py4ai data

---

A Python library defining data structures optimized for machine learning pipelines

What is it ?

py4ai-data is a Python package with modular design that provides powerful abstractions to build data ingestion pipelines and run end to end machine learning pipelines. The library offers lightweight object-oriented interface to MongoDB as well as Pandas based data structures. The aim of the library is to provide extensive support for developing machine learning based applications with a focus on practicing clean code and modular design. The detailed documentation for all its features is available here.

Features

Some cool features that we are proud to mention are:

Data layers

The Data Layer abstractions are designed to decouple the business layers from the detail of the persistence layer implementation. We currently implemented a few data layers based on some common abstractions:

1. Repository: class representing the access point to data. Specifies the methods to read/write data from/to persistence layer. It must be instantiated by passing it a Serializer, to specify how to read/write data from/to persistence layer to/from memory.
2. Serializer: class that specifies how to read/write data from/to persistence layer to/from memory.
3. Criteria: class implementing query method to filter data retrieved from repository. The data layer currently implemented are:

• file-system: access generic data stored in the file system as generic files, handling them in memory as generic indexed objects.
• pandas: access tabular data stored in the file system as csv files, handling them in memory as pandas objects.
• mongo: access data stored in MongoDB. Serializers for this layer should be implemented by the user according to its specific needs.
• sqlalchemy: access data stored in a database handled with sqlalchemy ORM. Serializers for this layer should be implemented by the user according to its specific needs.

Data Model

Offers the following data structures:

1. Document : Data structure specifically designed to work with NLP applications that parses a json-like document into a couple of uuid and dictionary of information.
2. Sample : Data structure representing an observation (a.k.a. sample) as used in machine learning applications
3. MultiFeatureSample : Data structure representing an observation defined by a nested list of arrays.
4. Dataset : Data structure designed to be used specifically for machine learning applications representing a collection of samples.

Installation

From pypi server

pip install py4ai-data

From source

git clone https://github.com/NicolaDonelli/py4ai-data
cd py4ai-data
make install

Tests

make tests

Checks

To run predefined checks (unit-tests, linting checks, formatting checks and static typing checks):

make checks

Examples

Data Layers

The Data Layer abstractions are designed to decouple the business layers from the detail of the persistence layer implementation. The basic abstraction that will make this possible is the Repository.

As an example, imagine to have a domain business object Entity that associates some string data to an integer id.

from pydantic import BaseModel

class Entity(BaseModel):
    my_id: int
    my_data: str

To start with, imagine we want to use csv files store on disk as a persistence layer. To do so, we will use the CsvRepository that uses pandas DataFrames stored in memory and written to the disk as csv. Thus, we need to write the business logic to serialize the Entity into a row of the pandas DataFrame, i.e. a pandas Series:

import pandas as pd
from py4ai.data.layer.common.serialiazer import DataSerializer

class EntitySerializer(DataSerializer[int, int, Entity, pd.Series]):
    def to_object(self, entity: Entity) -> pd.Series:
        return pd.Series(entity.dict())

    def to_entity(self, document: pd.Series) -> Entity:
        return Entity(**document)

    def to_object_key(self, key: int) -> int:
        return key

    def get_key(self, entity: Entity) -> int:
        return entity.my_id

We can now instantiate the repository class that has all the methods for reading and writing objects from the persistence layer.

from py4ai.data.layer.pandas.repository import CsvRepository

repo = CsvRepository(filename, EntitySerializer())

entity = Entity(my_id=1234, my_data="Important data")

# This will create the entity in the persistence layer
await repo.create(entity)

# Retrieving the entity
retrieved = repo.retrieve(key=1234)

# Retrieving all entities
all_entities = repo.list()

Imagine now that, given the data increase in size, we now would like to change the persistence layer with a proper backend into something more structured and scalable, such as a NoSQL document-based data platform, such as MongoDB. We only need to create a new business logic to serialize/deserialize our class into a json (represented in python by a dictionary):

from bson import ObjectId
from py4ai.data.layer.mongo.serializer import create_mongo_id
from py4ai.data.layer.common.serialiazer import DataSerializer

class MongoDataSerializer(DataSerializer[int, ObjectId, Entity, dict]):
    def get_key(self, entity: Entity) -> int:
        return entity.my_id

    def to_object(self, entity: Entity) -> dict:
        doc = entity.dict()
        doc["_id"] = self.to_object_key(self.get_key(entity))
        return doc

    def to_entity(self, document: dict) -> Entity:
        return Entity(**document)

    def to_object_key(self, key: int) -> ObjectId:
        # This converts a string into an hash compatible with MongoDB format
        return create_mongo_id(str(key))

A new repository based on the MongoDB persistence layer can now be created using

from py4ai.data.layer.mongo.repository import MongoRepository

repo = MongoRepository(collection, MongoDataSerializer())

This repository is compatible with the previous and can be used in place of the previous one, having the same signatures.

Abstracting Data Querying

The Repository abstraction also allow to retrieve data based on certain queries/filters:

entities = repo.retrieve_by_query(criteria)

However, the format of the query also depends on the type of the persistence layer and more specifically on how the data are organized. Therefore, in order to abstract and decouple the notion of the underlying persistence layer, we need to define a general class containing the possible queries for a certain database:

from typing import Generic
from abc import ABC, abstractmethod

from py4ai.data.layer.common.criteria import SearchCriteria


class EntityQueryFactory(ABC, Generic[Q]):
    @abstractmethod
    def by_id(self, id: int) -> SearchCriteria[Q]:
        ...

When considering a particular persistence layer, the querying business logic needs to be specified

from py4ai.data.layer.mongo.criteria import MongoSearchCriteria


class MongoQueryFactory(EntityQueryFactory[Dict[str, Any]]):

    def by_id(self, my_id: int) -> MongoSearchCriteria:
        return MongoSearchCriteria({"my_id": my_id})


criteria = MongoQueryFactory()

entities = repo.retrieve_by_query(criteria.by_id(1234))

Note that SearchCriteria can be also joined using logical operators:

entities = repo.retrieve_by_query(
    criteria.by_id(1234) & criteria.by_id(1235) 
)

entities = repo.retrieve_by_query(
    criteria.by_id(1234) | criteria.by_id(1235) 
)

Data Model

The data.model module contains data structures commonly used. The most relevant structures defined in the core submodule are CachedIterable and LazyIterable that represent iterable objects that are either completely stored in memory or accessed lazily. In the same submodule we also define Range and CompositeRange that represent time and numeric ranges, either continuous or not, with methods to compare two instances checking for overlaps or performing range union by sum.

The ml submodule on the other hand defines some data structures that are of common use in Machine Learning applications. The basic structure we define is the Sample, which is a simple serializable object characterized by a set of features, an optional set of labels and an optional name identifier:

class Sample(DillSerialization, Generic[FeatType, LabType]):
    """Base class for representing a sample/observation."""

    def __init__(
        self,
        features: FeatType,
        label: Optional[LabType] = None,
        name: Optional[Union[int, str, Any]] = None,
    ) -> None:
        """
        Return an object representing a single sample of a training or test set.

        :param features: features of the sample
        :param label: labels of the sample (optional)
        :param name: id of the sample (optional)
        """
        self.features: FeatType = features
        self.label: Optional[LabType] = label
        self.name: Optional[Union[str, int, Any]] = name

We also define a specialization of this class, called MultiFeatureSample, thought for Deep Learning applications, for which the features attribute is a list of numpy arrays. The most used objects defined in this module are the datasets, which represents collections of Samples in various fashions. All the datasets share some common method allowing to retrieve the type of the Samples, to check the names of the samples, to retrieve features and labels in a form to be chosen among pd.DataFrame, np.ndarray, dict, list, or a generator of samples. In particular, we defined:

• LazyDataset: a collection with lazy access to samples
• CachedDataset: a collection of samples fully stored in memory
• PandasDataset: a collection of samples, fully stored in memory, where features and labels are pd.DataFrames
• PandasTimeIndexedDataset: a collection of samples, fully stored in memory, where features and labels are pd.DataFrames with time index

Creating a PandasDataset object

import pandas as pd
import numpy as np
from py4ai.data.model.ml import PandasDataset

dataset = PandasDataset(features=pd.concat([pd.Series([1, np.nan, 2, 3], name="feat1"),
                                            pd.Series([1, 2, 3, 4], name="feat2")], axis=1),
                        labels=pd.Series([0, 0, 0, 1], name="Label"))

# access features as a pandas dataframe 
print(dataset.features.head())
# access labels as pandas dataframe 
print(dataset.labels.head())
# access features as a python dictionary 
dataset.getFeaturesAs('dict')
# access features as numpy array 
dataset.getFeaturesAs('array')

# indexing operations 
# access features and labels at the given index as a pandas dataframe  
print(dataset.loc([2]).features.head())
print(dataset.loc([2]).labels.head())

Creating a PandasTimeIndexedDataset object

import pandas as pd
import numpy as np
from py4ai.data.model.ml import PandasTimeIndexedDataset

dateStr = [str(x) for x in pd.date_range('2010-01-01', '2010-01-04')]
dataset = PandasTimeIndexedDataset(
    features=pd.concat([
        pd.Series([1, np.nan, 2, 3], index=dateStr, name="feat1"),
        pd.Series([1, 2, 3, 4], index=dateStr, name="feat2")
    ], axis=1))

How to contribute ?

We are very much willing to welcome any kind of contribution whether it is bug report, bug fixes, contributions to the existing codebase or improving the documentation.

Where to start ?

Please look at the Github issues tab to start working on open issues

Contributing to py4ai-data

Please make sure the general guidelines for contributing to the code base are respected

1. Fork the py4ai-data repository.
2. Create/choose an issue to work on in the Github issues page.
3. Create a new branch to work on the issue.
4. Commit your changes and run the tests to make sure the changes do not break any test.
5. Open a Pull Request on Github referencing the issue.
6. Once the PR is approved, the maintainers will merge it on the main branch.