Exercise with TensorFlow

My journey, Dominika Kalinowska

Background

As research continues in the development of self-driving cars, one of the key challenges is computer vision, allowing these cars to develop an understanding of their environment from digital images. In particular, this involves the ability to recognize and distinguish road signs – stop signs, speed limit signs, yield signs, and more.

In this project, you’ll use TensorFlow to build a neural network to classify road signs based on an image of those signs. To do so, you’ll need a labeled dataset: a collection of images that have already been categorized by the road sign represented in them.

Several such data sets exist, but for this project, we’ll use the German Traffic Sign Recognition Benchmark (GTSRB) dataset, which contains thousands of images of 43 different kinds of road signs.

Code

Model:

model = tf.keras.Sequential([

        keras.layers.experimental.preprocessing.Rescaling(1. / 255),
        tf.keras.layers.Conv2D(32, 3, padding="same", activation='relu', input_shape=(IMG_WIDTH, IMG_HEIGHT,3)),
        tf.keras.layers.MaxPooling2D(),  # Max pooling is then used to reduce the spatial dimensions
        tf.keras.layers.Conv2D(64, 3, padding="same", activation='relu'), # We then learn 64 filters
        tf.keras.layers.MaxPooling2D(),  
        tf.keras.layers.Conv2D(128, 3, padding="same", activation='relu'),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Flatten(),
        # Hidden layer with 128 neurons, Dense - fully conected
        tf.keras.layers.Dense(128, activation='relu'),
        # output
        keras.layers.Dense(NUM_CATEGORIES, activation="softmax")
    ])
    
    # Train neural network
    model.compile(optimizer='adam',
                  loss='categorical_crossentropy',  # not sparse_categorical_crossentropy bc 1D
                  metrics=['accuracy'])

Result

Epoch 1/10
500/500 [==============================] - 33s 61ms/step - loss: 2.1360 - accuracy: 0.3980
Epoch 2/10
500/500 [==============================] - 30s 60ms/step - loss: 0.5102 - accuracy: 0.8436
Epoch 3/10
500/500 [==============================] - 30s 60ms/step - loss: 0.1850 - accuracy: 0.9465
Epoch 4/10
500/500 [==============================] - 30s 60ms/step - loss: 0.0879 - accuracy: 0.9761
Epoch 5/10
500/500 [==============================] - 33s 67ms/step - loss: 0.0529 - accuracy: 0.9858
Epoch 6/10
500/500 [==============================] - 31s 63ms/step - loss: 0.0549 - accuracy: 0.9839
Epoch 7/10
500/500 [==============================] - 32s 65ms/step - loss: 0.0323 - accuracy: 0.9909
Epoch 8/10
500/500 [==============================] - 32s 64ms/step - loss: 0.0280 - accuracy: 0.9925
Epoch 9/10
500/500 [==============================] - 31s 62ms/step - loss: 0.0284 - accuracy: 0.9914
Epoch 10/10
500/500 [==============================] - 31s 61ms/step - loss: 0.0251 - accuracy: 0.9932
333/333 - 6s - loss: 0.1159 - accuracy: 0.9704

Tips

Start with small date: gtsrb-small which have NUM_CATEGORIES = 3!!! Remember to change that when you chane data.

When you load date you can see something. You want to see something to understand what's going on.

print(image.shape) to see that your image is: (30,30,3) When you print(image) you will see something like that:

[[[212 221 222]
  [249 244 249]
  [255 255 255]
  ...
  [ 97  97  97]
  [127 127 124]
  [183 198 197]]

 [[244 246 242]
  [253 250 250]
  [255 255 255]
  ...
  [100  93  93]
  [149 128 137]
  [238 232 241]]

 [[232 252 246]
  [254 240 238]
  [255 255 255]
  ...
  [ 97  93  93]
  [136 121 144]
  [234 230 248]]

 ...

 [[255 255 255]
  [255 255 251]
  [254 250 250]
  ...
  [100  98  99]
  [104 101 101]
  [105 101  99]]

 [[255 255 255]
  [255 255 255]
  [254 255 255]
  ...
  [ 97  95  96]
  [105  99 100]
  [102 100 101]]

 [[255 255 254]
  [255 255 255]
  [255 255 255]
  ...
  [ 95  95  96]
  [ 98  96  97]
  [ 97  96  98]]]

Not very funny. But:

plt.imshow(image, cmap=plt.cm.binary)
plt.show()

And you can see image. Remember to change BGR to RGB: image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) .

print("LEN OF IMAGES: ", len(images)) to see how many is images: gtsrb-small -> 840

You can:

plt.hist(labels)
plt.show()

to see more about images in categories.

If you can see some random images:

# Determine the (random) indexes of the images that you want to see
    traffic_signs = [100, 558, 650, 838]   
    # https://www.datacamp.com/community/tutorials/tensorflow-tutorial
    # Fill out the subplots with the random images that you defined
    for i in range(len(traffic_signs)):
        plt.subplot(1, 4, i + 1)
        plt.axis('off')
        plt.title("Label {0} ({1})".format(i, labels.count(i)))
        plt.imshow(images[traffic_signs[i]])
        plt.subplots_adjust(wspace=0.5)

    plt.show()