def lambdify(args,
expr,
modules=None,
printer=None,
use_imps=True,
dummify=True):
"""
Returns a lambda function for fast calculation of numerical values.
If not specified differently by the user, ``modules`` defaults to
``["numpy"]`` if NumPy is installed, and ``["math", "mpmath", "sympy"]``
if it isn't, that is, SymPy functions are replaced as far as possible by
either ``numpy`` functions if available, and Python's standard library
``math``, or ``mpmath`` functions otherwise. To change this behavior, the
"modules" argument can be used. It accepts:
- the strings "math", "mpmath", "numpy", "numexpr", "sympy", "tensorflow"
- any modules (e.g. math)
- dictionaries that map names of sympy functions to arbitrary functions
- lists that contain a mix of the arguments above, with higher priority
given to entries appearing first.
.. warning::
Note that this function uses ``eval``, and thus shouldn't be used on
unsanitized input.
The default behavior is to substitute all arguments in the provided
expression with dummy symbols. This allows for applied functions (e.g.
f(t)) to be supplied as arguments. Call the function with dummify=False if
dummy substitution is unwanted (and `args` is not a string). If you want
to view the lambdified function or provide "sympy" as the module, you
should probably set dummify=False.
For functions involving large array calculations, numexpr can provide a
significant speedup over numpy. Please note that the available functions
for numexpr are more limited than numpy but can be expanded with
implemented_function and user defined subclasses of Function. If specified,
numexpr may be the only option in modules. The official list of numexpr
functions can be found at:
https://github.com/pydata/numexpr#supported-functions
In previous releases ``lambdify`` replaced ``Matrix`` with ``numpy.matrix``
by default. As of release 1.0 ``numpy.array`` is the default.
To get the old default behavior you must pass in ``[{'ImmutableDenseMatrix':
numpy.matrix}, 'numpy']`` to the ``modules`` kwarg.
>>> from sympy import lambdify, Matrix
>>> from sympy.abc import x, y
>>> import numpy
>>> array2mat = [{'ImmutableDenseMatrix': numpy.matrix}, 'numpy']
>>> f = lambdify((x, y), Matrix([x, y]), modules=array2mat)
>>> f(1, 2)
matrix([[1],
[2]])
Usage
=====
(1) Use one of the provided modules:
>>> from sympy import sin, tan, gamma
>>> from sympy.utilities.lambdify import lambdastr
>>> from sympy.abc import x, y
>>> f = lambdify(x, sin(x), "math")
Attention: Functions that are not in the math module will throw a name
error when the lambda function is evaluated! So this would
be better:
>>> f = lambdify(x, sin(x)*gamma(x), ("math", "mpmath", "sympy"))
(2) Use some other module:
>>> import numpy
>>> f = lambdify((x,y), tan(x*y), numpy)
Attention: There are naming differences between numpy and sympy. So if
you simply take the numpy module, e.g. sympy.atan will not be
translated to numpy.arctan. Use the modified module instead
by passing the string "numpy":
>>> f = lambdify((x,y), tan(x*y), "numpy")
>>> f(1, 2)
-2.18503986326
>>> from numpy import array
>>> f(array([1, 2, 3]), array([2, 3, 5]))
[-2.18503986 -0.29100619 -0.8559934 ]
(3) Use a dictionary defining custom functions:
>>> def my_cool_function(x): return 'sin(%s) is cool' % x
>>> myfuncs = {"sin" : my_cool_function}
>>> f = lambdify(x, sin(x), myfuncs); f(1)
'sin(1) is cool'
Examples
========
>>> from sympy.utilities.lambdify import implemented_function
>>> from sympy import sqrt, sin, Matrix
>>> from sympy import Function
>>> from sympy.abc import w, x, y, z
>>> f = lambdify(x, x**2)
>>> f(2)
4
>>> f = lambdify((x, y, z), [z, y, x])
>>> f(1,2,3)
[3, 2, 1]
>>> f = lambdify(x, sqrt(x))
>>> f(4)
2.0
>>> f = lambdify((x, y), sin(x*y)**2)
>>> f(0, 5)
0.0
>>> row = lambdify((x, y), Matrix((x, x + y)).T, modules='sympy')
>>> row(1, 2)
Matrix([[1, 3]])
Tuple arguments are handled and the lambdified function should
be called with the same type of arguments as were used to create
the function.:
>>> f = lambdify((x, (y, z)), x + y)
>>> f(1, (2, 4))
3
A more robust way of handling this is to always work with flattened
arguments:
>>> from sympy.utilities.iterables import flatten
>>> args = w, (x, (y, z))
>>> vals = 1, (2, (3, 4))
>>> f = lambdify(flatten(args), w + x + y + z)
>>> f(*flatten(vals))
10
Functions present in `expr` can also carry their own numerical
implementations, in a callable attached to the ``_imp_``
attribute. Usually you attach this using the
``implemented_function`` factory:
>>> f = implemented_function(Function('f'), lambda x: x+1)
>>> func = lambdify(x, f(x))
>>> func(4)
5
``lambdify`` always prefers ``_imp_`` implementations to implementations
in other namespaces, unless the ``use_imps`` input parameter is False.
Usage with Tensorflow module:
>>> import tensorflow as tf
>>> f = Max(x, sin(x))
>>> func = lambdify(x, f, 'tensorflow')
>>> result = func(tf.constant(1.0))
>>> result # a tf.Tensor representing the result of the calculation
<tf.Tensor 'Maximum:0' shape=() dtype=float32>
>>> sess = tf.Session()
>>> sess.run(result) # compute result
1.0
>>> var = tf.Variable(1.0)
>>> sess.run(tf.global_variables_initializer())
>>> sess.run(func(var)) # also works for tf.Variable and tf.Placeholder
1.0
>>> tensor = tf.constant([[1.0, 2.0], [3.0, 4.0]]) # works with any shape tensor
>>> sess.run(func(tensor))
array([[ 1., 2.],
[ 3., 4.]], dtype=float32)
"""
from sympy.core.symbol import Symbol
from sympy.utilities.iterables import flatten
# If the user hasn't specified any modules, use what is available.
module_provided = True
if modules is None:
module_provided = False
try:
_import("numpy")
except ImportError:
# Use either numpy (if available) or python.math where possible.
# XXX: This leads to different behaviour on different systems and
# might be the reason for irreproducible errors.
modules = ["math", "mpmath", "sympy"]
else:
modules = ["numpy"]
# Get the needed namespaces.
namespaces = []
# First find any function implementations
if use_imps:
namespaces.append(_imp_namespace(expr))
# Check for dict before iterating
if isinstance(modules, (dict, str)) or not hasattr(modules, '__iter__'):
namespaces.append(modules)
else:
# consistency check
if _module_present('numexpr', modules) and len(modules) > 1:
raise TypeError("numexpr must be the only item in 'modules'")
namespaces += list(modules)
# fill namespace with first having highest priority
namespace = {}
for m in namespaces[::-1]:
buf = _get_namespace(m)
namespace.update(buf)
if hasattr(expr, "atoms"):
#Try if you can extract symbols from the expression.
#Move on if expr.atoms in not implemented.
syms = expr.atoms(Symbol)
for term in syms:
namespace.update({str(term): term})
if printer is None:
if _module_present('mpmath', namespaces):
from sympy.printing.pycode import MpmathPrinter as Printer
elif _module_present('numpy', namespaces):
from sympy.printing.pycode import NumPyPrinter as Printer
elif _module_present('numexpr', namespaces):
from sympy.printing.lambdarepr import NumExprPrinter as Printer
elif _module_present('tensorflow', namespaces):
from sympy.printing.lambdarepr import TensorflowPrinter as Printer
elif _module_present('sympy', namespaces):
from sympy.printing.pycode import SymPyPrinter as Printer
else:
from sympy.printing.pycode import PythonCodePrinter as Printer
user_functions = {}
for m in namespaces[::-1]:
if isinstance(m, dict):
for k in m:
user_functions[k] = k
printer = Printer({
'fully_qualified_modules': False,
'inline': True,
'user_functions': user_functions
})
# Get the names of the args, for creating a docstring
if not iterable(args):
args = (args, )
names = []
# Grab the callers frame, for getting the names by inspection (if needed)
callers_local_vars = inspect.currentframe().f_back.f_locals.items()
for n, var in enumerate(args):
if hasattr(var, 'name'):
names.append(var.name)
else:
# It's an iterable. Try to get name by inspection of calling frame.
name_list = [
var_name for var_name, var_val in callers_local_vars
if var_val is var
]
if len(name_list) == 1:
names.append(name_list[0])
else:
# Cannot infer name with certainty. arg_# will have to do.
names.append('arg_' + str(n))
# Create lambda function.
lstr = lambdastr(args, expr, printer=printer, dummify=dummify)
flat = '__flatten_args__'
imp_mod_lines = []
for mod, keys in (getattr(printer, 'module_imports', None) or {}).items():
for k in keys:
if k not in namespace:
imp_mod_lines.append("from %s import %s" % (mod, k))
for ln in imp_mod_lines:
exec_(ln, {}, namespace)
if flat in lstr:
namespace.update({flat: flatten})
# Provide lambda expression with builtins, and compatible implementation of range
namespace.update({'builtins': builtins, 'range': range})
func = eval(lstr, namespace)
# For numpy lambdify, wrap all input arguments in arrays.
# This is a fix for gh-11306.
if module_provided and _module_present('numpy', namespaces):
def array_wrap(funcarg):
@wraps(funcarg)
def wrapper(*argsx, **kwargsx):
asarray = namespace['asarray']
newargs = [
asarray(i) if isinstance(i, integer_types +
(float, complex)) else i
for i in argsx
]
return funcarg(*newargs, **kwargsx)
return wrapper
func = array_wrap(func)
# Apply the docstring
sig = "func({0})".format(", ".join(str(i) for i in names))
sig = textwrap.fill(sig, subsequent_indent=' ' * 8)
expr_str = str(expr)
if len(expr_str) > 78:
expr_str = textwrap.wrap(expr_str, 75)[0] + '...'
func.__doc__ = ("Created with lambdify. Signature:\n\n"
"{sig}\n\n"
"Expression:\n\n"
"{expr}\n\n"
"Source code:\n\n"
"{src}\n\n"
"Imported modules:\n\n"
"{imp_mods}").format(sig=sig,
expr=expr_str,
src=lstr,
imp_mods='\n'.join(imp_mod_lines))
return func
def lambdify(args,
expr,
modules=None,
printer=None,
use_imps=True,
dummify=False):
"""
Translates a SymPy expression into an equivalent numeric function
For example, to convert the SymPy expression ``sin(x) + cos(x)`` to an
equivalent NumPy function that numerically evaluates it:
>>> from sympy import sin, cos, symbols, lambdify
>>> import numpy as np
>>> x = symbols('x')
>>> expr = sin(x) + cos(x)
>>> expr
sin(x) + cos(x)
>>> f = lambdify(x, expr, 'numpy')
>>> a = np.array([1, 2])
>>> f(a)
[1.38177329 0.49315059]
The primary purpose of this function is to provide a bridge from SymPy
expressions to numerical libraries such as NumPy, SciPy, NumExpr, mpmath,
and tensorflow. In general, SymPy functions do not work with objects from
other libraries, such as NumPy arrays, and functions from numeric
libraries like NumPy or mpmath do not work on SymPy expressions.
``lambdify`` bridges the two by converting a SymPy expression to an
equivalent numeric function.
The basic workflow with ``lambdify`` is to first create a SymPy expression
representing whatever mathematical function you wish to evaluate. This
should be done using only SymPy functions and expressions. Then, use
``lambdify`` to convert this to an equivalent function for numerical
evaluation. For instance, above we created ``expr`` using the SymPy symbol
``x`` and SymPy functions ``sin`` and ``cos``, then converted it to an
equivalent NumPy function ``f``, and called it on a NumPy array ``a``.
.. warning::
This function uses ``exec``, and thus shouldn't be used on unsanitized
input.
Arguments
=========
The first argument of ``lambdify`` is a variable or list of variables in
the expression. Variable lists may be nested. Variables can be Symbols,
undefined functions, or matrix symbols. The order and nesting of the
variables corresponds to the order and nesting of the parameters passed to
the lambdified function. For instance,
>>> from sympy.abc import x, y, z
>>> f = lambdify([x, (y, z)], x + y + z)
>>> f(1, (2, 3))
6
The second argument of ``lambdify`` is the expression, list of
expressions, or matrix to be evaluated. Lists may be nested. If the
expression is a list, the output will also be a list.
>>> f = lambdify(x, [x, [x + 1, x + 2]])
>>> f(1)
[1, [2, 3]]
If it is a matrix, an array will be returned (for the NumPy module).
>>> from sympy import Matrix
>>> f = lambdify(x, Matrix([x, x + 1]))
>>> f(1)
[[1]
[2]]
Note that the argument order here, variables then expression, is used to
emulate the Python ``lambda`` keyword. ``lambdify(x, expr)`` works
(roughly) like ``lambda x: expr`` (see :ref:`lambdify-how-it-works` below).
The third argument, ``modules`` is optional. If not specified, ``modules``
defaults to ``["scipy", "numpy"]`` if SciPy is installed, ``["numpy"]`` if
only NumPy is installed, and ``["math", "mpmath", "sympy"]`` if neither is
installed. That is, SymPy functions are replaced as far as possible by
either ``scipy`` or ``numpy`` functions if available, and Python's
standard library ``math``, or ``mpmath`` functions otherwise.
``modules`` can be one of the following types
- the strings ``"math"``, ``"mpmath"``, ``"numpy"``, ``"numexpr"``,
``"scipy"``, ``"sympy"``, or ``"tensorflow"``. This uses the
corresponding printer and namespace mapping for that module.
- a module (e.g., ``math``). This uses the global namespace of the
module. If the module is one of the above known modules, it will also
use the corresponding printer and namespace mapping (i.e.,
``modules=numpy`` is equivalent to ``modules="numpy"``).
- a dictionary that maps names of SymPy functions to arbitrary functions
(e.g., ``{'sin': custom_sin}``).
- a list that contains a mix of the arguments above, with higher priority
given to entries appearing first (e.g., to use the NumPy module but
override the ``sin`` function with a custom version, you can use
``[{'sin': custom_sin}, 'numpy']``).
The ``dummify`` keyword argument controls whether or not the variables in
the provided expression that are not valid Python identifiers are
substituted with dummy symbols. This allows for undefined functions like
``Function('f')(t)`` to be supplied as arguments. By default, the
variables are only dummified if they are not valid Python identifiers. Set
``dummify=True`` to replace all arguments with dummy symbols (if ``args``
is not a string) - for example, to ensure that the arguments do not
redefine any built-in names.
.. _lambdify-how-it-works:
How it works
============
When using this function, it helps a great deal to have an idea of what it
is doing. At its core, lambdify is nothing more than a namespace
translation, on top of a special printer that makes some corner cases work
properly.
To understand lambdify, first we must properly understand how Python
namespaces work. Say we had two files. One called ``sin_cos_sympy.py``,
with
.. code:: python
# sin_cos_sympy.py
from sympy import sin, cos
def sin_cos(x):
return sin(x) + cos(x)
and one called ``sin_cos_numpy.py`` with
.. code:: python
# sin_cos_numpy.py
from numpy import sin, cos
def sin_cos(x):
return sin(x) + cos(x)
The two files define an identical function ``sin_cos``. However, in the
first file, ``sin`` and ``cos`` are defined as the SymPy ``sin`` and
``cos``. In the second, they are defined as the NumPy versions.
If we were to import the first file and use the ``sin_cos`` function, we
would get something like
>>> from sin_cos_sympy import sin_cos # doctest: +SKIP
>>> sin_cos(1) # doctest: +SKIP
cos(1) + sin(1)
On the other hand, if we imported ``sin_cos`` from the second file, we
would get
>>> from sin_cos_numpy import sin_cos # doctest: +SKIP
>>> sin_cos(1) # doctest: +SKIP
1.38177329068
In the first case we got a symbolic output, because it used the symbolic
``sin`` and ``cos`` functions from SymPy. In the second, we got a numeric
result, because ``sin_cos`` used the numeric ``sin`` and ``cos`` functions
from NumPy. But notice that the versions of ``sin`` and ``cos`` that were
used was not inherent to the ``sin_cos`` function definition. Both
``sin_cos`` definitions are exactly the same. Rather, it was based on the
names defined at the module where the ``sin_cos`` function was defined.
The key point here is that when function in Python references a name that
is not defined in the function, that name is looked up in the "global"
namespace of the module where that function is defined.
Now, in Python, we can emulate this behavior without actually writing a
file to disk using the ``exec`` function. ``exec`` takes a string
containing a block of Python code, and a dictionary that should contain
the global variables of the module. It then executes the code "in" that
dictionary, as if it were the module globals. The following is equivalent
to the ``sin_cos`` defined in ``sin_cos_sympy.py``:
>>> import sympy
>>> module_dictionary = {'sin': sympy.sin, 'cos': sympy.cos}
>>> exec('''
... def sin_cos(x):
... return sin(x) + cos(x)
... ''', module_dictionary)
>>> sin_cos = module_dictionary['sin_cos']
>>> sin_cos(1)
cos(1) + sin(1)
and similarly with ``sin_cos_numpy``:
>>> import numpy
>>> module_dictionary = {'sin': numpy.sin, 'cos': numpy.cos}
>>> exec('''
... def sin_cos(x):
... return sin(x) + cos(x)
... ''', module_dictionary)
>>> sin_cos = module_dictionary['sin_cos']
>>> sin_cos(1)
1.38177329068
So now we can get an idea of how ``lambdify`` works. The name "lambdify"
comes from the fact that we can think of something like ``lambdify(x,
sin(x) + cos(x), 'numpy')`` as ``lambda x: sin(x) + cos(x)``, where
``sin`` and ``cos`` come from the ``numpy`` namespace. This is also why
the symbols argument is first in ``lambdify``, as opposed to most SymPy
functions where it comes after the expression: to better mimic the
``lambda`` keyword.
``lambdify`` takes the input expression (like ``sin(x) + cos(x)``) and
1. Converts it to a string
2. Creates a module globals dictionary based on the modules that are
passed in (by default, it uses the NumPy module)
3. Creates the string ``"def func({vars}): return {expr}"``, where ``{vars}`` is the
list of variables separated by commas, and ``{expr}`` is the string
created in step 1., then ``exec``s that string with the module globals
namespace and returns ``func``.
In fact, functions returned by ``lambdify`` support inspection. So you can
see exactly how they are defined by using ``inspect.getsource``, or ``??`` if you
are using IPython or the Jupyter notebook.
>>> f = lambdify(x, sin(x) + cos(x))
>>> import inspect
>>> print(inspect.getsource(f))
def _lambdifygenerated(x):
return (sin(x) + cos(x))
This shows us the source code of the function, but not the namespace it
was defined in. We can inspect that by looking at the ``__globals__``
attribute of ``f``:
>>> f.__globals__['sin']
<ufunc 'sin'>
>>> f.__globals__['cos']
<ufunc 'cos'>
>>> f.__globals__['sin'] is numpy.sin
True
This shows us that ``sin`` and ``cos`` in the namespace of ``f`` will be
``numpy.sin`` and ``numpy.cos``.
Note that there are some convenience layers in each of these steps, but at
the core, this is how ``lambdify`` works. Step 1 is done using the
``LambdaPrinter`` printers defined in the printing module (see
:mod:`sympy.printing.lambdarepr`). This allows different SymPy expressions
to define how they should be converted to a string for different modules.
You can change which printer ``lambdify`` uses by passing a custom printer
in to the ``printer`` argument.
Step 2 is augmented by certain translations. There are default
translations for each module, but you can provide your own by passing a
list to the ``modules`` argument. For instance,
>>> def mysin(x):
... print('taking the sin of', x)
... return numpy.sin(x)
...
>>> f = lambdify(x, sin(x), [{'sin': mysin}, 'numpy'])
>>> f(1)
taking the sin of 1
0.8414709848078965
The globals dictionary is generated from the list by merging the
dictionary ``{'sin': mysin}`` and the module dictionary for NumPy. The
merging is done so that earlier items take precedence, which is why
``mysin`` is used above instead of ``numpy.sin``.
If you want to modify the way ``lambdify`` works for a given function, it
is usually easiest to do so by modifying the globals dictionary as such.
In more complicated cases, it may be necessary to create and pass in a
custom printer.
Finally, step 3 is augmented with certain convenience operations, such as
the addition of a docstring.
Understanding how ``lambdify`` works can make it easier to avoid certain
gotchas when using it. For instance, a common mistake is to create a
lambdified function for one module (say, NumPy), and pass it objects from
another (say, a SymPy expression).
For instance, say we create
>>> from sympy.abc import x
>>> f = lambdify(x, x + 1, 'numpy')
Now if we pass in a NumPy array, we get that array plus 1
>>> import numpy
>>> a = numpy.array([1, 2])
>>> f(a)
[2 3]
But what happens if you make the mistake of passing in a SymPy expression
instead of a NumPy array:
>>> f(x + 1)
x + 2
This worked, but it was only by accident. Now take a different lambdified
function:
>>> from sympy import sin
>>> g = lambdify(x, x + sin(x), 'numpy')
This works as expected on NumPy arrays:
>>> g(a)
[1.84147098 2.90929743]
But if we try to pass in a SymPy expression, it fails
>>> try:
... g(x + 1)
... # NumPy release after 1.17 raises TypeError instead of
... # AttributeError
... except (AttributeError, TypeError):
... raise AttributeError() # doctest: +IGNORE_EXCEPTION_DETAIL
Traceback (most recent call last):
...
AttributeError:
Now, let's look at what happened. The reason this fails is that ``g``
calls ``numpy.sin`` on the input expression, and ``numpy.sin`` does not
know how to operate on a SymPy object. **As a general rule, NumPy
functions do not know how to operate on SymPy expressions, and SymPy
functions do not know how to operate on NumPy arrays. This is why lambdify
exists: to provide a bridge between SymPy and NumPy.**
However, why is it that ``f`` did work? That's because ``f`` doesn't call
any functions, it only adds 1. So the resulting function that is created,
``def _lambdifygenerated(x): return x + 1`` does not depend on the globals
namespace it is defined in. Thus it works, but only by accident. A future
version of ``lambdify`` may remove this behavior.
Be aware that certain implementation details described here may change in
future versions of SymPy. The API of passing in custom modules and
printers will not change, but the details of how a lambda function is
created may change. However, the basic idea will remain the same, and
understanding it will be helpful to understanding the behavior of
lambdify.
**In general: you should create lambdified functions for one module (say,
NumPy), and only pass it input types that are compatible with that module
(say, NumPy arrays).** Remember that by default, if the ``module``
argument is not provided, ``lambdify`` creates functions using the NumPy
and SciPy namespaces.
Examples
========
>>> from sympy.utilities.lambdify import implemented_function
>>> from sympy import sqrt, sin, Matrix
>>> from sympy import Function
>>> from sympy.abc import w, x, y, z
>>> f = lambdify(x, x**2)
>>> f(2)
4
>>> f = lambdify((x, y, z), [z, y, x])
>>> f(1,2,3)
[3, 2, 1]
>>> f = lambdify(x, sqrt(x))
>>> f(4)
2.0
>>> f = lambdify((x, y), sin(x*y)**2)
>>> f(0, 5)
0.0
>>> row = lambdify((x, y), Matrix((x, x + y)).T, modules='sympy')
>>> row(1, 2)
Matrix([[1, 3]])
``lambdify`` can be used to translate SymPy expressions into mpmath
functions. This may be preferable to using ``evalf`` (which uses mpmath on
the backend) in some cases.
>>> import mpmath
>>> f = lambdify(x, sin(x), 'mpmath')
>>> f(1)
0.8414709848078965
Tuple arguments are handled and the lambdified function should
be called with the same type of arguments as were used to create
the function:
>>> f = lambdify((x, (y, z)), x + y)
>>> f(1, (2, 4))
3
The ``flatten`` function can be used to always work with flattened
arguments:
>>> from sympy.utilities.iterables import flatten
>>> args = w, (x, (y, z))
>>> vals = 1, (2, (3, 4))
>>> f = lambdify(flatten(args), w + x + y + z)
>>> f(*flatten(vals))
10
Functions present in ``expr`` can also carry their own numerical
implementations, in a callable attached to the ``_imp_`` attribute. This
can be used with undefined functions using the ``implemented_function``
factory:
>>> f = implemented_function(Function('f'), lambda x: x+1)
>>> func = lambdify(x, f(x))
>>> func(4)
5
``lambdify`` always prefers ``_imp_`` implementations to implementations
in other namespaces, unless the ``use_imps`` input parameter is False.
Usage with Tensorflow:
>>> import tensorflow as tf
>>> from sympy import Max, sin, lambdify
>>> from sympy.abc import x
>>> f = Max(x, sin(x))
>>> func = lambdify(x, f, 'tensorflow')
After tensorflow v2, eager execution is enabled by default.
If you want to get the compatible result across tensorflow v1 and v2
as same as this tutorial, run this line.
>>> tf.compat.v1.enable_eager_execution()
If you have eager execution enabled, you can get the result out
immediately as you can use numpy.
If you pass tensorflow objects, you may get an ``EagerTensor``
object instead of value.
>>> result = func(tf.constant(1.0))
>>> print(result)
tf.Tensor(1.0, shape=(), dtype=float32)
>>> print(result.__class__)
<class 'tensorflow.python.framework.ops.EagerTensor'>
You can use ``.numpy()`` to get the numpy value of the tensor.
>>> result.numpy()
1.0
>>> var = tf.Variable(2.0)
>>> result = func(var) # also works for tf.Variable and tf.Placeholder
>>> result.numpy()
2.0
And it works with any shape array.
>>> tensor = tf.constant([[1.0, 2.0], [3.0, 4.0]])
>>> result = func(tensor)
>>> result.numpy()
[[1. 2.]
[3. 4.]]
Notes
=====
- For functions involving large array calculations, numexpr can provide a
significant speedup over numpy. Please note that the available functions
for numexpr are more limited than numpy but can be expanded with
``implemented_function`` and user defined subclasses of Function. If
specified, numexpr may be the only option in modules. The official list
of numexpr functions can be found at:
https://numexpr.readthedocs.io/en/latest/user_guide.html#supported-functions
- In previous versions of SymPy, ``lambdify`` replaced ``Matrix`` with
``numpy.matrix`` by default. As of SymPy 1.0 ``numpy.array`` is the
default. To get the old default behavior you must pass in
``[{'ImmutableDenseMatrix': numpy.matrix}, 'numpy']`` to the
``modules`` kwarg.
>>> from sympy import lambdify, Matrix
>>> from sympy.abc import x, y
>>> import numpy
>>> array2mat = [{'ImmutableDenseMatrix': numpy.matrix}, 'numpy']
>>> f = lambdify((x, y), Matrix([x, y]), modules=array2mat)
>>> f(1, 2)
[[1]
[2]]
- In the above examples, the generated functions can accept scalar
values or numpy arrays as arguments. However, in some cases
the generated function relies on the input being a numpy array:
>>> from sympy import Piecewise
>>> from sympy.utilities.pytest import ignore_warnings
>>> f = lambdify(x, Piecewise((x, x <= 1), (1/x, x > 1)), "numpy")
>>> with ignore_warnings(RuntimeWarning):
... f(numpy.array([-1, 0, 1, 2]))
[-1. 0. 1. 0.5]
>>> f(0)
Traceback (most recent call last):
...
ZeroDivisionError: division by zero
In such cases, the input should be wrapped in a numpy array:
>>> with ignore_warnings(RuntimeWarning):
... float(f(numpy.array([0])))
0.0
Or if numpy functionality is not required another module can be used:
>>> f = lambdify(x, Piecewise((x, x <= 1), (1/x, x > 1)), "math")
>>> f(0)
0
"""
from sympy.core.symbol import Symbol
# If the user hasn't specified any modules, use what is available.
if modules is None:
try:
_import("scipy")
except ImportError:
try:
_import("numpy")
except ImportError:
# Use either numpy (if available) or python.math where possible.
# XXX: This leads to different behaviour on different systems and
# might be the reason for irreproducible errors.
modules = ["math", "mpmath", "sympy"]
else:
modules = ["numpy"]
else:
modules = ["numpy", "scipy"]
# Get the needed namespaces.
namespaces = []
# First find any function implementations
if use_imps:
namespaces.append(_imp_namespace(expr))
# Check for dict before iterating
if isinstance(modules,
(dict, string_types)) or not hasattr(modules, '__iter__'):
namespaces.append(modules)
else:
# consistency check
if _module_present('numexpr', modules) and len(modules) > 1:
raise TypeError("numexpr must be the only item in 'modules'")
namespaces += list(modules)
# fill namespace with first having highest priority
namespace = {}
for m in namespaces[::-1]:
buf = _get_namespace(m)
namespace.update(buf)
if hasattr(expr, "atoms"):
#Try if you can extract symbols from the expression.
#Move on if expr.atoms in not implemented.
syms = expr.atoms(Symbol)
for term in syms:
namespace.update({str(term): term})
if printer is None:
if _module_present('mpmath', namespaces):
from sympy.printing.pycode import MpmathPrinter as Printer
elif _module_present('scipy', namespaces):
from sympy.printing.pycode import SciPyPrinter as Printer
elif _module_present('numpy', namespaces):
from sympy.printing.pycode import NumPyPrinter as Printer
elif _module_present('numexpr', namespaces):
from sympy.printing.lambdarepr import NumExprPrinter as Printer
elif _module_present('tensorflow', namespaces):
from sympy.printing.tensorflow import TensorflowPrinter as Printer
elif _module_present('sympy', namespaces):
from sympy.printing.pycode import SymPyPrinter as Printer
else:
from sympy.printing.pycode import PythonCodePrinter as Printer
user_functions = {}
for m in namespaces[::-1]:
if isinstance(m, dict):
for k in m:
user_functions[k] = k
printer = Printer({
'fully_qualified_modules': False,
'inline': True,
'allow_unknown_functions': True,
'user_functions': user_functions
})
# Get the names of the args, for creating a docstring
if not iterable(args):
args = (args, )
names = []
# Grab the callers frame, for getting the names by inspection (if needed)
callers_local_vars = inspect.currentframe().f_back.f_locals.items()
for n, var in enumerate(args):
if hasattr(var, 'name'):
names.append(var.name)
else:
# It's an iterable. Try to get name by inspection of calling frame.
name_list = [
var_name for var_name, var_val in callers_local_vars
if var_val is var
]
if len(name_list) == 1:
names.append(name_list[0])
else:
# Cannot infer name with certainty. arg_# will have to do.
names.append('arg_' + str(n))
# Create the function definition code and execute it
funcname = '_lambdifygenerated'
if _module_present('tensorflow', namespaces):
funcprinter = _TensorflowEvaluatorPrinter(printer, dummify)
else:
funcprinter = _EvaluatorPrinter(printer, dummify)
funcstr = funcprinter.doprint(funcname, args, expr)
# Collect the module imports from the code printers.
imp_mod_lines = []
for mod, keys in (getattr(printer, 'module_imports', None) or {}).items():
for k in keys:
if k not in namespace:
ln = "from %s import %s" % (mod, k)
try:
exec_(ln, {}, namespace)
except ImportError:
# Tensorflow 2.0 has issues with importing a specific
# function from its submodule.
# https://github.com/tensorflow/tensorflow/issues/33022
ln = "%s = %s.%s" % (k, mod, k)
exec_(ln, {}, namespace)
imp_mod_lines.append(ln)
# Provide lambda expression with builtins, and compatible implementation of range
namespace.update({'builtins': builtins, 'range': range})
funclocals = {}
global _lambdify_generated_counter
filename = '<lambdifygenerated-%s>' % _lambdify_generated_counter
_lambdify_generated_counter += 1
c = compile(funcstr, filename, 'exec')
exec_(c, namespace, funclocals)
# mtime has to be None or else linecache.checkcache will remove it
linecache.cache[filename] = (len(funcstr), None, funcstr.splitlines(True),
filename)
func = funclocals[funcname]
# Apply the docstring
sig = "func({0})".format(", ".join(str(i) for i in names))
sig = textwrap.fill(sig, subsequent_indent=' ' * 8)
expr_str = str(expr)
if len(expr_str) > 78:
expr_str = textwrap.wrap(expr_str, 75)[0] + '...'
func.__doc__ = ("Created with lambdify. Signature:\n\n"
"{sig}\n\n"
"Expression:\n\n"
"{expr}\n\n"
"Source code:\n\n"
"{src}\n\n"
"Imported modules:\n\n"
"{imp_mods}").format(sig=sig,
expr=expr_str,
src=funcstr,
imp_mods='\n'.join(imp_mod_lines))
return func