tensorflow实现验证码识别（三）

搭积木

数据都准备好了之后，就剩下定义网络结构了，定义网络结构说白了就是搭积木。我这边的积木是酱紫：

可以看出来我搭了4层卷积+1层FC+1层softmax的辣鸡积木，搭积木的代码是酱紫：

#推理
    def __inference(self, images, keep_prob):
        images = tf.reshape(images, (-1, utility.IMG_HEIGHT, utility.IMG_WIDTH, 1))
        # 用于tensorboard中可视化原图
        tf.summary.image('src_img', images, 5)

        with tf.variable_scope('conv1') as scope:
            kernel = self.__weight_variable('weights_1', shape=[5, 5, 1, 64])
            biases = self.__bias_variable('biases_1', [64])
            pre_activation = tf.nn.bias_add(self.__conv2d(images, kernel), biases)
            conv1 = tf.nn.relu(pre_activation, name=scope.name)
            tf.summary.histogram('conv1/weights_1', kernel)
            tf.summary.histogram('conv1/biases_1', biases)

            kernel_2 = self.__weight_variable('weights_2', shape=[5, 5, 64, 64])
            biases_2 = self.__bias_variable('biases_2', [64])
            pre_activation = tf.nn.bias_add(self.__conv2d(conv1, kernel_2), biases_2)
            conv2 = tf.nn.relu(pre_activation, name=scope.name)
            tf.summary.histogram('conv1/weights_2', kernel_2)
            tf.summary.histogram('conv1/biases_2', biases_2)


            # 用于可视化第一层卷积后的图像
            conv1_for_show1 = tf.reshape(conv1[:, :, :, 1], (-1, 60, 160, 1))
            conv1_for_show2 = tf.reshape(conv1[:, :, :, 2], (-1, 60, 160, 1))
            conv1_for_show3 = tf.reshape(conv1[:, :, :, 3], (-1, 60, 160, 1))
            tf.summary.image('conv1_for_show1', conv1_for_show1, 5)
            tf.summary.image('conv1_for_show2', conv1_for_show2, 5)
            tf.summary.image('conv1_for_show3', conv1_for_show3, 5)

            # max pooling
            pool1 = self.__max_pool_2x2(conv1, name='pool1')



        with tf.variable_scope('conv2') as scope:
            kernel = self.__weight_variable('weights_1', shape=[5, 5, 64, 64])
            biases = self.__bias_variable('biases_1', [64])
            pre_activation = tf.nn.bias_add(self.__conv2d(pool1, kernel), biases)
            conv2 = tf.nn.relu(pre_activation, name=scope.name)
            tf.summary.histogram('conv2/weights_1', kernel)
            tf.summary.histogram('conv2/biases_1', biases)

            kernel_2 = self.__weight_variable('weights_2', shape=[5, 5, 64, 64])
            biases_2 = self.__bias_variable('biases_2', [64])
            pre_activation = tf.nn.bias_add(self.__conv2d(conv2, kernel_2), biases_2)
            conv2 = tf.nn.relu(pre_activation, name=scope.name)
            tf.summary.histogram('conv2/weights_2', kernel_2)
            tf.summary.histogram('conv2/biases_2', biases_2)


            # 用于可视化第二层卷积后的图像
            conv2_for_show1 = tf.reshape(conv2[:, :, :, 1], (-1, 30, 80, 1))
            conv2_for_show2 = tf.reshape(conv2[:, :, :, 2], (-1, 30, 80, 1))
            conv2_for_show3 = tf.reshape(conv2[:, :, :, 3], (-1, 30, 80, 1))
            tf.summary.image('conv2_for_show1', conv2_for_show1, 5)
            tf.summary.image('conv2_for_show2', conv2_for_show2, 5)
            tf.summary.image('conv2_for_show3', conv2_for_show3, 5)

            # max pooling
            pool2 = self.__max_pool_2x2(conv2, name='pool2')



        with tf.variable_scope('conv3') as scope:
            kernel = self.__weight_variable('weights', shape=[3, 3, 64, 64])
            biases = self.__bias_variable('biases', [64])
            pre_activation = tf.nn.bias_add(self.__conv2d(pool2, kernel), biases)
            conv3 = tf.nn.relu(pre_activation, name=scope.name)
            tf.summary.histogram('conv3/weights', kernel)
            tf.summary.histogram('conv3/biases', biases)

            kernel_2 = self.__weight_variable('weights_2', shape=[3, 3, 64, 64])
            biases_2 = self.__bias_variable('biases_2', [64])
            pre_activation = tf.nn.bias_add(self.__conv2d(conv3, kernel_2), biases_2)
            conv3 = tf.nn.relu(pre_activation, name=scope.name)
            tf.summary.histogram('conv3/weights_2', kernel_2)
            tf.summary.histogram('conv3/biases_2', biases_2)

            conv3_for_show1 = tf.reshape(conv3[:, :, :, 1], (-1, 15, 40, 1))
            conv3_for_show2 = tf.reshape(conv3[:, :, :, 2], (-1, 15, 40, 1))
            conv3_for_show3 = tf.reshape(conv3[:, :, :, 3], (-1, 15, 40, 1))
            tf.summary.image('conv3_for_show1', conv3_for_show1, 5)
            tf.summary.image('conv3_for_show2', conv3_for_show2, 5)
            tf.summary.image('conv3_for_show3', conv3_for_show3, 5)

            pool3 = self.__max_pool_2x2(conv3, name='pool3')


        with tf.variable_scope('conv4') as scope:
            kernel = self.__weight_variable('weights', shape=[3, 3, 64, 64])
            biases = self.__bias_variable('biases', [64])
            pre_activation = tf.nn.bias_add(self.__conv2d(pool3, kernel), biases)
            conv4 = tf.nn.relu(pre_activation, name=scope.name)
            tf.summary.histogram('conv4/weights', kernel)
            tf.summary.histogram('conv4/biases', biases)

            conv4_for_show1 = tf.reshape(conv4[:, :, :, 1], (-1, 8, 20, 1))
            conv4_for_show2 = tf.reshape(conv4[:, :, :, 2], (-1, 8, 20, 1))
            conv4_for_show3 = tf.reshape(conv4[:, :, :, 3], (-1, 8, 20, 1))
            tf.summary.image('conv4_for_show1', conv4_for_show1, 5)
            tf.summary.image('conv4_for_show2', conv4_for_show2, 5)
            tf.summary.image('conv4_for_show3', conv4_for_show3, 5)

            pool4 = self.__max_pool_2x2(conv4, name='pool4')



        #全连接层
        with tf.variable_scope('local1') as scope:
            reshape = tf.reshape(pool4, [images.get_shape()[0].value, -1])
            weights = self.__weight_variable('weights', shape=[4*10*64, 1024])
            biases = self.__bias_variable('biases', [1024])
            local1 = tf.nn.relu(tf.matmul(reshape, weights) + biases, name=scope.name)
            tf.summary.histogram('local1/weights', kernel)
            tf.summary.histogram('local1/biases', biases)

            local1_drop = tf.nn.dropout(local1, keep_prob)
            tf.summary.tensor_summary('local1/dropout', local1_drop)

        #输出层
        with tf.variable_scope('softmax_linear') as scope:
            weights = self.__weight_variable('weights', shape=[1024, self.__CHARS_NUM * self.__CLASSES_NUM])
            biases = self.__bias_variable('biases', [self.__CHARS_NUM * self.__CLASSES_NUM])
            result = tf.add(tf.matmul(local1_drop, weights), biases, name=scope.name)

        reshaped_result = tf.reshape(result, [-1, self.__CHARS_NUM, self.__CLASSES_NUM])
        return reshaped_result

训练

第二篇已经说过数据是dump成tfrecords，然后用队列来读取tfrecords，所以训练的时候肯定是要运行队列，让队列从tfrecords里不断的拿数据。但是如果单线程来拿数据的话肯定不是个好主意，所以应该是主线程创建一个专门往队列里塞数据的子线程。主线程在需要的时候只要从队列里拿数据训练就行了。这些线程间同步的操作tf已经封装好了，tf.train.Coordinator()能实例化一个线程协调器，tf.train.start_queue_runners()会把graph里的所有队列run起来，并返回管理队列的对应的子线程。所以代码时酱紫：

        with tf.Session() as sess:
            tf.global_variables_initializer().run()
            tf.local_variables_initializer().run()
            writer = tf.summary.FileWriter(utility.LOG_DIR, sess.graph)
            coord = tf.train.Coordinator()
            threads = tf.train.start_queue_runners(sess=sess, coord=coord)
            try:
                step = 0
                while not coord.should_stop():
                    start_time = time.time()
                    _, loss_value, performance, summaries = sess.run([train_op, loss, accuracy, summary_op])
                    duration = time.time() - start_time
                    if step % 10 == 0:
                        print('>> 已训练%d个批次: loss = %.2f (%.3f sec), 该批正确数量 = %d' % (step, loss_value, duration, performance))
                    if step % 100 == 0:
                        writer.add_summary(summaries, step)
                        saver.save(sess, utility.MODEL_DIR, global_step=step)
                    step += 1
            except tf.errors.OutOfRangeError:
                print('训练结束')
                saver.save(sess, utility.MODEL_DIR, global_step=step)
                coord.request_stop()
            finally:
                coord.request_stop()
            coord.join(threads)

---

demo的完整代码在：https://github.com/aolingwen/fuck_verifycode
代码写的比较搓，欢迎各位大佬指正。如果对你有帮助，还望fork或star一把，谢谢！