MAC下deeplab_V3+测试自带例程“local_test_mobilenetv2.sh”

deeplab_V3自带测试shell（如下图），现在我们测试下mobilentv2网络。

1. 修改“local_test_mobilenetv2.sh”第75行的迭代次数（原来是10），10的情况下训练的网络非常差，vis测试时识别的都是乱七八糟的东西。这里我改成了10000，识别出的效果会好很多。

   # Train 10 iterations.
   NUM_ITERATIONS=10000

    

2.  改“local_test_mobilenetv2.sh”第45行执行另一个shell

   sh download_and_convert_voc2012.sh

   我们进入到这个shell中（再datasets文件夹里），然后屏蔽掉第60行

   # Download the images.
   BASE_URL="http://host.robots.ox.ac.uk/pascal/VOC/voc2012/"
   FILENAME="VOCtrainval_11-May-2012.tar"
   
   #download_and_uncompress "${BASE_URL}" "${FILENAME}"
   

   因为这是在线下载并解压，数据包有2个G大小，太慢了，所以屏蔽掉后我们手动下载来解压。

3. 我把数据分享到了网盘，需要的可以下载。链接: https://pan.baidu.com/s/1iaGVUMY6Dh5RSzlcUbNSYQ  密码: o39j ，然后按照“build_voc2012_data.py”提示，在datasets下创建文件夹pascal_voc_seg，然后在pascal_voc_seg里创建文件夹VOCdevkit，把解压的文件夹VOC2012移动到这里。

   PASCAL VOC 2012 dataset is expected to have the following directory structure:
   
     + pascal_voc_seg
       - build_data.py
       - build_voc2012_data.py (current working directory).
       + VOCdevkit
         + VOC2012
           + JPEGImages
           + SegmentationClass
           + ImageSets
             + Segmentation
       + tfrecord

    

4. 运行“local_test_mobilenetv2.sh”就可以了，训练结果如下：

   INFO:tensorflow:global step 9890: loss = 0.6183 (3.877 sec/step)
   INFO:tensorflow:global step 9900: loss = 0.6939 (3.913 sec/step)
   INFO:tensorflow:global step 9910: loss = 0.7819 (3.866 sec/step)
   INFO:tensorflow:global step 9920: loss = 0.8424 (3.875 sec/step)
   INFO:tensorflow:global step 9930: loss = 0.3926 (3.896 sec/step)
   INFO:tensorflow:global step 9940: loss = 0.6908 (3.869 sec/step)
   INFO:tensorflow:global step 9950: loss = 0.5328 (3.899 sec/step)
   INFO:tensorflow:Recording summary at step 9958.
   INFO:tensorflow:global step 9960: loss = 0.3902 (3.894 sec/step)
   INFO:tensorflow:global step 9970: loss = 0.5666 (3.880 sec/step)
   INFO:tensorflow:global step 9980: loss = 0.7343 (3.884 sec/step)
   INFO:tensorflow:global step 9990: loss = 0.7285 (3.885 sec/step)
   INFO:tensorflow:global step 10000: loss = 0.6550 (3.875 sec/step)
   INFO:tensorflow:Stopping Training.
   INFO:tensorflow:Finished training! Saving model to disk.
   

   评估结果如下：

   INFO:tensorflow:Starting evaluation at 2020-09-07-20:17:50
   INFO:tensorflow:Evaluation [144/1449]
   INFO:tensorflow:Evaluation [288/1449]
   INFO:tensorflow:Evaluation [432/1449]
   INFO:tensorflow:Evaluation [576/1449]
   INFO:tensorflow:Evaluation [720/1449]
   INFO:tensorflow:Evaluation [864/1449]
   INFO:tensorflow:Evaluation [1008/1449]
   INFO:tensorflow:Evaluation [1152/1449]
   INFO:tensorflow:Evaluation [1296/1449]
   INFO:tensorflow:Evaluation [1440/1449]
   INFO:tensorflow:Evaluation [1449/1449]
   INFO:tensorflow:Finished evaluation at 2020-09-07-20:20:35
   miou_1.0[0.548890591]
   

   可视化输出如下：

   

   当然，更改训练迭代次数越多结果会越好。我这里很明显有些图片并不理想。可能迭代次数改到30000就好了吧。