02_採用CNN實現CIFAR-10數據集

文章導讀

—在前面的介紹中，使用卷積神經網絡對MNIST數據集做了應用，然而MNIST數據集僅限於對手寫數字的識別，而且手寫數字相對於自然物體和圖片非常簡單，也缺少相應的噪聲和變換。—本文將使用CNN對CIFAR-10數據集進行驗證，同時會比較不同參數作用下卷積神經網絡對準確率產生的影響

CIFAR-10數據集

從網站首頁可以看到，這裏提供10個分類的現實物體的圖片，與前面所講的成熟的人工手寫識別相比，現實物體識別挑戰巨大，而且圖片中含有大量特徵、噪聲，識別物體比例不一，也加大了識別的難度，使其非常具有挑戰性

數據結構介紹

可以看到，數據集中的數據分成了兩部分：第一部分是特徵部分，使用一個[10000,3072的uint8的矩陣進行存儲，每一行向量都是3X3大小的3通道圖片，構成的格式類似於[3,3,3]；第二部分爲標籤部分，使用一個10000數據的list進行存儲，每個list對應的是0-9中的一個數字，對應於物品分類。另外對於python的數據集，還有一個標籤爲“label_names”，例如label_names[0] == “airplane”等。
對於具體的數據讀取，官網上也提供了相應的代碼

def unpickle(file):
    import pickle
    with open(file, 'rb') as fo:
        dict = pickle.load(fo, encoding='bytes')
    return dict

代碼實例

前面說到，label是一個包含0-9的list列表，根據之前我們用到的one-hot方法，採用稀疏性列表法，即10個列表數字中只有對應的那個值爲1，其他值都爲0，因此需要將list格式轉化成對應的one-hot矩陣。

def unpickle(filename):
    with open(filename, 'rb') as f:
        d = pickle.load(f, encoding='latin1')
        return d

def onehot(labels):
    '''one-hot 編碼'''
    n_sample = len(labels)
    n_class = max(labels) + 1
    onehot_labels = np.zeros((n_sample, n_class))
    onehot_labels[np.arange(n_sample), labels] = 1
    return onehot_labels

# 訓練數據集
data1 = unpickle('cifar10-dataset/data_batch_1')
data2 = unpickle('cifar10-dataset/data_batch_2')
data3 = unpickle('cifar10-dataset/data_batch_3')
data4 = unpickle('cifar10-dataset/data_batch_4')
data5 = unpickle('cifar10-dataset/data_batch_5')
X_train = np.concatenate((data1['data'], data2['data'], data3['data'], data4['data'], data5['data']), axis=0)
y_train = np.concatenate((data1['labels'], data2['labels'], data3['labels'], data4['labels'], data5['labels']), axis=0)
y_train = onehot(y_train)
# 測試數據集
test = unpickle('cifar10-dataset/test_batch')
X_test = test['data'][:5000, :]
y_test = onehot(test['labels'])[:5000, :]

print('Training dataset shape:', X_train.shape)
print('Training labels shape:', y_train.shape)
print('Testing dataset shape:', X_test.shape)
print('Testing labels shape:', y_test.shape)

模型參數

learning_rate = 1e-3
training_iters = 200
batch_size = 50
display_step = 5
n_features = 3072  # 32*32*3
n_classes = 10
n_fc1 = 384
n_fc2 = 192

模型構建

# 構建模型
x = tf.placeholder(tf.float32, [None, n_features])
y = tf.placeholder(tf.float32, [None, n_classes])

W_conv = {
    'conv1': tf.Variable(tf.truncated_normal([5, 5, 3, 32], stddev=0.0001)),
    'conv2': tf.Variable(tf.truncated_normal([5, 5, 32, 64],stddev=0.01)),
    'fc1': tf.Variable(tf.truncated_normal([8*8*64, n_fc1], stddev=0.1)),
    'fc2': tf.Variable(tf.truncated_normal([n_fc1, n_fc2], stddev=0.1)),
    'fc3': tf.Variable(tf.truncated_normal([n_fc2, n_classes], stddev=0.1))
}
b_conv = {
    'conv1': tf.Variable(tf.constant(0.0, dtype=tf.float32, shape=[32])),
    'conv2': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[64])),
    'fc1': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[n_fc1])),
    'fc2': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[n_fc2])),
    'fc3': tf.Variable(tf.constant(0.0, dtype=tf.float32, shape=[n_classes]))
}

x_image = tf.reshape(x, [-1, 32, 32, 3])
# 卷積層 1
conv1 = tf.nn.conv2d(x_image, W_conv['conv1'], strides=[1, 1, 1, 1], padding='SAME')
conv1 = tf.nn.bias_add(conv1, b_conv['conv1'])
conv1 = tf.nn.relu(conv1)
# 池化層 1
pool1 = tf.nn.avg_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')
# LRN層，Local Response Normalization
norm1 = tf.nn.lrn(pool1, 4, bias=1.0, alpha=0.001/9.0, beta=0.75)
# 卷積層 2
conv2 = tf.nn.conv2d(norm1, W_conv['conv2'], strides=[1, 1, 1, 1], padding='SAME')
conv2 = tf.nn.bias_add(conv2, b_conv['conv2'])
conv2 = tf.nn.relu(conv2)
# LRN層，Local Response Normalization
norm2 = tf.nn.lrn(conv2, 4, bias=1.0, alpha=0.001/9.0, beta=0.75)
# 池化層 2
pool2 = tf.nn.avg_pool(norm2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')
reshape = tf.reshape(pool2, [-1, 8*8*64])

fc1 = tf.add(tf.matmul(reshape, W_conv['fc1']), b_conv['fc1'])
fc1 = tf.nn.relu(fc1)
# 全連接層 2
fc2 = tf.add(tf.matmul(fc1, W_conv['fc2']), b_conv['fc2'])
fc2 = tf.nn.relu(fc2)
# 全連接層 3, 即分類層
fc3 = tf.nn.softmax(tf.add(tf.matmul(fc2, W_conv['fc3']), b_conv['fc3']))

# 定義損失
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=fc3, labels=y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(loss)
# 評估模型
correct_pred = tf.equal(tf.argmax(fc3, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

init = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init)
    c = []
    total_batch = int(X_train.shape[0] / batch_size)
#    for i in range(training_iters):
    start_time = time.time()
    for i in range(200):
        for batch in range(total_batch):
            batch_x = X_train[batch*batch_size : (batch+1)*batch_size, :]
            batch_y = y_train[batch*batch_size : (batch+1)*batch_size, :]
            sess.run(optimizer, feed_dict={x: batch_x, y: batch_y})
        acc = sess.run(accuracy, feed_dict={x: batch_x, y: batch_y})
        print(acc)
        c.append(acc)
        end_time = time.time()
        print('time: ', (end_time - start_time))
        start_time = end_time
        print("---------------%d onpech is finished-------------------",i)
    print("Optimization Finished!")

    # Test
    test_acc = sess.run(accuracy, feed_dict={x: X_test, y: y_test})
    print("Testing Accuracy:", test_acc)
    plt.plot(c)
    plt.xlabel('Iter')
    plt.ylabel('Cost')
    plt.title('lr=%f, ti=%d, bs=%d, acc=%f' % (learning_rate, training_iters, batch_size, test_acc))
    plt.tight_layout()
    plt.savefig('cnn-tf-cifar10-%s.png' % test_acc, dpi=200)

運行結果

根據計算機的運行速率不同，在筆者的計算機，大概10s運行一個週期（GPU：GTX1060 6G），訓練200個週期，準確率爲0.62，測試集準確率0.498。

完整代碼

# coding: utf-8

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import _pickle as pickle
import time


def unpickle(filename):
    with open(filename, 'rb') as f:
        d = pickle.load(f, encoding='latin1')
        return d


def onehot(labels):
    '''one-hot 編碼'''
    n_sample = len(labels)
    n_class = max(labels) + 1
    onehot_labels = np.zeros((n_sample, n_class))
    onehot_labels[np.arange(n_sample), labels] = 1
    return onehot_labels


# 訓練數據集
data1 = unpickle('cifar10-dataset/data_batch_1')
data2 = unpickle('cifar10-dataset/data_batch_2')
data3 = unpickle('cifar10-dataset/data_batch_3')
data4 = unpickle('cifar10-dataset/data_batch_4')
data5 = unpickle('cifar10-dataset/data_batch_5')
X_train = np.concatenate((data1['data'], data2['data'], data3['data'], data4['data'], data5['data']), axis=0)
y_train = np.concatenate((data1['labels'], data2['labels'], data3['labels'], data4['labels'], data5['labels']), axis=0)
y_train = onehot(y_train)
# 測試數據集
test = unpickle('cifar10-dataset/test_batch')
X_test = test['data'][:5000, :]
y_test = onehot(test['labels'])[:5000, :]

print('Training dataset shape:', X_train.shape)
print('Training labels shape:', y_train.shape)
print('Testing dataset shape:', X_test.shape)
print('Testing labels shape:', y_test.shape)


# 模型參數
learning_rate = 1e-3
training_iters = 200
batch_size = 50
display_step = 5
n_features = 3072  # 32*32*3
n_classes = 10
n_fc1 = 384
n_fc2 = 192

# 構建模型
x = tf.placeholder(tf.float32, [None, n_features])
y = tf.placeholder(tf.float32, [None, n_classes])

W_conv = {
    'conv1': tf.Variable(tf.truncated_normal([5, 5, 3, 32], stddev=0.0001)),
    'conv2': tf.Variable(tf.truncated_normal([5, 5, 32, 64],stddev=0.01)),
    'fc1': tf.Variable(tf.truncated_normal([8*8*64, n_fc1], stddev=0.1)),
    'fc2': tf.Variable(tf.truncated_normal([n_fc1, n_fc2], stddev=0.1)),
    'fc3': tf.Variable(tf.truncated_normal([n_fc2, n_classes], stddev=0.1))
}
b_conv = {
    'conv1': tf.Variable(tf.constant(0.0, dtype=tf.float32, shape=[32])),
    'conv2': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[64])),
    'fc1': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[n_fc1])),
    'fc2': tf.Variable(tf.constant(0.1, dtype=tf.float32, shape=[n_fc2])),
    'fc3': tf.Variable(tf.constant(0.0, dtype=tf.float32, shape=[n_classes]))
}

x_image = tf.reshape(x, [-1, 32, 32, 3])
# 卷積層 1
conv1 = tf.nn.conv2d(x_image, W_conv['conv1'], strides=[1, 1, 1, 1], padding='SAME')
conv1 = tf.nn.bias_add(conv1, b_conv['conv1'])
conv1 = tf.nn.relu(conv1)
# 池化層 1
pool1 = tf.nn.avg_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')
# LRN層，Local Response Normalization
norm1 = tf.nn.lrn(pool1, 4, bias=1.0, alpha=0.001/9.0, beta=0.75)
# 卷積層 2
conv2 = tf.nn.conv2d(norm1, W_conv['conv2'], strides=[1, 1, 1, 1], padding='SAME')
conv2 = tf.nn.bias_add(conv2, b_conv['conv2'])
conv2 = tf.nn.relu(conv2)
# LRN層，Local Response Normalization
norm2 = tf.nn.lrn(conv2, 4, bias=1.0, alpha=0.001/9.0, beta=0.75)
# 池化層 2
pool2 = tf.nn.avg_pool(norm2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')
reshape = tf.reshape(pool2, [-1, 8*8*64])

fc1 = tf.add(tf.matmul(reshape, W_conv['fc1']), b_conv['fc1'])
fc1 = tf.nn.relu(fc1)
# 全連接層 2
fc2 = tf.add(tf.matmul(fc1, W_conv['fc2']), b_conv['fc2'])
fc2 = tf.nn.relu(fc2)
# 全連接層 3, 即分類層
fc3 = tf.nn.softmax(tf.add(tf.matmul(fc2, W_conv['fc3']), b_conv['fc3']))

# 定義損失
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=fc3, labels=y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(loss)
# 評估模型
correct_pred = tf.equal(tf.argmax(fc3, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

init = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init)
    c = []
    total_batch = int(X_train.shape[0] / batch_size)
#    for i in range(training_iters):
    start_time = time.time()
    for i in range(200):
        for batch in range(total_batch):
            batch_x = X_train[batch*batch_size : (batch+1)*batch_size, :]
            batch_y = y_train[batch*batch_size : (batch+1)*batch_size, :]
            sess.run(optimizer, feed_dict={x: batch_x, y: batch_y})
        acc = sess.run(accuracy, feed_dict={x: batch_x, y: batch_y})
        print(acc)
        c.append(acc)
        end_time = time.time()
        print('time: ', (end_time - start_time))
        start_time = end_time
        print("---------------%d onpech is finished-------------------",i)
    print("Optimization Finished!")

    # Test
    test_acc = sess.run(accuracy, feed_dict={x: X_test, y: y_test})
    print("Testing Accuracy:", test_acc)
    plt.plot(c)
    plt.xlabel('Iter')
    plt.ylabel('Cost')
    plt.title('lr=%f, ti=%d, bs=%d, acc=%f' % (learning_rate, training_iters, batch_size, test_acc))
    plt.tight_layout()
    plt.savefig('cnn-tf-cifar10-%s.png' % test_acc, dpi=200)