Spellbook

This repo aims to optimize usability of ML techniques by implementing a standardized code piece suitable to various ML tasks such as classification, object detection, regression, etc.

By directly importing the code piece the learning canvas may be built rather fast and allow engineers to focus on task-specific issues rather than cross-task, reusable actions.

The goal is to build a machine learning pipeline, where you would only change global variables contained in the globalVariables.py file and start the whole data engineering and model/s training process in either magicClassification.py or magicDetection.py files. Data preprocessing is done differently for each project/competition beforehand, that is before it is fed to the pipeline.

Spellbook Contents

The table below contains the descriptions for each of the scripts this repo provides:

Script	Description
callbacks.py	Custom callbacks for Keras library
globalVariables.py	Global variables and parameters such as data shapes, learning rate decay, etc.
helpers.py	Miscellaneous functions
layers.py	Custom / Specific layers.
losses.py	Custom loss functions
magicClassification.py	Main wrapper used to fit given models with given parameters and data. Used for classification/regression tasks.
magicDetection.py	Main wrapper used to fit given models with given parameters and data. Used for object localization tasks.
metrics.py	Custom metrics
models.py	Custom models as well as pre-trained Tensorflow models (such as ImageNet)
optimizers.py	Custom optimizers
permutationFunctions.py	Functions for data permutations
prepareTrainDataset.py	Functions to prepare datasets (initialization of tf.data.Dataset object with optional permutations)
preprocessingFunctions.py	Functions for data preprocessing
templates.py	Templates and formats on how to structure and document
train.py	Collection of train steps and full complete train cycles