, b: TensorLike, axes: int | Sequence[Sequence[int]] = 2
) -> TensorVariable:
"""
Compute tensor dot product along specified axes.
Implementation is mostly taken from numpy version 1.26.0
Given two tensors, `a` and `b`, and a sequence object containing
two sequence objects, ``(a_axes, b_axes)``, sum the products of
`a`'s and `b`'s elements (components) over the axes specified by
``a_axes`` and ``b_axes``. The third argument can be a single non-negative
integer_like scalar, ``N``; if it is such, then the last ``N`` dimensions
of `a` and the first ``N`` dimensions of `b` are summed over.
Parameters
----------
a, b : tensor_like
Tensors to "dot".
axes : int or (2,) array_like
* integer_like
If an int N, sum over the last N axes of `a` and the first N axes
of `b` in order. The sizes of the corresponding axes must match.
* (2,) array_like
Or, a list of axes to be summed over, first sequence applying to `a`,
second to `b`. Both elements array_like must be of the same length.
Returns
-------
output : TensorVariable
The tensor dot product of the input.
Its shape will be equal to the concatenation of `a` and `b` shapes
(ignoring the dimensions that were summed over given in ``a_axes``
and ``b_axes``)
Examples
--------
It may be helpful to consider an example to see what tensordot does.
PyTensor's implementation is identical to NumPy's. Here ``a`` has shape (2, 3, 4)
and ``b`` has shape (5, 6, 4, 3). The axes to sum over are [[1, 2], [3, 2]] --
note that a.shape[1] == b.shape[3] and a.shape[2] == b.shape[2]; these axes
are compatible. The resulting tensor will have shape (2, 5, 6) -- the
dimensions that are not being summed:
>>> a = np.random.random((2,3,4))
>>> b = np.random.random((5,6,4,3))
#tensordot
>>> c = np.tensordot(a, b, [[1,2],[3,2]])
#loop replicating tensordot
>>> a0, a1, a2 = a.shape
>>> b0, b1, _, _ = b.shape
>>> cloop = np.zeros((a0,b0,b1))
#loop over non-summed indices -- these exist
#in the tensor product.
>>> for i in range(a0):
... for j in range(b0):
... for k in range(b1):
... #loop over summed indices -- these don't exist
... #in the tensor product.
... for l in range(a1):
... for m in range(a2):
... cloop[i,j,k] += a[i,l,m] * b[j,k,m,l]
>>> np.allclose(c, cloop)
true
This specific implementation avoids a loop by transposing a and b such that
the summed axes of ``a`` are last and the summed axes of ``b`` are first. The
resulting arrays are reshaped to 2 dimensions and a matrix dot product is taken.
The result is reshaped back to the required output dimensions.
In an extreme case, no axes may be specified. The resulting tensor
will have shape equal to the concatenation of the shapes of a and b:
>>> c = np.tensordot(a, b, 0)
>>> print(a.shape)
(2,3,4)
>>> print(b.shape)
(5,6,4,3)
>>> print(c.shape)
(2,3,4,5,6,4,3)
See the documentation of numpy.tensordot for more examples.
"""
try:
iter(axes)
except Exception:
axes_a = list(range(-axes, 0))
axes_b = list(range(0, axes))
else:
axes_a, axes_b = axes
try:
na = len(axes_a)
axes_a = list(axes_a)
except TypeError:
axes_a = [axes_a]
na = 1
try:
nb = len(axes_b)
axes_b = list(axes_b)
except TypeError:
axes_b = [axes_b]
nb = 1
a = as_tensor_variable(a)
b = as_tensor_variable(b)
runtime_shape_a = a.shape
bcast_a = a.broadcastable
static_shape_a = a.type.shape
ndim_a = a.ndim
runtime_shape_b = b.shape
bcast_b = b.broadcastable
static_shape_b = b.type.shape
ndim_b = b.ndim
if na != nb:
raise ValueError(
"The number of axes supplied for tensordot must be equal for each tensor. "
f"Got {na} and {nb} respectively."
)
axes_a = list(normalize_axis_tuple(axes_a, ndim_a))
axes_b = list(normalize_axis_tuple(axes_b, ndim_b))
must_assert_runtime = False
for k in range(na):
ax_a = axes_a[k]
ax_b = axes_b[k]
if (bcast_a[ax_a] != bcast_b[ax_b]) or (
static_shape_a[ax_a] is not None
and static_shape_b[ax_b] is not None
and static_shape_a[ax_a] != static_shape_b[ax_b]
):
raise ValueError(
"Input arrays have inconsistent broadcastable pattern or type shape along the axes "
"that are to be reduced with tensordot."
)
elif static_shape_a[ax_a] is None or static_shape_b[ax_b] is None:
if must_assert_runtime:
a = Assert(
"Input array shape along reduced axes of tensordot are not equal"
)(a, eq(a.shape[ax_a], b.shape[ax_b]))
must_assert_runtime = True
# Move the axes to sum over to the end of "a"
# and to the front of "b"
notin = [k for k in range(ndim_a) if k not in axes_a]
newaxes_a = notin + axes_a
N2 = 1
for axis in axes_a:
N2 *= runtime_shape_a[axis]
newshape_a = (-1, N2)
olda = [runtime_shape_a[axis] for axis in notin]
notin = [k for k in range(ndim_b) if k not in axes_b]
newaxes_b = axes_b + notin
N2 = 1
for axis in axes_b:
N2 *= runtime_shape_b[axis]
newshape_b = (N2, -1)
oldb = [runtime_shape_b[axis] for axis in notin]
at = a.transpose(newaxes_a).reshape(newshape_a)
bt = b.transpose(newaxes_b).reshape(newshape_b)
res = _dot(at, bt)
return res.reshape(olda + oldb)
def outer(x, y):
"""Return vector-vector outer product.
If an input isn't a vector, we flatten it first.
"""
if x.ndim != 1:
x = x.flatten()
if y.ndim != 1:
y = y.flatten()
return dot(x.dimshuffle(0, "x"), y.dimshuffle("x", 0))
class All(FixedOpCAReduce):
"""Applies `logical and` to all the values of a tensor along the
specified axis(es).
"""
nfunc_spec = ("all", 1, 1)
def __init__(self, axis=None):
super().__init__(ps.and_, axis)
def _output_dtype(self, idtype):
return "bool"
def make_node(self, input):
input = as_tensor_variable(input)
if input.dtype != "bool":
input = neq(input, 0)
ret = super().make_node(input)
return ret
def grad(self, inp, grads):
(x,) = inp
return [x.zeros_like(config.floatX)]
def clone(self, **kwargs):
axis = kwargs.get("axis", self.axis)
return type(self)(axis=axis)
class Any(FixedOpCAReduce):
"""Applies `bitwise or` to all the values of a tensor along the
specified axis(es).
"""
nfunc_spec = ("any", 1, 1)
def __init__(self, axis=None):
super().__init__(ps.or_, axis)
def _output_dtype(self, idtype):
return "bool"
def make_node(self, input):
input = as_tensor_variable(input)
if input.dtype != "bool":
input = neq(input, 0)
ret = super().make_node(input)
return ret
def grad(self, inp, grads):
(x,) = inp
return [x.zeros_like(config.floatX)]
def clone(self, **kwargs):
axis = kwargs.get("axis", self.axis)
return type(self)(axis=axis)
class Sum(FixedOpCAReduce):
"""
Sums all the values of a tensor along the specified axis(es).
Equivalent to `CAReduce(scalar.add, axis=axis, dtype=dtype)`,
with the difference that this defines the gradient of sum wrt its
tensor input.
"""
nfunc_spec = ("sum", 1, 1)
def __init__(self, axis=None, dtype=None, acc_dtype=None):
super().__init__(
ps.add,
axis=axis,
dtype=dtype,
acc_dtype=acc_dtype,
upcast_discrete_output=True,
)
def L_op(self, inp, out, grads):
(x,) = inp
if out[0].dtype not in continuous_dtypes:
return [x.zeros_like(dtype=config.floatX)]
(gz,) = grads
gz = as_tensor_variable(gz)
axis = self.axis
if axis is None:
axis = list(range(x.type.ndim))
if axis == ():
return (gz,)
new_dims = []
i = 0
for j, _ in enumerate(x.type.broadcastable):
if j in axis:
new_dims.append("x")
else:
new_dims.append(i)
i += 1
ds_op = DimShuffle(gz.type.broadcastable, new_dims)
gx = Elemwise(ps.second)(x, ds_op(gz))
return [gx]
def R_op(self, inputs, eval_points):
# There is just one element in inputs and eval_points, the axis are
# part of self
if None in eval_points:
return [None]
return self(*eval_points, return_list=True)
def clone(self, **kwargs):
axis = kwargs.get("axis", self.axis)
dtype = kwargs.get("dtype", self.dtype)
acc_dtype = kwargs.get("acc_dtype", self.acc_dtype)
return type(self)(axis=axis, dtype=dtype, acc_dtype=acc_dtype)
[docs]
def sum(input, axis=None, dtype=None, keepdims=False, acc_dtype=None):
"""
Computes the sum along the given axis(es) of a tensor `input`.
When axis is None (the default value), the sum is performed
over the flattened tensor.
For full documentation see `Sum`.
In particular please pay attention to the important warning when using
a custom acc_dtype.
Parameters
----------
keepdims: bool
If this is set to True, the axes which are reduced are left in
the result as dimensions with size one. With this option, the result
will broadcast correctly against the original tensor.
"""
out = Sum(axis=axis, dtype=dtype, acc_dtype=acc_dtype)(input)
if keepdims:
out = makeKeepDims(input, out, axis)
return out
pprint.assign(Sum, printing.FunctionPrinter(["sum"], ["axis"]))
class Prod(FixedOpCAReduce):
"""
Multiplies all the values of a tensor along the specified axis(es).
Equivalent to `CAReduce(scalar.mul, axis = axis)`, with the
difference that this defines the gradient of prod wrt its tensor
input.
"""
__props__ = ("scalar_op", "axis", "dtype", "acc_dtype", "no_zeros_in_input")
nfunc_spec = ("prod", 1, 1)
def __init__(self, axis=None, dtype=None, acc_dtype=None, no_zeros_in_input=False):
super().__init__(
ps.mul,
axis=axis,
dtype=dtype,
acc_dtype=acc_dtype,
upcast_discrete_output=True,
)
self.no_zeros_in_input = no_zeros_in_input
def L_op(self, inp, out, grads):
"""
The grad of this Op could be very easy, if it is was not for the case
where zeros are present in a given "group" (ie. elements reduced
together to form the product).
If no zeros are found in the elements of the product, then the
partial derivative of the product relative to one of the elements
(one of the inputs) is simply the product of the other elements.
That's easy to see from the chain rule.
Now the trick (with no zeros) is to take the overall product, then
for every original element, the partial derivative is given by
this product divided by the element itself (which equals the product
of the other terms). This is easy to do by broadcasting the original
product.
(Note that we also need to broadcast-multiply by the
"incoming gradient", ie. the gradient of the cost relative to the
output/product).
With zeros, things get more complicated. For a given group, we have 3
cases:
* No zeros in the group. Use previous trick.
* If only one zero is present, then the gradient for that element is
non-zero, but is zero for all others.
* If more than one zero is present, then all the derivatives are zero.
For the last two cases (with 1 or more zeros), we can't use the
division trick, as this gives divisions by 0.
Implementing that case-by-case logic is not as trivial, so a bunch of
hacks are piled down here to do it. Notably, for the "only one zero"
case, there's a special Op that computes the product of the elements
in the group, minus the zero (see `ProdWithoutZeros`). The trick is then
to use the division trick for groups with no zero, to use the
`ProdWithoutZeros` op where there's only one zero, and to output a
derivative of zero for any element part of a group with more than
one zero.
I do this by first counting the number of zeros in each group (see the
`at.eq` bits), then taking this or that behavior (see `at.switch`)
based on the result of this count.
"""
(prod_in,) = inp
(gz,) = grads
if out[0].dtype in discrete_dtypes or self.acc_dtype in discrete_dtypes:
# There is an int conversion in the way
return [prod_in.zeros_like(dtype=config.floatX)]
# Prepare the broadcasting that is used everywhere to broadcast
# over the original groups (ie. broadcast over the elements of a given
# product)
gz = as_tensor_variable(gz)
axis = self.axis
if axis is None:
axis = list(range(prod_in.type.ndim))
if axis == ():
return (gz,)
new_dims = []
i = 0
for j, _ in enumerate(prod_in.type.broadcastable):
if j in axis:
new_dims.append("x")
else:
new_dims.append(i)
i += 1
# result of the product, broadcastable over groups
prod_out = self(prod_in).dimshuffle(new_dims)
# incoming gradient, broadcastable over groups
gz = gz.dimshuffle(new_dims)
# division trick if we don't have zeros. This will contain
# NaNs to be eliminated in the `at.switch` if we do have zeros.
grad_case_without_zeros = gz * prod_out / prod_in
if self.no_zeros_in_input:
# this handles inputs with zeros, but only certain input shapes
return [grad_case_without_zeros]
else:
where_zeros = eq(prod_in, 0.0)
sum_where_zeros = sum(where_zeros, axis=self.axis)
groups_with_single_zero = eq(sum_where_zeros, 1).dimshuffle(new_dims)
# tensor with 0 everywhere except for those places where
# a 0 part of a group with a single zero was to be found
where_single_zero = groups_with_single_zero * where_zeros
# further optimization to avoid computing ProdWithoutZeros
# if the incoming gradient is 0
where_gz_not_zero = neq(gz, 0.0)
# only take ProdWithoutZeros for the groups with single zeros
# with non-null incoming gradient
where_to_take_prod_without_zeros = (
groups_with_single_zero * where_gz_not_zero
)
# preprocess the original input so that we set 0 everywhere
# except for groups that contain a single zero, to avoid computing
# multiplications on other groups
prod_without_zeros_in = where_to_take_prod_without_zeros * prod_in
# TODO: put lazy switch here, if it'd work
# this is pretty efficient already (no multiplication if 0), but
# it'd be even better if we had a lazy if per element
prod_without_zeros = ProdWithoutZeros(axis=self.axis)(prod_without_zeros_in)
prod_without_zeros = prod_without_zeros.dimshuffle(new_dims)
groups_without_zeros = eq(sum_where_zeros, 0).dimshuffle(new_dims)
final_grad = switch(
groups_without_zeros,
grad_case_without_zeros,
switch(where_single_zero, prod_without_zeros, 0.0) * gz,
)
return [final_grad]
def c_code_cache_version(self):
return (1,)
def clone(self, **kwargs):
axis = kwargs.get("axis", self.axis)
dtype = kwargs.get("dtype", self.dtype)
acc_dtype = kwargs.get("acc_dtype", self.acc_dtype)
no_zeros_in_input = kwargs.get("no_zeros_in_input", self.no_zeros_in_input)
return type(self)(
axis=axis,
dtype=dtype,
acc_dtype=acc_dtype,
no_zeros_in_input=no_zeros_in_input,
)
def __str__(self):
if self.no_zeros_in_input:
return f"{super().__str__()[:-1]}, no_zeros_in_input}})"
return super().__str__()
def __repr__(self):
return f"{super().__repr__()[:-1]}, no_zeros_in_input={self.no_zeros_in_input})"
[docs]
def prod(
input,
axis=None,
dtype=None,
keepdims=False,
acc_dtype=None,
no_zeros_in_input=False,
):
"""
Computes the product along the given axis(es) of a tensor `input`.
When axis is None (the default value), the product is performed
over the flattened tensor.
For full documentation see ``tensor.elemwise.Prod``.
Parameters
----------
keepdims: bool
If this is set to True, the axes which are reduced are left in
the result as dimensions with size one. With this option, the result
will broadcast correctly against the original tensor.
"""
out = Prod(
axis, dtype=dtype, acc_dtype=acc_dtype, no_zeros_in_input=no_zeros_in_input
)(input)
if keepdims:
out = makeKeepDims(input, out, axis)
return out
class MulWithoutZeros(BinaryScalarOp):
# "identity" here is zero, as in Reduce we don't want to start
# with reducing (1, something_else): this leads to the erroneous
# case where a vector of zeros is reduced by binary reductions
# of (1, 0), which always ends up as 1 (ie. the result for
# the c version, for the product of [0,0,0], is 1.0)
identity = 0.0
commutative = True
associative = True
def impl(self, x, y):
if x == 0:
return y
if y == 0:
return x
return x * y
def c_code(self, node, name, inp, out, sub):
x, y = inp
(z,) = out
return (
"%(z)s = ((%(x)s == 0) ? (%(y)s) : "
+ "((%(y)s == 0) ? (%(x)s) : ((%(y)s)*(%(x)s))) );"
) % locals()
def c_code_cache_version(self):
return (1,)
mul_without_zeros = MulWithoutZeros(ps.upcast_out, name="mul_without_zeros")
class ProdWithoutZeros(FixedOpCAReduce):
def __init__(self, axis=None, dtype=None, acc_dtype=None):
super().__init__(
mul_without_zeros,
axis=axis,
dtype=dtype,
acc_dtype=acc_dtype,
upcast_discrete_output=True,
)
def grad(self, inp, grads):
from pytensor.gradient import grad_not_implemented
(a,) = inp
a_grad = grad_not_implemented(
self,
0,
a,
"2nd derivatives of `product(a)` is not currently supported."
"If `a` is guaranteed to contains no zeros, use "
"`product(a, no_zeros_in_input=True)`.",
)
return [a_grad]
def clone(self, **kwargs):
axis = kwargs.get("axis", self.axis)
dtype = kwargs.get("dtype", self.dtype)
acc_dtype = kwargs.get("acc_dtype", self.acc_dtype)
return type(self)(axis=axis, dtype=dtype, acc_dtype=acc_dtype)
def any(x, axis=None, keepdims=False):
out = Any(axis)(x)
if keepdims:
out = makeKeepDims(x, out, axis)
return out
def all(x, axis=None, keepdims=False):
out = All(axis)(x)
if keepdims:
out = makeKeepDims(x, out, axis)
return out
def ptp(a, axis=None):
"""
Range of values (maximum - minimum) along an axis.
The name of the function comes from the acronym for peak to peak.
Parameters
----------
a
Input tensor.
axis
Axis along which to find the peaks. By default, flatten the array.
Returns
-------
array
A new array holding the result.
"""
a = as_tensor_variable(a)
out = max(a, axis) - min(a, axis)
return out
def power(x, y):
return x**y
[docs]
def logaddexp(*xs):
"""Logarithm of the sum of exponentiations of the inputs.
See ``numpy.logaddexp``.
Parameters
----------
xs : symbolic tensors
Input
Returns
-------
tensor
"""
return log(add(*[exp(x) for x in xs]))
[docs]
def logsumexp(x, axis=None, keepdims=False):
"""Compute the log of the sum of exponentials of input elements.
See ``scipy.special.logsumexp``.
Parameters
----------
x : symbolic tensor
Input
axis : None or int or tuple of ints, optional
Axis or axes over which the sum is taken. By default axis is None,
and all elements are summed.
keepdims : bool, optional
If this is set to True, the axes which are reduced are left in the
result as dimensions with size one. With this option, the result will
broadcast correctly against the original array.
Returns
-------
tensor
"""
return log(sum(exp(x), axis=axis, keepdims=keepdims))
_matrix_matrix_matmul = Blockwise(
_dot,
signature="(m,k),(k,n)->(m,n)",
gufunc_spec=("numpy.matmul", 2, 1),
)
[docs]
def matmul(x1: "ArrayLike", x2: "ArrayLike", dtype: Optional["DTypeLike"] = None):
"""Compute the matrix product of two tensor variables.
Parameters
----------
x1, x2
Input arrays, scalars not allowed.
dtype
The desired data-type for the array. If not given, then the type will
be determined as the minimum type required to hold the objects in the
sequence.
Returns
-------
out : ndarray
The matrix product of the inputs. This is a scalar only when both
`x1`, `x2` are 1-d vectors.
Raises
------
ValueError
If the last dimension of `x1` is not the same size as the second-to-last
dimension of `x2`. If a scalar value is passed in.
Notes
-----
The behavior depends on the arguments in the following way.
- If both arguments are 2-D they are multiplied like conventional matrices.
- If either argument is N-D, N > 2, it is treated as a stack of matrices
residing in the last two indexes and broadcast accordingly.
- If the first argument is 1-D, it is promoted to a matrix by prepending a
1 to its dimensions. After matrix multiplication the prepended 1 is removed.
- If the second argument is 1-D, it is promoted to a matrix by appending a
1 to its dimensions. After matrix multiplication the appended 1 is removed.
`matmul` differs from `dot` in two important ways:
- Multiplication by scalars is not allowed, use `mul` instead.
- Stacks of matrices are broadcast together as if the matrices were elements,
respecting the signature ``(n, k), (k, m) -> (n, m)``:
"""
x1 = as_tensor_variable(x1)
x2 = as_tensor_variable(x2)
if x1.type.ndim == 0 or x2.type.ndim == 0:
raise ValueError("matmul operand cannot be scalar")
if x1.type.ndim == 1 and x2.type.ndim == 1:
out = _dot(x1, x2)
elif x1.type.ndim == 1:
out = _matrix_matrix_matmul(x1[None], x2).squeeze(-2)
elif x2.type.ndim == 1:
out = _matrix_matrix_matmul(x1, x2[:, None]).squeeze(-1)
else:
out = _matrix_matrix_matmul(x1, x2)
if dtype is not None:
out = out.astype(dtype)
return out
@_vectorize_node.register(Dot)
def vectorize_node_dot_to_matmul(op, node, batched_x, batched_y):
old_x, old_y = node.inputs
if old_x.type.ndim == 2 and old_y.type.ndim == 2:
# If original input is equivalent to a matrix-matrix product,
# return specialized Matmul Op to avoid unnecessary new Ops.
return matmul(batched_x, batched_y).owner
else:
return vectorize_node_fallback(op, node, batched_x, batched_y)
# NumPy logical aliases
square = sqr
bitwise_and = and_
bitwise_or = or_
bitwise_xor = xor
bitwise_not = invert
greater = gt
greater_equal = ge
less = lt
less_equal = le
equal = eq
not_equal = neq
__all__ = [
"max_and_argmax",
"max",
"matmul",
"argmax",
"min",
"argmin",
"smallest",
"largest",
"lt",
"less",
"gt",
"greater",
"le",
"less_equal",
"ge",
"greater_equal",
"eq",
"equal",
"neq",
"not_equal",
"isnan",
"isinf",
"allclose",
"isclose",
"and_",
"bitwise_and",
"or_",
"bitwise_or",
"xor",
"bitwise_xor",
"invert",
"bitwise_not",
"abs",
"exp",
"exp2",
"expm1",
"neg",
"reciprocal",
"log",
"log2",
"log10",
"log1p",
"sgn",
"sign",
"ceil",
"floor",
"trunc",
"iround",
"round",
"round_half_to_even",
"round_half_away_from_zero",
"sqr",
"square",
"cov",
"sqrt",
"deg2rad",
"rad2deg",
"cos",
"arccos",
"sin",
"arcsin",
"tan",
"arctan",
"arctan2",
"cosh",
"arccosh",
"sinh",
"arcsinh",
"tanh",
"arctanh",
"erf",
"erfc",
"erfcx",
"erfinv",
"erfcinv",
"owens_t",
"gamma",
"gammaln",
"psi",
"digamma",
"tri_gamma",
"polygamma",
"chi2sf",
"gammainc",
"gammaincc",
"gammau",
"gammal",
"gammaincinv",
"gammainccinv",
"j0",
"j1",
"jv",
"i0",
"i1",
"iv",
"ive",
"sigmoid",
"expit",
"softplus",
"log1pexp",
"log1mexp",
"betainc",
"betaincinv",
"real",
"imag",
"angle",
"complex",
"conj",
"conjugate",
"complex_from_polar",
"sum",
"prod",
"mean",
"var",
"std",
"std",
"maximum",
"minimum",
"divmod",
"add",
"sub",
"mul",
"true_div",
"int_div",
"floor_div",
"ceil_intdiv",
"mod",
"pow",
"clip",
"dot",
"dense_dot",
"tensordot",
"outer",
"any",
"all",
"ptp",
"power",
"logaddexp",
"logsumexp",
"hyp2f1",
]