# Open Source Projects

I have made contributions to Torch7 deep nets library and ROS. Here are some of my codes, tutorials and contributions.

**Recurrent Batch Normalization of LSTMS**

**Recurrent Batch Normalization of LSTMS**

My implementation of Cooijman et. al's **Recurrent Batch Normalization** in Torch7. Available in Torch 7's Element Research Inc. RNN Library.

**Sorensen-Dice Coefficients in Torch 7**

**Sorensen-Dice Coefficients in Torch 7**

The Sorensen-Dice index measures the degree of similarity between two sample sets. Geiven targets X and Y in two sample datasets, the quotient of similarity is calculated as .

where X and Y are the two samples. X n Y denote the intersection where the elements of X and Y are equal. The resulting quotient is a measure of the similarity between the two samples. It ranges between 0 and 1. If it is 1, the two images are perfectly similar. Otherwise, they are perfectly dissimilar. The input tensor and output tensor are expected to be of the same size when calling forward(input, target) and backward(input, target).

**Savitzky-Golay Filters in C++**

**Savitzky-Golay Filters in C++**

Nicely computes the Vandermonde matrix, Savitzky-Golay differentiation filters and smoothing coefficients for a noisy, sequential signal. It is a textbook implementation of the Savitzky-Golay filter algorithm following Sophocles Orfanidis treatment in his book.

**Fully Automated Deep Recurrent Neural Networks for Nonlinear Identification and Control **

**Fully Automated Deep Recurrent Neural Networks for Nonlinear Identification and Control**

Using stacked layers of neural networks, I trained multilayer perceptrons, simple recurrent neural networks, long short-term memory cells, fast LSTMS (with gates computed in one fell-swoop) to model the relationship between input and output data.

Please check out the **"soft-robot" tag** only. Other branches for other identification datasets are actively being developed.

**Identifying and Segmenting Nerve Structures in Ultrasound Images of the Neck**

**Identifying and Segmenting Nerve Structures in Ultrasound Images of the Neck**

This was my code entry for Kaggle's ultrasound data science challenge of August 2016. The problem is well-described on the competition page. This code was developed in Torch 7. My convolutional architecture is strongly inspired by U-Net, Alec Radford and Soumith Chintala's DcGAN and the all-convolutional net of Springenberg et al.

Code here**Linear Regression in Torch [Torch Demos]**

**Linear Regression in Torch [Torch Demos]**

Part of the Torch demos/tutorials

**Adaptive Monte Carlo Localization and Mapping for the Navigation of an Unmanned Ground Vehicle**

**Adaptive Monte Carlo Localization and Mapping for the Navigation of an Unmanned Ground Vehicle**

**ROS Image Transport Subscriber to Kinect2 ROS Drivers**

**ROS Image Transport Subscriber to Kinect2 ROS Drivers**

My implemention of approximate sync policies within the message filters class of ROS Image Transport communication protocol. This code subscribes to the color_rect, depth_rect and creates a point cloud visualizer from the depth and colored images.