pytorch-tabular 0.6.0

A standard framework for using Deep Learning for tabular data

PyTorch Tabular aims to make Deep Learning with Tabular data easy and accessible to real-world cases and research alike. The core principles behind the design of the library are:

• Low Resistance Useability
• Easy Customization
• Scalable and Easier to Deploy

It has been built on the shoulders of giants like PyTorch(obviously), and PyTorch Lightning.

Table of Contents

• Installation
• Documentation
• Available Models
• Usage
• Blogs
• Citation

Installation

Although the installation includes PyTorch, the best and recommended way is to first install PyTorch from here, picking up the right CUDA version for your machine.

Once, you have got Pytorch installed, just use:

 pip install pytorch_tabular[all]

to install the complete library with extra dependencies.

And :

 pip install pytorch_tabular

for the bare essentials.

The sources for pytorch_tabular can be downloaded from the Github repo_.

You can either clone the public repository:

git clone git://github.com/manujosephv/pytorch_tabular

Once you have a copy of the source, you can install it with:

python setup.py install

Documentation

For complete Documentation with tutorials visit []

Available Models

• FeedForward Network with Category Embedding is a simple FF network, but with an Embedding layers for the categorical columns.
• Neural Oblivious Decision Ensembles for Deep Learning on Tabular Data is a model presented in ICLR 2020 and according to the authors have beaten well-tuned Gradient Boosting models on many datasets.
• TabNet: Attentive Interpretable Tabular Learning is another model coming out of Google Research which uses Sparse Attention in multiple steps of decision making to model the output.
• Mixture Density Networks is a regression model which uses gaussian components to approximate the target function and provide a probabilistic prediction out of the box.
• AutoInt: Automatic Feature Interaction Learning via Self-Attentive Neural Networks is a model which tries to learn interactions between the features in an automated way and create a better representation and then use this representation in downstream task

To implement new models, see the How to implement new models tutorial. It covers basic as well as advanced architectures.

Usage

from pytorch_tabular import TabularModel
from pytorch_tabular.models import CategoryEmbeddingModelConfig
from pytorch_tabular.config import DataConfig, OptimizerConfig, TrainerConfig, ExperimentConfig

data_config = DataConfig(
    target=['target'], #target should always be a list. Multi-targets are only supported for regression. Multi-Task Classification is not implemented
    continuous_cols=num_col_names,
    categorical_cols=cat_col_names,
)
trainer_config = TrainerConfig(
    auto_lr_find=True, # Runs the LRFinder to automatically derive a learning rate
    batch_size=1024,
    max_epochs=100,
    gpus=1, #index of the GPU to use. 0, means CPU
)
optimizer_config = OptimizerConfig()

model_config = CategoryEmbeddingModelConfig(
    task="classification",
    layers="1024-512-512",  # Number of nodes in each layer
    activation="LeakyReLU", # Activation between each layers
    learning_rate = 1e-3
)

tabular_model = TabularModel(
    data_config=data_config,
    model_config=model_config,
    optimizer_config=optimizer_config,
    trainer_config=trainer_config,
)
tabular_model.fit(train=train, validation=val)
result = tabular_model.evaluate(test)
pred_df = tabular_model.predict(test)
tabular_model.save_model("examples/basic")
loaded_model = TabularModel.load_from_checkpoint("examples/basic")

Blogs

PyTorch Tabular – A Framework for Deep Learning for Tabular Data Neural Oblivious Decision Ensembles(NODE) – A State-of-the-Art Deep Learning Algorithm for Tabular Data Mixture Density Networks: Probabilistic Regression for Uncertainty Estimation

Future Roadmap(Contributions are Welcome)

1. Add GaussRank as Feature Transformation
2. Add ability to use custom activations in CategoryEmbeddingModel
3. Add differential dropouts(layer-wise) in CategoryEmbeddingModel
4. Add Fourier Encoding for cyclic time variables
5. Integrate Optuna Hyperparameter Tuning
6. Add Text and Image Modalities for mixed modal problems
7. Integrate Wide and Deep model
8. Integrate TabTransformer

Citation

If you use PyTorch Tabular for a scientific publication, we would appreciate citations to the published software and the following paper:

• arxiv Paper

@misc{joseph2021pytorch,
      title={PyTorch Tabular: A Framework for Deep Learning with Tabular Data}, 
      author={Manu Joseph},
      year={2021},
      eprint={2104.13638},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

• Zenodo Software Citation

@article{manujosephv_2021, 
    title={manujosephv/pytorch_tabular: v0.5.0-alpha}, 
    DOI={10.5281/zenodo.4732773}, 
    abstractNote={<p>First Alpha Release</p>}, 
    publisher={Zenodo}, 
    author={manujosephv}, 
    year={2021}, 
    month={May}
}

History

0.0.1 (2021-01-26)

• First release on PyPI.

0.2.0 (2021-02-07)

• Fixed an issue with torch.clip and torch version
• Fixed an issue with gpus parameter in TrainerConfig, by setting default value to None for CPU
• Added feature to use custom sampler in the training dataloader
• Updated documentation and added a new tutorial for imbalanced classification

0.3.0 (2021-03-02)

• Fixed a bug on inference

0.4.0 (2021-03-18)

• Added AutoInt Model
• Added Mixture Density Networks
• Refactored the classes to separate backbones from the head of the models
• Changed the saving and loading model to work for custom parameters that you pass in fit

0.5.0 (2021-03-18)

• Added more documentation
• Added Zenodo citation

0.6.0 (2021-06-21)

• Upgraded versions of PyTorch Lightning to 1.3.6
• Changed the way gpus parameter is handled to avoid confusion. None is CPU, -1 is all GPUs, int is number of GPUs
• Added a few more Trainer Params like deterministic, auto_select_gpus
• Some bug fixes and changes to docs
• Added seed_everything to the fit method to ensure reproducibility
• Refactored data_aware_initialization to be part of the BaseModel. Inherited Models can override the method to implement data aware initialization techniques