Updated! My 99.68% solution to Udacity Nanodegree project P2 (Traffic Sign Classification)

There is no change in the previous network structure. However there are some minor changes that leads to the improvement:

  1. Fix one bug in the code and add L2 regularisation. (This was left out in my previous code)
  2. Change last dropout to 0.9 and change data augmentation parameters.

Here is a preliminary draft of my report:
https://github.com/hengck23-udacity/udacity-driverless-car-nd-p2/blob/master/submission(notebook%2Bhtml)/002/writeup_report.pdf

And some interesting observations from my report:

  1. Loss curve

2. Expected probability per class on test set

3. The 40 (0.32%) wrong predictions on the test set

4. Results on internet images

5. Results of network’s data preprocessing layer

Now that I have completed P2, the next project will be P3, Driver Behavioural Cloning. Do stay tuned!