1
0
mirror of https://github.com/gsi-upm/senpy synced 2024-11-24 09:02:28 +00:00
senpy/docs/plugins.rst
J. Fernando Sánchez 21a5a3f201 Macro commit
* Fixed Options for extra_params in UI
* Enhanced meta-programming for models
* Plugins can be imported from a python file if they're named
`senpy_<whatever>.py>` (no need for `.senpy` anymore!)
* Add docstings and tests to most plugins
* Read plugin description from the docstring
* Refactor code to get rid of unnecessary `.senpy`s
* Load models, plugins and utils into the main namespace (see __init__.py)
* Enhanced plugin development/experience with utils (easy_test, easy_serve)
* Fix bug in check_template that wouldn't check objects
* Make model defaults a private variable
* Add option to list loaded plugins in CLI
* Update docs
2018-01-06 21:03:20 +01:00

315 lines
11 KiB
ReStructuredText

Developing new plugins
----------------------
This document contains the minimum to get you started with developing new analysis plugin.
For an example of conversion plugins, see :doc:`conversion`.
For a description of definition files, see :doc:`plugins-definition`.
A more step-by-step tutorial with slides is available `here <https://lab.cluster.gsi.dit.upm.es/senpy/senpy-tutorial>`__
.. contents:: :local:
What is a plugin?
=================
A plugin is a python object that can process entries. Given an entry, it will modify it, add annotations to it, or generate new entries.
What is an entry?
=================
Entries are objects that can be annotated.
In general, they will be a piece of text.
By default, entries are `NIF contexts <http://persistence.uni-leipzig.org/nlp2rdf/ontologies/nif-core/nif-core.html>`_ represented in JSON-LD format.
It is a dictionary/JSON object that looks like this:
.. code:: python
{
"@id": "<unique identifier or blank node name>",
"nif:isString": "input text",
"sentiments": [ {
...
}
],
...
}
Annotations are added to the object like this:
.. code:: python
entry = Entry()
entry.vocabulary__annotationName = 'myvalue'
entry['vocabulary:annotationName'] = 'myvalue'
entry['annotationNameURI'] = 'myvalue'
Where vocabulary is one of the prefixes defined in the default senpy context, and annotationURI is a full URI.
The value may be any valid JSON-LD dictionary.
For simplicity, senpy includes a series of models by default in the ``senpy.models`` module.
What are annotations?
=====================
They are objects just like entries.
Senpy ships with several default annotations, including: ``Sentiment``, ``Emotion``, ``EmotionSet``...jk bb
What's a plugin made of?
========================
When receiving a query, senpy selects what plugin or plugins should process each entry, and in what order.
It also makes sure the every entry and the parameters provided by the user meet the plugin requirements.
Hence, two parts are necessary: 1) the code that will process the entry, and 2) some attributes and metadata that will tell senpy how to interact with the plugin.
In practice, this is what a plugin looks like, tests included:
.. literalinclude:: ../senpy/plugins/example/rand_plugin.py
:emphasize-lines: 5-11
:language: python
The lines highlighted contain some information about the plugin.
In particular, the following information is mandatory:
* A unique name for the class. In our example, Rand.
* The subclass/type of plugin. This is typically either `SentimentPlugin` or `EmotionPlugin`. However, new types of plugin can be created for different annotations. The only requirement is that these new types inherit from `senpy.Analysis`
* A description of the plugin. This can be done simply by adding a doc to the class.
* A version, which should get updated.
* An author name.
Plugins Code
============
The basic methods in a plugin are:
* analyse_entry: called in every user requests. It takes two parameters: ``Entry``, the entry object, and ``params``, the parameters supplied by the user. It should yield one or more ``Entry`` objects.
* activate: used to load memory-hungry resources. For instance, to train a classifier.
* deactivate: used to free up resources when the plugin is no longer needed.
Plugins are loaded asynchronously, so don't worry if the activate method takes too long. The plugin will be marked as activated once it is finished executing the method.
How does senpy find modules?
============================
Senpy looks for files of two types:
* Python files of the form `senpy_<NAME>.py` or `<NAME>_plugin.py`. In these files, it will look for: 1) Instances that inherit from `senpy.Plugin`, or subclasses of `senpy.Plugin` that can be initialized without a configuration file. i.e. classes that contain all the required attributes for a plugin.
* Plugin definition files (see :doc:`advanced-plugins`)
Defining additional parameters
==============================
Your plugin may ask for additional parameters from the users of the service by using the attribute ``extra_params`` in your plugin definition.
It takes a dictionary, where the keys are the name of the argument/parameter, and the value has the following fields:
* aliases: the different names which can be used in the request to use the parameter.
* required: if set to true, users need to provide this parameter unless a default is set.
* options: the different acceptable values of the parameter (i.e. an enum). If set, the value provided must match one of the options.
* default: the default value of the parameter, if none is provided in the request.
.. code:: python
"extra_params":{
"language": {
"aliases": ["language", "lang", "l"],
"required": True,
"options": ["es", "en"],
"default": "es"
}
}
Loading data and files
======================
Most plugins will need access to files (dictionaries, lexicons, etc.).
These files are usually heavy or under a license that does not allow redistribution.
For this reason, senpy has a `data_folder` that is separated from the source files.
The location of this folder is controlled programmatically or by setting the `SENPY_DATA` environment variable.
Plugins have a convenience function `self.open` which will automatically prepend the data folder to relative paths:
.. code:: python
import os
class PluginWithResources(AnalysisPlugin):
file_in_data = <FILE PATH>
file_in_sources = <FILE PATH>
def activate(self):
with self.open(self.file_in_data) as f:
self._classifier = train_from_file(f)
file_in_source = os.path.join(self.get_folder(), self.file_in_sources)
with self.open(file_in_source) as f:
pass
It is good practice to specify the paths of these files in the plugin configuration, so the same code can be reused with different resources.
Docker image
============
Add the following dockerfile to your project to generate a docker image with your plugin:
.. code:: dockerfile
FROM gsiupm/senpy
Once you make sure your plugin works with a specific version of senpy, modify that file to make sure your build will work even if senpy gets updated.
e.g.:
.. code:: dockerfile
FROM gsiupm/senpy:1.0.1
This will copy your source folder to the image, and install all dependencies.
Now, to build an image:
.. code:: shell
docker build . -t gsiupm/exampleplugin
And you can run it with:
.. code:: shell
docker run -p 5000:5000 gsiupm/exampleplugin
If the plugin uses non-source files (:ref:`loading data and files`), the recommended way is to use `SENPY_DATA` folder.
Data can then be mounted in the container or added to the image.
The former is recommended for open source plugins with licensed resources, whereas the latter is the most convenient and can be used for private images.
Mounting data:
.. code:: bash
docker run -v $PWD/data:/data gsiupm/exampleplugin
Adding data to the image:
.. code:: dockerfile
FROM gsiupm/senpy:1.0.1
COPY data /
F.A.Q.
======
What annotations can I use?
???????????????????????????
You can add almost any annotation to an entry.
The most common use cases are covered in the :doc:`apischema`.
Why does the analyse function yield instead of return?
??????????????????????????????????????????????????????
This is so that plugins may add new entries to the response or filter some of them.
For instance, a chunker may split one entry into several.
On the other hand, a conversion plugin may leave out those entries that do not contain relevant information.
If I'm using a classifier, where should I train it?
???????????????????????????????????????????????????
Training a classifier can be time time consuming. To avoid running the training unnecessarily, you can use ShelfMixin to store the classifier. For instance:
.. code:: python
from senpy.plugins import ShelfMixin, AnalysisPlugin
class MyPlugin(ShelfMixin, AnalysisPlugin):
def train(self):
''' Code to train the classifier
'''
# Here goes the code
# ...
return classifier
def activate(self):
if 'classifier' not in self.sh:
classifier = self.train()
self.sh['classifier'] = classifier
self.classifier = self.sh['classifier']
def deactivate(self):
self.close()
By default the ShelfMixin creates a file based on the plugin name and stores it in that plugin's folder.
However, you can manually specify a 'shelf_file' in your .senpy file.
Shelves may get corrupted if the plugin exists unexpectedly.
A corrupt shelf prevents the plugin from loading.
If you do not care about the data in the shelf, you can force your plugin to remove the corrupted file and load anyway, set the 'force_shelf' to True in your plugin and start it again.
How can I turn an external service into a plugin?
?????????????????????????????????????????????????
This example ilustrate how to implement a plugin that accesses the Sentiment140 service.
.. code:: python
class Sentiment140Plugin(SentimentPlugin):
def analyse_entry(self, entry, params):
text = entry.text
lang = params.get("language", "auto")
res = requests.post("http://www.sentiment140.com/api/bulkClassifyJson",
json.dumps({"language": lang,
"data": [{"text": text}]
}
)
)
p = params.get("prefix", None)
polarity_value = self.maxPolarityValue*int(res.json()["data"][0]
["polarity"]) * 0.25
polarity = "marl:Neutral"
neutral_value = self.maxPolarityValue / 2.0
if polarity_value > neutral_value:
polarity = "marl:Positive"
elif polarity_value < neutral_value:
polarity = "marl:Negative"
sentiment = Sentiment(id="Sentiment0",
prefix=p,
marl__hasPolarity=polarity,
marl__polarityValue=polarity_value)
sentiment.prov(self)
entry.sentiments.append(sentiment)
yield entry
Can I activate a DEBUG mode for my plugin?
???????????????????????????????????????????
You can activate the DEBUG mode by the command-line tool using the option -d.
.. code:: bash
senpy -d
Additionally, with the ``--pdb`` option you will be dropped into a pdb post mortem shell if an exception is raised.
.. code:: bash
python -m pdb yourplugin.py
Where can I find more code examples?
????????????????????????????????????
See: `<http://github.com/gsi-upm/senpy-plugins-community>`_.