README.md

PyTorch TabNet tool(v0.1.0-dev0)

This tool uses tabnet, a deep learning model designed for tabular data structured in rows and columns. TabNet is suitable for classification, regression, and multi-task learning.

Inputs:

Input data:

The input tabular data that need to be trained. This plugin supports .csv, .featherand .arrow file formats

Details:

PyTorch-TabNet can be employed for:.

1. TabNetClassifier: For binary and multi-class classification problems
2. TabNetRegressor: For simple and multi-task regression problems
3. TabNetMultiTaskClassifier: multi-task multi-classification problems

Building

To build the Docker image for the conversion plugin, run ./build-docker.sh.

Install WIPP Plugin

If WIPP is running, navigate to the plugins page and add a new plugin. Paste the contents of plugin.json into the pop-up window and submit. For more information on WIPP, visit the official WIPP page.

Options

This plugin takes 38 input arguments and one output argument:

Name	Description	I/O	Type
--inpdir	Input tabular data	Input	genericData
--filePattern	Pattern to parse tabular files	Input	string
--testSize	Proportion of the dataset to include in the test set	Input	number
--nD	Width of the decision prediction layer	Input	integer
--nA	Width of the attention embedding for each mask	Input	integer
--nSteps	Number of steps in the architecture	Input	integer
--gamma	Coefficient for feature reuse in the masks	Input	number
--catEmbDim	List of embedding sizes for each categorical feature	Input	integer
--nIndependent	Number of independent Gated Linear Unit layers at each step	Input	integer
--nShared	Number of shared Gated Linear Unit layers at each step	Input	integer
--epsilon	Constant value	Input	number
--seed	Random seed for reproducibility	Input	integer
--momentum	Momentum for batch normalization	Input	number
--clipValue	Clipping of the gradient value	Input	number
--lambdaSparse	Extra sparsity loss coefficient	Input	number
--optimizerFn	Pytorch optimizer function	Input	enum
--lr	learning rate for the optimizer	Input	number
--schedulerFn	Parameters used initialize the optimizer	Input	enum
--stepSize	Parameter to apply to the scheduler_fn	Input	integer
--deviceName	Platform used for training	Input	enum
--maskType	A masking function for feature selection	Input	enum
--groupedFeatures	Allow the model to share attention across features within the same group	Input	integer
--nSharedDecoder	Number of shared GLU block in decoder	Input	integer
--nIndepDecoder	Number of independent GLU block in decoder	Input	integer
--evalMetric	Metrics utilized for early stopping evaluation	Input	enum
--maxEpochs	Maximum number of epochs for training	Input	integer
--patience	Consecutive epochs without improvement before early stopping	Input	integer
--weights	Sampling parameter only for TabNetClassifier	Input	integer
--lossFn	Loss function	Input	enum
--batchSize	Batch size	Input	integer
--virtualBatchSize	Size of mini-batches for Ghost Batch Normalization	Input	integer
--numWorkers	Number or workers used in torch.utils.data.Dataloader	Input	integer
--dropLast	Option to drop incomplete last batch during training	Input	boolean
--warmStart	For scikit-learn compatibility, enabling fitting the same model twice	Input	boolean
--targetVar	Target feature containing classification labels	Input	string
--computeImportance	Compute feature importance	Input	boolean
--classifier	Pytorch tabnet Classifier for training	Input	enum
--preview	Generate JSON file of sample outputs	Input	boolean
--outdir	Output collection	Output	genericData