Using spaCy v2.1 with spacyr

A major update of spaCy (v2.1) was released recently. spaCy is one of the best and fastest tools for tokenization, part-of-speech tagging, dependency parsing, and entity recognition. In this post, I will discuss how it works with our spacyr package along with some tips on having multiple versions of spaCy using conda environments.

Good news: It works

Our package spacyr is an R wrapper to the spaCy Python library. To work with the spacyr package, users have to prepare a Python environment with spaCy installed. This may not be difficult if you are familiar with both R and Python, but that may not be necessarily the case for R-only users. To help such users, we implemented a function called spacy_install() which provides a one-step solution.

The gist of what the function does is:

1. Create a new conda environment (with new Python executable)
2. pip-install spacy in the new environment

The implementation of the function owes a lot to a similar function in the tensorflow package. If everything works as intended, the installation process is truly one-step. And for Mac and Linux users, spacy_install() even downloads Miniconda when it does not exist in the system. This seems to be working for most cases. However, as you can see from the issues in our GitHub repository, some users have troubles in installing it (see for example this issue). Typically, the problem is caused by not having installed a C++ compiler, as some operating systems (pretty much all except Linux) do not include one by default.

Assuming that you have an old spaCy installed through spacyr, you can try updating to the new spaCy by our other function:

spacy_upgrade()

In my environment (Mac OS Mojave), it did not work and the installation stopped in the middle. The error was fixable^[The error message I got was: ImportError: Something is wrong with the numpy installation. While importing we detected an older version of numpy in [’…/anaconda/envs/spacy_condaenv/lib/python3.6/site-packages/numpy’]. One method of fixing this is to repeatedly uninstall numpy until none is found, then reinstall this version.], but it requires some work in console and does not fit to the purpose of our one-step solution through spacy_install().

So I decided to use an easy solution (which I recoomend if you experience the same issue): Delete the current conda environment and re-install

spacy_uninstall() 
spacy_install()

Once we did this on macOS, everything worked fine:

library(spacyr)
spacy_initialize(model = "en_core_web_sm")
## Finding a python executable with spaCy installed...
## spaCy (language model: en_core_web_sm) is installed in C:\Users\Mayazure\AppData\Local\Programs\Python\Python37\python.exe
## successfully initialized (spaCy Version: 2.3.2, language model: en_core_web_sm)
## (python options: type = "python_executable", value = "C:\Users\Mayazure\AppData\Local\Programs\Python\Python37\python.exe")
spacy_parse("hello world")
##   doc_id sentence_id token_id token lemma   pos entity
## 1  text1           1        1 hello hello  INTJ       
## 2  text1           1        2 world world PROPN

If you have your original environment (e.g. a custom language model you trained) and do not want to mess up the setup, you can test a new version by creating another environment as described below.

Another good news: Tokenization is faster and you can feel it in spacyr

The tokenization is really fast in spaCy v2.1. That can be seen from spacy_tokenize(). The benchmark comparison looks like this:

## Unit: milliseconds
##     expr       min        lq      mean    median        uq       max neval
##   v2.1.3  134.6249  143.3171  149.7733  151.7219  155.2604  163.8899    10
##  v2.0.18 1472.2077 1487.7420 1536.3897 1496.1156 1523.6667 1869.4669    10

This is based on tokenizing a fairly small corpus of 14 documents containing a total of about 54,000 words (the quanteda::data_corpus_irishbudget2010 corpus). Due to the impossiblity of unloading Python (see the last section of this post), each benchmark had to be run in a separate R session and combined afterwards. (See this gist for details.)

The difference is massive: spaCy v2.1 is about 10 (!) times faster in tokenization called from spacyr.

So this is great. However, there is a caveat in this performance gain. spacy_tokenize() is fast under limited conditions. By default, spaCy has four pipeline components: tokenizer, tagger, parser and ner. In version v2.1, the tokenizer became really, really fast. So, if you run only tokenizer, which is the first component of the pipeline (i.e. run spacy_tokenize() with all default options), it is very fast. However, if you need to conduct more feature rich tokenization (e.g. spacy_tokenize(remove_numbers = TRUE)), the later components of the pipeline have to be run and it will take longer to finish.

data(data_corpus_irishbudget2010, package = "quanteda.textmodels")
microbenchmark::microbenchmark(
    "remove_numbers = TRUE" = spacy_tokenize(data_corpus_irishbudget2010, remove_numbers = TRUE),
    "remove_numbers = FALSE" = spacy_tokenize(data_corpus_irishbudget2010),
    times = 1
)
## Unit: milliseconds
##                    expr       min        lq      mean    median        uq
##   remove_numbers = TRUE 3873.1447 3873.1447 3873.1447 3873.1447 3873.1447
##  remove_numbers = FALSE  106.2504  106.2504  106.2504  106.2504  106.2504
##        max neval
##  3873.1447     1
##   106.2504     1

(I didn’t check whether or not this slowdown is caused by our code in either R or Python. We will test this more thoroughly in the future.)

A note on having multiple versions of spaCy in your spacyr installation

The current setup of spacy_install() creates a new conda environment isolated from other environments, and thus, if you want to have multiple versions of spaCy, this is easily possible.

One thing to remember for doing that is you need to restart R when switching from one spaCy to another. That is because of the difficulty of unloading the Python environment. In the backend, we use the wonderful reticulate package by RStudio for seamless integration of R with Python. The developer of the package has made it clear that unloading Python is technically difficult and reticulate does not support that.^[spacyr has a function called spacy_finalize(). This function deletes all objects created by spaCy in the Python space but does not delete the Python space itself.]

Having said that, here is the way to install two versions of spaCy:

library(spacyr)

## install the latest version of spaCy (will be installed in "spacy_condaenv")
spacy_install()

## install an older version of spaCy (in an enviroment "spacy_old")
spacy_install(version = "2.0.18", envname = "spacy_old")

To use these environments, you can specify the version when you call spacy_initialize().

## to use latest version
spacy_initialize(refresh_settings = TRUE)

## to use the older version
spacy_initialize(condaenv = "spacy_old", refresh_settings = TRUE)

The first line spacy_initialize(refresh_settings = TRUE) will use spacy_condaenv as that’s the first thing spacyr will check when initializing spaCy. spacyr searches possible locations of spaCy installation when spacy_initialize() is called for the first time (not in this session, but in the entire history). After successfully initializing spaCy, spacyr will remember the location and use the same setting thereafter. Now that you are switching between two environments, you need to have spacyr ignoring the saved setting. refresh_settings = TRUE in spacy_initialize() will do the job.

Summary

In this post, I have discussed a few features of spacyr related to the new release of spaCy. In summary:

1. spaCy v2.1 works with spacyr;
2. spaCy tokenization is much faster in v2.1 than the tokenization in previous versions; and
3. spacyr can help you to set up multiple Python environments.