Articles

Behavioral cloning

This is my work for the Behavioral Cloning project of the Udacity Self-driving Cars Nanodegree. A ConvNet-based model was built to autonomously predict steering of a simulated car. The architecture of this model is based on Nvidia's End to End Learning for Self-Driving Cars. However, several adjustments have been made on the architecture to improve the performance on my dataset, and were documented in this documented. Although Udacity has provided some data, I decided to collect my own dataset using Udacity's car simulator. In this project, I have generated a very small dataset of 2700 images on purpose, and augmented it with several data augmentation strategies to train my model. My model has been tested on the basic track provided in the simulator, indicating that data augmentatation can stretch the limit of a small dataset. Results of the model are shown below. This model does not make too far on the challenge track, which is longer and with more sharp turnings. Data augmentataion cannot help on this complex problem, and a much larger dataset are required.

➟ Read more

Traffic sign recognition with ConvNet

In this project, a classifier to recognize traffic sign was built based on Convolutional Neural Networks, which is a project of the Udacity Self-driving Cars Nanodegree program. The model was trained on a fraction of the German Traffic Sign dataset, which is provided as pickled files by the Self-driving car program. The final validation accuracy of our model was 97%, and the test accuracy was 98%. Six images downloaded from the internet were also used to test the model, and a accuracy of 66.7% was achieved.

➟ Read more

Setting up a jupyter notebook server on paperspace

A brief tutorial on how to set up a Jupyter Notebook server on Paperspace. I have been using AWS for Kaggle and for problems of my personal interests. The AWS spot instances are convenient and affordable for individuals in most situations except for deep learning. The costs easily ramp up to $1.00 per hour when I ran a g2.2xlarge instance set up with the deep learning AMI and with sufficient storage. In this case, Paperspace is a good and affordable alternative to AWS.

➟ Read more