deeplab环境安装

2021年12月24日 tf最高版本2.6，对应最新的models文件。在research文件中的deeplab代码是tf1版本 ，因此环境推荐安装

  1. tensorflow = 1.15
  2. cudatoolkit = 10.0.130
  3. cudnn = 7.6.5
  4. opencv-python = 4.5.3.56
  5. labelme = 3.9
  6. pip install labelme -i https://baidu.com/pypi/simple

剩下的缺什么装什么，tf安装好后别的库版本都会自动适配。
测试环境通过后，开始制作数据集。

制作数据集

首先对数据集进行标注，使用labelme,标注时注意标签一致（区分大小写）。
在jpg文件夹下会出现对应的同名json文件、
image.png
首先对每个类别的数据，进行重命名，同时检验标签是否一致：

  1. import os
  2. import json
  3. from PIL import Image
  5. path = "/media/neal/neal/Dataset/medical2021.11.13/YF2021.10.30"
  6. output_path = "/media/neal/neal/Dataset/medical2021.11.13/YF"
  7. class_name = "yf"
  8. count = os.listdir(path)
  9. cnt = 0
  10. for i in range(0, len(count)):
  11.     if count[i].endswith("jpg"):
  12.         cnt += 1
  13.         jpg_path = os.path.join(path, count[i])
  14.         img = Image.open(jpg_path)
  15.         new_img_name =  class_name + '_' + '%05d' % int(cnt) + '.jpg'
  16.         img.save(output_path + '/' + new_img_name)
  18.         json_path = path + "/" + count[i].split('.')[0] + '.json'
  19.         with open(json_path , 'rb') as f:
  20.             params = json.load(f)
  21.             if params["shapes"][0]["label"] != class_name:
  22.                 print("dataset label has error !!!")
  23.             params['imagePath'] = new_img_name
  24.             new_json_name = os.path.join(output_path, new_img_name.replace("jpg", "json"))
  25.             with open(new_json_name, 'w') as r:
  26.                 json.dump(params, r, indent = 4)
  27.                 r.close()
  28.             f.close()

image.png
然后将所有类别的jpg和json文件一起放在jpg_json_dir文件夹中，具体文件夹格式见代码，运行代码：

  1. import os
  2. from PIL import Image
  3. import numpy as np
  4. """
  5. input :  workspace_dir,jpg_json_dir ,class_name.txt
  6. output: png
  7. required pkgs: pip install labelme 
  9. + workspace_dir         root目录
  10.     + jpg_json_dir      在root目录下的子文件夹，存放原始jpg图片和labelme标注后生成的json文件
  11.       - *.jpg
  12.       - *.json
  13.       + labelme_output    所有图片的labelme输出文件夹
  14.     + jpg               所有jpg图片
  15.     + png               用于最终训练的全景图，单通道
  16.     - class_name.txt    按行存放所有的标签，第一行 _background_
  17. """
  18. # 文件路径处理
  19. workspace_dir = "/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical"
  20. assert (os.path.exists(workspace_dir)), "please check workspace_dir folder"
  21. jpg_json_dir = os.path.join(workspace_dir, "jf_and_yf_jpg_json_file")
  22. assert (os.path.exists(jpg_json_dir)), "please check jpg_json_dir folder"
  23. jpg_dir = os.path.join(workspace_dir, "jpg")
  24. png_dir = os.path.join(workspace_dir, "png")
  25. class_name_path = workspace_dir + "/class_name.txt"
  26. if not os.path.exists(class_name_path): print("clsaa_name.txt is not found")
  27. if not os.path.exists(jpg_dir): os.mkdir(jpg_dir)
  28. if not os.path.exists(png_dir): os.mkdir(png_dir)
  30. def generate_labelme_output_file():
  31.     json_file = os.listdir(jpg_json_dir)
  32.     for file in json_file:
  33.         if file.split('.')[1] == 'json':
  34.             os.system("labelme_json_to_dataset  %s" % (jpg_json_dir + '/' + file))  #
  35. def main():
  37.     # 读取原文件夹
  38.     count = os.listdir(jpg_json_dir)
  39.     for i in range(0, len(count)):
  40.         # 如果里的文件以jpg结尾,保存到jpg文件夹,寻找它对应的png
  41.         if count[i].endswith("jpg"):
  42.             jpg_path = os.path.join(jpg_json_dir, count[i])
  43.             img = Image.open(jpg_path)
  44.             img.save(os.path.join(jpg_dir, count[i]))
  45.             # 找到对应的png
  46.             labelme_png_path = jpg_json_dir + "/" + count[i].split(".")[0] + "_json/label.png"
  47.             img = Image.open(labelme_png_path)
  48.             with open(jpg_json_dir + "/" + count[i].split(".")[0] + "_json/label_names.txt", "r") as f:
  49.                 new = Image.new("P", (img.width, img.height))
  50.                 label = f.read().splitlines()[1]
  51.                 # label = f.read().splitlines()
  52.                 # class_txt = open(class_name_path, "r")
  53.                 # class_name = class_txt.read().splitlines()
  54.                 # # 找到局部的类在全局中的类的序号
  55.                 # for name in label:
  56.                 #     # index_json是x_json文件里存在的类label_names.txt，局部类
  57.                 #     index_json = label.index(name)
  58.                 #     # index_all是全局的类,
  59.                 #     index_all = class_name.index(name)
  60.                 #     new = new + (index_all * (np.array(img) == index_json))
  62.                 if label == "jf":
  63.                     new = new + (1 * (np.array(img) == 1))
  64.                 elif label == "yf":
  65.                     new = new + (2 * (np.array(img) == 1))
  66.                 else:
  67.                     print(" label_names has error -->{}".format(labelme_png_path))
  68.             # # 找到全局的类
  69.             # class_txt = open(class_name_path, "r")
  70.             # class_name = class_txt.read().splitlines()
  71.             # print(class_name)
  72.             # # ["bk","cat","dog"] 全局的类
  73.             # # 打开x_json文件里面存在的类，称其为局部类
  74.             # with open(output_dir + "/" + count[i].split(".")[0] + "_json/label_names.txt", "r") as f:
  75.                 # names = f.read().splitlines()
  76.                 # # ["bk","dog"] 局部的类
  77.                 # # 新建一张空白图片, 单通道
  78.                 # new = Image.new("P", (img.width, img.height))
  79.                 #
  80.                 # # 找到局部的类在全局中的类的序号
  81.                 # for name in names:
  82.                 #     # index_json是x_json文件里存在的类label_names.txt，局部类
  83.                 #     index_json = names.index(name)
  84.                 #     # index_all是全局的类,
  85.                 #     index_all = class_name.index(name)
  86.                 #
  87.                 #     # 将局部类转换成为全局类
  88.                 #     # 将原图img中像素点的值为index_json的像素点乘以其在全局中的像素点的所对应的类的序号 得到 其实际在数据集中像素点的值
  89.                 #     # 比如dog,在局部类（output/x_json/label_names）中它的序号为1,dog在原图中的像素点的值也为1.
  90.                 #     # 但是在全局的类（before/classes.txt）中其对应的序号为2，所以在新的图片中要将局部类的像素点的值*全局类的序号，从而得到标签文件
  91.                 #     new = new + (index_all * (np.array(img) == index_json))
  93.             new = Image.fromarray(np.uint8(new))
  94.             new = new.convert("P")
  95.             new.save(os.path.join(png_dir, count[i].replace("jpg", "png")))
  96.             # 找到新的标签文件中像素点值的最大值和最小值，最大值为像素点对应的类在class_name.txt中的序号，最小值为背景，即0
  97.             print(np.max(new), np.min(new), count[i].replace("jpg", "png"))
  100. if __name__ == '__main__':
  101.     #generate_labelme_output_file()
  102.     main()

出现对应的全局mask（或者叫全景图）训练文件，我保存在png文件夹下（注意是单通道图），
image.png

使用showimage文件可以查看单张图片中具体位置的像素值：

  1. import cv2
  2. img= cv2.imread('/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/png/jf_10.png')          #定义图片位置
  3. #img= cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)  #转化为灰度图 因为我读取的直接是灰度标签图片就不用转化了
  4. def onmouse(event, x, y, flags, param):   #标准鼠标交互函数
  5.   if event==cv2.EVENT_MOUSEMOVE:      #当鼠标移动时
  6.     print(img[y,x])           #显示鼠标所在像素的数值，注意像素表示方法和坐标位置的不同
  7. def main():
  8.   cv2.namedWindow("img")          #构建窗口
  9.   cv2.setMouseCallback("img", onmouse)   #回调绑定窗口
  10.   while True:               #无限循环
  11.     cv2.imshow("img",img)        #显示图像
  12.     if cv2.waitKey() == ord('q'):break  #按下‘q'键，退出
  13.   cv2.destroyAllWindows()         #关闭窗口
  14. if __name__ == '__main__':          #运行
  15.   main()

然后，制作index标签：

  1. # 文件名写入txt
  2. import random
  3. import glob
  5. img_path = glob.glob('/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/jf_and_yf_jpg_json_file/*.jpg')
  6. all_txt = '/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/all.txt'
  7. train_txt = "/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/index/train.txt"
  8. trainval_txt = "/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/index/trainval.txt"
  9. val_txt = "/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/index/val.txt"
  10. def clear_file():
  11.     with open(all_txt, 'a') as f:
  12.         f.truncate(0)
  13.     with open(train_txt, 'a') as f:
  14.         f.truncate(0)
  15.     with open(trainval_txt, 'a') as f:
  16.         f.truncate(0)
  17.     with open(val_txt, 'a') as f:
  18.         f.truncate(0)
  20. def split_train_val():
  21.     for each in img_path:
  22.         with open(all_txt, 'a') as f:
  23.             f.write(each.split("/")[-1][:-4] + '\n')  # 切片换成自己路径对应的文件名
  24.     with open('/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/all.txt', 'r') as f:
  25.         lines = f.readlines()
  26.         g = [i for i in range(1, len(img_path)+1)]  # 设置文件总数
  27.         random.shuffle(g)
  28.         # 设置需要的文件数
  29.         train = g[:len(img_path)*7//10]
  30.         print("train num --> {}".format(len(train)))
  31.         trainval = g[len(img_path)*7//10:len(img_path)*9//10]
  32.         print("trainval num --> {}".format(len(trainval)))
  33.         val = g[len(img_path)*9//10:]
  34.         print("val num --> {}".format(len(val)))
  36.         for index, line in enumerate(lines, 1):
  37.             if index in train:
  38.                 with open(train_txt, 'a') as trainf:
  39.                     trainf.write(line)
  40.             elif index in trainval:
  41.                 with open(trainval_txt, 'a') as trainvalf:
  42.                     trainvalf.write(line)
  43.             elif index in val:
  44.                 with open(val_txt, 'a') as valf:
  45.                     valf.write(line)
  47. if __name__ == '__main__':
  48.     clear_file()
  49.     split_train_val()

下一步生成TFRecord，这个代码是官网带的，我在前面添加了一些注释，以及命令行运行的代码，直接复制后命令行运行。

  1. # Lint as: python2, python3
  2. # Copyright 2018 The TensorFlow Authors All Rights Reserved.
  3. #
  4. # Licensed under the Apache License, Version 2.0 (the "License");
  5. # you may not use this file except in compliance with the License.
  6. # You may obtain a copy of the License at
  7. #
  8. #     http://www.apache.org/licenses/LICENSE-2.0
  9. #
  10. # Unless required by applicable law or agreed to in writing, software
  11. # distributed under the License is distributed on an "AS IS" BASIS,
  12. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. # See the License for the specific language governing permissions and
  14. # limitations under the License.
  15. # ==============================================================================
  17. """Converts PASCAL VOC 2012 data to TFRecord file format with Example protos.
  19. PASCAL VOC 2012 dataset is expected to have the following directory structure:
  21.   + pascal_voc_seg
  22.     - build_data.py
  23.     - build_voc2012_data.py (current working directory).
  24.     + VOCdevkit
  25.       + VOC2012
  26.         + JPEGImages
  27.         + SegmentationClass
  28.         + ImageSets
  29.           + Segmentation
  30.     + tfrecord
  32. Image folder:
  33.   ./VOCdevkit/VOC2012/JPEGImages
  35. Semantic segmentation annotations:
  36.   ./VOCdevkit/VOC2012/SegmentationClass
  38. list folder:
  39.   ./VOCdevkit/VOC2012/ImageSets/Segmentation
  41. This script converts data into sharded data files and save at tfrecord folder.
  43. The Example proto contains the following fields:
  45.   image/encoded: encoded image content.
  46.   image/filename: image filename.
  47.   image/format: image file format.
  48.   image/height: image height.
  49.   image/width: image width.
  50.   image/channels: image channels.
  51.   image/segmentation/class/encoded: encoded semantic segmentation content.
  52.   image/segmentation/class/format: semantic segmentation file format.
  53. """
  54. from __future__ import absolute_import
  55. from __future__ import division
  56. from __future__ import print_function
  57. import math
  58. import os.path
  59. import sys
  60. import build_data
  61. from six.moves import range
  62. import tensorflow as tf
  63. # image_format 这个参数只能在命令行给 ，没有仔细研究，先用命令行吧
  64. """
  65. python ./build_voc2012_data.py \
  66.   --image_folder="/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/jpg" \
  67.   --semantic_segmentation_folder="/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/png" \
  68.   --list_folder="/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/index" \
  69.   --image_format="jpg" \
  70.   --output_dir="/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/TFRecord"
  71. """
  73. FLAGS = tf.app.flags.FLAGS
  75. tf.app.flags.DEFINE_string('image_folder',
  76.                            '/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/jpg',
  77.                            'Folder containing images.')
  79. tf.app.flags.DEFINE_string(
  80.     'semantic_segmentation_folder',
  81.     '/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/png',
  82.     'Folder containing semantic segmentation annotations.')
  84. tf.app.flags.DEFINE_string(
  85.     'list_folder',
  86.     '/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/index',
  87.     'Folder containing lists for training and validation')
  89. tf.app.flags.DEFINE_string(
  90.     'output_dir',
  91.     '/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/TFRecord',
  92.     'Path to save converted SSTable of TensorFlow examples.')
  95. _NUM_SHARDS = 4
  98. def _convert_dataset(dataset_split):
  99.   """Converts the specified dataset split to TFRecord format.
  101.   Args:
  102.     dataset_split: The dataset split (e.g., train, test).
  104.   Raises:
  105.     RuntimeError: If loaded image and label have different shape.
  106.   """
  107.   dataset = os.path.basename(dataset_split)[:-4]
  108.   sys.stdout.write('Processing ' + dataset)
  109.   filenames = [x.strip('\n') for x in open(dataset_split, 'r')]
  110.   num_images = len(filenames)
  111.   num_per_shard = int(math.ceil(num_images / _NUM_SHARDS))
  113.   image_reader = build_data.ImageReader('jpeg', channels=3)
  114.   label_reader = build_data.ImageReader('png', channels=1)
  116.   for shard_id in range(_NUM_SHARDS):
  117.     output_filename = os.path.join(
  118.         FLAGS.output_dir,
  119.         '%s-%05d-of-%05d.tfrecord' % (dataset, shard_id, _NUM_SHARDS))
  120.     with tf.python_io.TFRecordWriter(output_filename) as tfrecord_writer:
  121.       start_idx = shard_id * num_per_shard
  122.       end_idx = min((shard_id + 1) * num_per_shard, num_images)
  123.       for i in range(start_idx, end_idx):
  124.         sys.stdout.write('\r>> Converting image %d/%d shard %d' % (
  125.             i + 1, len(filenames), shard_id))
  126.         sys.stdout.flush()
  127.         # Read the image.
  128.         image_filename = os.path.join(
  129.             FLAGS.image_folder, filenames[i] + '.' + FLAGS.image_format)
  130.         image_data = tf.gfile.GFile(image_filename, 'rb').read()
  131.         height, width = image_reader.read_image_dims(image_data)
  132.         # Read the semantic segmentation annotation.
  133.         seg_filename = os.path.join(
  134.             FLAGS.semantic_segmentation_folder,
  135.             filenames[i] + '.' + FLAGS.label_format)
  136.         seg_data = tf.gfile.GFile(seg_filename, 'rb').read()
  137.         seg_height, seg_width = label_reader.read_image_dims(seg_data)
  138.         if height != seg_height or width != seg_width:
  139.           raise RuntimeError('Shape mismatched between image and label.')
  140.         # Convert to tf example.
  141.         example = build_data.image_seg_to_tfexample(
  142.             image_data, filenames[i], height, width, seg_data)
  143.         tfrecord_writer.write(example.SerializeToString())
  144.     sys.stdout.write('\n')
  145.     sys.stdout.flush()
  148. def main(unused_argv):
  149.   dataset_splits = tf.gfile.Glob(os.path.join(FLAGS.list_folder, '*.txt'))
  150.   for dataset_split in dataset_splits:
  151.     _convert_dataset(dataset_split)
  154. if __name__ == '__main__':
  155.   tf.app.run()

生成TFRecode文件后，数据集制作完成。

注册数据集

data_generator.py

在/models/research/deeplab/datasets 路径的data_generator.py 第93行：

  1. _MYDATA_INFORMATION = DatasetDescriptor(
  2.   splits_to_sizes={
  3.   'train':44,  # num of samples in images/training  train.txt的行数
  4.   'val':27,  # num of samples in images/validation val.txt的行数
  5.   },
  6.   num_classes=6,  # 我的标签是5类（包括background），加上ignore_label总共六类
  7.   ignore_label=255,
  8. )

然后找到 _DATASETS_INFORMATION ，加上mydata….一行

  1. _DATASETS_INFORMATION = {
  3.    'cityscapes': _CITYSCAPES_INFORMATION,
  5.     'pascal_voc_seg': _PASCAL_VOC_SEG_INFORMATION,
  7.     'ade20k': _ADE20K_INFORMATION,
  9.     'mydata': _MYDATA_INFORMATION,    #注册上面的数据集 加粗部分一致，前面的mydata随意取
  11. }

同样在models/research/deeplab/deprecated下的segmentation_dataset.py文件进行同样的操作

train_utils.py

在models/research/deeplab/utils/train_utils.py的213行左右

  1. exclude_list = ['global_step','logits'] #本来只有global_step ，现在加上 logits，表示不加载逻辑层的参数
  2. if not initialize_last_layer:
  3. exclude_list.extend(last_layers)

train.py

在models/research/deeplab/train.py里 156行左右

  1. # Set to False if one does not want to re-use the trained classifier weights.
  2. flags.DEFINE_boolean('initialize_last_layer', False, 'Initialize the last layer.') #这个本来是True设置为False
  3. flags.DEFINE_boolean('last_layers_contain_logits_only', True, 'Only consider logits as last layers or not.')#这个设置为True

image.png

input_preprocess.py

  1. models/research/deeplab/input_preprocess.py  128行左右
  3. # Randomly crop the image and label.
  5. if is_trainingand labelis not None:
  7.   processed_image, label = preprocess_utils.random_crop([processed_image, label], crop_height, crop_width)

为这个if加一个else，下面的代码显示得不好，不过在链接里也有
参考链接：https://github.com/tensorflow/models/issues/3695

  1. else:
  3.   rr = tf.minimum(tf.cast(crop_height, tf.float32) / tf.cast(image_height, tf.float32), \
  5.   tf.cast(crop_width, tf.float32) / tf.cast(image_width, tf.float32))
  7.   newh = tf.cast(tf.cast(image_height, tf.float32) * rr, tf.int32)
  9.   neww = tf.cast((tf.cast(image_width, tf.float32) * rr), tf.int32)
  11.   processed_image = tf.image.resize_images(processed_image, (newh, neww), method=tf.image.ResizeMethod.BILINEAR,           align_corners=True)
  13.   processed_image = preprocess_utils.pad_to_bounding_box(processed_image, 0, 0, crop_height, crop_width, mean_pixel)

不加这一段在运行vis.py和eval的时候会报错
类似
InvalidArgumentError (see above for traceback): padded_shape[1]=128 is not divisible by block_shape[1]=12
Invalid argument: padded_shape[1]=69 is not divisible by block_shape[1]=2
这种错基本都是这样解决

训练

  1. python train.py \
  2.     --logtostderr \
  3.     --training_number_of_steps=100000 \
  4.     --train_split="train" \
  5.     --model_variant="xception_65" \
  6.     --atrous_rates=6 \
  7.     --atrous_rates=12 \
  8.     --atrous_rates=18 \
  9.     --output_stride=16 \
  10.     --decoder_output_stride=4 \
  11.     --train_crop_size="513,513" \
  12.     --train_batch_size=4 \
  13.     --dataset="mydata" \
  14.     --tf_initial_checkpoint='/media/neal/neal/deeplab/models/research/deeplab/my_utils/pt/deeplabv3_xception_2018_01_04/xception/model.ckpt' \
  15.     --train_logdir='/media/neal/neal/deeplab/models/research/deeplab/my_utils/result' \
  16.     --dataset_dir='/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/TFRecord'


命令行提示找不到python模块，在python文件中导入包前加入

  1. import sys
  2. sys.path.append("项目根目录")

可视化

  1. python vis.py \
  2.     --logtostderr \
  3.     --vis_split="val" \
  4.     --model_variant="xception_65" \
  5.     --atrous_rates=6 \
  6.     --atrous_rates=12 \
  7.     --atrous_rates=18 \
  8.     --output_stride=16 \
  9.     --decoder_output_stride=4 \
  10.     --vis_crop_size="513,513" \
  11.     --dataset="mydata" \
  12.     --colormap_type="pascal" \
  13.     --checkpoint_dir='/media/neal/neal/deeplab/models/research/deeplab/my_utils/result' \
  14.     --vis_logdir='/media/neal/neal/deeplab/models/research/deeplab/my_utils/result_vis' \
  15.     --dataset_dir='/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/TFRecord'

验证

  1. python eval.py \
  2.     --logtostderr \
  3.     --eval_split="val" \
  4.     --model_variant="xception_65" \
  5.     --atrous_rates=6 \
  6.     --atrous_rates=12 \
  7.     --atrous_rates=18 \
  8.     --output_stride=16 \
  9.     --decoder_output_stride=4 \
  10.     --eval_crop_size="513,513" \
  11.     --dataset="mydata" \
  12.     --checkpoint_dir='/media/neal/neal/deeplab/models/research/deeplab/my_utils/result' \
  13.     --eval_logdir='/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/eval' \
  14.     --dataset_dir='/media/neal/neal/deeplab/models/research/deeplab/my_utils/medical/TFRecord'

Tensorboard

  1. python /home/neal/Python/anaconda/envs/tf115/lib/python3.7/site-packages/tensorboard/main.py \
  2. --logdir="/media/neal/neal/deeplab/models/research/deeplab/my_utils/eval"
  3. python /home/neal/Python/anaconda/envs/tf115/lib/python3.7/site-packages/tensorboard/main.py \
  4. --logdir="/media/neal/neal/deeplab/models/research/deeplab/my_utils/result"
  6. http://neal:6006/

制作数据集，deeplab的过程,deeplab训练自己的数据集，链接1,链接2