每次识别都会重新加载模型 about captcha_recognize HOT 2 OPEN

with tf.variable_scope('conv1') as scope:
    kernel = _weight_variable('weights', shape=[3, 3, 1, 64])
    biases = _bias_variable('biases', [64])
    pre_activation = tf.nn.bias_add(_conv2d(images, kernel), biases)
    conv1 = tf.nn.relu(pre_activation, name=scope.name)

pool1 = _max_pool_2x2(conv1, name='pool1')

with tf.variable_scope('conv2') as scope:
    kernel = _weight_variable('weights', shape=[3, 3, 64, 64])
    biases = _bias_variable('biases', [64])
    pre_activation = tf.nn.bias_add(_conv2d(pool1, kernel), biases)
    conv2 = tf.nn.relu(pre_activation, name=scope.name)

pool2 = _max_pool_2x2(conv2, name='pool2')

with tf.variable_scope('conv3') as scope:
    kernel = _weight_variable('weights', shape=[3, 3, 64, 64])
    biases = _bias_variable('biases', [64])
    pre_activation = tf.nn.bias_add(_conv2d(pool2, kernel), biases)
    conv3 = tf.nn.relu(pre_activation, name=scope.name)

pool3 = _max_pool_2x2(conv3, name='pool3')

with tf.variable_scope('conv4') as scope:
    kernel = _weight_variable('weights', shape=[3, 3, 64, 64])
    biases = _bias_variable('biases', [64])
    pre_activation = tf.nn.bias_add(_conv2d(pool3, kernel), biases)
    conv4 = tf.nn.relu(pre_activation, name=scope.name)

pool4 = _max_pool_2x2(conv4, name='pool4')

with tf.variable_scope('local1') as scope:
    batch_size = 1  # images.get_shape()[0].value
    reshape = tf.reshape(pool4, [batch_size, -1])
    # dim = reshape.get_shape()[1].value
    # dim = 512  # for 26x60
    dim = 2048 # for 52x120
    # dim = 3200 # for 70x160
    weights = _weight_variable('weights', shape=[dim, 1024])
    biases = _bias_variable('biases', [1024])
    local1 = tf.nn.relu(tf.matmul(reshape, weights) + biases, name=scope.name)

local1_drop = tf.nn.dropout(local1, keep_prob)

with tf.variable_scope('softmax_linear') as scope:
    weights = _weight_variable('weights', shape=[1024, CHARS_NUM * CLASSES_NUM])
    biases = _bias_variable('biases', [CHARS_NUM * CLASSES_NUM])
    softmax_linear = tf.add(tf.matmul(local1_drop, weights), biases, name=scope.name)

return tf.reshape(softmax_linear, [-1, CHARS_NUM, CLASSES_NUM])