AlgebraicRelations.jl is a Julia library built to provide an intuitive and elegant method for generating and querying a scientific database. This package provides tooling for defining database schemas, generating query visualizations, and connecting directly up to a PostgreSQL server. This package is built on top of Catlab.jl which is the powerhouse behind its functions.
The functions of this library may be best explained by showing an example of how it can be used. This will be done in the steps of Defining a Schema, Creating Queries, and Connecting to PostgreSQL.
Within this library, we define database schemas based on the presentation of a workflow (more generally, the presentation of a symmetric monoidal category). The presentation of a workflow includes the data types of products in the workflow (objects in an SMC) and the processes that transform these products (homomorphisms in an SMC). We will give an example of defining the schema of a traditional computer vision workflow. This involves extracting images from a file, performing a test/train split on images, training a neural network on images, and finally evaluating a network on images. This example is also presented in this notebook.
In order to define types for the presentation, we need to provide the name of
the type (e.g. File
for compressed files of images) and then the Julia
datatype which can store this type (The filename can be stored uniquely as a
String
). The definition of all types that we will need for our example is as
follows:
# Initialize presentation object
present = Presentation()
# Add types to presentation
File, Images, NeuralNet,
Accuracy, Metadata = add_types!(present, [(:File, String),
(:Images, String),
(:NeuralNet, String),
(:Accuracy, Real),
(:Metadata, String)]);
To define processes that operate on these types, we need three pieces of
information. First, we need the name of the processes (extract
for the
process that extracts images from files), the input types (File
for the file
to extract) and the output types (Images
for the images which were
extracted). The symbol ⊗
(monoidal product) joins two types, allowing for multiple types
in the inputs and outputs of processes. To the schema, this means nothing more than that,
for the process train
there are two objects need for the input, the first of
type NeuralNet
and the second of type Images
.
# Add Processes to presentation
extract, split, train,
evaluate = add_processes!(present, [(:extract, File, Images),
(:split, Images, Images⊗Images),
(:train, NeuralNet⊗Images, NeuralNet⊗Metadata),
(:evaluate, NeuralNet⊗Images, Accuracy⊗Metadata)]);
Once this presentation is defined, the database schema can be generated as follows:
# Convert to Schema
TrainDB = present_to_schema(present);
print(generate_schema_sql(TrainDB()))
CREATE TABLE evaluate (NeuralNet1 text, Images2 text, Accuracy3 real, Metadata4 text);
CREATE TABLE extract (File1 text, Images2 text);
CREATE TABLE split (Images1 text, Images2 text, Images3 text);
CREATE TABLE train (NeuralNet1 text, Images2 text, NeuralNet3 text, Metadata4 text);
In order to create queries, we use the @query
macro (based on the @relation
macro in Catlab). For this, we must specify a list of objects to get as results
of the query, list of all objects used in the query, and finally a list of
relationships between these objects (based on the primitives defined for the
workflow). In this case, the relationships between objects are the processes
from the presentation and the types of objects are the types defined in the
presentation. Following is an example workflow
q = @query TrainDB() (im_train, nn, im_test, acc, md2) where (im_train, im_test, nn,
nn_trained, acc, md,
md2, _base_acc, im) begin
train(nn, im_train, nn_trained, md)
evaluate(nn_trained, im_test, acc, md2)
split(im, im_train, im_test)
>=(acc, _base_acc)
end
print(to_sql(q))
This produces the following query:
SELECT t1.Images2 AS im_train, t1.NeuralNet1 AS nn, t2.Images2 AS im_test, t2.Accuracy3 AS acc, t2.Metadata4 AS md2
FROM train AS t1, evaluate AS t2, split AS t3
WHERE t2.NeuralNet1=t1.NeuralNet3 AND t3.Images2=t1.Images2 AND t3.Images3=t2.Images2 AND t2.Accuracy3>=$1
The connection to PostgreSQL is fairly straightforward. We first create a connection using the LibPQ.jl library:
conn = Connection("dbname=test_db");
We then can prepare statements and run them with arguments like:
statement = prepare(conn,q)
execute(statement, [0.6])
which will obtain all of the rows from the previous query which contain an accuracy of greater than 0.6.
The execute
function will return a DataFrame
object (from the
DataFrames.jl
library)
Some excellent resources for understanding how Bicategories of Relations relate to SQL queries (and inspiriation for this library) are as follows:
- "Knowledge Representation in Bicategories of Relations"
- This work does an excellent job of elucidating operations on the Bicategories of Relations and how that relates to methods of knowledge representation like SQL
- "The operad of wiring diagrams: formalizing a graphical language for databases, recursion, and plug-and-play circuits"
- This work presents the concepts behind converting undirected wiring diagrams to queries (as well as the limitations present in this conversion)
- Category Theory for Scientists by Spivak
- While generally a very useful introduction to Category Theory, this book elaborates on the categorization of databases in Chapter 3 (in the online version)