takahish / clispy

Common Lisp Interpreter written in Python

CLisPy

Overview

Common Lisp interpreter written in Python (CLisPy) inspired from Peter Norvig, lispy and lispy.py. This interpreter will be implemented to satisfy　ANSI Common Lisp (Syntax, Macro System, CLOS etc).

Motivation

Comon lisp s-expression have good compatibility with some deep learning frameworks to stack layers. And to build the machine learning model by using scikit-learn, the s-expression also have compatibility for data and model pipline.

Required libraries

• Python 3.x
• numpy 1.14+

Examples

$ python
Python 3.7.1 (default, Dec 14 2018, 19:28:38) 
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import clispy; clispy.repl()
CL-USER=> (setq pi 3.14)
3.14

CL-USER=> (defun area (r) (* pi r r))
AREA

CL-USER=> (area 3)
28.26

CL-USER=> (defmacro when (test-form &rest body)
            `(if ,test-form
               (progn ,@body)))
WHEN

CL-USER=> (when t 3)
3

CL-USER=> (when nil 3)
NIL

CL-USER=> (car '(1 2 3 4 5))
1

CL-USER=> (cdr '(1 2 3 4 5))
(2 3 4 5)

CL-USER=> (cons 'a nil)
(A)

CL-USER=> (cons 'a 'b)
(A . B)

CL-USER=> (quit)
Although never is often better than *right* now.

>>> 

Python bridge

Clispy supports python built-in Functions.

• https://docs.python.org/3/library/functions.html

CL-USER=> (py:abs -100)
100

CL-USER=> (py:sorted '(9 8 7 6 5 4 3 2 1 0))
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

CL-USER=> (type-of (py:sorted '(9 8 7 6 5 4 3 2 1 0)))
PYTHON-OBJECT

CL-USER=> (coerce (py:sorted '(9 8 7 6 5 4 3 2 1 0)) 'cons)
(0 1 2 3 4 5 6 7 8 9)

CL-USER=>

Tensorflow bridge

There are python object manipulation functions.

• py:import Imports python modules
• py:attr Gets an attribute of python object.
• py:call Calls a method of python object.
• py:item Gets an item of python sequence object, this function is run with index or py:slice.

CL-USER=> (py:import tf "tensorflow"
                     keras "tensorflow.keras")
KERAS

CL-USER=> (setq mnist (py:call keras "datasets.mnist.load_data"))
...

CL-USER=> (setq x-train (py:item (py:item mnist 0) 0)
                y-train (py:item (py:item mnist 0) 1)
                x-test (py:item (py:item mnist 1) 0)
                y-test (py:item (py:item mnist 1) 1))
...

CL-USER=> (setq x-train (/ x-train 255.0)
                x-test (/ x-test 255.0))
...

CL-USER=> (setq model (py:call keras "models.Sequential"
                        (list (py:call keras "layers.Flatten" :input-shape '(28 28))
                              (py:call keras "layers.Dense" 512 :activation (py:attr tf "nn.relu"))
                              (py:call keras "layers.Dropout" 0.2)
                              (py:call keras "layers.Dense" 10 :activation (py:attr tf "nn.softmax")))))
<tensorflow.python.keras.engine.sequential.Sequential object at 0x000001E70CFEF470>

CL-USER=> (py:call model "compile" :optimizer "adam"
                                   :loss "sparse_categorical_crossentropy"
                                   :metrics '("accuracy"))
None

CL-USER=> (py:call model "fit" x-train y-train :epochs 5)
Epoch 1/5
60000/60000 [==============================] - 11s 189us/sample - loss: 0.2195 - acc: 0.9353
Epoch 2/5
60000/60000 [==============================] - 11s 184us/sample - loss: 0.0972 - acc: 0.9704
Epoch 3/5
60000/60000 [==============================] - 11s 184us/sample - loss: 0.0689 - acc: 0.9785
Epoch 4/5
60000/60000 [==============================] - 11s 187us/sample - loss: 0.0534 - acc: 0.9827
Epoch 5/5
60000/60000 [==============================] - 11s 187us/sample - loss: 0.0403 - acc: 0.9871
<tensorflow.python.keras.callbacks.History object at 0x000001E71E9687B8>

CL-USER=> (py:call model "evaluate" x-test y-test)
10000/10000 [==============================] - 1s 56us/sample - loss: 0.0675 - acc: 0.9792
[0.06751626054281369, 0.9792]

References

• lispy
• lispy.py
• Common Lisp HyperSpec
• https://github.com/sbcl/sbcl
• https://github.com/armedbear/abcl
• https://docs.python.org/3/library/functions.html
• https://www.tensorflow.org/

License

Apache License 2.0

Acknowledgements

I wish to thank Peter Norvig, lispy and lispy.py. I also wish to thank sbcl and abcl for referencing how do they implement common lisp interpreter.