Keras exists to make coding deep neural networks simpler. To demonstrate just how easy it is, you’re going to use Keras to build a convolutional neural network in a few dozen lines of code.
You’ll be connecting the concepts from the previous lessons to the methods that Keras provides.
The network you'll build with Keras is similar to the example that you can find in Keras’s GitHub repository that builds out a convolutional neural network for MNIST.
However, instead of using the MNIST dataset, you're going to use the German Traffic Sign Recognition Benchmark dataset that you've used previously.
You can download pickle files with sanitized traffic sign data here.
Here are the steps you'll take to build the network:
Keep an eye on the network’s accuracy over time. Once the accuracy reaches the 98% range, you can be confident that you’ve built and trained an effective model.
In [1]:
# TODO: Implement load the data here.
# Load pickled data
import pickle
import csv
import os
# TODO: fill this in based on where you saved the training and testing data
training_file = '../../traffic-signs/traffic-signs-data/train.p'
testing_file = '../../traffic-signs/traffic-signs-data/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)
X_train, y_train = train['features'], train['labels']
X_test, y_test = test['features'], test['labels']
# Make dictionary of sign names from CSV file
with open('../../traffic-signs/signnames.csv', 'r') as csvfile:
reader = csv.reader(csvfile)
next(reader, None) # skip the headers
sign_names = dict((int(n),label) for n, label in reader)
cls_numbers, cls_names = zip(*sign_names.items())
n_classes = len(set(y_train))
flat_img_size = 32*32*3
# STOP: Do not change the tests below. Your implementation should pass these tests.
assert(X_train.shape[0] == y_train.shape[0]), "The number of images is not equal to the number of labels."
assert(X_train.shape[1:] == (32,32,3)), "The dimensions of the images are not 32 x 32 x 3."
In [2]:
import numpy as np
from PIL import Image
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelBinarizer
from sklearn.utils import resample
from tqdm import tqdm_notebook
from zipfile import ZipFile
import time
from datetime import timedelta
import math
import tensorflow as tf
In [3]:
# TODO: Implement data normalization here.
def normalize_color(image_data):
"""
Normalize the image data with Min-Max scaling to a range of [0.1, 0.9]
:param image_data: The image data to be normalized
:return: Normalized image data
"""
a = -0.5
b = +0.5
Xmin = 0.0
Xmax = 255.0
norm_img = np.empty_like(image_data, dtype=np.float32)
norm_img = a + (image_data - Xmin)*(b-a)/(Xmax - Xmin)
return norm_img
X_train = normalize_color(X_train)
X_test = normalize_color(X_test)
# STOP: Do not change the tests below. Your implementation should pass these tests.
assert(round(np.mean(X_train)) == 0), "The mean of the input data is: %f" % np.mean(X_train)
assert(np.min(X_train) == -0.5 and np.max(X_train) == 0.5), "The range of the input data is: %.1f to %.1f" % (np.min(X_train), np.max(X_train))
The code you've written so far is for data processing, not specific to Keras. Here you're going to build Keras-specific code.
Build a two-layer feedforward neural network, with 128 neurons in the fully-connected hidden layer.
To get started, review the Keras documentation about models and layers.
The Keras example of a Multi-Layer Perceptron network is similar to what you need to do here. Use that as a guide, but keep in mind that there are a number of differences.
In [4]:
from keras.models import Sequential
from keras.layers.core import Dense, Activation
from keras.optimizers import Adam
from keras.utils import np_utils
In [5]:
# TODO: Build a two-layer feedforward neural network with Keras here.
model = Sequential()
model.add(Dense(128, input_shape=(flat_img_size,), name='hidden1'))
model.add(Activation('relu'))
model.add(Dense(43, name='output'))
model.add(Activation('softmax'))
# STOP: Do not change the tests below. Your implementation should pass these tests.
assert(model.get_layer(name="hidden1").input_shape == (None, 32*32*3)), "The input shape is: %s" % model.get_layer(name="hidden1").input_shape
assert(model.get_layer(name="output").output_shape == (None, 43)), "The output shape is: %s" % model.get_layer(name="output").output_shape
Compile and train the network for 2 epochs. Use the adam optimizer, with categorical_crossentropy loss.
Hint 1: In order to use categorical cross entropy, you will need to one-hot encode the labels.
Hint 2: In order to pass the input images to the fully-connected hidden layer, you will need to reshape the input.
Hint 3: Keras's .fit() method returns a History.history object, which the tests below use. Save that to a variable named history.
In [6]:
# One-Hot encode the labels
Y_train = np_utils.to_categorical(y_train, n_classes)
Y_test = np_utils.to_categorical(y_test, n_classes)
# Reshape input for MLP
X_train_mlp = X_train.reshape(-1, flat_img_size)
X_test_mlp = X_test.reshape(-1, flat_img_size)
In [7]:
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
history = model.fit(X_train_mlp, Y_train, batch_size=128, nb_epoch=10,
validation_data=(X_test_mlp, Y_test), verbose=1)
# STOP: Do not change the tests below. Your implementation should pass these tests.
assert(history.history['acc'][0] > 0.5), "The training accuracy was: %.3f" % history.history['acc'][0]
Split the training data into a training and validation set.
Measure the validation accuracy of the network after two training epochs.
Hint: Use the train_test_split() method from scikit-learn.
In [8]:
# Get randomized datasets for training and validation
X_train, X_val, Y_train, Y_val = train_test_split(
X_train,
Y_train,
test_size=0.25,
random_state=0xdeadbeef)
X_val_mlp = X_val.reshape(-1, flat_img_size)
print('Training features and labels randomized and split.')
# STOP: Do not change the tests below. Your implementation should pass these tests.
assert(round(X_train.shape[0] / float(X_val.shape[0])) == 3), "The training set is %.3f times larger than the validation set." % (X_train.shape[0] / float(X_val.shape[0]))
assert(history.history['val_acc'][0] > 0.6), "The validation accuracy is: %.3f" % history.history['val_acc'][0]
In [15]:
loss, acc = model.evaluate(X_val.reshape(-1, flat_img_size), Y_val, verbose=1)
print('\nValidation accuracy : {0:>6.2%}'.format(acc))
Validation Accuracy: 95.92%
Build a new network, similar to your existing network. Before the hidden layer, add a 3x3 convolutional layer with 32 filters and valid padding.
Then compile and train the network.
Hint 1: The Keras example of a convolutional neural network for MNIST would be a good example to review.
Hint 2: Now that the first layer of the network is a convolutional layer, you no longer need to reshape the input images before passing them to the network. You might need to reload your training data to recover the original shape.
Hint 3: Add a Flatten() layer between the convolutional layer and the fully-connected hidden layer.
In [23]:
from keras.layers import Convolution2D, MaxPooling2D, Dropout, Flatten
In [17]:
# TODO: Re-construct the network and add a convolutional layer before the first fully-connected layer.
model = Sequential()
model.add(Convolution2D(16, 5, 5, border_mode='same', input_shape=(32, 32, 3)))
model.add(Activation('relu'))
model.add(Flatten())
model.add(Dense(128, input_shape=(flat_img_size,), name='hidden1'))
model.add(Activation('relu'))
model.add(Dense(43, name='output'))
model.add(Activation('softmax'))
# TODO: Compile and train the model.
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
history = model.fit(X_train, Y_train, batch_size=128, nb_epoch=10,
validation_data=(X_val, Y_val), verbose=1)
# STOP: Do not change the tests below. Your implementation should pass these tests.
assert(history.history['val_acc'][0] > 0.9), "The validation accuracy is: %.3f" % history.history['val_acc'][0]
Validation Accuracy: 96.98%
Re-construct your network and add a 2x2 pooling layer immediately following your convolutional layer.
Then compile and train the network.
In [22]:
# TODO: Re-construct the network and add a pooling layer after the convolutional layer.
model = Sequential()
model.add(Convolution2D(16, 5, 5, border_mode='same', input_shape=(32, 32, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Flatten())
model.add(Dense(128, input_shape=(flat_img_size,), name='hidden1'))
model.add(Activation('relu'))
model.add(Dense(43, name='output'))
model.add(Activation('softmax'))
# TODO: Compile and train the model.
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
history = model.fit(X_train, Y_train, batch_size=128, nb_epoch=10,
validation_data=(X_val, Y_val), verbose=1)
# STOP: Do not change the tests below. Your implementation should pass these tests.
## Fixed bug
assert(history.history['val_acc'][-1] > 0.9), "The validation accuracy is: %.3f" % history.history['val_acc'][0]
Validation Accuracy: 97.36%
Re-construct your network and add dropout after the pooling layer. Set the dropout rate to 50%.
In [24]:
# TODO: Re-construct the network and add dropout after the pooling layer.
model = Sequential()
model.add(Convolution2D(16, 5, 5, border_mode='same', input_shape=(32, 32, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2,2)))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(128, input_shape=(flat_img_size,), name='hidden1'))
model.add(Activation('relu'))
model.add(Dense(43, name='output'))
model.add(Activation('softmax'))
# TODO: Compile and train the model.
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
history = model.fit(X_train, Y_train, batch_size=128, nb_epoch=10,
validation_data=(X_val, Y_val), verbose=1)
# STOP: Do not change the tests below. Your implementation should pass these tests.
assert(history.history['val_acc'][-1] > 0.9), "The validation accuracy is: %.3f" % history.history['val_acc'][0]
Validation Accuracy: 97.75%
Congratulations! You've built a neural network with convolutions, pooling, dropout, and fully-connected layers, all in just a few lines of code.
Have fun with the model and see how well you can do! Add more layers, or regularization, or different padding, or batches, or more training epochs.
What is the best validation accuracy you can achieve?
In [28]:
pool_size = (2,2)
model = Sequential()
model.add(Convolution2D(16, 5, 5, border_mode='same', input_shape=(32, 32, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Convolution2D(64, 3, 3))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Dropout(0.5))
model.add(Convolution2D(128, 3, 3))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Dropout(0.5))
model.add(Flatten())
model.add(Dense(256, input_shape=(flat_img_size,), name='hidden1'))
model.add(Activation('relu'))
model.add(Dense(43, name='output'))
model.add(Activation('softmax'))
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
In [33]:
history = model.fit(X_train, Y_train, batch_size=128, nb_epoch=50,
validation_data=(X_val, Y_val), verbose=1)
Best Validation Accuracy: 99.65%
Once you've picked out your best model, it's time to test it.
Load up the test data and use the evaluate() method to see how well it does.
Hint 1: After you load your test data, don't forget to normalize the input and one-hot encode the output, so it matches the training data.
Hint 2: The evaluate() method should return an array of numbers. Use the metrics_names() method to get the labels.
In [35]:
# with open('./test.p', mode='rb') as f:
# test = pickle.load(f)
# X_test = test['features']
# y_test = test['labels']
# X_test = X_test.astype('float32')
# X_test /= 255
# X_test -= 0.5
# Y_test = np_utils.to_categorical(y_test, 43)
model.evaluate(X_test, Y_test)
Out[35]:
In [40]:
model.save('test-acc-9716-epoch50.h5')
from keras.models import load_model
model2 = load_model('test-acc-9716-epoch50.h5')
# model2.evaluate(X_test, Y_test)
model2.summary()
Test Accuracy: 97.15%