Tensoflow sess.run导致的内存溢出

       下面是调用模型进行批量测试的代码(出现溢出)，开始以为导致溢出的原因是数据读入方式问题引起的，用了tf , PIL和cv等方式读入图片数据，发现越来越慢，内存占用飙升，调试时发现是sess.run这里出了问题（随着for循环进行速度越来越慢）。

    # Creates graph from saved GraphDef
    create_graph(pb_path)

    # Init tf Session
    config = tf.ConfigProto()
    config.gpu_options.allow_growth = True
    sess = tf.Session(config=config)
    init = tf.global_variables_initializer()
    sess.run(init)


    input_image_tensor = sess.graph.get_tensor_by_name("create_inputs/batch:0") 
    output_tensor_name = sess.graph.get_tensor_by_name("conv6/out_1:0")  


    for filename in os.listdir(image_dir):
        image_path = os.path.join(image_dir, filename)
  
        start = time.time()
        image_data = cv2.imread(image_path)
        image_data = cv2.resize(image_data, (w, h))
        image_data_1 = image_data - IMG_MEAN
        input_image = np.expand_dims(image_data_1, 0)

        raw_output_up = tf.image.resize_bilinear(output_tensor_name, size=[h, w], align_corners=True) 
        raw_output_up = tf.argmax(raw_output_up, axis=3)
        

        predict_img = sess.run(raw_output_up, feed_dict={input_image_tensor: input_image})       # 1，height，width
        predict_img = np.squeeze(predict_img)     #  height， width 

        voc_palette = visual.make_palette(3)
        masked_im = visual.vis_seg(image_data, predict_img, voc_palette)
        cv2.imwrite("%s_pred.png" % (save_dir + filename.split(".")[0]), masked_im)


        print(time.time() - start)

    print(">>>>>>Done")

下面是解决溢出问题的代码（将部分代码放在for循环外）

    # Creates graph from saved GraphDef
    create_graph(pb_path)

    # Init tf Session
    config = tf.ConfigProto()
    config.gpu_options.allow_growth = True
    sess = tf.Session(config=config)
    init = tf.global_variables_initializer()
    sess.run(init)

    input_image_tensor = sess.graph.get_tensor_by_name("create_inputs/batch:0") 
    output_tensor_name = sess.graph.get_tensor_by_name("conv6/out_1:0")  
    
##############################################################################################################
    raw_output_up = tf.image.resize_bilinear(output_tensor_name, size=[h, w], align_corners=True) 
    raw_output_up = tf.argmax(raw_output_up, axis=3)
##############################################################################################################

    for filename in os.listdir(image_dir):
        image_path = os.path.join(image_dir, filename)
  
        start = time.time()
        image_data = cv2.imread(image_path)
        image_data = cv2.resize(image_data, (w, h))
        image_data_1 = image_data - IMG_MEAN
        input_image = np.expand_dims(image_data_1, 0)
        
        predict_img = sess.run(raw_output_up, feed_dict={input_image_tensor: input_image})       # 1，height，width
        predict_img = np.squeeze(predict_img)     #  height， width 

        voc_palette = visual.make_palette(3)
        masked_im = visual.vis_seg(image_data, predict_img, voc_palette)
        cv2.imwrite("%s_pred.png" % (save_dir + filename.split(".")[0]), masked_im)
        print(time.time() - start)

    print(">>>>>>Done")

 总结:

1. 在迭代过程中, 在sess.run的for循环中不要加入tensorflow一些op操作，会增加图节点，否则随着迭代的进行，tf的图会越来越大，最终导致溢出；
2. 建议不要使用tf.gfile.FastGFile(image_path, 'rb').read()读入数据（有可能会造成溢出），用opencv之类读取。