py4cl2

[Last update: v2.9.0]

Introduction

py4cl is a package by Ben Dudson, aimed at making python libraries availble in Common Lisp using streams to communicate with a separate python process - this is the approach taken by cl4py and is different to the CFFI approach used by burgled-batteries, but has the same goal.

py4cl2 is intended to be an improvement over the original py4cl - importing of arguments posed a critical problem with backwards-compatibility. (See Highlights and Limitations.)

Please report the issues on github: py4cl2 or py4cl).

Highlights and Limitations of py4cl

  • Speed: About 6500 (pycall "int" "5") instructions per second @ 1GHz intel 8750H. This shouldn't be a bottleneck if you're planning to run "long" processes in python. (For example, deep learning :). )
  • Virtual environments: pycmd (*python-command* in py4cl): Choose which python binary to use. Works with miniconda.
  • Multiple python processes (not documented here) - parallel execution?
  • CI for SBCL, CCL, and ECL
  • v2.6.0 worked with ABCL 1.7.1 (Java 1.8) despite several limitations (see the documentation included in that release).
  • While CI has not been set up for windows (a PR is welcome!), from v2.7.0, things will be tested locally before releasing.
  • No support for inheriting python classes - should require MOP
  • Mac users should update to a later version of bash. See this stackoverflow question. Thanks to byulparen for pointing out!
  • Embeddable into lisp-image - the code from py4cl.py is copied into *python-code* and heredocs are used. This is made to happen for unix and not for windows; until someone gets into Windows heredocs. (Maintainer note: This entails not using single-quote character ' in py4cl.py.) Also note that this still requires the developer (= py4cl2 user) to supply the python libraries and binaries along with the lisp image to the (very-)end-user.
  • Improvements in large array transfer speed numpy-file-format (see initialize); while this does not beat remote-objects in existence since py4cl, it does provide a faster way to send array data from lisp to python and can be beneficial while offloading the work to python process, eg. deep learning.
  • Ability to interrupt the python process using (pyinterrupt)


Upcoming Sources of Possible Code Breakage

  • *lispifiers* and with-lispifiers symbols are currently unstable introduced in this very version. Avoid using in production code until stability.

Changes over py4cl

Backwards Incompatible Changes

Unavoidable

  • Arguments are imported; submodules can be imported with an option to defpymodule. However, this is only possible for python3. In fact, with python2's end-of-life, no support for python2 is provided even.
  • Asynchronous Printing: use (with-python-output &body body) to capture python output; in py4cl, (with-output-to-stream (*standard-output*) &body body) works. The separate macro with-python-output was necessitated due to the asyncronous printing in py4cl2.
  • Semantics of nil: see Type Mapping and Pythonize

Avoidable

  • Several (but not all) names have been shorted from python- to py; remote-objects have been changed to with-remote-object(s). Personal preference for these names stems from:
    • defpyfun/module reminds of the equivalent in burgled-batteries and cffi
    • pynames are shorter
    • with-remote seems more appropriate
    • chain and chain* with more "uniformity"

IMO Backwards Portable Changes

  • defpymodule (previously import-module) works "as-expected" with asdf / defpackage
  • Embedding py4cl.py into lisp image
  • importing submodules
  • process startup errors

Other Notes

  • Argument ordering can be wrong with ABCL, CCL or ECL. I've not used it extensively at anywhere other than SBCL. Basic tests concerning argument orders do work on ABCL, CCL and ECL; since, in most cases, you are good with keyword args. Early adopters are welcome :D!
Feature / Implementation (default: linux) SBCL SBCL (Windows) CCL ECL ABCL
Basic Functionality ? ? ?
Interrupt
with-python-output
array-element-type preservation
Fast Large Array Transfer

Installation

Dependencies

This fork is possible due to the following (and therefore, depends on):

On the CL side:

  • alexandria
  • bordeaux-threads
  • cl-json
  • trivial-garbage
  • iterate
  • numpy-file-format
  • parse-number
  • uiop (some implementations have an older version of uiop; support for launch-program is needed for asynchronous processes)

On python side:

  • numpy (recommended for arrays)

(other packages should be available in a standard python distribution - tested with CPython.)

Installation

Besides installing directly from quicklisp, download the (latest) release from the Releases and untar/unzip into ~/quicklisp/local-projects/ or any other location where it can be discovered by quicklisp:

wget -qO- https://github.com/digikar99/py4cl2/archive/v2.6.0.tar.gz | tar xvz - -C ~/quicklisp/local-projects

Load into REPL with

(ql:quickload :py4cl2)

Tests

See py4cl2-tests

Setting up

initialize

On loading this library for the first time, run initialize and provide the necessary details.

(py4cl2:initialize)

(You may want to note the printed information, about the location of config-file. Of course, you can call this function again, but be sure to refill the values.)

The library uses (temporary) pickled .npy files for transferring large numpy arrays efficiently between lisp and python. This process is IO intensive, writing as much as 100MB or even a GB each time, and therefore, using a ram-disk is recommended for this purpose would be recommended. (How to create a ram disk on Linux?)

*config* / config-var

These values can also be accessed using *config* and config-var:

CL-USER> py4cl2:*config*
((PY4CL2:PRINT-PYTHON-TRACEBACK . T)
 (PY4CL2:PYCMD . "/home/user/miniconda3/bin/python")
 (PY4CL2:NUMPY-PICKLE-LOCATION . "/home/user/ram-disk/_numpy_pickle.npy")
 (PY4CL2:NUMPY-PICKLE-LOWER-BOUND . 100000))
CL-USER> (py4cl2:config-var 'py4cl2:numpy-pickle-location)
"/home/user/ram-disk/_numpy_pickle.npy"
CL-USER> (setf (config-var 'py4cl2:pycmd) "python")
"python"

Complementary to config-var are save-config and load-config. The latter is called on startup, the config-file exists. (setf config-var) calls the former unless it is pycmd, as well as asks the python process to load the config, from the config file. (The exception for pycmd exists so as to let the users set up project-local environments.)

*internal-features* / *warn-on-unavailable-feature-usage*

This lists the features available on your system/implementation. Values in the list may include:

  • :ARRAYS: requires numpy, and can be seen only after a call to (pystart)
  • :TYPED-ARRAYS: requires support for specialized-arrays on lisp side; ABCL 1.7.1 does not provide this
  • :FAST-LARGE-ARRAY-TRANSFER: requires support for numpy-file-format
  • :INTERRUPT: requires system support for SIGINT (and some more things not working on ECL)
  • :WITH-PYTHON-OUTPUT

See also: the feature comparison table.

Attempt to use a feature that is not available will signal a warning and has undefined consequences. To avoid the warning, set the value of *warn-on-unavailable-feature-usage* to nil.

Examples and Documentation

CL-USER> (use-package :py4cl2)

Python Processes

It all starts with a python process (actually, more than one as well - however, this use hasn't been documented here.).

pycmd

CL-USER> (config-var 'pycmd)
"python"

Also see config-var.

pyversion-info

CL-USER> (pyversion-info)
(3 7 3 "final" 0)

pyinterrupt

(pyinterrupt &optional process)

A simple C-c C-c only interrupts the lisp process from slime - the python process keeps running. (pyinterrupt) can be used in these cases to send a SIGINT (2) to the python process.

Also note that if pyinterrupt is not called before sending the next form to eval or exec, the input-output would go out of sync. A known way to get out is to (pystop) the python-process.

Therefore, you may want to have (pyinterrupt) called on the reception of SIGINT in SLIME:

(when (find-package :swank)
  (defvar swank-simple-break)
  (setf (fdefinition 'swank-simple-break)
        (fdefinition (find-symbol "SIMPLE-BREAK" :swank)))
  (defun swank:simple-break
      (&optional (datum "Interrupt from Emacs") &rest args)
    (py4cl2:pyinterrupt)
    (apply (fdefinition 'swank-simple-break) datum args)))

However, I have been unable to get the code to work (by adding to do-after-load with as well as without SLIME. Further, people may not like a library to fiddle with their environments - so it might be better to leave it up to the user to set it.

py-cd

(py-cd path)

Equivalent of slime-cd, since python is a separate process.

Other useful functions and variables

pystart
pystop
python-alive-p
python-start-if-not-alive
*defpymodule-silent-p*

Doing arbitrary things in python

Unlike lisp, python (and most other languages) make a distinction between statements and expressions: see Quora or stackoverflow.

A general rule of thumb from there is: if you can print it, or assign it to a variable, then it's an expression, otherwise it's a statement.

Both pyeval and pyexec take any type of arguments. The arg is pythonized if the arg is not a string, or it is a string that can be read into a real.

raw-pyeval

(raw-pyeval &rest strings)

Concatenates the strings and sends them to the python process for evaluation. The concatenation should be a valid python expression. Returns the result of evaluating the expression.

raw-pyexec

(raw-pyexec &rest strings)

Concatenates the strings and sends them to the python process for execuation. The concatenation should be a valid python statement. Returns nil.

Note that one limitation of pyexec is that modules imported on the top-level (of python) are not available inside some things. These "some things" include functions.

The following should illustrate this point:

CL-USER> (pyexec "import time")
NIL
CL-USER> (pyeval "time.time()")
1.5623434e9
CL-USER> (pyexec "
def foo():
  return time.time()")
NIL
CL-USER> (pyeval "foo()")
; Evaluation aborted on #<PYERROR {100C24DF03}> ;; says 'time' is not defined
CL-USER> (pyeval "time.time()")
1.5623434e9

THe workaround in this case is to import inside the def.

Often times, the two commands above would be tedious - since you'd need to convert objects into their string representations every time. To avoid this hassle, there are the following useful functions.

pyeval

(pyeval &rest args)

For python expressions

CL-USER> (pyeval 4 "+" 3)
7

There's also (setf pyeval), which unlike (pyexec), can return non-nil values.

CL-USER> (setf (pyeval "a") "5")
"5"
CL-USER> (pyeval "a")
"5"

pyeval (and pyexec) treats the string as a python string, if it can be parsed into a number. In fact, in accordance with an internal function pythonizep.

CL-USER> (pyeval "1.0")
"1.0"
CL-USER> (pyeval "hello")
; Evaluation aborted on #<PYERROR {1003AC0183}>.

See also Doing arbitrary things in python.

pyexec

(pyexec &rest args)

For python statements

CL-USER> (pyexec "
if True:
  print(5)
else:
  print(10)")
; 5
NIL

pyexec (and pyeval) treats the string as a python string, if it can be parsed into a number). In fact, in accordance with an internal function pythonizep. (See pyeval.)

See also Doing arbitrary things in python to learn about pyeval and pyexec.

Defining python functions and modules

Rather, we define functions that call python functions.

Names are lispified by converting underscores hyphens, and converting CamelCase to camel-case. Also see Name Mapping.

defpyfun

(defpyfun fun-name &optional pymodule-name &key
  (as fun-name) (lisp-fun-name (lispify-name as))
  (lisp-package *package*)
  (safety t))

lisp-fun-name is the name of the symbol that would be fboundped to the function [that calls the python function].

Example Usage:

CL-USER> (defpyfun "Input" "keras.layers" :lisp-fun-name "INP")
INP

CL-USER> (inp :shape '(1 2))
##S(PY4CL2::PYTHON-OBJECT
   :TYPE "<class 'tensorflow.python.framework.ops.Tensor'>"
   :HANDLE 1849)

safety takes care to import the required function from the required module after python process restarts for some reason. However, this affects speed.

Refer (describe 'defpyfun).

defpymodule

(defpymodule (pymodule-name
              &optional (import-submodules nil)
              &key (cache t)
                (continue-ignoring-errors t)
                (lisp-package (lispify-name pymodule-name) lisp-package-supplied-p)
                (reload t)
                (recompile-on-change nil)
                (safety t)
                (silent *defpymodule-silent-p*)))

lisp-package is the name of the symbol that the package would be bound to.

Example Usage:

CL-USER> (defpymodule "keras.layers" t :lisp-package "KL")
Defining KL for accessing python package kl...
Defining KL.ADVANCED-ACTIVATIONS for accessing python package kl.advanced_activations...
Defining KL.CONVOLUTIONAL for accessing python package kl.convolutional...
Defining KL.CONVOLUTIONAL-RECURRENT for accessing python package kl.convolutional_recurrent...
Defining KL.CORE for accessing python package kl.core...
Defining KL.CUDNN-RECURRENT for accessing python package kl.cudnn_recurrent...
Defining KL.EMBEDDINGS for accessing python package kl.embeddings...
Defining KL.LOCAL for accessing python package kl.local...
Defining KL.MERGE for accessing python package kl.merge...
Defining KL.NOISE for accessing python package kl.noise...
Defining KL.NORMALIZATION for accessing python package kl.normalization...
Defining KL.POOLING for accessing python package kl.pooling...
Defining KL.RECURRENT for accessing python package kl.recurrent...
Defining KL.WRAPPERS for accessing python package kl.wrappers...
T

CL-USER> (kl:input/1 :shape '(1 2))
#S(PY4CL2::PYTHON-OBJECT
   :TYPE "<class 'tensorflow.python.framework.ops.Tensor'>"
   :HANDLE 816)

CL-USER> (pycall (kl.advanced-activations:softmax/class :input-shape '(1 2))
                 (kl:input/1 :shape '(1 2)))
#S(PY4CL2::PYTHON-OBJECT
   :TYPE "<class 'tensorflow.python.framework.ops.Tensor'>"
   :HANDLE 144)

Note that unlike Common Lisp, python has a single namespace. Therefore, currently, to call a callable (in Python) object, but not defined as a function in Common Lisp, you'd need to use something like pycall.

defpyfuns

(Undocumented here.)

Customizing Type Mapping

*lispifiers*

NOTE: This is a new feature and hence unstable; recommended to avoid in production code.

Each entry in the alist *lispifiers* maps from a lisp-type to a single-argument lisp function. This function takes as input the "default" lisp objects and is expected to appropriately parse it to the corresponding lisp object.

with-lispifiers

(with-lispifiers ((&rest overriding-lispifiers) &body body))

NOTE: This is a new feature and hence unstable; recommended to avoid in production code.

Each entry of overriding-lispifiers is a two-element list of the form

  (type lispifier)

Here, type is unevaluated, while lispifier will be evaluated; the lispifier is expected to take a default-lispified object (see lisp-python types translation table in docs) and return the appropriate object user expects.

For example,

  (pyeval "[1, 2, 3]") ;=> #(1 2 3) ; the default lispified object
  (with-lispifiers ((vector (lambda (x) (coerce (print x) 'list))))
    (print (pyeval "[1,2,3]"))
    (print (pyeval 5)))
  ; #(1 2 3) ; default lispified object
  ; (1 2 3)  ; coerced to LIST by the lispifier
  ; 5        ; lispifier uncalled for non-VECTOR
  5

*pythonizers*

NOTE: This is a new feature and hence unstable; recommended to avoid in production code.

Each entry in the alist *pythonizers* maps from a lisp-type to a single-argument PYTHON-FUNCTION-DESIGNATOR. This python function takes as input the "default" python objects and is expected to appropriately convert it to the corresponding python object.

with-pythonizers

(with-pythonizers (&rest overriding-pythonizers) &body body)

NOTE: This is a new feature and hence unstable; recommended to avoid in production code.

Each entry of overriding-pythonizers is a two-element list of the form

  (type pythonizer)

Here, type is unevaluated, while pythonizer will be evaluated; the pythonizer is expected to take a default-pythonized object (see lisp-python types translation table in docs) and return the appropriate object user expects.

For example,

  (pyeval "[1, 2, 3]") ;=> #(1 2 3) ; the default object
  (with-pythonizers ((vector "tuple"))
    (print (pyeval "[1,2,3]"))
    (print (pyeval 5)))
  ; #(1 2 3) ; default object
  ; (1 2 3)  ; coerced to tuple by the pythonizer, which then translates to list
  ; 5        ; pythonizer uncalled for non-VECTOR
  5

Errors

pyerror

Signalled if there an error in a running python process

python-process-startup-error

Signalled if pystart fails to start a python process

python-eof-and-dead

Signalled (from dispatch-messages) if the lisp process is unable to read/write from/to a python process and the process has died. This should provide a cleaner error than CL:END-OF-FILE.

python-eof-but-alive

Signalled (from dispatch-messages) if the lisp process is unable to read/write from/to a python process but the process is alive. This should provide a cleaner error than CL:END-OF-FILE.

Using functions and methods

pycall

(pycall fun-name &rest args)

Equivalent to the lisp (funcall function &rest arguments). Call a python (or lisp! See generators and lambdas) function.

CL-USER> (py4cl2:pycall "print" "hello")
;; hello
"None"
CL-USER> (py4cl2:pycall #'+ 2 3 1)
6

Note that fun-name can be a name (see Name Mapping), a function, or a [callable] python-object. See the example in defpymodule.

pymethod

(pymethod obj method-name &rest args)

pymethod always pythonizes; method-name is name mapped to Python names.

SEQ2SEQ> (pymethod model 'summary) ;; for some "ready" model
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to
==================================================================================================
input_1 (InputLayer)            (None, None, 43)     0
__________________________________________________________________________________________________
input_2 (InputLayer)            (None, None, 64)     0
__________________________________________________________________________________________________
lstm_1 (LSTM)                   [(None, 256), (None, 307200      input_1[0][0]
__________________________________________________________________________________________________
lstm_2 (LSTM)                   [(None, None, 256),  328704      input_2[0][0]
                                                                 lstm_1[0][1]
                                                                 lstm_1[0][2]
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, None, 64)     16448       lstm_2[0][0]
==================================================================================================
Total params: 652,352
Trainable params: 652,352
Non-trainable params: 0
__________________________________________________________________________________________________
"None"

See pymethod-list.

pyslot-value

(pyslot-value object slot-name)

CL-USER> (pyslot-value model 'input-shape)
#((NIL NIL 43) (NIL NIL 64))

See pyslot-list

Also see Name Mapping.

export-function

(export-funtion function python-name)

Lisp functions can be made available to python code using export-function:

(py4cl:python-exec "from scipy.integrate import romberg")

(py4cl:export-function (lambda (x) (/ (exp (- (* x x)))
                                      (sqrt pi))) "gaussian")

(py4cl:python-eval "romberg(gaussian, 0.0, 1.0)") ; => 0.4213504

pyhelp

(pyhelp python-object)

Calls python's help function on python-object. (NOTE: some descriptions, especially for modules, are too big to be transferred in a reasonable time.)

Generators and Lambdas

pygenerator

(pygenerator function stop-value)

CL-USER> (let ((a 0)) (defun foo () (incf a)))
FOO

CL-USER> (pyeval "[x for x in " (pygenerator #'foo 3) "]")
#(1 2)

lambdas

Lisp functions are pythonized to LispCallbackObjects. As the name suggests, python can call LispCallbackObjects (and therefore, lisp functions), just like it is any other python callable (which it is!).

CL-USER> (py4cl::pythonize #'car)
"_py4cl_LispCallbackObject(4)"

CL-USER> (pycall (lambda (string) (concatenate 'string string " - from Lisp"))
                 "hello")
"hello - from Lisp"

Slot and Method Lists

Currently, all the python objects are grouped under the class python-object. The list of methods and slots associated with these objects can be obtained using the following two functions.

pyslot-list

(pyslot-list python-object &key as-vector)

CL-USER> (defpyfun "Model" "keras")
NIL

CL-USER> (pyslot-list (model))
("__class__" "__delattr__" "__dict__" "__doc__" "__eq__" "__ge__"
 "__getattribute__" "__gt__" "__hash__" "__le__" "__lt__" "__module__" "__ne__"
 "__repr__" "__str__" "__weakref__" "_built" "_expects_training_arg"
 "_inbound_nodes" "_initial_weights" "_is_compiled" "_is_graph_network"
 "_layers" "_losses" "_outbound_nodes" "_per_input_losses" "_per_input_updates"
 "_updates" "_uses_inputs_arg" "built" "input_spec" "inputs" "layers" "losses"
 "name" "non_trainable_weights" "optimizer" "outputs" "state_updates"
 "stateful" "supports_masking" "trainable" "trainable_weights" "updates"
 "uses_learning_phase" "weights")

CL-USER> (pyeval (model) ".inputs")
"None"

Optionally, see pyslot-value

pymethod-list

(pymethod-list python-object &key as-vector)

CL-USER> (pymethod-list (model))
("__call__" "__class__" "__delattr__" "__dir__" "__eq__" "__format__" "__ge__"
 "__getattribute__" "__getstate__" "__gt__" "__hash__" "__init__"
 "__init_subclass__" "__le__" "__lt__" "__ne__" "__new__" "__reduce__"
 "__reduce_ex__" "__repr__" "__setattr__" "__setstate__" "__sizeof__" "__str__"
 "__subclasshook__" "_add_inbound_node" "_base_init"
 "_check_trainable_weights_consistency" "_get_node_attribute_at_index"
 "_init_graph_network" "_init_subclassed_network" "_make_predict_function"
 "_make_test_function" "_make_train_function" "_node_key" "_set_inputs"
 "_standardize_user_data" "_updated_config" "_uses_dynamic_learning_phase"
 "add_loss" "add_update" "add_weight" "assert_input_compatibility" "build"
 "call" "compile" "compute_mask" "compute_output_shape" "count_params"
 "evaluate" "evaluate_generator" "fit" "fit_generator" "from_config"
 "get_config" "get_input_at" "get_input_mask_at" "get_input_shape_at"
 "get_layer" "get_losses_for" "get_output_at" "get_output_mask_at"
 "get_output_shape_at" "get_updates_for" "get_weights" "load_weights" "predict"
 "predict_generator" "predict_on_batch" "reset_states" "run_internal_graph"
 "save" "save_weights" "set_weights" "summary" "test_on_batch" "to_json"
 "to_yaml" "train_on_batch")

Optionally, see pymethod.

chain(*)

(chain &rest chain)

This is inspired by the chain in parenscript, discussed in this issue.

In python it is quite common to apply a chain of method calls, data member access, and indexing operations to an object. To make this work smoothly in Lisp, there is the chain macro (Thanks to @kat-co and parenscript for the inspiration). This consists of a target object, followed by a chain of operations to apply. For example

(chain "hello {0}" (format "world") (capitalize)) ; => "Hello world"

which is converted to python return "hello {0}".format("world").capitalize().

chain has two variants: chain is a macro, and chain* is a function.

A few examples are as follows:

(chain (slice 3) stop) ; => 3
(let ((format-str "hello {0}")
      (argument "world"))
 (py4cl2:chain* format-str `(format ,argument))) ; => "hello world"

Arguments to methods are lisp, since only the top level forms in chain are treated specially:

CL-USER> (chain (slice 3) stop)
3
CL-USER> (let ((format-str "hello {0}")
               (argument "world"))
           (chain* format-str `(format ,argument)))
"hello world"
CL-USER> (chain* "result: {0}" `(format ,(+ 1 2)))
"result: 3"
CL-USER> (chain (aref "hello" 4))
"o"
CL-USER> (chain (aref "hello" (slice 2 4)))
"ll"
CL-USER> (chain (aref #2A((1 2 3) (4 5 6)) (slice 0 2)))
#2A((1 2 3) (4 5 6))
CL-USER> (chain (aref #2A((1 2 3) (4 5 6))  1 (slice 0 2)))
#(4 5)
CL-USER> (pyexec "class TestClass:
      def doThing(self, value = 42):
        return value")
CL-USER> (chain ("TestClass") ("doThing" :value 31))
31

There is also (setf chain):

CL-USER> (pyeval
          (with-remote-object (array (np:zeros '(2 2)))
            (setf (chain* `(aref ,array 0 1)) 1.0
                  (chain* `(aref ,array 1 0)) -1.0)
            array))
#2A((0.0 1.0) (-1.0 0.0))

Note that this modifies the value in python, so the above example only works because =array= is a handle to a python object, rather than an array which is stored in lisp. The following therefore does not work:

CL-USER> (let ((array (np:zeros '(2 2))))
           (setf (chain* `(aref ,array 0 1)) 1.0
                 (chain* `(aref ,array 1 0)) -1.0)
           array)
#2A((0.0 0.0) (0.0 0.0))

with-remote-objects(*)

`(with-remote-objects &body body)

If a sequence of python functions and methods are being used to manipulate data, then data may be passed between python and lisp. This is fine for small amounts of data, but inefficient for large datasets.

The with-remote-objects and with-remote-objects* macros provide unwind-protect environments in which all python functions return handles rather than values to lisp. This enables python functions to be combined without transferring much data.

(with-remote-objects (py4cl:python-eval "1+2"))
; => #S(PY4CL::PYTHON-OBJECT :TYPE "<class 'int'>" :HANDLE 0)

with-remote-objects* evaluates the last result, instead of merely returning a handle

(with-remote-objects* (py4cl:python-eval "1+2")) ; => 3

The advantage comes when dealing with large arrays or other datasets:

CL-USER> (time (let ((arr (make-array 1000000
                                      :element-type 'single-float
                                      :initial-element 2.0)))
                 (np:sum (np:add arr arr))))
;  0.258 seconds of real time
;  8,065,504 bytes consed
4000000.0
CL-USER> (time (with-remote-objects
                 (let ((arr (make-array 1000000
                                        :element-type 'single-float
                                        :initial-element 2.0)))
                   (np:sum (np:add arr arr)))))
;  0.100 seconds of real time
;  4,065,456 bytes consed
4000000.0

Note that this requires you to solely use python functions and methods. So, do not expect something like this to work:

(with-remote-objects (print (aref (np:ones :shape '(10000000)) 0)))
; Error

to work.

Besides this, see Setting up for using ram-disk and numpy-file-format to combine lisp and python functions. .

python-getattr

(python-getattr object slot-name)

Lisp structs and class objects can be passed to python, put into data structures and returned:

(defpyfun "dict") ; Makes python dictionaries

(defstruct test-struct
    x y)

(let ((map (dict :key (make-test-struct :x 1 :y 2))))  ; Make a dictionary, return as hash-map
  ;; Get the struct from the hash-map, and get the Y slot
  (test-struct-y
    (chain* `(aref ,map "key"))))  ; => 2

In python this is handled using an object of class UnknownLispObject, which contains a handle. The lisp object is stored in a hash map *lisp-objects*. When the python object is deleted, a message is sent to remove the object from the hash map.

To enable python to access slots, or call methods on a struct or class, a handler function needs to be registered. This is done by providing a method for generic function python-getattr. This function will be called when a python function attempts to access attributes of an object (__getattr__ method).

;; Define a class with some slots
(defclass test-class ()
  ((value :initarg :value)))

;; Define a method to handle calls from python
(defmethod python-getattr ((object test-class) slot-name)
  (cond
    ((string= slot-name "value") ; data member
      (slot-value object 'value))
    ((string= slot-name "func")  ; method, return a function
      (lambda (arg) (* 2 arg)))
    (t (call-next-method)))) ; Otherwise go to next method

(let ((instance (make-instance 'test-class :value 21)))
  ;; Get the value from the slot, call the method
  ;; python: instance.func(instance.value)
  (chain* `((@ ,instance func) (@ ,instance value))))  ; => 42

Inheritance then works as usual with CLOS methods:

;; Class inheriting from test-class
(defclass child-class (test-class)
  ((other :initarg :other)))

;; Define method which passes to the next method if slot not recognised
(defmethod py4cl:python-getattr ((object child-class) slot-name)
  (cond
    ((string= slot-name "other")
     (slot-value object 'other))
    (t (call-next-method))))

(let ((object (make-instance 'child-class :value 42 :other 3)))
  (list
    (chain* object 'value) ; Call TEST-CLASS getattr method via CALL-NEXT-METHOD
    (chain* object 'other))) ;=> (42 3)

python-setattr

Type Mapping and Pythonize

Data is passed between python and lisp as text. The python function lispify converts values to a form which can be read by the lisp reader; the lisp function pythonize outputs strings which can be eval'd in python. The following type conversions are done:

| Lisp type               | Python type           |
|-------------------------+-----------------------|
| NIL                     | False                 |
| integer                 | int                   |
| ratio                   | fraction.Fractions    |
| real                    | float                 |
| complex                 | complex float         |
| string                  | str                   |
| hash map                | dict                  |
| list                    | tuple                 |
| unspecialized vector    | list                  |
| (un)specialized array   | NumPy array*          |
| single-float            | numpy.float32         |
| double-float            | float                 |
| symbol                  | Symbol class          |
| function                | function              |

*The currently supported numpy types include: ((un)signed-byte XX) where XX can be from (08 16 32 64), and arrays of bits. Raise an issue if you want more types supported or this functionality exposed to the user.

Special conversion rules include:

nil     False
t       True
"None"  None
"()"    ()        # empty tuple

Because pyeval and pyexec evaluate strings as python expressions, strings passed to them are not escaped or converted as other types are. To pass a string to python as an argument, call py4cl:pythonize

CL-USER> (py4cl:pythonize "string")
"\"string\""
CL-USER> (py4cl:pythonize #'identity)
"_py4cl_LispCallbackObject(1)"
CL-USER> (py4cl:pythonize 3.0)
"3.0"
CL-USER> (py4cl:pythonize (model)) ;; keras.Model
"_py4cl_objects[1918]"

If python objects cannot be converted into a lisp value, then they are stored and a handle is returned to lisp. This handle can be used to manipulate the object, and when it is garbage collected the python object is also deleted (using the trivial-garbage package).

Customizing Type-Mapping

Unstable feature

Since version 2.8.0, values returned from the python process are wrapped inside a call to py4cl2::customize. This function essentially does the following:

(defun customize (object)
  (loop :for (type . lispifier) :in *lispifiers*
        :if (typep object type)
          :do (return-from customize (funcall lispifier object)))
  object)

with-lispifiers provides a convenient wrapper to bind the dynamic variable *lispifiers* for executing its body.

Name Mapping

The arguments passed to pycall are parsed as follows: the lisp keywords are converted to their python equivalents. This only entails downcasing the symbol-name of the keywords and replacing hyphens with underscores. If the symbol-name contained capital letters, then, if all the letters are capitals, the symbol-name is downcased; else it stays as it is

CL-USER> (pyexec "
def foo(A, b):
  return True")
CL-USER> (pycall 'foo :*A* 4 :b 3)
T
CL-USER> (pycall 'foo :a 4 :b 3)
; Evaluation aborted on #<PYERROR {100E2AF473}>.
;; unexpected keyword argument 'a'
CL-USER> (pycall 'foo 4 3)
T

Lispfication of python names is done by defpyfun, in import-export.lisp. Both CamelCase and joint_words are converted to camel-case and joint-words; the actual names of the arguments are substituted:

CL-USER> (macroexpand-1 '(defpyfun "foo"))
(DEFUN FOO (&KEY (A 'NIL) (B 'NIL))
  "None"
  NIL
  (PYTHON-START-IF-NOT-ALIVE)
  (RAW-PYEVAL "foo" "(" "A" "=" (PY4CL2::PYTHONIZE A) "," "b" "="
              (PY4CL2::PYTHONIZE B) "," ")"))
T

The format of the calling expression does depend on the signature of the function.

What remains?

Feel free to create an Issue on Github.

Future Work

In no order of priority:

  • adding/documenting proper multithreaded support
  • [ABANDON since python2 has reached end-of-life] finding equivalent of inspect._empty in python2 (unable to google)
  • importing python classes, and methods, may be, as subclasses of 'python-object; to be able to use make-instance and slot-value might require knowledge of MOP, to make python-object at the same level as standard-object
  • should return value of defpyfun matter - so as to indicate success or failure? failure is anyways indicated by errors
  • ability to define customized arg-lists, documentation, and calling methods for functions: this can serve as a community project to cover up some naming and arg-list idiosyncrasies
  • cleaning up documentation while defining functions - many python functions have documentation intended for use directly in md/rst files

Also check out

The Common Lisp Cookbook