def build_model(model_path):
  model_input = tf.placeholder('float32', [1, IMAGE_HEIGHT, IMAGE_WIDTH, COLOR_CHANNELS])
  vec1 = ...
  ...... 
  return model_input,vec1
  
def get_style_represent(vec):
  # 一些tf计算操作
  return new_vec

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  img_input,vec1 = build_model(VGG19_MODEL)    # 加载模型
  for cur_img_path in imgs_path_list:   # 遍历图片
    cur_img = load_image(cur_img_path)
    vec1_out = sess.run(vec1, feed_dict = {img_input:cur_img})   # 正向传播输出模型中的vec1
    # 对vec1进行一些处理，此处在遍历图片时额外使用了tensorflow的计算节点，导致在内存中遗留信息
    new_vec = get_style_represent(vec1_out)  

def build_model(model_path):
  model_input = tf.placeholder('float32', [1, IMAGE_HEIGHT, IMAGE_WIDTH, COLOR_CHANNELS])
  vec1 = ...
  ...... 
  new_vec = ...    # 将get_style_represent计算操作定义在模型中
  return model_input,vec1,new_vec

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  img_input,vec1,new_vec = build_model(VGG19_MODEL)
  for cur_img_path in imgs_path_list:
    cur_img = load_image(cur_img_path)
    # 一次正向传播将处理后的vec1也得到了，避免在每次图片正向传播时留下额外信息
    vec1_out,new_vec_out = sess.run([vec1,new_vec], feed_dict = {img_input:cur_img})