JNA libpython bindings to the tech ecosystem.

Version Info

• Note - Please avoid deprecated versions such as [cnuernber/libpython-clj "1.36"] (note name change).

• API Documentation
• This library, which has received the efforts of many excellent people, is built mainly upon cnuernber/dtype-next and techascent/tech.jna. It supports zerocopy transfer of data to/from numpy, dynamically creating Clojure namespaces from Python modules, and has a deep intergration with the Clojure language.

libpython-clj features

• 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. Scope based resource contexts are also available.
• Finding the python libraries is done dynamically allowing one system to run on multiple versions of python.
• REPL oriented design means fast, smooth, iterative development.
• Carin Meier has written excellent posts on plotting and advanced text generation. She also has some great examples.

Vision

We aim to integrate Python into Clojure at a deep level. This means that we want to be able to load/use python modules almost as if they were Clojure namespaces. We also want to be able to use Clojure to extend Python objects. I gave a talk at Clojure Conj 2019 that outlines more of what is going on.

This code is a concrete example that generates an embedding for faces:

(ns facial-rec.face-feature
  (:require [libpython-clj.require :refer [require-python]]
            [libpython-clj.python :refer [py. py.. py.-] :as py]
            [tech.v2.datatype :as dtype]))



(require-python 'mxnet
                '(mxnet ndarray module io model))
(require-python 'cv2)
(require-python '[numpy :as np])


(defn load-model
  [& {:keys [model-path checkpoint]
      :or {model-path "models/recognition/model"
           checkpoint 0}}]
  (let [[sym arg-params aux-params] (mxnet.model/load_checkpoint model-path checkpoint)
        all-layers (py. sym get_internals)
        target-layer (py/get-item all-layers "fc1_output")
        model (mxnet.module/Module :symbol target-layer
                                   :context (mxnet/cpu)
                                   :label_names nil)]
    (py. model bind :data_shapes [["data" [1 3 112 112]]])
    (py. model set_params arg-params aux-params)
    model))

(defonce model (load-model))


(defn face->feature
  [img-path]
  (py/with-gil-stack-rc-context
    (if-let [new-img (cv2/imread img-path)]
      (let [new-img (cv2/cvtColor new-img cv2/COLOR_BGR2RGB)
            new-img (np/transpose new-img [2 0 1])
            input-blob (np/expand_dims new-img :axis 0)
            data (mxnet.ndarray/array input-blob)
            batch (mxnet.io/DataBatch :data [data])]
        (py. model forward batch :is_train false)
        (-> (py. model get_outputs)
            first
            (py. asnumpy)
            (#(dtype/make-container :java-array :float32 %))))
      (throw (Exception. (format "Failed to load img: %s" img-path))))))

Usage

Config namespace

(ns my-py-clj.config
  (:require [libpython-clj.python :as py]))

;; When you use conda, it should look like this.
(py/initialize! :python-executable "/opt/anaconda3/envs/my_env/bin/python3.7"
                :library-path "/opt/anaconda3/envs/my_env/lib/libpython3.7m.dylib")

Update project.clj

{...
  ;; This namespace going to run when the REPL is up. 
  :repl-options {:init-ns my-py-clj.config}
...}

user> (require '[libpython-clj.require :refer [require-python]])
...logging info....
nil
user> (require-python '[numpy :as np])
nil
user> (def test-ary (np/array [[1 2][3 4]]))
#'user/test-ary
user> test-ary
[[1 2]
 [3 4]]

We have a document on all the features but beginning usage is pretty simple. Import your modules, use the things from Clojure. We have put effort into making sure things like sequences and ranges transfer between the two languages.

Environments

One very complimentary aspect of Python with respect to Clojure is it's integration with cutting edge native libraries. Our support isn't perfect so some understanding of the mechanism is important to diagnose errors and issues.

Current, we launch the python3 executable and print out various different bits of configuration as json. We parse the json and use the output to attempt to find the libpython3.Xm.so shared library so for example if we are loading python 3.6 we look for libpython3.6m.so on Linux or libpython3.6m.dylib on the Mac.

This pathway has allowed us support Conda albeit with some work. For examples using Conda, check out the facial rec repository above or look into how we build our test docker containers.

Further Information

• Clojure Conj 2019 video and slides.
• development discussion forum
• design documentation
• scope and garbage collection docs
• examples
• docker setup
• pandas bindings (!!)
• nextjournal notebooks
• scicloj video
• Clojure/Python interop technical blog post
• persistent datastructures in python

New To Clojure

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.

Resources

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

To install jar to local .m2 :

$ lein install

Deploy to clojars

$ lein deploy clojars

This command will sign jar before deploy, using your gpg key. (see dev/src/build.clj for signing options)

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.