GitLocker: The Coding Marketplace

Description:

basenetapi 1.9.3

BaseNet: A simpler way to build AI models.

Basenet API Package - 1.9.3: (2.0.0 Is coming!!)
This package implements an API over Keras and Tensorflow to build Deep Learning models easily without losing the
framework flexibility. BaseNet API tries to implement almost everything from a few lines of code.

Disclaimer: This API is under development. This means that there isn't a stable release yet.
However, some features are stable and can be used. Check the tutorials for more info.
Please, check the roadmap below to see the future updates.

Any feature request is wellcome and appreciated; and has a high probability to be implemented.

Any issue is wellcome and appreciated; and it will often be inspected.

About
Author: A.Palomo-Alonso (a.palomo@uah.es)
Universidad de Alcalá.
Escuela Politécnica Superior.
Departamento de Teoría De la Señal y Comunicaciones (TDSC).
ISDEFE Chair of Research.

Features
MACHINE LEARNING

Feature 01: Database train, validation and test automatic and random segmentation for DeepLearning.
Feature 02: Solving optimization problems for MetaHeuristic models.

EFFICIENCY

Feature 03: Real multiprocessing training process (CPU usage optimization).
Feature 04: Automatic and custom GPU usage and assignment.

SIMPLICITY

Feature 05: Easy-to-use API.
Feature 06: API documentation.
Feature 07: JuPyter Notebooks tutorials included.
Feature 08: Python Packaging and PyPi indexing.

MONITORIZATION

Feature 09: Real-time logging.
Feature 10: Dashboards included.
Feature 11: Model printing and easy debugging.

CONNECTIVITY

Feature 12: Model merging and multiple model inputs.
Feature 13: Computational cluster linking.
Feature 14: The different parts of the API are designed to interact.
Feature 15: It allows to create dynamic databases for data ingestion and math problems that require synthetic data.

RELIABILITY

Feature 16: Depends on huge and reliable frameworks: KERAS, TENSORFLOW, RAY.
Feature 17: Code updating and active support.

Cons:
FLEXIBILITY

An API must look for a balance between simplicity and flexibility. In this case, we bet on simplicity; but the API
is still highly flexible.
The API is designed for high level research and design. But it is not optimal for low-level research.

DEPENDENCE

This API is built over highly reliable frameworks, however, the API depends on those frameworks to run the models.

DEPLOYMENT

The API can not deploy models by itself yet in this current version. But I am planning!

Roadmap:
Discipline coverage:

Database building.
Supervised learning.
MetaHeuristic optimization.
Unsupervised learning. (Planning in 3.0 release)
Reinforcement learning. (Planning in 4.0 release)
Computer Vision. (Planning in 5.0 release)
Natural Language Processing. (Will be in development from 3.0 to 6.0)

Feature roadmap:

Feeder databases.

Computational clustering.

Monitoring.

BaseNetDeployment:

Deploy your model in your infrastructure with high connectivity and scalability.
With preprocess, model, postprocess and front-end sector.
Automatic workload balance (probably with Ray or Spark clusters).
Dynamic training after deployment. The model will still be learning from input data if desired.

Reinforcement learning:

This will be hard to add in deployment.
Custom environments.
Connectivity with BaseNetCompiler: It will work with RL policy and DL!
State-of-the-art RL.

Computer Vision:

It is already possible to be implemented in DeepLearning API; but can be improved.
Database visualization.
Database size optimization.
Computer vision utils.

Natural Language Processing:

Hard to tell here, this discipline is very ad-hoc, and it is hard to pack it in an API.
Transformers, LSTM and Word Vectorization will be included.
Preprocessing will be a must.
Probably will use HuggingFace package.
Will be in constant development.

What's new?
< 0.1.0

BaseNetModel included.
BaseNetDatabase included.
BaseNetCompiler included.
Inheritance from CorNetAPI project.
Multi-processing fitting.
Tensorboard launching.

0.2.0

BaseNetResults included (working).
Now the model is callable.
Switched print to logging.
Project documentation.

1.0.0 - 1.0.3

Python packaging
1.0.x: Upload bug solving.

1.1.0

Functional package.
PyPi indexing.

1.2.0:

Loss results included in the BaseNetResults while multiprocessing.
GPU auto set up to avoid TensorFlow memory errors.
Method BaseNetCompiler.set_up_devices() configures the GPUs according to the free RAM to be used in the API.

1.3.0

Included WindowDiff to the project scope.

1.4.0

Solved python packaging problems.
Included force stop callback in the BaseNetModel.fit_stop() method.

1.5.0

BaseNetDatabase now has the attributes BaseNetDatabase.size and BaseNetDatabase.distribution.
Solved forced stopping bugs with multiprocessing in the method BaseNetDatabase.fit_stop().
BaseNetModel._threshold() private method now takes a set of outputs instead only one.
This was only for optimization.
Solved wrong BaseNetModel.recover().
Auto recover implemented, now BaseNetModel.recover() is a private method: BaseNetModel._recover().
Now the used does not need to recover it. The model recovers by itself. -- Hans Niemann 2022.
NOTE: RECOVER IS NECESARY WHEN THE MODEL IS EARLY STOPPED; CONSIDER RECOVERING ALWAYS THE MODEL.

1.5.1 - 1.5.3

Solved a bug where BaseNetDatabase modified the incoming list of instances in the database; avoiding checkpoints
for large database generators.
Exception handler for nvml library if NVIDIA Drivers are not installed`in the machine.

1.5.4

Added some BaseNetDatabase utils: merge and split databases.
Added BaseNetDatabase equality check.
Added a BaseNetDatabase._reversion(), BaseNetCompiler._reversion() and BaseNetModel.__version__. Which
rebuilds the Classes to the current version of the API.

1.6.0

Start to develop the second branch of the API: BaseNetHeuristic
Start to create JuPyter Notebook tutorials of the API.
Included BaseNetDatabase binarization and normalization of databases.
Included in BaseNetDatabase to read TensorFlow and Pandas databases.
Some rework was done for bug-fixing and providing more logging information.

1.7.0 - 1.9.1

Reworked BaseNetDatabase for minor bug fixing. (1.7.0)
Added the BaseNetFeeder Class to generate dynamic BaseNetDatabases. (1.7.0)
Jupyter tutorials. (1.8.0)
Added the BaseNetDeployment Class to deploy scalable Machine Learning models (in alpha, do not use until 3.0).
(1.9.0)
Added BaseNetHeuristic algorithms that implements MetaHeuristic methods such as PSO. (1.8.0)

1.9.2

Added BaseNetLMSE, solving linear problems with matrix multiplication.

1.9.3

Added CassandraDatabase.

Basic and fast usage
There are Jupyter Notebooks with usage tutorials! Refer to them
HERE. You should run your notebook creating a
virtual environment in this README path.
Cite as
Please, cite this library as:
@misc{basenetapi,
title={BaseNet: A simpler way to build AI models.},
author={A. Palomo-Alonso},
booktitle={PhD in TIC: Machine Learning and NLP.},
year={2022}
}