GITHUB PROJECTS
Datasets and Utilities
nba_pbp_video_dataset
This repo allows you to acquire a nearly 2TB video dataset of aproximately 500,000 small 10-15 second NBA play mp4 video files. Each clip comes from a regular season or playoff game over the past two seasons and has an accompanying XML label describing the type of event, players involved, and time of occurrence. To find out more about this dataset, you can check out this blog post.
distrib_training_scripts
This repo contains a set of scripts and Keras example training modules to set-up and ultimately run a distributed training session across multiple AWS nodes. This repo requires a system with a multitude of frameworks including Horovod, OpenMPI, Tensorflow, Keras, NCCL, and CUDA. You can install these components yourself or follow this blog post for detailed information on quickly setting up a distributed cluster on AWS.
dl_utilities
This repo features a collection of common functionality for deep learning on Keras. It includes (but is not limited to) code for data augmentation, custom batch generation, new/unsupported DL layers/activations, and evaluation functions for model results. You can find the code for some of the repo's custom Keras functionality with an accompanying explanation in this blog post.
Famous DNN Implementations
state_of_art_cnns
This repo holds a set of famous CNN implementations including ones for DenseNet, PolyNet, and Dual Path Networks. It also features a 'tester.py' script to play around with configuring and running each implemented model.
video_processing_networks
This repo contains a single video processing DNN based on the Predictive Correct Networks paper. While I have elected to not yet release the other custom video processing models, they may be released in the coming months.
time_series_models
This repo has implementations for models for processing sequential data input such as music, words, letters, or even images in a video. Some of the models include the Adaptive Compute Time (by Alex Graves), ByteNet, WaveNet, and Recurrent Highway Network. Like in the 'state_of_art_cnns' repo, there is a 'tester.py' script to allow users to quickly run a majority of these models with various configurations.
vqa_models
This repo contains a single Visual Question Anwsering (VQA) model based off an excellent DeepMind paper called "Relational Reasoning Networks". At some future point, this repo will likely be renamed and refactored to feature implementations of recent extensions of the original paper's high level idea (including the "Temporal Relational Reasoning in Videos" and "Visual Interaction Network" papers).
Miscellaneous
final_udacity_capstone
This repo includes the code and research results from my Udacity Nanodegree capstone project. It covers the effects of using multiple loss functions and optimizers to inject more diversity in local minima of snapshot models in a Snapshot Ensemble.
menu_calculator
This repo has a Java project for restaurant owners to calculate the exact cost of items/dishes on their menus (using the dishes' recipes). In the repo, you'll find source code and a pre-compiled executable jar to quickly have access to the GUI.