sealion 3.7

SeaLion is a comprehensive machine learning and data science library for beginners and ml-engineers alike.

SeaLion

SeaLion is designed to teach aspiring data scientists the popular machine learning concepts of today in a way that gives both intution and ways of application. We do this through documentation that explains the models, how they work, and how to use them with examples on familiar datasets like iris, breast cancer, MNIST, etc. Through this library we hope new students will learn common algorithms and ways of implementation for things like regression, unsupervised clusering, dimensionality reduction, ensemble learning, bayesian models, and even deep neural networks. Those more experienced will appreciate the extra morsels of functionality.

Installation

The package is available on PyPi. Install like such :

pip install sealion

SeaLion can only support Python 3 (preferable 3.8), so please make sure you are on the newest version.

Documentation

All documentation is currently being put on a website. However useful it may be, I highly recommend you check the examples posted on GitHub here to see the usage of the APIs and how it works.

General Information

SeaLion was built by Anish Lakkapragada starting in Thanksgiving of 2020 and has continued onto early 2021. The library is meant for beginners to use when solving the standard libraries like iris, breast cancer, swiss roll, the moons dataset, MNIST, etc. The source code is not as much as most other ML libraries (only 4000 lines) and has much more relaxed security with the MIT license so feel free to read it or contribute. I am open to any suggestions.

Updates for v3.0.0!

SeaLion v3.0 and up has had a lot of major milestones.

The first thing is that all the code examples (in jupyter notebooks) for basically all of the modules in sealion are put into the examples directory. Most of them go over using actual datasets like iris, breast cancer, moons, blobs, MNIST, etc. These were all built using v3.0.8 - hopefully that clears up any confusion. I hope you enjoy them.

Perhaps the biggest change in v3.0 is how we have changed the Cython compilation. A quick primer on Cython if you are unfamiliar - you take your python code (in .py files), change it and add some return types and type declarations, put that in a .pyx file, and compile it to a .so file. The .so file is then imported in the python module which you use.

The main bug fixed was that the .so file is actually specific to the architecture of the user. I use macOS and compiled all my files in .so, so prior v3.0 I would just give those .so files to anybody else. However other architectures and OSs like Ubuntu would not be able to recognize those files. Instead what we do know is just store the .pyx files (universal for all computers) in the source code, and the first time you import sealion all of those .pyx files will get compiled into .so files (so they will work for whatever you are using.) This means the first import will take about 40 seconds, but after that it will be as quick as any other import.

Machine Learning Algorithms

The machine learning algorithms of SeaLion are listed below. Please note that the stucture of the listing isn't meant to resemble that of SeaLion's APIs. Of course, new algorithms are being made right now.

1. Deep Neural Networks
   • Optimizers
     • Gradient Descent (and mini-batch gradient descent)
     • Momentum Optimization w/ Nesterov Accelerated Gradient
     • Stochastic gradient descent (w/ momentum + nesterov)
     • AdaGrad
     • RMSprop
     • Adam
     • Nadam
   • Layers
     • Flatten (turn 2D+ data to 2D matrices)
     • Dense (fully-connected layers)
   • Regularization
     • Dropout
   • Activations
     • ReLU
     • Tanh
     • Sigmoid
     • Softmax
     • Leaky ReLU
     • ELU
     • SELU
     • Swish
   • Loss Functions
     • MSE (for regression)
     • CrossEntropy (for classification)
   • Transfer Learning
     • Save weights (in a pickle file)
     • reload them and then enter them into the same neural network
     • this is so you don't have to start training from scratch
2. Regression
   • Linear Regression (Normal Equation, closed-form)
   • Ridge Regression (L2 regularization, closed-form solution)
   • Lasso Regression (L1 regularization)
   • Elastic-Net Regression
   • Logistic Regression
   • Softmax Regression
   • Exponential Regression
   • Polynomial Regression
3. Dimensionality Reduction
   • Principal Component Analysis (PCA)
   • t-distributed Stochastic Neighbor Embedding (tSNE)
4. Unsupervised Clustering
   • KMeans (w/ KMeans++)
   • DBSCAN
5. Naive Bayes
   • Multinomial Naive Bayes
   • Gaussian Naive Bayes
6. Trees
   • Decision Tree (with max_branches, min_samples regularization + CART training)
7. Ensemble Learning
   • Random Forests
   • Ensemble/Voting Classifier
8. Nearest Neighbors
   • k-nearest neighbors
9. Utils
   • one_hot encoder function (one_hot())
   • plot confusion matrix function (confusion_matrix())
   • revert one hot encoding to 1D Array (revert_one_hot())
   • revert softmax predictions to 1D Array (revert_softmax())

Algorithms in progress

Some of the algorithms we are working on right now.

1. Batch Normalization
2. Binomial Naive Bayes
3. Gaussian Mixture Models
4. Barnes Hut t-SNE (please, please contribute for this one)

Contributing

If you feel you can do something better than how it is right now in SeaLion, please do! Believe me, you will find great joy in simplifying my code (probably using numpy) and speeding it up. The major problem right now is speed, some algorithms like PCA can handle 10000+ data points, whereas tSNE is unscalable with O(n^2) time complexity. We have solved this problem with Cython + parallel processing (thanks joblib), so algorithms (aside from neural networks) are working well with <1000 points. Getting to the next level will need some help.

Most of the modules I use are numpy, pandas, joblib, and tqdm. I prefer using less dependencies in the code, so please keep it down to a minimum.

Other than that, thanks for contributing!

Acknowledgements

Plenty of articles and people helped me a long way. Some of the tougher questions I dealt with were Automatic Differentiation in neural networks, in which this tutorial helped me. I also got some help on the O(n^2) time complexity problem of the denominator of t-SNE from this article and understood the mathematical derivation for the gradients (original paper didn't go over it) from here. Lastly special thanks to Evan M. Kim and Peter Washington for helping make the normal equation and cauchy distribution in tSNE make sense. Also thanks to @KentoNishi for helping me understand open-source.

Feedback, comments, or questions

If you have any feedback or something you would like to tell me, please do not hesitate to share! Feel free to comment here on github or reach out to me through [e-mail]!

©Anish Lakkapragada 2021