模型結構與代碼實現
第一層：32個feature map 5x5卷積、步長爲2、最大值池化 局部相應歸一化處理（LRN） 第二層：64個feature map 3x3卷積、步長爲1、沒有池化 第三層：128個feature map 3x3卷積、步長爲1、最大值池化 局部相應歸一化處理（LRN） 扁平層操作12x12x128個神經元 輸出層操作2個神經元輸出、sigmoid激活函數 卷積層採用relu作爲激活函數。
模型解釋
卷積層深度不斷加深，用以補償分辨率下降帶來的信息損失、 LRN提升神經元競爭能力，增強最終模型的泛化能力。
通過上述簡單的卷積神經網絡，對25000張的貓狗圖像進行訓練，對卷積層1、3後面使用局部響應歸一化處理(LRN), 最終輸出二分類圖像。從測試集選擇測試圖像進行分類預測，計算準確率。
網絡模型代碼實現
def inference(input_tensor):
# -----------------------第一層----------------------------
with tf.variable_scope('layer1-conv1'):
# 初始化權重conv1_weights爲可保存變量，大小爲5x5,3個通道（RGB），數量爲32個
conv1_weights = tf.get_variable("weight", [5, 5, 3, 32],
initializer=tf.truncated_normal_initializer(stddev=0.1))
conv1_biases = tf.get_variable("bias", [32], initializer=tf.constant_initializer(0.0))
conv1 = tf.nn.conv2d(input_tensor, conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases))
print(relu1)
with tf.name_scope("layer2-pool1"):
# 池化計算，調用tensorflow的max_pool函數，strides=[1,2,2,1]，表示池化邊界，2個對一個生成，padding="VALID"表示不操作。
pool1 = tf.nn.max_pool(relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="VALID")
norm1 = tf.nn.lrn(pool1, depth_radius=5, bias=2.0, alpha=1e-3, beta=0.75, name='norm1')
# -----------------------第二層----------------------------
with tf.variable_scope("layer3-conv2"):
# 同上，不過參數的有變化，根據卷積計算和通道數量的變化，64個feature maps
conv2_weights = tf.get_variable("weight", [3, 3, 32, 64],
initializer=tf.truncated_normal_initializer(stddev=0.1))
conv2_biases = tf.get_variable("bias", [64], initializer=tf.constant_initializer(0.0))
conv2 = tf.nn.conv2d(norm1, conv2_weights, strides=[1, 2, 2, 1], padding='SAME')
relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases))
print(relu2)
# Local Response Normalization (parameters from paper)
# 128個 feature maps
with tf.variable_scope("layer4-conv3"):
conv3_weights = tf.get_variable("weight", [3, 3, 64, 128],
initializer=tf.truncated_normal_initializer(stddev=0.1))
conv3_biases = tf.get_variable("bias", [128], initializer=tf.constant_initializer(0.0))
conv3 = tf.nn.conv2d(relu2, conv3_weights, strides=[1, 1, 1, 1], padding='SAME')
relu3 = tf.nn.relu(tf.nn.bias_add(conv3, conv3_biases))
print(relu3)
with tf.name_scope("layer5-pool2"):
pool2 = tf.nn.max_pool(relu3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID')
print(pool2)
norm2 = tf.nn.lrn(pool2, depth_radius=5, bias=2.0, alpha=1e-3, beta=0.75, name='norm2')
# -----------------------全連接層----------------------------
with tf.variable_scope("fc1"):
fc1 = tf.layers.flatten(norm2)
fc2 = tf.layers.dense(fc1, 2, activation=tf.nn.sigmoid)
return fc2數據加載與訓練
對下載的訓練數據集根據名稱排序，分爲兩個目錄
- 文件夾0，所有貓的圖像
- 文件夾1，所有狗的圖像
使用one-hot編碼標籤 [0, 1] 表示貓 [1, 0] 表示狗
加載所有圖像數據與標籤的代碼如下：
def get_filelist():
images = []
labels = []
for root, dirs, files in os.walk('D:/images/train_data/train_img/0'):
for file in files:
file = 'D:/images/train_data/train_img/0/' + file
images.append(file)
labels.append([0, 1])
for root, dirs, files in os.walk('D:/images/train_data/train_img/1'):
for file in files:
file = 'D:/images/train_data/train_img/1/' + file
images.append(file)
labels.append([1, 0])
return np.asarray(images), np.asarray(labels, np.int32)
def get_data(file_list, index, batch_size, label_list):
images = []
labels = []
for i in range(index * batch_size, (1 + index) * batch_size):
i = i % (len(file_list))
img = io.imread(file_list[i])
img = transform.resize(img, (100, 100))
images.append(img)
labels.append(label_list[i])
return np.asarray(images, np.float32), np.asarray(labels, np.int32)每個batch=64張圖像進行訓練，輸入圖像大小resize爲100x100x3, RGB三通道彩色圖像 訓練時候輸入圖像與標籤定義代碼如下：
# 兩個佔位符 x = tf.placeholder(tf.float32, shape=[None, 100, 100, 3], name='x') y_ = tf.placeholder(tf.float32, shape=[None, 2], name='y_')
計算損失採用交叉熵損失，使用Adam優化器進行優化，代碼實現如下：
logits = inference(x)
cross_loss = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=y_)
loss = tf.reduce_mean(cross_loss)
tf.add_to_collection('losses', loss)
# 設置整體學習率爲α爲0.001
train_vars = tf.trainable_variables()
train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss, var_list=train_vars)在1050ti GPU上運行10000次迭代，會保存最後的檢查點文件、訓練與保存檢查點代碼如下：
# 設置爲gpu
tf.device('/gpu:0')
print("training start")
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train_acc = 0
test_acc = 0
train_loss = 0
for epoch in range(n_epoch):
start_time = time.time()
feed_img, feed_label = get_data(x_train, epoch, batch_size, y_train)
_, err, ac = sess.run([train_op, loss, acc], feed_dict={x: feed_img, y_: feed_label})
if epoch % 100 == 0:
print("epoch %d, loss: %.2f, ac : %.2f"%(epoch, err, ac))
saver.save(sess, "./dog_and_cat.model", global_step=10000)使用模型進行預測
定義預測結果代碼如下：
prediction = tf.cast(tf.argmax(logits, 1), tf.float32)
對保存好的檢查點進行恢復，加載隨機測試圖像數據，調用模型進行測試，代碼如下：
with tf.Session() as sess:
saver.restore(sess, tf.train.latest_checkpoint('.'))
cat_path = "D:/images/train_data/test_img/0/"
dog_path = "D:/images/train_data/test_img/1/"
dogs = os.listdir(dog_path)
cats = os.listdir(cat_path)
count = 0
for f in dogs:
if os.path.isfile(os.path. join(dog_path, f)):
image = io.imread(os.path.join(dog_path, f))
copy = np.copy(image)
image = transform.resize(image, (100, 100))
image = np.float32(image)
image_tensor = np.expand_dims(image, 0)
digit = sess.run(prediction, feed_dict={x: image_tensor})
print("predict digit : %d., actual digit : %s"%(digit[0], 0))
if digit[0] == 0:
count = count + 1
cv.putText(copy, "dog", (20, 50), cv.FONT_HERSHEY_SCRIPT_SIMPLEX, 1.0, (0, 0, 255), 2, 8)
cv.imshow("Image Classification", copy)
cv.waitKey(0)
for f in cats:
if os.path.isfile(os.path.join(cat_path, f)):
image = io.imread(os.path.join(cat_path, f))
copy = np.copy(image)
image = transform.resize(image, (100, 100))
image = np.float32(image)
image_tensor = np.expand_dims(image, 0)
digit = sess.run(prediction, feed_dict={x: image_tensor})
print("predict digit : %d., actual digit : %s"%(digit[0], 1))
if digit[0] == 1:
count = count + 1
cv.putText(copy, "cat", (20, 50), cv.FONT_HERSHEY_SCRIPT_SIMPLEX, 1.0, (0, 0, 255), 2, 8)
cv.imshow("Image Classification", copy)
cv.waitKey(0)
print("correct precent: %f"%(count/(len(cats)+len(dogs))))測試運行截圖如下：
