In this notebook, a template is provided for you to implement your functionality in stages which is required to successfully complete this project. If additional code is required that cannot be included in the notebook, be sure that the Python code is successfully imported and included in your submission, if necessary. Sections that begin with 'Implementation' in the header indicate where you should begin your implementation for your project. Note that some sections of implementation are optional, and will be marked with 'Optional' in the header.
In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.
Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. In addition, Markdown cells can be edited by typically double-clicking the cell to enter edit mode.
In [1]:
# Load pickled data
import pickle
import csv
import cv2
import numpy as np
import math
import matplotlib.pyplot as plt
signnames = []
with open("signnames.csv", 'r') as f:
next(f)
reader = csv.reader(f)
signnames = list(reader)
n_classes = len(signnames)
training_file = "./train.p"
testing_file = "./test.p"
with open(training_file, mode='rb') as f:
train = pickle.load(f)
with open(testing_file, mode='rb') as f:
test = pickle.load(f)
In [2]:
from sklearn import cross_validation
X_train, X_test = [], []
y_train, y_test = [], test['labels']
for i, img in enumerate(train['features']):
img = cv2.resize(img,(48, 48), interpolation = cv2.INTER_CUBIC)
X_train.append(img)
y_train.append(train['labels'][i])
# Adaptive Histogram (CLAHE)
imgLab = cv2.cvtColor(img, cv2.COLOR_RGB2Lab)
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
l, a, b = cv2.split(imgLab)
l = clahe.apply(l)
imgLab = cv2.merge((l, a, b))
imgLab = cv2.cvtColor(imgLab, cv2.COLOR_Lab2RGB)
X_train.append(imgLab)
y_train.append(train['labels'][i])
# Rotate -15
M = cv2.getRotationMatrix2D((24, 24), -15.0, 1)
imgL = cv2.warpAffine(img, M, (48, 48))
X_train.append(imgL)
y_train.append(train['labels'][i])
# Rotate 15
M = cv2.getRotationMatrix2D((24, 24), 15.0, 1)
imgR = cv2.warpAffine(img, M, (48, 48))
X_train.append(imgR)
y_train.append(train['labels'][i])
for img in test['features']:
X_test.append(cv2.resize(img,(48, 48), interpolation = cv2.INTER_CUBIC))
X_train, X_validation, y_train, y_validation = cross_validation.train_test_split(X_train, y_train, test_size=0.2, random_state=7)
from sklearn.utils import shuffle
X_train, y_train = shuffle(X_train, y_train)
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 IMAGESComplete the basic data summary below.
In [3]:
n_train = len(X_train)
n_test = len(X_test)
image_shape = X_train[0].shape
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)
print("Number of X_train = ", len(X_train))
print("Number of X_validation = ", len(X_validation))
print("Number of y_train = ", len(y_train))
print("Number of y_validation = ", len(y_validation))
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.
In [4]:
import random
# Visualizations will be shown in the notebook.
%matplotlib inline
index = random.randint(0, len(X_train))
image = X_train[index].squeeze()
plt.figure(figsize=(1,1))
plt.imshow(image)
print(y_train[index], signnames[y_train[index]][1])
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:
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!
In [5]:
import tensorflow as tf
from tensorflow.contrib.layers import flatten
EPOCHS = 15
BATCH_SIZE = 128
def ConvNet(x):
mu = 0
sigma = 0.1
# Layer 1: Convolutional. Input = 48x48x3. Output = 42x42x100.
c1_W = tf.Variable(tf.truncated_normal([7, 7, 3, 100], mean=mu, stddev=sigma))
c1_b = tf.Variable(tf.zeros(100))
c1 = tf.nn.conv2d(x, c1_W, strides=[1, 1, 1, 1], padding='VALID')
c1 = tf.nn.bias_add(c1, c1_b)
c1 = tf.nn.relu(c1)
# Layer 2: Max Pooling. Input = 42x42x100. Output = 21x21x100.
s2 = tf.nn.max_pool(c1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# Layer 3: Convolutional. Input = 21x21x100. Output = 18x18x150.
c3_W = tf.Variable(tf.truncated_normal([4, 4, 100, 150], mean=mu, stddev=sigma))
c3_b = tf.Variable(tf.zeros(150))
c3 = tf.nn.conv2d(s2, c3_W, strides=[1, 1, 1, 1], padding='VALID')
c3 = tf.nn.bias_add(c3, c3_b)
c3 = tf.nn.relu(c3)
# Layer 4: Max Pooling. Input = 18x18x150. Output = 9x9x150
s4 = tf.nn.max_pool(c3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# Layer 5: Convolutional. Input = 9x9x150. Output = 6x6x250.
c5_W = tf.Variable(tf.truncated_normal([4, 4, 150, 250], mean=mu, stddev=sigma))
c5_b = tf.Variable(tf.zeros(250))
c5 = tf.nn.conv2d(s4, c5_W, strides=[1, 1, 1, 1], padding='VALID')
c5 = tf.nn.bias_add(c5, c5_b)
c5 = tf.nn.relu(c5)
# Layer 6: Max Pooling. Input = 6x6x250. Output = 3x3x250.
s6 = tf.nn.max_pool(c5, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
# Layer 6: Flatten. Input = 3x3x250. Output = 2250
s6 = flatten(s6)
# Layer 7: Fully Connected. Input = 2250. Output = 300.
fc7_W = tf.Variable(tf.truncated_normal([2250, 300], mean=mu, stddev=sigma))
fc7_b = tf.Variable(tf.zeros(300))
fc7 = tf.add(tf.matmul(s6, fc7_W), fc7_b)
fc7 = tf.nn.relu(fc7)
# Layer 8: Fully Connected. Input = 300. Output = 43.
fc8_W = tf.Variable(tf.truncated_normal([300, 43], mean=mu, stddev=sigma))
fc8_b = tf.Variable(tf.zeros(43))
fc8 = tf.add(tf.matmul(fc7, fc8_W), fc8_b)
return fc8
In [6]:
x = tf.placeholder(tf.float32, (None, 48, 48, 3))
y = tf.placeholder(tf.int32, (None))
one_hot_y = tf.one_hot(y, n_classes)
In [7]:
rate = 0.001
logits = ConvNet(x)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits, one_hot_y)
loss_operation = tf.reduce_mean(cross_entropy)
optimizer = tf.train.AdamOptimizer(learning_rate = rate)
training_operation = optimizer.minimize(loss_operation)
In [8]:
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})
total_accuracy += (accuracy * len(batch_x))
return total_accuracy / num_examples
In [9]:
with tf.Session() 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)
for offset in range(0, num_examples, BATCH_SIZE):
end = offset + BATCH_SIZE
batch_x, batch_y = X_train[offset:end], y_train[offset:end]
sess.run(training_operation, feed_dict={x: batch_x, y: batch_y})
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, 'convnet')
print("Model saved")
In [10]:
with tf.Session() as sess:
loader = tf.train.import_meta_graph("convnet.meta")
loader.restore(sess, tf.train.latest_checkpoint('./'))
test_accuracy = evaluate(X_test, y_test)
print("Test Accuracy = {:.3f}".format(test_accuracy))
Answer:
I did not preprocess the data.
Answer:
I generate additional data using 3 methods:
1. Adaptive histogram
2. Rotate -15 degree
3. Rotate 15 degree
Additionally, I resize the images from 32x32 to 48x48 I took the action according to the paper Multi-Column Deep Neural Network for Traffic Sign Classification (for Adaptive histogram) and Traffic Sign Recognition with Multi-Scale Convolutional Networks (for rotation).
Finally, I split the training data into 2 parts following 80:20 for cross validation.
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:
I uses the architecture from Multi-Column Deep Neural Network for Traffic Sign Classification by Ciresan et. al. The following lines briefly describes the layers:
Input: 48x48x3 images
Layer 1: Convolutional layer with 7x7 kernel which output 100 maps of 42x42 neurons
Layer 2: Max-pooling layer with 2x2 kernel and 2 strides which output 100 maps of 21x21 neurons
Layer 3: Convolutional layer with 4x4 kernel which output 150 maps of 18x18 neurons
Layer 4: Max-pooling layer with 2x2 kernel and 2 strides which output 100 maps of 9x9 neurons
Layer 5: Convolutional layer with 4x4 kernel which output 250 maps of 6x6 neurons
Layer 6: Max-pooling layer with 2x2 kernel and 2 strides which output 100 maps of 3x3 neurons
Layer 7: Fully-connected layer outputing 300 neurons
Layer 8: Fully-connected layer outputing 43 neurons/logits
Answer:
Optimizer: Adam-optimizer Batch size: 128 Epochs: 10 Hyperparameters: mu = 0, sigma = 0.1 Learning rate: 0.001
In the review, the reviewer suggested to look into tuning the number of Epochs. From the result of the cross-validation in my first submission, I used 10 Epoch which ended up with the accuracy around 0.9. However, at 10 Epochs, the model seems to be still learning better. Therefore, I increased the from 10 Epochs to 20 Epochs.
In the second iteration (after increasing the number of epoch), the results plateau from 15th epoch at 0.960 and then decreased after 19th epoch until 0.935. The accuracy of model evaluation on test data is only 0.9 (slight increase from first iteration at 0.898). I decided to decrease the epoch to 15 epochs, fearing it might overfit the training data.
The accuracy of third iteration (at 15 epochs) remain at 0.96. However, when tested on test data, the accuracy increases to 0.947 which explain early-stopping helps the model to generalise better by not overfitting the training data too much.
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:
I tried Lenet-5 prior to the current architecture. The evaluation result is about 0.8. After changing the architecture, the result improved to above 0.9. I decided to add additional training images after viewing the images. Some of them are darkened, and some of them off-centered. Therefore, I took cue from 2 papers above to generate additional data by using adaptive histogram and rotation. The results improved.
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.
In [11]:
from PIL import Image
# Visualizations will be shown in the notebook.
%matplotlib inline
new_images = []
new_labels = np.array([4, 17, 26, 28, 14])
fig = plt.figure()
for i in range(1, 6):
subplot = fig.add_subplot(2,3,i)
img = cv2.imread("./dataset/{}.png".format(i))
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
img = cv2.resize(img,(48, 48), interpolation = cv2.INTER_CUBIC)
subplot.set_title(signnames[new_labels[i-1]][1],fontsize=8)
subplot.imshow(img)
new_images.append(img)
Answer:
In [12]:
with tf.Session() as sess:
loader = tf.train.import_meta_graph("convnet.meta")
loader.restore(sess, tf.train.latest_checkpoint('./'))
new_pics_classes = sess.run(logits, feed_dict={x: new_images})
test_accuracy = evaluate(new_images, new_labels)
print("Test Accuracy = {:.3f}".format(test_accuracy))
top3 = sess.run(tf.nn.top_k(new_pics_classes, k=3, sorted=True))
for i in range(len(top3[0])):
labels = list(map(lambda x: signnames[x][1], top3[1][i]))
print("Image {} predicted labels: {} with probabilities: {}".format(i+1, labels, top3[0][i]))
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:
The accuracy is 80%. With only 1 minor mistake which the model mistaken 70 as 20 which is incredible narrow because 2 is very similar to 7 in some perspective.
There is one particular image that is difficult to predict which is Speed limit (70km/h). It predicted as Speed limit (20km/h), only the value of the speed limit differs. There is several factors that influenced this result. It might probably due to the fact that the shape of 7 looks like 2 in some perspective.
On another attempt after retrain the model, the model still failed to predict the label correctly. It predicted it as Speed limit (100km/h)The second reason can be because there are a total of 8 speed limit signs in the training data, there might not be enough sign for 70km/h which made the prediction for the particular sign more difficult.
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)
tf.nn.top_k will return the values and indices (class ids) of the top k predictions. So if k=3, for each sign, it'll return the 3 largest probabilities (out of a possible 43) and the correspoding class ids.
Take this numpy array as an example:
# (5, 6) array
a = np.array([[ 0.24879643, 0.07032244, 0.12641572, 0.34763842, 0.07893497,
0.12789202],
[ 0.28086119, 0.27569815, 0.08594638, 0.0178669 , 0.18063401,
0.15899337],
[ 0.26076848, 0.23664738, 0.08020603, 0.07001922, 0.1134371 ,
0.23892179],
[ 0.11943333, 0.29198961, 0.02605103, 0.26234032, 0.1351348 ,
0.16505091],
[ 0.09561176, 0.34396535, 0.0643941 , 0.16240774, 0.24206137,
0.09155967]])Running it through sess.run(tf.nn.top_k(tf.constant(a), k=3)) produces:
TopKV2(values=array([[ 0.34763842, 0.24879643, 0.12789202],
[ 0.28086119, 0.27569815, 0.18063401],
[ 0.26076848, 0.23892179, 0.23664738],
[ 0.29198961, 0.26234032, 0.16505091],
[ 0.34396535, 0.24206137, 0.16240774]]), indices=array([[3, 0, 5],
[0, 1, 4],
[0, 5, 1],
[1, 3, 5],
[1, 4, 3]], dtype=int32))Looking just at the first row we get [ 0.34763842, 0.24879643, 0.12789202], you can confirm these are the 3 largest probabilities in a. You'll also notice [3, 0, 5] are the corresponding indices.
Answer:
Looking at all the Top 3, the model predicted Image 2 with high confidence whereas the rest are very close. It is astonishing to see the model recognize Image 1 as a speed limit sign just that it wrongly predict the speed limit (70->20).
Using the tf.nn.top to analyse the prediction failure, I was surprised to see that the model correctly identify the Speed limit 70km/h as a speed limit sign. Although the prediction for the Top 3 are wrong, but all the Top 3 predictions are all speed limit with different value. All 3 predictions have similar probablities, which infers that the model tried to guess the value of the sign (but it failed to capture 70km/h). It might response better if I increase the number of training data for every speed limit sign so that the model can recognize the value of the speed limit better.
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.