• django-ai version:0.0.2.1
• Django version:2.0.10
• Python version:3.6
• Operating System:Windows 10

Description

I have created a Django project and then trying to integrate django-ai with it by following below link-
https://django-ai.readthedocs.io/en/latest/installation.html

What I Did

When I ran 'python manage.py makemigrations 'command as mentioned in above ink I got following error: ModuleNotFoundError: No module named 'bayesian_networks'.

my settings.py snippet is as below-
`DEBUG = True

ALLOWED_HOSTS = []

INSTALLED_APPS = [
'django.contrib.admin',
'django.contrib.auth',
'django.contrib.contenttypes',
'django.contrib.sessions',
'django.contrib.messages',
'django.contrib.staticfiles',

# Dependencies
'nested_admin',

# django-ai apps
'django_ai.base',
'django_ai.bayesian_networks',
'django_ai.supervised_learning',
'django_ai.systems.spam_filtering',

# optional but highly recommended
'django_ai.examples',

]`

I tried to search in internet but I have not found any solution so I am asking here. Let me know if anything I am missing here.

Update

This issue was posted in the wrong repo because of a confusión with names, it can be kindly ignored

Description

As of version 5.5.2 of django-ai-core, I believe there is a little room for improvement on this model:

• It uses default=now on created_at and lastmodified_at fields, instead of the good old auto_now=True and auto_now_add=True, which is cleaner and does not make these fields visible/editable on modelForms by default, which is ideal for audit purposes.
• The created_by and lastmodified_by fields should have the editable flag set to False so they are fully automatic and not visible on forms by default.

The current settings are not wrong, but I believe the ones I'm proposing suit more the typical use case for Django Models.

How I believe the fields should look like

created_at = models.DateTimeField(_("Created at"), auto_now_add=True, db_index=True)
created_by = models.ForeignKey(settings.AUTH_USER_MODEL, verbose_name=_("Created by"), blank=True, null=True,
                                   related_name="%(app_label)s_%(class)s_created", on_delete=models.SET_NULL, editable=False)
lastmodified_at = models.DateTimeField(_("Last modified at"), auto_now=True, db_index=True)
lastmodified_by = models.ForeignKey(settings.AUTH_USER_MODEL, verbose_name=_("Last modified by"), blank=True,
                                        null=True, related_name="%(app_label)s_%(class)s_lastmodified",
                                        on_delete=models.SET_NULL, editable=False)

Currently Deterministic Nodes are not handled in the initialization of the Engine Object in Nodes

(this a Low-Hanging-Fruit for Bayesian Networks app)

Currently there is only support for *args in the functions passed as Node parameters, support for **kwargs is needed.

See https://github.com/math-a3k/django-ai/blob/master/django_ai/bayesian_networks/utils.py

Observable Nodes should have the possibility of linking to several columns / Django Models Fields in order to have observations of R^d.

See http://bayespy.org/examples/gmm.html where d=2.

Catch BayesPy's exceptions and handle them

Port the examples in http://bayespy.org/examples/examples.html to the Bayesian Network Examples App.

Add model validation to BN Nodes for Stochastic and Deterministic types and reflect it in the frontend.

Because of how BayesPy is designed, the nodes involved in a network must be the same objects - in the construction / initialization of each and when the inference engine is called - because they pass messages between each other. If the objects are recreated, i.e., on another request cycle, that won't work.

Currently Bayesian Networks support "one level of depth" graphs, it should support all kinds of graphs. This means that all the nodes in the DAG must be created accordingly - with their dependencies - ""at once"" / ""in the sample place"".

One way could be start initializing the parent-less nodes, then "go down" to the children (which may have other parents, but they are already initialized)

See https://github.com/math-a3k/django-ai/blob/master/django_ai/bayesian_networks/models.py#L75

Currently only floats and node names are supported as parameters for distribution and deterministic.

It should also support other python expressions like matrices and/or a set of functions.

The goal of this issue is to add support for Linear Regression using the scikit-learn engine for django-ai.

Its scope is limited only for Ordinary Least Squares, setting the building block for incorporating the Generalized Linear Models provided in the scikit-learn module.

Insights

• Thoroughly read the documentation on django-ai's API - specially on "Design Goals", "Engine related" and "Supervised Learning Technique"
• Create a class on the appropriate place that inherits from SupervisedLearningTechique and implements Regression.
• Implement on that class all the methods in the django-ai's API accordingly for incorporating this technique implementation - a reference implementation can be found here, which is the analogous for SVM using the same engine.
  • Note: The class and all fields should be documented in a way that Sphinx can generate the documentation
• Create a simple test which proves the correct functioning of the implementation and covers all the essential parts of the implementation, analogous to this one.
• Create the admin front-end for the technique, like is done here.
• Create the documentation for the technique, like is done here
• Provide an example of the technique like is done here.
  • Note that the example is not about a Linear Regression, but how to incorporate the use of a Linear Regression into a Django application to improve its functionality or solve an issue seamlessly.
• You may have to create tests accordingly so the general testing coverage does not descend, see here.