新電腦的所有項目資料都沒了，現在想搞一個簡單的CNN多分類問題都還得重新寫好麻煩，簡單記錄一下實現步驟，用於今後備用。
1、以服裝分類爲例，在百度圖片上搜了各類衣服的圖片，裁成方塊，分成5類，每一類都放到一個文件夾，爲了方便取名0-4
2、裁剪成固定尺寸並且鏡面旋轉後閾值化操作（這裏採用的是自適應閾值，也可不使用閾值化，灰度圖亦可）重新保存。
（原圖路徑D:/photo，保存路徑D:/photos）
import cv2
import os
def Threshold(img):
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
GrayImage = cv2.medianBlur(img, 5) # 中值濾波
th = cv2.adaptiveThreshold(GrayImage,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C,\
cv2.THRESH_BINARY_INV,3,5) # 閾值化
return th
def process(path, save_path):
x = -1
for root, dirs, files in os.walk(path):
savepath = save_path + '/' + str(x)
x = x + 1
for index, file in enumerate(files):
img_path = root + '/' + file
img = cv2.imread(img_path, 1)
img_L = cv2.resize(img, (64, 64))
y = 100 + index
img_R = cv2.flip(img_L, 1, dst=None)
th1 = Threshold(img_L)
th2 = Threshold(img_R)
img_saving_path1 = savepath + '/' + str(index) + '.jpg'
img_saving_path2 = savepath + '/' + str(y) + '.jpg'
cv2.imwrite(img_saving_path1, th1)
cv2.imwrite(img_saving_path2, th2)
if __name__ == '__main__':
path = 'D:/photo'
save = 'D:/photos'
process(path, save)
print('Finished validation!')
附上結果：
3、寫一個get_batch函數，用於訓練時獲取數據，新建一個Date.py文件用來存儲
（按照4：1的比例分割訓練集和測試集）
import numpy as np
import os
from PIL import Image
class Conversion:
def __init__(self, batch_size, data_path, ratio=0.8):
"""
:param batch_size:
:param data_path:
"""
self.batch_size = batch_size
self.data_path = data_path
self.labels = []
self.image_names = []
self.count1 = 0
self.count2 = 0
self.dir_names = os.listdir(self.data_path)
for name in self.dir_names:
image_path = os.path.join(self.data_path, name)
image_names = os.listdir(image_path)
for image1 in image_names:
self.labels.append(name)
path = os.path.join(image_path, image1)
self.image_names.append(path)
self.shuffle_set()
s = np.int(len(self.labels) * ratio)
self.x_train = self.image_names[:s]
self.y_train = self.labels[:s]
self.x_test = self.image_names[s:]
self.y_test = self.labels[s:]
print("Data Loading finished!")
def shuffle_set(self):
np.random.seed(12)
np.random.shuffle(self.image_names)
np.random.seed(12)
np.random.shuffle(self.labels)
def get_batch_data(self):
if self.count1 + self.batch_size >= len(self.x_train):
self.count1 = 0
if self.count2 + self.batch_size >= len(self.x_test):
self.count2 = 0
train_name_batch = self.x_train[self.count1: (self.count1 + self.batch_size)]
train_labels = self.y_train[self.count1: (self.count1 + self.batch_size)]
test_name_batch = self.x_test[self.count2: (self.count2 + self.batch_size)]
test_labels = self.y_test[self.count2: (self.count2 + self.batch_size)]
images_train = []
images_test = []
for images_path in train_name_batch:
image = Image.open(images_path).convert('L')
image = np.array(image) / 255.0
images_train.append(image)
datas_train = np.array(images_train)
datas_train = datas_train.reshape((-1, 64, 64, 1))
train_labels = np.array(train_labels)
for images_path in test_name_batch:
image = Image.open(images_path).convert('L')
image = np.array(image) / 255.0
images_test.append(image)
datas_test = np.array(images_test)
datas_test = datas_test.reshape((-1, 64, 64, 1))
test_labels = np.array(test_labels)
self.count1 = self.count1 + self.batch_size
self.count2 = self.count2 + self.batch_size
return datas_train, train_labels, datas_test, test_labels
4、寫一個神經網絡模型，模型都是在網上搜到的代碼，在稍加修改。以下的模型是LeNet5，與train和test函數一起封裝爲對象。文件名稱：Model.py
import tensorflow as tf
import numpy as np
class LeNet5:
def __init__(self, num_channels, label_num):
self.num_channels = num_channels
self.label_num = label_num
self.conv1_size = 5
self.conv1_deep = 32
self.conv2_size = 5
self.conv2_deep = 64
self.fc_size = 512
self.loss_data = []
self.acctr = []
self.accte = []
def inference(self, input, train, regularizer=None):
with tf.variable_scope('layer1-conv1'):
conv1_w = tf.get_variable('w', [self.conv1_size, self.conv1_size, self.num_channels, self.conv1_deep],
initializer=tf.truncated_normal_initializer(stddev=0.1))
conv1_b = tf.get_variable('b', [self.conv1_deep], initializer=tf.constant_initializer(0.0))
self.conv1 = tf.nn.conv2d(input, conv1_w, strides=[1,1,1,1], padding='SAME')
self.relu1 = tf.nn.relu(tf.nn.bias_add(self.conv1, conv1_b))
with tf.name_scope('layer2-pool1'):
self.pool1 = tf.nn.max_pool(self.relu1, ksize=[1, 2, 2, 1],strides=[1,2,2,1],padding='SAME')
with tf.variable_scope('layer3-conv2'):
conv2_w = tf.get_variable('w', [self.conv2_size, self.conv2_size, self.conv1_deep, self.conv2_deep],
initializer=tf.truncated_normal_initializer(stddev=0.1))
conv2_b = tf.get_variable('b', [self.conv2_deep], initializer=tf.constant_initializer(0.0))
self.conv2 = tf.nn.conv2d(self.pool1, conv2_w, strides=[1,1,1,1], padding='SAME')
self.relu2 = tf.nn.relu(tf.nn.bias_add(self.conv2, conv2_b))
#池化層 7*7*64
with tf.name_scope('layer4-pool2'):
self.pool2 = tf.nn.max_pool(self.relu2, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')
pool_shape = self.pool2.get_shape()
nodes = pool_shape[1] * pool_shape[2] * pool_shape[3]
reshaped = tf.reshape(self.pool2, [-1, int(nodes)])
with tf.variable_scope('layer5-fc1'):
fc1_w = tf.get_variable('w', [nodes, self.fc_size],
initializer=tf.truncated_normal_initializer(stddev=0.1))
fc1_b = tf.get_variable('b', [self.fc_size], initializer=tf.constant_initializer(0.1))
if regularizer != None:
tf.add_to_collection('losses', regularizer(fc1_w))
self.fc1 = tf.nn.relu(tf.matmul(reshaped, fc1_w)+fc1_b)
if train == True:
self.fc1 = tf.nn.dropout(self.fc1, 0.5)
with tf.variable_scope('layer6-fc2'):
fc2_w = tf.get_variable('w', [self.fc_size, self.label_num], initializer=tf.truncated_normal_initializer(stddev=0.1))
fc2_b = tf.get_variable('b', [self.label_num], initializer=tf.constant_initializer(0.1))
if regularizer != None:
tf.add_to_collection('losses', regularizer(fc2_w))
self.logis = tf.matmul(self.fc1, fc2_w)+fc2_b
return self.logis
def train(self, data, steps, learning_rate, save_dir, regular_rate=0.0001, moving_deacy=0.99):
self.learing_rate = learning_rate
self.save_dir = save_dir
x = tf.placeholder(tf.float32, [None, 64, 64, 1], name='x-input')
y = tf.placeholder(tf.int64, [None, 5], name='y-input')
regularizer = tf.contrib.layers.l2_regularizer(regular_rate)
y_ = self.inference(x, True, regularizer)
global_step = tf.Variable(0, trainable=False)
variable_averages = tf.train.ExponentialMovingAverage(moving_deacy, global_step)
variable_averages_op = variable_averages.apply(tf.trainable_variables())
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_, labels=y))
loss = cross_entropy + tf.add_n(tf.get_collection('losses'))
train_step = tf.train.GradientDescentOptimizer(learning_rate=self.learing_rate).minimize(loss,
global_step=global_step)
correct_prediction = tf.equal(tf.arg_max(y_, 1), tf.arg_max(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
with tf.control_dependencies([train_step, variable_averages_op]):
train_op = tf.no_op(name='train')
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(steps):
x_train, y_train, x_test, y_test = data.get_batch_data()
train_y = sess.run(tf.one_hot(y_train, self.label_num))
test_y = sess.run(tf.one_hot(y_test, self.label_num))
_, loss_value, step, acc = sess.run([train_op, loss, global_step, accuracy],
feed_dict={x: x_train, y: train_y})
acc_test = sess.run(accuracy, feed_dict={x: x_test, y: test_y})
if i == 0:
print('%d輪的損失率爲%g, 訓練集的正確率爲%.2f%%, 測試集的正確率爲%.2f%%' % (
i, loss_value, acc * 100.0, acc_test * 100.0))
self.loss_data.append(loss_value)
self.acctr.append(acc)
self.accte.append(acc_test)
if (i + 1) % 10 == 0:
print('%d輪之後的損失率爲%g, 訓練集的正確率爲%.2f%%, 測試集的正確率爲%.2f%%' % (
i + 1, loss_value, acc * 100.0, acc_test * 100.0))
self.loss_data.append(loss_value)
self.acctr.append(acc)
self.accte.append(acc_test)
saver.save(sess, self.save_dir)
def test(self, image, model_path):
x = tf.placeholder(tf.float32, [None, 64, 64, 1], name='x-input')
y_ = self.inference(x, False)
image = np.array(image) / 255.0
image = np.reshape(image, (-1, 64, 64, 1))
logit = tf.arg_max(y_, 1)
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
saver.restore(sess, model_path)
y, label = sess.run((y_, logit), feed_dict={x:image})
return label
5、一個新的文件中調用，實現訓練和測試，命名爲main.py
import Process
import Model
import Data
import cv2
def train(model, batch_size, path, steps, learning_rate, save_dir):
data = Data.Conversion(batch_size, path)
model.train(data=data, steps=steps, learning_rate=learning_rate, save_dir=save_dir)
def train_save():
logs_train_dir = "D:/model/model/model.ckpt"
model = Model.LeNet5(1, 5)
train(model, batch_size=20, path='D:/photos/', steps=500, learning_rate=0.05, save_dir=logs_train_dir)
def restore_test():
image_path = "D:/image/test/0.jpg"
image_data = cv2.imread(image_path, 1)
img_size = cv2.resize(image_data, (64, 64))
image = Process.Threshold(img_size)
model = Model.LeNet5(1, 5)
path = "D:/model/model/model.ckpt"
sort = model.test(image, path)
print("識別結果爲：" + str(sort))
if __name__ == "__main__":
#train_save()
restore_test()
6、先運行train_save函數，再運行restore_test函數，模型保存位置爲"D:/model/model/model.ckpt"