libpython-clj

JNA libpython bindings to the tech ecosystem.

• Bridge between JVM objects and Python objects easily; use Python in your Java and use some Java in your Python.
• Python objects are linked to the JVM GC such that when they are no longer reachable from the JVM their references are released.
• The exact same binary can run top of on multiple version of python reducing version dependency chain management issues.
• Development of new functionality is faster because it can be done from purely from the REPL.

We have a video up of a scicloj community discussion with demos.

TechAscent has a blog post up about how the inter-language bindings actually work at a very low level.

A walkthough using libpython-clj to render some graphs via matplotlib is on nextjournal.

If you want to quickly start using python from clojure your fastest path is probably the panthera library in addition to learning how the primitives in this library work.

New to Clojure or the JVM? Try remixing the nextjournal entry and playing around there. For more resources on learning and getting more comfortable with Clojure, we have an introductory document.

Usage

Python objects are essentially two dictionaries, one for 'attributes' and one for 'items'. When you use python and use the '.' operator, you are referencing attributes. If you use the '[]' operator, then you are referencing items. Attributes are built in, item access is optional and happens via the __getitem__ and __setitem__ attributes. This is important to realize in that the code below doesn't look like python because we are referencing the item and attribute systems by name and not via '.' or '[]'.

sudo apt install libpython3.6-dev
# numpy and pandas are required for unit tests.  Numpy is required for
# zero copy support.
python3.6 -m pip install numpy pandas --user

Initialize python

user>

user> (require '[libpython-clj.python
                 :refer [as-python as-jvm
                         ->python ->jvm
                         get-attr call-attr call-attr-kw
                         get-item att-type-map
                         call call-kw initialize!
                         as-numpy as-tensor ->numpy
                         run-simple-string
                         add-module module-dict
                         import-module
                         python-type]])
nil


user> (initialize!)
Jun 30, 2019 4:47:39 PM clojure.tools.logging$eval7369$fn__7372 invoke
INFO: executing python initialize!
Jun 30, 2019 4:47:39 PM clojure.tools.logging$eval7369$fn__7372 invoke
INFO: Library python3.6m found at [:system "python3.6m"]
Jun 30, 2019 4:47:39 PM clojure.tools.logging$eval7369$fn__7372 invoke
INFO: Reference thread starting
:ok

This dynamically finds the python shared library and loads it. If you desire a different shared library you can override here.

Execute Some Python

*out* and *err* capture python stdout and stderr respectively.

user> (run-simple-string "print('hey')")
hey
{:globals
 {'__name__': '__main__', '__doc__': None, '__package__': None, '__loader__': <class '_frozen_importlib.BuiltinImporter'>, '__spec__': None, '__annotations__': {}, '__builtins__': <module 'builtins' (built-in)>},
 :locals
 {'__name__': '__main__', '__doc__': None, '__package__': None, '__loader__': <class '_frozen_importlib.BuiltinImporter'>, '__spec__': None, '__annotations__': {}, '__builtins__': <module 'builtins' (built-in)>}}

The results have been 'bridged' into java meaning they are still python objects but there are java wrappers over the top of them. For instance, Object.toString forwards its implementation to the python function __str__.

(instance? java.util.Map (:globals bridged))
true
user> (:globals bridged)
{'__name__': '__main__', '__doc__': None, '__package__': None, '__loader__': <class '_frozen_importlib.BuiltinImporter'>, '__spec__': None, '__annotations__': {}, '__builtins__': <module 'builtins' (built-in)>}

We can get and set global variables here. If we run another string, these are in the environment. The globals map itself is the global dict of the main module:

(def main-globals (-> (add-module "__main__")
                            (module-dict)))
#'user/main-globals

user> main-globals
{'__name__': '__main__', '__doc__': None, '__package__': None, '__loader__': <class '_frozen_importlib.BuiltinImporter'>, '__spec__': None, '__annotations__': {}, '__builtins__': <module 'builtins' (built-in)>}
user> (keys main-globals)
("__name__"
 "__doc__"
 "__package__"
 "__loader__"
 "__spec__"
 "__annotations__"
 "__builtins__")
user> (get main-globals "__name__")
"__main__"
user> (.put main-globals "my_var" 200)
nil

user> (run-simple-string "print('your variable is:' + str(my_var))")
your variable is:200
{:globals
 {'__name__': '__main__', '__doc__': None, '__package__': None, '__loader__': <class '_frozen_importlib.BuiltinImporter'>, '__spec__': None, '__annotations__': {}, '__builtins__': <module 'builtins' (built-in)>, 'my_var': 200},
 :locals
 {'__name__': '__main__', '__doc__': None, '__package__': None, '__loader__': <class '_frozen_importlib.BuiltinImporter'>, '__spec__': None, '__annotations__': {}, '__builtins__': <module 'builtins' (built-in)>, 'my_var': 200}}

Running python isn't ever really necessary, however, although it may at times be convenient. You can call attributes from clojure easily:

user> (def np (import-module "numpy"))
#'user/np
user> (def ones-ary (call-attr np "ones" [2 3]))
#'user/ones-ary
user> ones-ary
[[1. 1. 1.]
 [1. 1. 1.]]
user> (call-attr ones-ary "__len__")
2
user> (vec ones-ary)
[[1. 1. 1.] [1. 1. 1.]]
user> (type (first *1))
:pyobject
user> (get-attr ones-ary "shape")
(2, 3)
user> (vec (get-attr ones-ary "shape"))
[2 3]

user> (att-type-map ones-ary)
{"T" :ndarray,
 "__abs__" :method-wrapper,
 "__add__" :method-wrapper,
 "__and__" :method-wrapper,
 "__array__" :builtin-function-or-method,
 "__array_finalize__" :none-type,
 "__array_function__" :builtin-function-or-method,
 "__array_interface__" :dict,
 "__array_prepare__" :builtin-function-or-method,
 "__array_priority__" :float,
 "__array_struct__" :py-capsule,
 "__array_ufunc__" :builtin-function-or-method,
 "__array_wrap__" :builtin-function-or-method,
 "__bool__" :method-wrapper,
 "__class__" :type,
 "__complex__" :builtin-function-or-method,
 "__contains__" :method-wrapper,
 ...
 "std" :builtin-function-or-method,
 "strides" :tuple,
 "sum" :builtin-function-or-method,
 "swapaxes" :builtin-function-or-method,
 "take" :builtin-function-or-method,
 "tobytes" :builtin-function-or-method,
 "tofile" :builtin-function-or-method,
 "tolist" :builtin-function-or-method,
 "tostring" :builtin-function-or-method,
 "trace" :builtin-function-or-method,
 "transpose" :builtin-function-or-method,
 "var" :builtin-function-or-method,
 "view" :builtin-function-or-method}

att-type-map

It can be extremely helpful to print out the attribute name->attribute type map:

user> (att-type-map ones-ary)
{"T" :ndarray,
 "__abs__" :method-wrapper,
 "__add__" :method-wrapper,
 "__and__" :method-wrapper,
 "__array__" :builtin-function-or-method,
 "__array_finalize__" :none-type,
 "__array_function__" :builtin-function-or-method,
 "__array_interface__" :dict,
 ...
  "real" :ndarray,
 "repeat" :builtin-function-or-method,
 "reshape" :builtin-function-or-method,
 "resize" :builtin-function-or-method,
 "round" :builtin-function-or-method,
 "searchsorted" :builtin-function-or-method,
 "setfield" :builtin-function-or-method,
 "setflags" :builtin-function-or-method,
 "shape" :tuple,
 "size" :int,
 "sort" :builtin-function-or-method,
 ...
}

Errors

Errors are caught and an exception is thrown. The error text is saved verbatim in the exception:

user> (run-simple-string "print('syntax errrr")
Execution error (ExceptionInfo) at libpython-clj.python.interpreter/check-error-throw (interpreter.clj:260).
  File "<string>", line 1
    print('syntax errrr
                      ^
SyntaxError: EOL while scanning string literal

Numpy

Speaking of numpy, you can move data between numpy and java easily.

user> (def tens-data (as-tensor ones-ary))
#'user/tens-data
user> (println tens-data)
#tech.v2.tensor<float64>[2 3]
[[1.000 1.000 1.000]
 [1.000 1.000 1.000]]
nil


user> (require '[tech.v2.datatype :as dtype])
nil
user> (def ignored (dtype/copy! (repeat 6 5) tens-data))
#'user/ignored
user> (.put main-globals "ones_ary" ones_ary)
Syntax error compiling at (*cider-repl cnuernber/libpython-clj:localhost:39019(clj)*:191:7).
Unable to resolve symbol: ones_ary in this context
user> (.put main-globals "ones_ary" ones-ary)
nil

user> (run-simple-string "print(ones_ary)")
[[5. 5. 5.]
 [5. 5. 5.]]
{:globals
 {'__name__': '__main__', '__doc__': None, '__package__': None, '__loader__': <class '_frozen_importlib.BuiltinImporter'>, '__spec__': None, '__annotations__': {}, '__builtins__': <module 'builtins' (built-in)>, 'my_var': 200, 'ones_ary': array([[5., 5., 5.],
       [5., 5., 5.]])},
 :locals
 {'__name__': '__main__', '__doc__': None, '__package__': None, '__loader__': <class '_frozen_importlib.BuiltinImporter'>, '__spec__': None, '__annotations__': {}, '__builtins__': <module 'builtins' (built-in)>, 'my_var': 200, 'ones_ary': array([[5., 5., 5.],
       [5., 5., 5.]])}}

So heavy data has a zero-copy route. Anything backed by a :native-buffer has a zero copy pathway to and from numpy. For more information on how this happens, please refer to the datatype library documentation.

Just keep in mind, careless usage of zero copy is going to cause spooky action at a distance.

Further Information

• design documentation
• examples
• docker setup
• pandas bindings (!!)
• nextjournal notebook
• scicloj video
• Clojure/Python interop technical blog post

Resources

• libpython C api
• create numpy from C ptr
• create C ptr from numpy

License

Copyright © 2019 Chris Nuernberger

This program and the accompanying materials are made available under the terms of the Eclipse Public License 2.0 which is available at http://www.eclipse.org/legal/epl-2.0.