suvadeepmaiti/lightning_factory

Very little code to make PyTorch Lightning models

Lightning Factory

PyTorch Lightning is great, but model building can be a bit...verbose.
Lightning Factory is a Python library designed to simplify the creation of PyTorch Lightning models for various types of neural networks. It follows the parameterized factory pattern and allows users to specify custom configurations or use common defaults for quick prototyping.

Usage

pip install lightning_factory

To create a feed-forward neural network model:

import lightning_factory as lf

model = lf.ffnn(layers=[5, 3, 3, 1])

Easily define the layer structure:

Set default parameters when constructing the factory:

from lightning_factory import LightningFactory
from lightning_factory import LossFunction
from lightning_factory import ActivationFunction

# setting the defaults.  These values will always be used unless otherwise specified
lf = LightningFactory(
    loss_function=LossFunction.MSE,
    batch_size=32,
    activation_function=ActivationFunction.Softplus
)
model1 = lf.ffnn(layers=[5,3,3,1])
model2 = lf.ffnn(layers=[5,8,4,2,1], activation_function=ActivationFunction.Tanh)

Full example of building, training, testing and predicting

Video tutorial

Example stock data for training used in the tutorial: Download

Full Example Code

import lightning_factory as lf
from lightning_factory import d_at

# Loading stock data built for NNs by D.AT
# download sample data at: https://d.at/example-data
d_at.load_data(
    'data/train.csv',  # Training data
    'data/test.csv',   # Data time-separated from training; used to get precision, accuracy, etc
    'data/latest.csv'  # The most recent data.  The model will be predicting the labels
)

# creating our model
model = lf.ffnn(layers=[30, 3, 3, 1])

# training with our model
d_at.train(model)

# precision, accuracy, p-value of precision, confusion matrix
d_at.print_statistics()

# stocks ordered by which are most likely to have a `true` label
d_at.print_predictions()

The LightningFactory class uses the following defaults when making a class:

default = {
    'layers': None,
    'learning_rate': 0.001,
    'max_epochs': 8,
    'batch_size': 64,
    'loss_function': 'BCE',
    'activation_function': 'ReLU',
    'optimizer': 'Adam',
    'dropout': 0,
    'l1_regularization': 0,
    'l2_regularization': 0,
    'weight_initialization': 'xavier_uniform'
}

Testing Coverage

Name	Stmts	Miss	Branch	BrPart	Cover	Missing
lightning_factory/LightningFactory.py	18	0	8	0	100%
lightning_factory/__init__.py	3	0	0	0	100%
lightning_factory/create_ffnn.py	41	29	14	0	22%	19-30, 34-36, 40-46, 50-59, 63-71
lightning_factory/enums.py	121	0	18	0	100%
lightning_factory/functions.py	4	2	0	0	50%	6-7
TOTAL	187	31	40	0	80%

To run a coverage report

coverage report --format=markdown

History

This code is born from the hyper-parameter tuning portion of Stock Prediction Neural Network and Machine Learning Examples

Future Work

This library is in early stages. Future work involves adding factory methods for LSTMs, RNNs and more.

If you have ideas for this, please fork and contribute!