Using Python Deep Learning Models In Common Lisp With a Web Services Interface

In older editions of this book I had an example of using the Java DeepLearning4J deep learning library using Armed Bear Common Lisp, implemented in Java. I no longer use hybrid Java and Common Lisp applications in my own work and I decided to remove this example and replace it with two projects that use simple Python web services that act as wrappers for state of the art deep learning models with Common Lisp clients in the subdirectories:

• src/spacy_web_client: use the spaCy deep learning models for general NLP. I sometimes use my own pure Common Lisp NLP libraries we saw in earlier chapters and sometimes I use a Common Lisp client calling deep learning libraries like spaCy and TensorFlow.
• src/coref_web_client: coreference or anaphora resolution is the act of replacing pronouns in text with the original nouns that they refer to. This has traditionally been a very difficult and only partially solved problem until recent advances in deep learning models like BERT.

Note: in the next chapter we will cover similar functionality but we will use the py4cl library to more directly use Python and libraries like spaCy by starting another Python process and using streams for communication.

You will need python and pip installed on your system. The source e code for the Python web services is found in the directory loving-common-lisp/python.

I assume that you have some familiarity with using Python. If not, you will still be able to follow these directions assuming that you have the utilities pip, and python installed. I recommend installing Python and Pip using Anaconda.

The server code is in the subdirectory python/python_spacy_nlp_server where you will work when performing a one time initialization. After the server is installed you can then run it from the command line from any directory on your laptop.

I recommend that you use virtual Python environments when using Python applications to separate the dependencies required for each application or development project. Here I assume that you are running in a Python version 3.6 or higher environment. First you must install the dependencies:

Then change directory to the subdirectory python/python_spacy_nlp_server in the git repo for this book and install the NLP server:

Once you install the server, you can run it from any directory on your laptop or server using:

I use deep learning models written in Python using TensorFlow or PyTorch and provide Python web services that can be used in applications I write in Haskell or Common Lisp using web client interfaces for the services written in Python. While it is possible to directly embed models in Haskell and Common Lisp, I find it much easier and developer friendly to wrap deep learning models I use a REST services as I have done here. Often deep learning models only require about a gigabyte of memory and using pre-trained models has lightweight CPU resource needs so while I am developing on my laptop I might have two or three models running and available as wrapped REST services. For production, I configure both the Python services and my Haskell and Common Lisp applications to start automatically on system startup.

This is not a Python programming book and I will not discuss the simple Python wrapping code but if you are also a Python developer you can easily read and understand the code.

I recommend that you use virtual Python environments when using Python applications to separate the dependencies required for each application or development project. Here I assume that you are running in a Python version 3.6 environment. First you should install the dependencies:

As I write this chapter the neuralcoref model and library require a slightly older version of SpaCy (the current latest version is 2.1.4).

Then change directory to the subdirectory python/python_coreference_anaphora_resolution_server in the git repo for this book and install the coref server:

Once you install the server, you can run it from any directory on your laptop or server using:

While. as we saw in the last example, it is possible to directly embed models in Haskell and Common Lisp, I find it much easier and developer friendly to wrap deep learning models I use a REST services as I have done here. Often deep learning models only require about a gigabyte of memory and using pre-trained models has lightweight CPU resource needs so while I am developing on my laptop I might have two or three models running and available as wrapped REST services. For production, I configure both the Python services and my Haskell and Common Lisp applications to start automatically on system startup.

This is not a Python programming book and I will not discuss the simple Python wrapping code but if you are also a Python developer you can easily read and understand the code.

Before looking at the code, I will show you typical output from running this example:

The client library is implemented in the file src/spacy_web_client/spacy-web-client.lisp:

On line 3 we define base URL for accessing the spaCy web service, assuming that it is running on your laptop and not a remote server. On line 5 we define a defstruct named spacy-data that has two fields: a list of entities in the input text and a list of word tokens in the input text.

The function spacy-client builds a query string on lines 10-12 that consists of the base-url and the input query text URL encoded. The drakma library, that we used before, is used to make a HTTP request from the Python spaCy server. Lines 14-15 uses the flexi-streams package to convert raw byte data to UTF8 characters. Lines 16-17 use the json package to parse the UTF8 encoded string, getting two lists of strings. I left the debug printout expression in line 18 so that you can see the results of parsing the JSON data. The function make-spacy-data was generated for us by the defstruct statement on line 5.

Let’s look at some typical output from this example, then we will look at the code:

Notice that pronouns in the input text are correctly replaced by the noun phrases that the pronoun refer to.

The implementation for the core client is in the file src/coref_web_client/coref.lisp:

This code is similar to the example in the last section for setting up a call to http-request but is simpler: here the Python coreference web service accepts a string as input and returns a string as output with pronouns replaced by the nouns or noun phrases that they refer to. The example in the last section had to parse returned JSON data, this example does not.

If you run Common Lisp in an IDE (for example in LispWorks’ IDE or VSCode with a Common Lisp plugin) make sure you start the IDE from the command line so your PATH environment variable will be set as it is in our bash or zsh shell.

Make sure you are starting your Common Lisp program or running a Common Lisp repl with the same Python installation (if you have Quicklisp installed, then you also have the package uiop installed):

Much of my professional work in the last five years involved deep learning models and currently most available software is written in Python. While there are available libraries for calling Python code from Common Lisp, these libraries tend to not work well for Python code using libraries like TensorFlow, spaCy, PyTorch, etc., especially if the Python code is configured to use GPUs via CUDA of special hardware like TPUs. I find it simpler to simply wrap functionality implemented in Python as a simple web service.