SymNet

SymNet is a deep learning pipeline with a focus on simplicity. Functionality is available through command-line options or as an API. The focus is on simplicity and getting quick results.

API Usage

Numeric data

The symnet.py file shows how to use the API for multi-class classification on a tabular (CSV) dataset. Start by creating a model:

model = NumericModel(csv_path, n_classes=3, label_column='target', task='classification')

Then, you can call fit and predict on the model, or find the loss and accuracy using the score method.

Image data

Image classifiers inherit from AbstractImageClassificationModel. Currently, only ResNet is implemented. See symnet.py for example usage. Like all models, you can call fit, predict, and score.

CLI Usage

You can use the symnet.py file to run classification on a tabular dataset. The available options are:

--task: One of 'classification' and 'regression'
--dataset: The CSV dataset.
--data-type: As of now, only 'numeric' and 'image' are supported.
--labels: The CSV column with labels
--num-classes: Number of classes (for classification)
--activation: The activation to use. Any of ('relu', 'elu', 'selu', 'sigmoid', 'softmax', 'linear', 'sbaf', 'arelu', 'softplus)
--no-header: Indicates that the CSV does not have a header row
--batch-size: The batch size to use
--train-split: The training data subset split size
--epochs: The number of epochs
--no-balance: Do not rebalance classes in classification problems
--no-augment: For image datasets, do not augment the data

Docker

The Dockerfile in symnet-docker sets up a minimal Debian Stretch system with Python 3.7 and required packages installed. Run it with

docker run -it -v [host-src:]/symnet symnet /bin/bash

This starts up an interactive terminal, mounts a volume at /symnet in the container, and runs a Bash shell. You can change the command run in the last argument.

Todo

Add DenseNet architecture
Add support for text datasets
Add support for image segmentation tasks
Resize and normalize images
For images, use LipschitzLR scheduler

Cite our work

SymNet uses the LipschitzLR learning rate policy: arXiv:1902.07399

BibTeX entry:

@article{yedida2019novel,
  title={A novel adaptive learning rate scheduler for deep neural networks},
  author={Yedida, Rahul and Saha, Snehanshu},
  journal={arXiv preprint arXiv:1902.07399},
  year={2019}
}

SymNet also implements the SBAF and A-ReLU activation functions: arXiv:1906.01975. If you use these, please cite:

@article{saha2019evolution,
  title={Evolution of Novel Activation Functions in Neural Network Training with Applications to Classification of Exoplanets},
  author={Saha, Snehanshu and Nagaraj, Nithin and Mathur, Archana and Yedida, Rahul},
  journal={arXiv preprint arXiv:1906.01975},
  year={2019}
}

sumedhpb/symnet