Self-Driving Car Engineer Nanodegree

Deep Learning

Project: Build a Traffic Sign Recognition Classifier

This notebook contains the code and accompanying explanation for a convoutional neural network (CNN) used to classify german traffic signs using the German Traffic Sign Dataset. The network is loosley based on the well-known LeNet architecture. ReLU activation functions are used throughout with three layers of convolution. In addition, 1x1 layers (to increase the depth), maxpooling, and dropout for regularization are used after each convolutional layer. These three convolutional layers are followed by 2 fully connected layers (with dropout) and finally the last layer to compute the logits.

In addition to the architecture, data augmentation (rotations, shadows, translations) are applied to the original data set to increase the size of the training data and to force the network to generalize

The CNN has an accuracy of 97% when applied to the test data.

---

Step 0: Load The Data

from tensorflow.examples.tutorials.mnist import input_data
from sklearn.model_selection import train_test_split
import pickle
import pdb
import numpy as np
# pickle file locations
training_file = 'train.p'
testing_file = 'test.p'

# load pickle data
with open(training_file, mode='rb') as f:
    train = pickle.load(f)
with open(testing_file, mode='rb') as f:
    test = pickle.load(f)
    
X_train, y_train = train['features'], train['labels']
X_test, y_test = test['features'], test['labels']

assert(len(X_train) == len(y_train))
assert(len(X_test) == len(y_test))

---

Step 1: Dataset Summary & Exploration

The pickled data is a dictionary with 4 key/value pairs:

• 'features' is a 4D array containing raw pixel data of the traffic sign images, (num examples, width, height, channels).
• 'labels' is a 2D array containing the label/class id of the traffic sign. The file signnames.csv contains id -> name mappings for each id.
• 'sizes' is a list containing tuples, (width, height) representing the the original width and height the image.
• 'coords' is a list containing tuples, (x1, y1, x2, y2) representing coordinates of a bounding box around the sign in the image. THESE COORDINATES ASSUME THE ORIGINAL IMAGE. THE PICKLED DATA CONTAINS RESIZED VERSIONS (32 by 32) OF THESE IMAGES

Complete the basic data summary below.

### Replace each question mark with the appropriate value.

# Number of training examples
n_train = len(y_train) # Before spliting training data into train and validation data. See cell below for post splitting lengths


# Number of testing examples.
n_test = len(y_test) 

# Shape of traffic sign image
image_shape = np.array(X_train.shape[1:4])#[32, 32, 3]

# Number of unique classes/labels there are in the dataset.
n_classes = np.unique(y_train).size #43

print("Number of training examples =", n_train)
print("Number of testing examples =", n_test)
print("Image data shape =", image_shape)
print("Number of classes =", n_classes)

Number of training examples = 39209
Number of testing examples = 12630
Image data shape = [32 32  3]
Number of classes = 43

Visualize the German Traffic Signs Dataset using the pickled file(s). This is open ended, suggestions include: plotting traffic sign images, plotting the count of each sign, etc.

The Matplotlib examples and gallery pages are a great resource for doing visualizations in Python.

NOTE: It's recommended you start with something simple first. If you wish to do more, come back to it after you've completed the rest of the sections.

### Data exploration visualization goes here.
### Feel free to use as many code cells as needed.
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
import random
import numpy as np
# Visualizations will be shown in the notebook.
%matplotlib inline

#plot a bunch of random examples
subplot_width = 10
plt.figure(figsize=(15,15))
for index in range(subplot_width**2):
    plt.subplot(subplot_width,subplot_width,index+1)
    index = random.randint(0, len(X_train))
    image = X_train[index].squeeze()
    plt.imshow(image)
    plt.xticks([])
    plt.yticks([])

# plot occurances
unique, counts = np.unique(y_train, return_counts=True)
df=pd.read_csv('/home/myles-iribarne/Documents/Self-Driving-Car-Nanodegree-Term1/CarND-Traffic-Sign-Classifier-Project/signnames.csv')
df_names = df['SignName']
df_name_array = df_names.values
# import pdb; pdb.set_trace()
# df_names = df('ID')
# y_pos = np.arange(len(objects))
# performance = [10,8,6,4,2,1]
plt.figure(figsize=(15,15))
plt.barh(unique,counts)
plt.yticks(unique, df_name_array)
plt.xlabel('Counts')
plt.title('Occurance of each image in Training Set')
 


# plt.figure(figsize=(1,1))
# plt.imshow(edges)
print(y_train[index])
print(X_train.shape)

6
(39209, 32, 32, 3)

---

Step 2: Design and Test a Model Architecture

Design and implement a deep learning model that learns to recognize traffic signs. Train and test your model on the German Traffic Sign Dataset.

There are various aspects to consider when thinking about this problem:

• Neural network architecture
• Play around preprocessing techniques (normalization, rgb to grayscale, etc)
• Number of examples per label (some have more than others).
• Generate fake data.

Here is an example of a published baseline model on this problem. It's not required to be familiar with the approach used in the paper but, it's good practice to try to read papers like these.

NOTE: The LeNet-5 implementation shown in the classroom at the end of the CNN lesson is a solid starting point. You'll have to change the number of classes and possibly the preprocessing, but aside from that it's plug and play!

My implementation

Preprocess Data

• Convert to Grayscale using: cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
• Equalize the histogram using: cv2.equalizeHist(img)
• Normalized the image using: cv2.normalize(image, image, alpha=-0.5, beta=0.5, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)

import cv2
from sklearn.utils import shuffle
import numpy as np

def grayscale(img):
    """Applies the Grayscale transform
    This will return an image with only one color channel
    but NOTE: to see the returned image as grayscale
    you should call plt.imshow(gray, cmap='gray')"""
    return cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
def norml(image):
    #normalize to between -0.5 and 0.5 and output in float
    cv2.normalize(image, image, alpha=-0.5, beta=0.5, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F) 
    return image

def preprocess(images):
    images = np.array([norml(cv2.equalizeHist(grayscale(image))) for image in images])
    images = images[..., np.newaxis]
    return images
    

X_train = preprocess(X_train)
X_test = preprocess(X_test)
# Covert to grayscale and normalize histogram and scale to 0.5 to 0.5
# X_train = np.array([preprocess(image) for image in X_train])
# X_train = X_train[..., np.newaxis]
# X_test = np.array([preprocess(image) for image in X_test])
# X_test = X_test[..., np.newaxis]


# shuffle input data
X_train, y_train = shuffle(X_train, y_train)

Question 1

Describe how you preprocessed the data. Why did you choose that technique?

Answer: To preprocess the data, I used the following pipeline:

• Convert to Grayscale using: cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
• Equalize the histogram using: cv2.equalizeHist(img)
• Normalized the image using: cv2.normalize(image, image, alpha=-0.5, beta=0.5, norm_type=cv2.NORM_MINMAX, dtype=cv2.CV_32F)

I chose to use grayscale because the published result "Traffic Sign Recognition with Multi-Scale Convolutional Networks" showed that it performed better than color. I also tested this myself and did not notice a significant difference in performance, so I chose the simpler of the two options

Equalizing the histogram and normalizing the data both serve the purpose of standardizing the images somewhat to make them more comparable to each other.

### and split the data into training/validation/testing sets here.
X_train, X_validation, y_train, y_validation = train_test_split(X_train, y_train, test_size = 0.2, random_state = 0)


assert(len(X_train) == len(y_train))
assert(len(X_validation) == len(y_validation))



# Number of training examples
n_train = len(y_train) 
n_validation = len(y_validation)

# Number of testing examples.
n_test = len(y_test) #12630

print("Number of training examples =", n_train)
print("Number of validation examples =", n_validation)
print("Number of testing examples =", n_test)

Number of training examples = 31367
Number of validation examples = 7842
Number of testing examples = 12630

Question 2

Describe how you set up the training, validation and testing data for your model. Optional: If you generated additional data, how did you generate the data? Why did you generate the data? What are the differences in the new dataset (with generated data) from the original dataset?

Answer: The training data and the validation data were both taken from training set, with 20% reserved for validation.

I do generate new data, but not in this code cell. When I train my model, I leveraged a function that Vivek Yadav wrote that uses cv2 transformations on the images to warp the image by translating it, shearing it, and rotating it. Each epoch takes the entire training set and applies a random set of transformations to this effect. The intent is to make the network robust to these transformations by feeding it a large variety of them. I should say that I did speak with Yadav and he was happy to have me use his code for the data augmentation.

It is also worth noting that the data visualization above shows that some traffic signs have many more training data points than others. It might be worth considering generating additional images to even out this distribution. This is future work for me.

### Define Architecture
import tensorflow as tf
from tensorflow.contrib.layers import flatten


def conv2d(x, W, b, strides=1):
    # Conv2D wrapper, with bias and relu activation
    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return x

def fully_connected(x,W,b):
    x = tf.add(
    tf.matmul(x, W),
    b)
    return x

def maxpool2d(x, k=2):
    return tf.nn.max_pool(
        x,
        ksize=[1, k, k, 1],
        strides=[1, k, k, 1],
        padding='VALID')

#parameters
input_depth = 1; # rgb = 3, grayscale = 1
output_depth =43; #43 traffic sign
    
# Hyperparameters
mu = 0 # initialization mean
sigma = 0.1 # initialization stddev

#placeholders
x = tf.placeholder(tf.float32, (None, 32, 32, input_depth)) # placeholder for a batch of input images
y = tf.placeholder(tf.int32, (None)) # placeholder for a batch of output labels.
one_hot_y = tf.one_hot(y, output_depth) #output labels one hot encoded
keep_prob_convlayers = tf.placeholder(tf.float32) #placeholder for convolution layer dropout
keep_prob_connected = tf.placeholder(tf.float32) #placeholder for fully connected dropout

# Layer Depths
layer_depth = {
    'layer_1': 32,
    'layer_1_1x1': 16,
    'layer_2': 64, 
    'layer_2_1x1': 32,
    'layer_3': 128, 
    'layer_3_1x1': 64,
    'fully_connected_1': 1024,
    'fully_connected_2': 512,
    'output_layer': output_depth,
}

weights = {
    'layer_1': tf.Variable(tf.truncated_normal(
        [3, 3, input_depth, layer_depth['layer_1']], mean = mu, stddev = sigma)), 
    'layer_1_1x1': tf.Variable(tf.truncated_normal(
        [1,1, layer_depth['layer_1'], layer_depth['layer_1_1x1']], mean = mu, stddev = sigma)), 
    'layer_2': tf.Variable(tf.truncated_normal(
        [3, 3, layer_depth['layer_1_1x1'], layer_depth['layer_2']], mean = mu, stddev = sigma)),
    'layer_2_1x1': tf.Variable(tf.truncated_normal(
        [1,1, layer_depth['layer_2'], layer_depth['layer_2_1x1']], mean = mu, stddev = sigma)), 
    'layer_3': tf.Variable(tf.truncated_normal(
        [3, 3, layer_depth['layer_2_1x1'], layer_depth['layer_3']], mean = mu, stddev = sigma)),
    'layer_3_1x1': tf.Variable(tf.truncated_normal(
        [1,1, layer_depth['layer_3'], layer_depth['layer_3_1x1']], mean = mu, stddev = sigma)), 
    'fully_connected_1': tf.Variable(tf.truncated_normal(
        [4*4*layer_depth['layer_3_1x1'], layer_depth['fully_connected_1']], mean = mu, stddev = sigma)),
    'fully_connected_2': tf.Variable(tf.truncated_normal(
        [layer_depth['fully_connected_1'], layer_depth['fully_connected_2']], mean = mu, stddev = sigma)),
    'output_layer': tf.Variable(tf.truncated_normal(
        [layer_depth['fully_connected_2'], layer_depth['output_layer']], mean = mu, stddev = sigma))
}
biases = {
    'layer_1': tf.Variable(tf.zeros(layer_depth['layer_1'])),
    'layer_1_1x1': tf.Variable(tf.zeros(layer_depth['layer_1_1x1'])),
    'layer_2': tf.Variable(tf.zeros(layer_depth['layer_2'])),
    'layer_2_1x1': tf.Variable(tf.zeros(layer_depth['layer_2_1x1'])),
    'layer_3': tf.Variable(tf.zeros(layer_depth['layer_3'])),
    'layer_3_1x1': tf.Variable(tf.zeros(layer_depth['layer_3_1x1'])),
    'fully_connected_1': tf.Variable(tf.zeros(layer_depth['fully_connected_1'])),
    'fully_connected_2': tf.Variable(tf.zeros(layer_depth['fully_connected_2'])),
    'output_layer': tf.Variable(tf.zeros(layer_depth['output_layer']))
}

def Net(x):
    # Layer 1: Convolutional. 3x3x32
    x = conv2d(x, weights['layer_1'], biases['layer_1'], strides=1)
    # Activation.
    x = tf.nn.relu(x)
    # Convolutional 1x1x16
    x = conv2d(x,weights['layer_1_1x1'], biases['layer_1_1x1'], strides=1)
    # Activation.
    x = tf.nn.relu(x)
    # Max Pooling. 2x2
    x = maxpool2d(x)
    # Dropout
    x = tf.nn.dropout(x, keep_prob_convlayers)

    # Layer 2: Convolutional. 3x3x64
    x = conv2d(x, weights['layer_2'], biases['layer_2'], strides=1) 
    # Activation.
    x = tf.nn.relu(x)
    # Convolutional 1x1x32
    x = conv2d(x,weights['layer_2_1x1'], biases['layer_2_1x1'], strides=1)
    # Activation.
    x = tf.nn.relu(x)
    # Max Pooling. 2x2
    x = maxpool2d(x)
    # use dropout
    x = tf.nn.dropout(x, keep_prob_convlayers)

    # Layer 3: Convolutional. 3x3x128
    x = conv2d(x, weights['layer_3'], biases['layer_3'], strides=1) 
    # Activation.
    x = tf.nn.relu(x)
    #1x1 convolution, Output 1x1x64
    x = conv2d(x,weights['layer_3_1x1'], biases['layer_3_1x1'], strides=1)
    # Activation.
    x = tf.nn.relu(x)
    # Max Pooling. 2x2.
    x = maxpool2d(x)
    # Dropout
    x = tf.nn.dropout(x, keep_prob_convlayers)



    # Flatten. Output is 1024
    x = flatten(x)

    # Layer 3: Fully Connected, output nuerons 1024

    x = fully_connected(x, weights['fully_connected_1'], biases['fully_connected_1'])
    # Activation.
    x = tf.nn.relu(x)
    # Dropout
    x = tf.nn.dropout(x, keep_prob_connected)

    # Layer 4: Fully Connected. output nuerons 512
    x = fully_connected(x, weights['fully_connected_2'], biases['fully_connected_2'])  
    # Activation.
    x = tf.nn.relu(x)
    # Dropout
    x = tf.nn.dropout(x, keep_prob_connected)

    # Layer 5: Fully Connected. output of size 43
    logits = fully_connected(x, weights['output_layer'], biases['output_layer'])  
    return logits

Question 3

What does your final architecture look like? (Type of model, layers, sizes, connectivity, etc.) For reference on how to build a deep neural network using TensorFlow, see Deep Neural Network in TensorFlow from the classroom.

Answer: My final architecture is outlined below. I use ReLU activation functions throughout, with three layers of convolution, each of these containing an additional 1x1 convolution to increase the depth, maxpooling, and dropout for regularization. These three convolutional layers are followed by 2 fully connected layers (with dropout) and finally the last layer to compute the logits.

Define Network Structure

**Layer 1: **

Convolutional. 3x3x32

Activation. ReLU

Convolutional. 1x1x16

Activation. ReLU

Maxpool. 2x2

Dropout. Keep_Prob = 0.6

---

**Layer 2: **

Convolutional. 3x3x64

Activation. ReLU

Convolutional. 1x1x32

Activation. ReLU

Maxpool. 2x2

Dropout. Keep_Prob = 0.6

---

**Layer 3: **

Convolutional. 3x3x128

Activation. ReLU

Convolutional. 1x1x64

Activation. ReLU

Maxpool. 2x2

Dropout. Keep_Prob = 0.6

---

**Layer 4: **

Fully Connected. 1024 outputs

Activation. ReLU

Dropout. Keep_Prob = 0.5

---

**Layer 5: **

Fully Connected. 512 outputs

Activation. ReLU

Dropout. Keep_Prob = 0.5

---

**Layer 6: **

Fully Connected (Logits). 43 outputs

Training Pipeline

Create a training pipeline that uses the model to classify data.

# Hyper Parameters
rate = 0.001 # learning rate
beta = 0.001 # regularization penalty

#define loss function
logits = Net(x)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits, one_hot_y)

# l2 regularization penalty
l2_reg = beta*tf.nn.l2_loss(weights['layer_1']) + beta*tf.nn.l2_loss(weights['layer_1_1x1']) + \
    beta*tf.nn.l2_loss(weights['layer_2']) + \
    beta*tf.nn.l2_loss(weights['layer_2_1x1']) + \
    beta*tf.nn.l2_loss(weights['layer_3']) + \
    beta*tf.nn.l2_loss(weights['fully_connected_1']) + \
    beta*tf.nn.l2_loss(weights['fully_connected_2']) + \
    beta*tf.nn.l2_loss(weights['output_layer']) + \
    beta*tf.nn.l2_loss(biases['layer_1']) + \
    beta*tf.nn.l2_loss(biases['layer_1_1x1']) + \
    beta*tf.nn.l2_loss(biases['layer_2']) + \
    beta*tf.nn.l2_loss(biases['layer_2_1x1']) + \
    beta*tf.nn.l2_loss(biases['layer_3']) + \
    beta*tf.nn.l2_loss(biases['fully_connected_1']) + \
    beta*tf.nn.l2_loss(biases['fully_connected_2']) + \
    beta*tf.nn.l2_loss(biases['output_layer'])
                               
loss_operation = tf.reduce_mean(cross_entropy + l2_reg)
# loss_operation = tf.reduce_mean(cross_entropy)
optimizer = tf.train.AdamOptimizer(learning_rate = rate)
training_operation = optimizer.minimize(loss_operation)

Model Evaluation

Evaluate how well the loss and accuracy of the model for a given dataset.

correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(one_hot_y, 1))
accuracy_operation = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

def evaluate(X_data, y_data):
    num_examples = len(X_data)
    total_accuracy = 0
    sess = tf.get_default_session()
    for offset in range(0, num_examples, BATCH_SIZE):
        batch_x, batch_y = X_data[offset:offset+BATCH_SIZE], y_data[offset:offset+BATCH_SIZE]
        accuracy = sess.run(accuracy_operation, feed_dict={x: batch_x, y: batch_y, keep_prob_convlayers: 1., keep_prob_connected: 1.})
        total_accuracy += (accuracy * len(batch_x))
    return total_accuracy / num_examples

Train the Model

Run the training data through the training pipeline to train the model.

Before each epoch: - shuffle the training set. - distort data by applying random rotatinos, translations, and shears

After each epoch, measure the loss and accuracy of the validation set.

Save the model after training.

# Hyperparameters
EPOCHS = 150
BATCH_SIZE = 256
def transform_image(img,ang_range,shear_range,trans_range,brightness=0):
    '''
    Function written by Vivek Yadav.
    
    This function transforms images to generate new images.
    The function takes in following arguments,
    1- Image
    2- ang_range: Range of angles for rotation
    3- shear_range: Range of values to apply affine transform to
    4- trans_range: Range of values to apply translations over.

    A Random uniform distribution is used to generate different parameters for transformation

    '''
    # Rotation

    ang_rot = np.random.uniform(ang_range)-ang_range/2
    rows,cols,ch = img.shape    
    Rot_M = cv2.getRotationMatrix2D((cols/2,rows/2),ang_rot,1)

    # Translation
    tr_x = trans_range*np.random.uniform()-trans_range/2
    tr_y = trans_range*np.random.uniform()-trans_range/2
    Trans_M = np.float32([[1,0,tr_x],[0,1,tr_y]])

    # Shear
    pts1 = np.float32([[5,5],[20,5],[5,20]])

    pt1 = 5+shear_range*np.random.uniform()-shear_range/2
    pt2 = 20+shear_range*np.random.uniform()-shear_range/2


    pts2 = np.float32([[pt1,5],[pt2,pt1],[5,pt2]])

    shear_M = cv2.getAffineTransform(pts1,pts2)

    img = cv2.warpAffine(img,Rot_M,(cols,rows))
    img = cv2.warpAffine(img,Trans_M,(cols,rows))
    img = cv2.warpAffine(img,shear_M,(cols,rows))

    return img
    


with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:
    sess.run(tf.global_variables_initializer())
    num_examples = len(X_train)

    print("Training...")
    print()
    for i in range(EPOCHS):
        
        X_train, y_train = shuffle(X_train, y_train)
        # distort image
        X_train1 = np.array([transform_image(image,15,5,2) for image in X_train])
        X_train1 = X_train1[..., np.newaxis]
        
        for offset in range(0, num_examples, BATCH_SIZE):
            end = offset + BATCH_SIZE
            batch_x, batch_y = X_train1[offset:end], y_train[offset:end]
            sess.run(training_operation, feed_dict={x: batch_x, y: batch_y, keep_prob_convlayers: 0.7, keep_prob_connected: 0.5})
        validation_accuracy = evaluate(X_validation, y_validation)
        print("EPOCH {} ...".format(i+1))
        print("Validation Accuracy = {:.3f}".format(validation_accuracy))
        print()
        
    try:
        saver
    except NameError:
        saver = tf.train.Saver()
    saver.save(sess, 'Traffic_Sign_Classifying_Net')
    
    print("Model saved")

Training...

EPOCH 1 ...
Validation Accuracy = 0.056

EPOCH 2 ...
Validation Accuracy = 0.056

EPOCH 3 ...
Validation Accuracy = 0.060

EPOCH 4 ...
Validation Accuracy = 0.056

EPOCH 5 ...
Validation Accuracy = 0.057

EPOCH 6 ...
Validation Accuracy = 0.058

EPOCH 7 ...
Validation Accuracy = 0.056

EPOCH 8 ...
Validation Accuracy = 0.057

EPOCH 9 ...
Validation Accuracy = 0.052

EPOCH 10 ...
Validation Accuracy = 0.056

EPOCH 11 ...
Validation Accuracy = 0.061

EPOCH 12 ...
Validation Accuracy = 0.139

EPOCH 13 ...
Validation Accuracy = 0.244

EPOCH 14 ...
Validation Accuracy = 0.302

EPOCH 15 ...
Validation Accuracy = 0.407

EPOCH 16 ...
Validation Accuracy = 0.517

EPOCH 17 ...
Validation Accuracy = 0.617

EPOCH 18 ...
Validation Accuracy = 0.680

EPOCH 19 ...
Validation Accuracy = 0.743

EPOCH 20 ...
Validation Accuracy = 0.760

EPOCH 21 ...
Validation Accuracy = 0.842

EPOCH 22 ...
Validation Accuracy = 0.858

EPOCH 23 ...
Validation Accuracy = 0.869

EPOCH 24 ...
Validation Accuracy = 0.895

EPOCH 25 ...
Validation Accuracy = 0.910

EPOCH 26 ...
Validation Accuracy = 0.919

EPOCH 27 ...
Validation Accuracy = 0.931

EPOCH 28 ...
Validation Accuracy = 0.934

EPOCH 29 ...
Validation Accuracy = 0.935

EPOCH 30 ...
Validation Accuracy = 0.944

EPOCH 31 ...
Validation Accuracy = 0.951

EPOCH 32 ...
Validation Accuracy = 0.941

EPOCH 33 ...
Validation Accuracy = 0.953

EPOCH 34 ...
Validation Accuracy = 0.960

EPOCH 35 ...
Validation Accuracy = 0.960

EPOCH 36 ...
Validation Accuracy = 0.960

EPOCH 37 ...
Validation Accuracy = 0.955

EPOCH 38 ...
Validation Accuracy = 0.971

EPOCH 39 ...
Validation Accuracy = 0.961

EPOCH 40 ...
Validation Accuracy = 0.969

EPOCH 41 ...
Validation Accuracy = 0.968

EPOCH 42 ...
Validation Accuracy = 0.973

EPOCH 43 ...
Validation Accuracy = 0.979

EPOCH 44 ...
Validation Accuracy = 0.973

EPOCH 45 ...
Validation Accuracy = 0.974

EPOCH 46 ...
Validation Accuracy = 0.976

EPOCH 47 ...
Validation Accuracy = 0.976

EPOCH 48 ...
Validation Accuracy = 0.980

EPOCH 49 ...
Validation Accuracy = 0.978

EPOCH 50 ...
Validation Accuracy = 0.975

EPOCH 51 ...
Validation Accuracy = 0.982

EPOCH 52 ...
Validation Accuracy = 0.980

EPOCH 53 ...
Validation Accuracy = 0.983

EPOCH 54 ...
Validation Accuracy = 0.981

EPOCH 55 ...
Validation Accuracy = 0.981

EPOCH 56 ...
Validation Accuracy = 0.982

EPOCH 57 ...
Validation Accuracy = 0.981

EPOCH 58 ...
Validation Accuracy = 0.984

EPOCH 59 ...
Validation Accuracy = 0.984

EPOCH 60 ...
Validation Accuracy = 0.987

EPOCH 61 ...
Validation Accuracy = 0.984

EPOCH 62 ...
Validation Accuracy = 0.984

EPOCH 63 ...
Validation Accuracy = 0.980

EPOCH 64 ...
Validation Accuracy = 0.985

EPOCH 65 ...
Validation Accuracy = 0.984

EPOCH 66 ...
Validation Accuracy = 0.987

EPOCH 67 ...
Validation Accuracy = 0.983

EPOCH 68 ...
Validation Accuracy = 0.986

EPOCH 69 ...
Validation Accuracy = 0.987

EPOCH 70 ...
Validation Accuracy = 0.986

EPOCH 71 ...
Validation Accuracy = 0.990

EPOCH 72 ...
Validation Accuracy = 0.985

EPOCH 73 ...
Validation Accuracy = 0.986

EPOCH 74 ...
Validation Accuracy = 0.988

EPOCH 75 ...
Validation Accuracy = 0.987

EPOCH 76 ...
Validation Accuracy = 0.988

EPOCH 77 ...
Validation Accuracy = 0.988

EPOCH 78 ...
Validation Accuracy = 0.990

EPOCH 79 ...
Validation Accuracy = 0.985

EPOCH 80 ...
Validation Accuracy = 0.988

EPOCH 81 ...
Validation Accuracy = 0.990

EPOCH 82 ...
Validation Accuracy = 0.992

EPOCH 83 ...
Validation Accuracy = 0.990

EPOCH 84 ...
Validation Accuracy = 0.990

EPOCH 85 ...
Validation Accuracy = 0.985

EPOCH 86 ...
Validation Accuracy = 0.990

EPOCH 87 ...
Validation Accuracy = 0.989

EPOCH 88 ...
Validation Accuracy = 0.991

EPOCH 89 ...
Validation Accuracy = 0.991

EPOCH 90 ...
Validation Accuracy = 0.989

EPOCH 91 ...
Validation Accuracy = 0.989

EPOCH 92 ...
Validation Accuracy = 0.987

EPOCH 93 ...
Validation Accuracy = 0.991

EPOCH 94 ...
Validation Accuracy = 0.985

EPOCH 95 ...
Validation Accuracy = 0.992

EPOCH 96 ...
Validation Accuracy = 0.988

EPOCH 97 ...
Validation Accuracy = 0.990

EPOCH 98 ...
Validation Accuracy = 0.991

EPOCH 99 ...
Validation Accuracy = 0.991

EPOCH 100 ...
Validation Accuracy = 0.991

EPOCH 101 ...
Validation Accuracy = 0.987

EPOCH 102 ...
Validation Accuracy = 0.990

EPOCH 103 ...
Validation Accuracy = 0.993

EPOCH 104 ...
Validation Accuracy = 0.992

EPOCH 105 ...
Validation Accuracy = 0.990

EPOCH 106 ...
Validation Accuracy = 0.990

EPOCH 107 ...
Validation Accuracy = 0.992

EPOCH 108 ...
Validation Accuracy = 0.992

EPOCH 109 ...
Validation Accuracy = 0.993

EPOCH 110 ...
Validation Accuracy = 0.991

EPOCH 111 ...
Validation Accuracy = 0.991

EPOCH 112 ...
Validation Accuracy = 0.992

EPOCH 113 ...
Validation Accuracy = 0.992

EPOCH 114 ...
Validation Accuracy = 0.991

EPOCH 115 ...
Validation Accuracy = 0.991

EPOCH 116 ...
Validation Accuracy = 0.989

EPOCH 117 ...
Validation Accuracy = 0.994

EPOCH 118 ...
Validation Accuracy = 0.986

EPOCH 119 ...
Validation Accuracy = 0.992

EPOCH 120 ...
Validation Accuracy = 0.991

EPOCH 121 ...
Validation Accuracy = 0.990

EPOCH 122 ...
Validation Accuracy = 0.992

EPOCH 123 ...
Validation Accuracy = 0.993

EPOCH 124 ...
Validation Accuracy = 0.993

EPOCH 125 ...
Validation Accuracy = 0.990

EPOCH 126 ...
Validation Accuracy = 0.993

EPOCH 127 ...
Validation Accuracy = 0.992

EPOCH 128 ...
Validation Accuracy = 0.990

EPOCH 129 ...
Validation Accuracy = 0.991

EPOCH 130 ...
Validation Accuracy = 0.994

EPOCH 131 ...
Validation Accuracy = 0.990

EPOCH 132 ...
Validation Accuracy = 0.993

EPOCH 133 ...
Validation Accuracy = 0.994

EPOCH 134 ...
Validation Accuracy = 0.994

EPOCH 135 ...
Validation Accuracy = 0.990

EPOCH 136 ...
Validation Accuracy = 0.993

EPOCH 137 ...
Validation Accuracy = 0.992

EPOCH 138 ...
Validation Accuracy = 0.993

EPOCH 139 ...
Validation Accuracy = 0.993

EPOCH 140 ...
Validation Accuracy = 0.992

EPOCH 141 ...
Validation Accuracy = 0.994

EPOCH 142 ...
Validation Accuracy = 0.993

EPOCH 143 ...
Validation Accuracy = 0.994

EPOCH 144 ...
Validation Accuracy = 0.991

EPOCH 145 ...
Validation Accuracy = 0.994

EPOCH 146 ...
Validation Accuracy = 0.995

EPOCH 147 ...
Validation Accuracy = 0.992

EPOCH 148 ...
Validation Accuracy = 0.990

EPOCH 149 ...
Validation Accuracy = 0.993

EPOCH 150 ...
Validation Accuracy = 0.993

Model saved

Question 4

How did you train your model? (Type of optimizer, batch size, epochs, hyperparameters, etc.)

Answer: I used the Adam optimizer with a learning rate of 0.001. In order to avoid overfitting, I added an l2 regularization penalty on all of the weights and biases with a beta of 0.001.

The batch size I used was 256 and I trained the model over 150 epochs.

As I mentioned before, I augment each of the images using a function written by Vivek Yadav before each epoch.This function warps the image by translating it, shearing it, and rotating it. Each epoch takes the entire training set and applies a random set of transformations to this effect. The intent is to make the network robust to these transformations by feeding it a large variety of them.

Question 5

What approach did you take in coming up with a solution to this problem? It may have been a process of trial and error, in which case, outline the steps you took to get to the final solution and why you chose those steps. Perhaps your solution involved an already well known implementation or architecture. In this case, discuss why you think this is suitable for the current problem.

Answer:

The approach I took on this problem was to start with the lenet architecture and make modifications to this to improve the accuracy. I added some depth, added dropout, added 1x1 convolutional layers.

Once I had achieved fairly good results, I read a about some published arquitectures for this data such as: http://yann.lecun.com/exdb/publis/pdf/sermanet-ijcnn-11.pdf https://chatbotslife.com/german-sign-classification-using-deep-learning-neural-networks-98-8-solution-d05656bf51ad#.d06fgw6jl

I also discussed with my mentor and classmates about different approaches. This led to me adding data augmentation to my training data, and adding l2 regularization to my cost function. I also increased the number of parameters in each layer significantly to detect more features.

My final test accuracy was 96.8%.

with tf.Session() as sess:
    loader = tf.train.import_meta_graph('Traffic_Sign_Classifying_Net.meta')
    loader.restore(sess, tf.train.latest_checkpoint('./'))

    test_accuracy = evaluate(X_test, y_test)
    print("Test Accuracy = {:.3f}".format(test_accuracy))

Test Accuracy = 0.968

---

Step 3: Test a Model on New Images

Take several pictures of traffic signs that you find on the web or around you (at least five), and run them through your classifier on your computer to produce example results. The classifier might not recognize some local signs but it could prove interesting nonetheless.

You may find signnames.csv useful as it contains mappings from the class id (integer) to the actual sign name.

### Load the images and plot them here.
### Feel free to use as many code cells as needed.
%matplotlib inline
def figplot(name):
    image_ = cv2.imread(name)
    image = cv2.resize(image_,(32,32))
    plt.figure(figsize=(3,3))
    plt.subplot(1,2,1)
    plt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
    plt.subplot(1,2,2)
    plt.imshow(cv2.cvtColor(image_, cv2.COLOR_BGR2RGB))
    return image

    
img1 = figplot("Image1_StopSign.jpg")
img2 = figplot("Image2_DoNotEnter.jpg")
img3 = figplot("Image3_DoNotEnter.jpg")
img4 = figplot("Image4_OneWay.jpg")
img5 = figplot("Image5_SpeedLimit.jpg")

[20 17 13 34  1]

Question 6

Choose five candidate images of traffic signs and provide them in the report. Are there any particular qualities of the image(s) that might make classification difficult? It could be helpful to plot the images in the notebook.

Answer: I chose 5 images of different american traffic signs offline. One of the first obvious challenges with these images is that they are not 32 by 32 pixels, and the images are not centered in the frame nor do they occupy the majority of the frame. Regardless, I am going to use the naive approach of coverting these to the correct size directoy with no further pre-processing. }

The naively scaled down image is on the left and the original is on the right.

Another obvious challenge is that I am using a network trained to classify german traffic signs to classify american traffic signs. This will have mixed results.

### Run the predictions here.
### Feel free to use as many code cells as needed.
images = preprocess([img1, img2, img3, img4, img5])
classifications = [14, 17, 13, 33, 1]
print("Correct Classifications are:")
print(classifications)
print()
classifications_one_hot = tf.one_hot(classifications, output_depth)

with tf.Session() as sess:
    loader = tf.train.import_meta_graph('Traffic_Sign_Classifying_Net.meta')
    loader.restore(sess, tf.train.latest_checkpoint('./'))
    my_logits = sess.run(logits, feed_dict={x: images, keep_prob_convlayers: 1., keep_prob_connected: 1.})
    predictions = tf.argmax(my_logits, 1)
    print("Predictions are:")
    print(predictions.eval())
    print()


    predict_correct = tf.equal(tf.argmax(my_logits, 1), tf.argmax(classifications_one_hot, 1))
    print("The following predictions are correct:")
    print(predict_correct.eval())
    print()
    my_accuracy = tf.reduce_mean(tf.cast(predict_correct, tf.float32))
    print("Which gives an accuracy of:")
    print(my_accuracy.eval())
    print()

Correct Classifications are:
[14, 17, 13, 33, 1]

Predictions are:
[20 17 13 34  1]

The following predictions are correct:
[False  True  True False  True]

Which gives an accuracy of:
0.6

Question 7

Is your model able to perform equally well on captured pictures when compared to testing on the dataset? The simplest way to do this check the accuracy of the predictions. For example, if the model predicted 1 out of 5 signs correctly, it's 20% accurate.

NOTE: You could check the accuracy manually by using signnames.csv (same directory). This file has a mapping from the class id (0-42) to the corresponding sign name. So, you could take the class id the model outputs, lookup the name in signnames.csv and see if it matches the sign from the image.

Answer:

No, my model is not able to perform equally well. It had an accuracy of 60%. I imagine the reasons are similar to the challenges I listed in Question 6. These are american traffic signs, not german signs the network was trained on. The one-way sign (4th image) does not exist in the german set, so this one was obviously predicted incorrectly. However, the prediction was turn left ahead which is a reasonable prediction because it means it detected that it was an arrow. The speed limit sign, too, is different because while it has numbers, it is significantly different from the german signs

I also naively scalled full sized images down to the 32x32 size without cropping into a square first, or cropping around the actual sign. This distorted the images and likely caused them to be harder to recognize

### Visualize the softmax probabilities here.
### Feel free to use as many code cells as needed.
print("Correct Classifications are:")
print(classifications)
print()
with tf.Session() as sess:
    loader = tf.train.import_meta_graph('Traffic_Sign_Classifying_Net.meta')
    loader.restore(sess, tf.train.latest_checkpoint('./'))
    probs = tf.nn.top_k(tf.nn.softmax(logits), k = 5)
    probabilities = sess.run(probs, feed_dict={x: images, keep_prob_convlayers: 1., keep_prob_connected: 1.})
    print(probabilities)

Correct Classifications are:
[14, 17, 13, 33, 1]

TopKV2(values=array([[  5.06562531e-01,   1.58400193e-01,   8.93689469e-02,
          4.82531600e-02,   3.23171206e-02],
       [  7.91175783e-01,   9.30441394e-02,   4.71178368e-02,
          3.87739390e-02,   5.54470252e-03],
       [  9.86915588e-01,   5.13651455e-03,   2.88560265e-03,
          2.23003351e-03,   4.97359550e-04],
       [  4.59732234e-01,   1.01718038e-01,   5.63225560e-02,
          3.09494156e-02,   3.00214235e-02],
       [  1.20476842e-01,   9.47792530e-02,   6.68851584e-02,
          6.26882538e-02,   5.72255999e-02]], dtype=float32), indices=array([[20, 28, 23, 38, 18],
       [17, 37, 14, 39, 34],
       [13, 11, 28, 20, 25],
       [34, 17,  9, 14, 13],
       [ 1, 34,  8,  2, 29]], dtype=int32))

Question 8

Use the model's softmax probabilities to visualize the certainty of its predictions, tf.nn.top_k could prove helpful here. Which predictions is the model certain of? Uncertain? If the model was incorrect in its initial prediction, does the correct prediction appear in the top k? (k should be 5 at most)

Answer: The predictions the model is most certain of is with the "Do Not Enter" and "Yield" signs. It is not certain at all about the speed sign, which I believe is mostly due to the distortion from the naive conversion from a high fidelity to 32x32 image. It is also incorrectly 50% certain about the stop sign which is suprising

Where the model was incorrect about its initial prediction, the correct prediction does not appear in the top 5.

While I am not surprised that the speed sign and the one-way sign were hard for the classifier to get right, since these were either highly distorted or not in the german sign set at all, I am suprised how poorly it did with the stop sign. The american stop sign is fairly similar to the german stop sign so it should have been able to identify it correctly. I imagine it performed poorly on this one, again, because of the distortion from directly scaling down a high resolution image from the web.

Note: Once you have completed all of the code implementations and successfully answered each question above, you may finalize your work by exporting the iPython Notebook as an HTML document. You can do this by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.