dummy-link

PyCall

Package to call Python functions from the Julia language

Readme

Calling Python functions from the Julia language

Build Status Build status Coverage Status

PyCall PyCall PyCall

This package provides the ability to directly call and fully interoperate with Python from the Julia language. You can import arbitrary Python modules from Julia, call Python functions (with automatic conversion of types between Julia and Python), define Python classes from Julia methods, and share large data structures between Julia and Python without copying them.

Installation

Within Julia, just use the package manager to run Pkg.add("PyCall") to install the files. Julia 0.5 or later is required.

The latest development version of PyCall is available from https://github.com/stevengj/PyCall.jl. If you want to switch to this after installing the package, run Pkg.checkout("PyCall"); Pkg.build("PyCall").

By default on Mac and Windows systems, Pkg.add("PyCall") or Pkg.build("PyCall") will use the Conda.jl package to install a minimal Python distribution (via Miniconda) that is private to Julia (not in your PATH). You can use the Conda Julia package to install more Python packages, and import Conda to print the Conda.PYTHONDIR directory where python was installed. On GNU/Linux systems, PyCall will default to using the python program (if any) in your PATH.

The advantage of a Conda-based configuration is particularly compelling if you are installing PyCall in order to use packages like PyPlot.jl or SymPy.jl, as these can then automatically install their Python dependencies. (To exploit this in your own packages, use the pyimport_conda function described below.)

Specifying the Python version

If you want to use a different version of Python than the default, you can change the Python version by setting the PYTHON environment variable to the path of the python executable and then re-running Pkg.build("PyCall"). In Julia:

ENV["PYTHON"] = "... path of the python program you want ..."
Pkg.build("PyCall")

Note also that you will need to re-run Pkg.build("PyCall") if your python program changes significantly (e.g. you switch to a new Python distro, or you switch from Python 2 to Python 3).

To force Julia to use its own Python distribution, via Conda, simply set ENV["PYTHON"] to the empty string "" and re-run Pkg.build("PyCall").

The current Python version being used is stored in the pyversion global variable of the PyCall module. You can also look at PyCall.libpython to find the name of the Python library or PyCall.pyprogramname for the python program name. If it is using the Conda Python, PyCall.conda will be true.

(Technically, PyCall does not use the python program per se: it links directly to the libpython library. But it finds the location of libpython by running python during Pkg.build.)

Subsequent builds of PyCall (e.g. when you update the package via Pkg.update) will use the same Python executable name by default, unless you set the PYTHON environment variable or delete the file Pkg.dir("PyCall","deps","PYTHON").

Note: If you use Python virtualenvs, then be aware that PyCall uses the virtualenv it was built with, even if you switch virtualenvs. If you want to switch PyCall to use a different virtualenv, then you should switch virtualenvs and run rm(Pkg.dir("PyCall","deps","PYTHON")); Pkg.build("PyCall").

Note: Usually, the necessary libraries are installed along with Python, but pyenv on MacOS requires you to install it with env PYTHON_CONFIGURE_OPTS="--enable-framework" pyenv install 3.4.3. The Enthought Canopy Python distribution is currently not supported. As a general rule, we tend to recommend the Anaconda Python distribution on MacOS and Windows, or using the Julia Conda package, in order to minimize headaches.

Usage

Here is a simple example to call Python's math.sin function and compare it to the built-in Julia sin:

using PyCall
@pyimport math
math.sin(math.pi / 4) - sin(pi / 4)  # returns 0.0

Type conversions are automatically performed for numeric, boolean, string, IO stream, date/period, and function types, along with tuples, arrays/lists, and dictionaries of these types. (Python functions can be converted/passed to Julia functions and vice versa!) Other types are supported via the generic PyObject type, below.

Python submodules must be imported by a separate @pyimport call, and in this case you must supply an identifier to to use in Julia. For example

@pyimport numpy.random as nr
nr.rand(3,4)

Multidimensional arrays exploit the NumPy array interface for conversions between Python and Julia. By default, they are passed from Julia to Python without making a copy, but from Python to Julia a copy is made; no-copy conversion of Python to Julia arrays can be achieved with the PyArray type below.

Keyword arguments can also be passed. For example, matplotlib's pyplot uses keyword arguments to specify plot options, and this functionality is accessed from Julia by:

@pyimport matplotlib.pyplot as plt
x = linspace(0,2*pi,1000); y = sin(3*x + 4*cos(2*x));
plt.plot(x, y, color="red", linewidth=2.0, linestyle="--")
plt.show()

However, for better integration with Julia graphics backends and to avoid the need for the show function, we recommend using matplotlib via the Julia PyPlot module.

Arbitrary Julia functions can be passed to Python routines taking function arguments. For example, to find the root of cos(x) - x, we could call the Newton solver in scipy.optimize via:

@pyimport scipy.optimize as so
so.newton(x -> cos(x) - x, 1)

A macro exists for mimicking Python's "with statement". For example:

@pywith pybuiltin("open")("file.txt","w") as f begin
    f[:write]("hello")
end

The type of f can be specified with f::T (for example, to override automatic conversion, use f::PyObject). Similarly, if the context manager returns a type which is automatically converted to a Julia type, you will have override this via @pywith EXPR::PyObject ....

Important: The biggest difference from Python is that object attributes/members are accessed with o[:attribute] rather than o.attribute, so that o.method(...) in Python is replaced by o[:method](...) in Julia. Also, you use get(o, key) rather than o[key]. (However, you can access integer indices via o[i] as in Python, albeit with 1-based Julian indices rather than 0-based Python indices.) (This is because Julia does not permit overloading the . operator yet.) See also the section on PyObject below, as well as the pywrap function to create anonymous modules that simulate . access (this is what @pyimport does). For example, using Biopython we can do:

@pyimport Bio.Seq as s
@pyimport Bio.Alphabet as a
my_dna = s.Seq("AGTACACTGGT", a.generic_dna)
my_dna[:find]("ACT")

whereas in Python the last step would have been my_dna.find("ACT").

Troubleshooting

Here are solutions to some common problems:

  • As mentioned above, use foo[:bar] and foo[:bar](...) rather than foo.bar and foo.bar(...), respectively, to access attributes and methods of Python objects.

  • By default, PyCall doesn't include the current directory in the Python search path. If you want to do that (in order to load a Python module from the current directory), just run unshift!(PyVector(pyimport("sys")["path"]), "").

Python object interfaces

The @pyimport macro is built on top of several routines for manipulating Python objects in Julia, via a type PyObject described below. These can be used to have greater control over the types and data passed between Julia and Python, as well as to access additional Python functionality (especially in conjunction with the low-level interfaces described later).

Types

PyObject

The PyCall module also provides a new type PyObject (a wrapper around PyObject* in Python's C API) representing a reference to a Python object.

Constructors PyObject(o) are provided for a number of Julia types, and PyCall also supplies convert(T, o::PyObject) to convert PyObjects back into Julia types T. Currently, the types supported are numbers (integer, real, and complex), booleans, strings, IO streams, dates/periods, and functions, along with tuples and arrays/lists thereof, but more are planned. (Julia symbols are converted to Python strings.)

Given o::PyObject, o[:attribute] is equivalent to o.attribute in Python, with automatic type conversion. To get an attribute as a PyObject without type conversion, do o["attribute"] instead. The keys(o::PyObject) function returns an array of the available attribute symbols.

Given o::PyObject, get(o, key) is equivalent to o[key] in Python, with automatic type conversion. To get as a PyObject without type conversion, do get(o, PyObject, key), or more generally get(o, SomeType, key). You can also supply a default value to use if the key is not found by get(o, key, default) or get(o, SomeType, key, default). Similarly, set!(o, key, val) is equivalent to o[key] = val in Python, and delete!(o, key) is equivalent to del o[key] in Python. For one or more integer indices, o[i] in Julia is equivalent to o[i-1] in Python.

You can call an o::PyObject via o(args...) just like in Python (assuming that the object is callable in Python). The explicit pycall form is still useful in Julia if you want to specify the return type.

Arrays and PyArray

From Julia to Python

Assuming you have NumPy installed (true by default if you use Conda), then a Julia a::Array of NumPy-compatible elements is converted by PyObject(a) into a NumPy wrapper for the same data, i.e. without copying the data. Julia arrays are stored in column-major order, and since NumPy supports column-major arrays this is not a problem.

However, the default ordering of NumPy arrays created in Python is row-major, and some Python packages will throw an error if you try to pass them column-major NumPy arrays. To deal with this, you can use PyReverseDims(a) to pass a Julia array as a row-major NumPy array with the dimensions reversed. For example, if a is a 3x4x5 Julia array, then PyReverseDims(a) passes it as a 5x4x3 NumPy row-major array (without making a copy of the underlying data).

A Vector{UInt8} object in Julia, by default, is converted to a Python bytearray object. If you want a bytes object instead, you can use the function pybytes(a).

From Python to Julia

Multidimensional NumPy arrays (ndarray) are supported and can be converted to the native Julia Array type, which makes a copy of the data.

Alternatively, the PyCall module also provides a new type PyArray (a subclass of AbstractArray) which implements a no-copy wrapper around a NumPy array (currently of numeric types or objects only). Just use PyArray as the return type of a pycall returning an ndarray, or call PyArray(o::PyObject) on an ndarray object o. (Technically, a PyArray works for any Python object that uses the NumPy array interface to provide a data pointer and shape information.)

Conversely, when passing arrays to Python, Julia Array types are converted to PyObject types without making a copy via NumPy, e.g. when passed as pycall arguments.

PyVector

The PyCall module provides a new type PyVector (a subclass of AbstractVector) which implements a no-copy wrapper around an arbitrary Python list or sequence object. (Unlike PyArray, the PyVector type is not limited to NumPy arrays, although using PyArray for the latter is generally more efficient.) Just use PyArray as the return type of a pycall returning a list or sequence object (including tuples), or call PyVector(o::PyObject) on a sequence object o.

A v::PyVector supports the usual v[index] referencing and assignment, along with delete! and pop! operations. copy(v) converts v to an ordinary Julia Vector.

PyDict

Similar to PyVector, PyCall also provides a type PyDict (a subclass of Association) that implements a no-copy wrapper around a Python dictionary (or any object implementing the mapping protocol). Just use PyDict as the return type of a pycall returning a dictionary, or call PyDict(o::PyObject) on a dictionary object o. By default, a PyDict is an Any => Any dictionary (or actually PyAny => PyAny) that performs runtime type inference, but if your Python dictionary has known, fixed types you can insteady use PyDict{K,V} given the key and value types K and V respectively.

Currently, passing Julia dictionaries to Python makes a copy of the Julia dictionary.

PyTextIO

Julia IO streams are converted into Python objects implementing the RawIOBase interface, so they can be used for binary I/O in Python. However, some Python code (notably unpickling) expects a stream implementing the TextIOBase interface, which differs from RawIOBase mainly in that read an readall functions return strings rather than byte arrays. If you need to pass an IO stream as a text-IO object, just call PyTextIO(io::IO) to convert it.

(There doesn't seem to be any good way to determine automatically whether Python wants a stream argument to return strings or binary data. Also, unlike Python, Julia does not open files separately in "text" or "binary" modes, so we cannot determine the conversion simply from how the file was opened.)

PyAny

The PyAny type is used in conversions to tell PyCall to detect the Python type at runtime and convert to the corresponding native Julia type. That is, pycall(func, PyAny, ...) and convert(PyAny, o::PyObject) both automatically convert their result to a native Julia type (if possible). This is convenient, but will lead to slightly worse performance (due to the overhead of runtime type-checking and the fact that the Julia JIT compiler can no longer infer the type).

Calling Python

In most cases, the @pyimport macro automatically makes the appropriate type conversions to Julia types based on runtime inspection of the Python objects. However greater control over these type conversions (e.g. to use a no-copy PyArray for a Python multidimensional array rather than copying to an Array) can be achieved by using the lower-level functions below. Using pycall in cases where the Python return type is known can also improve performance, both by eliminating the overhead of runtime type inference and also by providing more type information to the Julia compiler.

  • pycall(function::PyObject, returntype::Type, args...). Call the given Python function (typically looked up from a module) with the given args... (of standard Julia types which are converted automatically to the corresponding Python types if possible), converting the return value to returntype (use a returntype of PyObject to return the unconverted Python object reference, or of PyAny to request an automated conversion). For convenience, a macro @pycall exists which automatically converts @pycall function(args...)::returntype into pycall(function,returntype,args...).

  • pyimport(s): Import the Python module s (a string or symbol) and return a pointer to it (a PyObject). Functions or other symbols in the module may then be looked up by s[name] where name is a string (for the raw PyObject) or symbol (for automatic type-conversion). Unlike the @pyimport macro, this does not define a Julia module and members cannot be accessed with s.name.

  • py"..." evaluates "..." as a Python string and returns the result converted to PyAny. Alternatively, py"..."o returns the raw PyObject (which can then be manually converted if desired). You can interpolate Julia variables and other expressions into the Python code with $, which interpolates the value (converted to PyObject) of the given expression---data is not passed as a string, so this is different from ordinary Julia string interpolation. e.g. py"sum($([1,2,3]))" calls the Python sum function on the Julia array [1,2,3], returning 6. In contrast, if you use $$ before the interpolated expression, then the value of the expression is inserted as a string into the Python code, allowing you to generate Python code itself via Julia expressions. For example, if x="1+1" in Julia, then py"$x" returns the string "1+1", but py"$$x" returns 2. If you use py"""...""" to pass a multi-line string, the string can contain arbitrary Python code (not just a single expression) to be evaluated, but the return value is nothing; this is useful e.g. to define pure-Python functions. (If you define a Python global g in a multiline py"""...""" string, you can retrieve it in Julia by subsequently evaluating py"g".)

  • pybuiltin(s): Look up s (a string or symbol) among the global Python builtins. If s is a string it returns a PyObject, while if s is a symbol it returns the builtin converted to PyAny. (You can also use py"s" to look up builtins or other Python globas.)

  • pywrap(o::PyObject) returns a wrapper w that is an anonymous module which provides (read) access to converted versions of o's members as w.member. (For example, @pyimport module as name is equivalent to name = pywrap(pyimport("module")).) If the Python module contains identifiers that are reserved words in Julia (e.g. function), they cannot be accessed as w.member; one must instead use w.pymember(:member) (for the PyAny conversion) or w.pymember("member") (for the raw PyObject). pywrap is rather inefficient since it converts every member of o at once; you are generally encouraged to simply access members via o[:member] rather than using pywrap.

Occasionally, you may need to pass a keyword argument to Python that is a reserved word in Julia. For example, calling f(x, function=g) will fail because function is a reserved word in Julia. In such cases, you can use the lower-level Julia syntax f(x; :function=>g).

Defining Python Classes

@pydef creates a Python class whose methods are implemented in Julia. For instance,

@pyimport numpy.polynomial as P
@pydef type Doubler <: P.Polynomial
    __init__(self, x=10) = (self[:x] = x)
    my_method(self, arg1::Number) = arg1 + 20
    x2.get(self) = self[:x] * 2
    x2.set!(self, new_val) = (self[:x] = new_val / 2)
end
Doubler()[:x2]

is essentially equivalent to the following Python code:

import numpy.polynomial
class Doubler(numpy.polynomial.Polynomial):
    def __init__(self, x=10):
        self.x = x
    def my_method(self, arg1): return arg1 + 20
    @property
    def x2(self): return self.x * 2
    @x2.setter
    def x2(self, new_val):
        self.x = new_val / 2
Doubler().x2

The method arguments and return values are automatically converted between Julia and Python. All Python special methods are supported (__len__, __add__, etc.).

@pydef allows for multiple inheritance of Python classes:

@pydef type SomeType <: (BaseClass1, BaseClass2)
    ...
end

Here's another example using Tkinter:

using PyCall
@pyimport Tkinter as tk

@pydef type SampleApp <: tk.Tk
    __init__(self, args...; kwargs...) = begin
        tk.Tk[:__init__](self, args...; kwargs...)
        self[:label] = tk.Label(text="Hello, world!")
        self[:label][:pack](padx=10, pady=10)
    end
end

app = SampleApp()
app[:mainloop]()

GUI Event Loops

For Python packages that have a graphical user interface (GUI), notably plotting packages like matplotlib (or MayaVi or Chaco), it is convenient to start the GUI event loop (which processes things like mouse clicks) as an asynchronous task within Julia, so that the GUI is responsive without blocking Julia's input prompt. PyCall includes functions to implement these event loops for some of the most common cross-platform GUI toolkits: wxWidgets, GTK+ version 2 (via PyGTK) or version 3 (via PyGObject), and Qt (via the PyQt4 or PySide Python modules).

You can set a GUI event loop via:

  • pygui_start(gui::Symbol=pygui()). Here, gui is either :wx, :gtk, :gtk3, :tk, or :qt to start the respective toolkit's event loop. (:qt will use PyQt4 or PySide, preferring the former; if you need to require one or the other you can instead use :qt_pyqt4 or :qt_pyside, respectively.) It defaults to the return value of pygui(), which returns a current default GUI (see below). Passing a gui argument also changes the default GUI, equivalent to calling pygui(gui) below. You may start event loops for more than one GUI toolkit (to run simultaneously). Calling pygui_start more than once for a given toolkit does nothing (except to change the current pygui default).

  • pygui(): return the current default GUI toolkit (Symbol). If the default GUI has not been set already, this is the first of :tk, :qt, :wx, :gtk, or :gtk3 for which the corresponding Python package is installed. pygui(gui::Symbol) changes the default GUI to gui.

  • pygui_stop(gui::Symbol=pygui()): Stop any running event loop for gui (which defaults to the current return value of pygui). Returns true if an event loop was running, and false otherwise.

To use these GUI facilities with some Python libraries, it is enough to simply start the appropriate toolkit's event-loop before importing the library. However, in other cases it is necessary to explicitly tell the library which GUI toolkit to use and that an interactive mode is desired. To make this even easier, it is convenient to have wrapper modules around popular Python libraries, such as the PyPlot module for Julia.

Low-level Python API access

If you want to call low-level functions in the Python C API, you can do so using ccall.

  • Use @pysym(func::Symbol) to get a function pointer to pass to ccall given a symbol func in the Python API. e.g. you can call int Py_IsInitialized() by ccall(@pysym(:Py_IsInitialized), Int32, ()).

  • PyCall defines the typealias PyPtr for PythonObject* argument types, and PythonObject (see above) arguments are correctly converted to this type. PythonObject(p::PyPtr) creates a Julia wrapper around a PyPtr return value.

  • Use PythonObject and the convert routines mentioned above to convert Julia types to/from PythonObject* references.

  • If a new reference is returned by a Python function, immediately convert the PyPtr return values to PythonObject objects in order to have their Python reference counts decremented when the object is garbage collected in Julia. i.e. PythonObject(ccall(func, PyPtr, ...)). Important: for Python routines that return a borrowed reference, you should instead do pyincref(PyObject(...)) to obtain a new reference.

  • You can call pyincref(o::PyObject) and pydecref(o::PyObject) to manually increment/decrement the reference count. This is sometimes needed when low-level functions steal a reference or return a borrowed one.

  • The function pyerr_check(msg::AbstractString) can be used to check if a Python exception was thrown, and throw a Julia exception (which includes both msg and the Python exception object) if so. The Python exception status may be cleared by calling pyerr_clear().

  • The function pytype_query(o::PyObject) returns a native Julia type that o can be converted into, if possible, or PyObject if not.

  • pyisinstance(o::PyObject, t::Symbol) can be used to query whether o is of a given Python type (where t is the identifier of a global PyTypeObject in the Python C API), e.g. pyisinstance(o, :PyDict_Type) checks whether o is a Python dictionary. Alternatively, pyisinstance(o::PyObject, t::PyObject) performs the same check given a Python type object t. pytypeof(o::PyObject) returns the Python type of o, equivalent to type(o) in Python.

Using PyCall from Julia Modules

You can use PyCall from any Julia code, including within Julia modules. However, some care is required when using PyCall from precompiled Julia modules. The key thing to remember is that all Python objects (any PyObject) contain pointers to memory allocated by the Python runtime, and such pointers cannot be saved in precompiled constants. (When a precompiled library is reloaded, these pointers will not contain valid memory addresses.)

The solution is fairly simple:

  • Python objects that you create in functions called after the module is loaded are always safe.

  • If you want to store a Python object in a global variable that is initialized automatically when the module is loaded, then initialize the variable in your module's __init__ function. For a type-stable global constant, initialize the constant to PyNULL() at the top level, and then use the copy! function in your module's __init__ function to mutate it to its actual value.

For example, suppose your module uses the scipy.optimize module, and you want to load this module when your module is loaded and store it in a global constant scipy_opt. You could do:

__precompile__() # this module is safe to precompile
module MyModule
using PyCall

const scipy_opt = PyNULL()

function __init__()
    copy!(scipy_opt, pyimport_conda("scipy.optimize", "scipy"))
end

end

Then you can access the scipy.optimize functions as scipy_opt[:newton] and so on.

Here, instead of pyimport, we have used the function pyimport_conda. The second argument is the name of the Anaconda package that provides this module. This way, if importing scipy.optimize fails because the user hasn't installed scipy, it will either (a) automatically install scipy and retry the pyimport if PyCall is configured to use the Conda Python install (or any other Anaconda-based Python distro for which the user has installation privileges), or (b) throw an error explaining that scipy needs to be installed, and explain how to configure PyCall to use Conda so that it can be installed automatically. More generally, you can call pyimport(module, package, channel) to specify an optional Anaconda "channel" for installing non-standard Anaconda packages.

(Note that you cannot use @pyimport safely with precompilation, because that declares a global constant that internally has a pointer to the module. You can use pywrap(pyimport(...)) in your __init__ function to a assign a global variable using the . notation like @pyimport, however, albeit without the type stability of the global const as above.)

Author

This package was written by Steven G. Johnson.

First Commit

02/15/2013

Last Touched

2 days ago

Commits

565 commits