CarND_Project_2

Contains the files from Deep Learning Project

Data Set Summary & Exploration

The data is loaded using the pickle package. The train and test data was provided seperately. The code on how to read those is provided in the first cell of the notebook.

The size of the training set: 39209
The size of the testing set: 12630
The shape of a traffic sign image: (32,32,3)
The number of unique classes/labels in the data set: 43

Visualization

Visualizing some of the traffic signs. The code is in cell 6 of the jupyter notebook.

Below is the histogram showing the frequeancy of each of the classes present in the dataset.

Design and Test a Model Architechture

First the training set has been split into a training subset and a validation subset using sklearn's train_test_split() function. The code for that is given below.

#shuffle and then split
from sklearn.utils import shuffle
X_train_gray, y_train = shuffle(X_train_gray, y_train)

from sklearn.model_selection import train_test_split
xTrain,xVal, yTrain,yVal = train_test_split(X_train_gray, y_train, test_size = 0.2, random_state = 144)

The size of the training set: 31367
The size of the validation set: 7842

For data preprocessing, I used the keras library's ImageGenerator.

Below is the code segment which achieves this.

We first center the data according to the suggestion in CS231n course offered at Stanford University. See here.

from keras.preprocessing.image import ImageDataGenerator

image_datagen = ImageDataGenerator(rotation_range=15.,
                                  zoom_range = 0.2,
                                  width_shift_range=0.1,
                                  height_shift_range =0.1,
                                  horizontal_flip = True)

Below is the visualization of a particular traffic sign (randomly selected) which goes through the ImageGenerator.

Neural Net architechture

Number of epochs: 60
Batch Size : 128
mu : 0
sigma : np.sqrt(2.0/n_train) [Following a suggestion at the CS231n course referred earlier]

1.  Convolutional Layer 1: 5 x 5 Convolution (in: 32 x 32 x 1, out: 28 x 28 x 1)
2.  ReLU  
3 . max_pooling (2 x 2)
4.  Convolutional Layer 2: 5 x 5 Convolution (in: 28 x 28 x 1, out: 14 x 14 x 1)
5.  ReLU  
6.  max_pooling (2 x 2)
8.  Flatten layers (flatter layer 1 and flatten layer 2) concatenated 1176 + 400 = 1576
9.  Fully Connected layer 1: (in: 1576, out: 120)
10.  Dropout layer
11.  Fully Connected layer 2: (in: 120, out: 84)
12.  Dropout layer
13. Outerlayer (in: 84, out: 43(no. of classes))
14. Softmax

For training the model, I used the above architecture for the neural net. For optimzation I have used Adam optimizer. I have used regularization in the form of Dropout with keep-probability = 0.5

The final architecture is based on LeNet model that was described in the class videos. However, i have modified the architechture a bit and played around with it a bit to optimize it. In the architecture, I have also used a notion of attaching some of the early layers during the flattening layer (as borrowed from the inception model shown in the class videos on deep learning).

I had tinkered wit the model a lot. But unfortunately, couldn't capture the level of accuracy that the others have obtained or the LeNet architecture has attained.

Training Accuracy; 88.5%
Validation Accuracy: 87.6%

Test

I have used a set of 5 (five) traffic signs that I have donwloaded from the internet and applied my mdoel.

The accuracy on the new test set: 60%

The five pictures are visualized below.

Below is a visualization of the probabilities associated with each prediction.

Comments:

As can be seen, the model isn't performed well. I was expecting a higher accuracy than the validation accuracy (see above) I have obtained. So there is lot of scope to play arround with the model and imrove it. I hope to work on it in future.

scigeek72/CarND_Project_2