Thanks to Sully Chen for the autopilot dataset. https://github.com/SullyChen/driving-datasets The approach to predict steering angle using CNN is based on paper "End to End Learning for Self-Driving Cars" from Nvidia in 2016 https://arxiv.org/pdf/1604.07316.pdf
-
Interactive Colab Notebook which any one can try.
-
Based on latest Tensorflow 2.
-
Dataset augmentation to balance the labels.
-
Loading Dataset of around 20k images into memory to perform quick training.
-
With whole dataset into memory enhance performance using caching or pre-loading (prefetch)
-
Plot Histogrom with converting tensorflow dataset into numpy to save time.
-
Use Large dataset with tfrecord format and shards (distribute shards among multiple workers).
-
Use the image preprocessing functions from tensorflow instead of tf.py_function() which is slow and doesn't use GPU.