Mask R-CNN 训练自己的数据集—踩坑与填坑

tensorflow/keras版（项目地址https://github.com/matterport）。

服务器cuda为8.0版本，尝试升级成9.1没成功，之后再降级回8.0了，python版本一开始为py36，提示错误libcublas.so.8.0:cannot open shared object file:No such file or directory，经过一番倒腾，tensorflow/keras降级等等，依然无效，最后把py36降级为py35反倒成功了，测试了py36以下版本都能兼容cuda-8。 以下是在使用Mask-RCNN训练自己数据时遇到的各种坑的记录，哈哈，大数据scala写习惯了，突然转到深度学习来，还有点吃不消。
 

一些步骤：

安装anaconda(Python 3.4+)

1.下载installer清华镜像https://mirrors.tuna.tsinghua.edu.cn/anaconda/archive/
2.进入下载好文件的文件夹，在命令行中输入：
bash Anaconda2-5.0.1-Linux-x86_64.sh
3.查看现有环境
conada list  

新建环境：
conda create --name py35 python=3.5
source activate py35

需要软件列表：
numpy
scipy
Pillow
cython
matplotlib
scikit-image
tensorflow>=1.3.0
keras>=2.0.8
opencv-python
h5py
imgaug
IPython

pip安装tensorflow太慢，可以下载安装：
版本选择网址（https://mirrors.tuna.tsinghua.edu.cn/help/tensorflow/）
pip install \
 -i https://pypi.tuna.tsinghua.edu.cn/simple/ \
 https://mirrors.tuna.tsinghua.edu.cn/tensorflow/linux/gpu/tensorflow_gpu-1.4.1-cp35-cp35m-linux_x86_64.whl

安装cuda的方法：
sudo dpkg -i cuda-repo-ubuntu1604-9-1-local_9.1.85-1_amd64.deb
sudo apt-key add /var/cuda-repo-9-1-local/7fa2af80.pub 
sudo apt-get update
sudo apt-get install cuda （sudo apt-get upgrade cuda）
在.bashrc中写入：
export CUDA_HOME=/usr/local/cuda （对cuda-8.0做了软链接）
export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/usr/local/cuda/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

打印cuda版本信息：
nvcc -V   

（0）
###################################--（项目过程）--##########################################

（1）resize图片成统一大小
（2）labelme标注图片，得到json标识文件
（3）生成一些文件，
     (py35) byz@ubuntu:~/Mask/Mask_RCNN/net_img$ python json_to_dataset.py rgb_5.json 
     # 移动生成文件到指定目录
     (py35) byz@ubuntu:~/Mask/Mask_RCNN/net_img$ mv rgb_*_json json/
     (py35) byz@ubuntu:~/Mask/Mask_RCNN/net_img$ mv rgb*.jpg rgb/
（4）将掩码标签 label.png 16位转8位，存到./mask/下
     (py35) byz@ubuntu:~/Mask/Mask_RCNN/net_img$ python 16_to_8.py 
（5）运行程序
     (py35) byz@ubuntu:~/Mask/Mask_RCNN$ python train_rcnn_test.py 


遇到的问题：

（一）
###################--（制作训练所需数据集：label.png、info.yaml）--#########################

byz@ubuntu:~/app/coco/tmp$ labelme_json_to_dataset rgb_1.jpg

问题：
/home/byz/anaconda3/lib/python3.6/site-packages/labelme/cli/json_to_dataset.py:14: UserWarning: This script is aimed to demonstrate how to convert the
JSON file to a single image dataset, and not to handle
multiple JSON files to generate a real-use dataset.
  warnings.warn("This script is aimed to demonstrate how to convert the\n"
Traceback (most recent call last):
  File "/home/byz/anaconda3/bin/labelme_json_to_dataset", line 11, in
    sys.exit(main())
  File "/home/byz/anaconda3/lib/python3.6/site-packages/labelme/cli/json_to_dataset.py", line 33, in main
    data = json.load(open(json_file))
  File "/home/byz/anaconda3/lib/python3.6/json/__init__.py", line 296, in load
    return loads(fp.read(),
  File "/home/byz/anaconda3/lib/python3.6/codecs.py", line 321, in decode
    (result, consumed) = self._buffer_decode(data, self.errors, final)
UnicodeDecodeError: 'utf-8' codec can't decode byte 0xff in position 0: invalid start byte

解决：
自己写json_to_dataset.py

import argparse
import json
import os
import os.path as osp
import warnings
import PIL.Image
import yaml
from labelme import utils
def main():
   warnings.warn("This script is aimed to demonstrate how to convert the\n"
         "JSON file to a single image dataset, and not to handle\n"
         "multiple JSON files to generate a real-use dataset.")
   parser = argparse.ArgumentParser()
   parser.add_argument('json_file')
   parser.add_argument('-o', '--out', default=None)
   args = parser.parse_args()
   json_file = args.json_file
   if args.out is None:
out_dir = osp.basename(json_file).replace('.', '_')
out_dir = osp.join(osp.dirname(json_file), out_dir)
   else:
out_dir = args.out
   if not osp.exists(out_dir):
os.mkdir(out_dir)
   data = json.load(open(json_file))
   print("json---load")
   img = utils.img_b64_to_arr(data['imageData'])
   label_name_to_value = {'_background_': 0}
   for shape in data['shapes']:
label_name = shape['label']
if label_name in label_name_to_value:
   label_value = label_name_to_value[label_name]
else:
   label_value = len(label_name_to_value)
   label_name_to_value[label_name] = label_value
   # label_values must be dense
   label_values, label_names = [], []
   for ln, lv in sorted(label_name_to_value.items(), key=lambda x: x[1]):
label_values.append(lv)
label_names.append(ln)
   assert label_values == list(range(len(label_values)))
   lbl = utils.shapes_to_label(img.shape, data['shapes'], label_name_to_value)
   captions = ['{}: {}'.format(lv, ln)
       for ln, lv in label_name_to_value.items()]
   lbl_viz = utils.draw_label(lbl, img, captions)
   PIL.Image.fromarray(img).save(osp.join(out_dir, 'img.png'))
   PIL.Image.fromarray(lbl).save(osp.join(out_dir, 'label.png'))
   PIL.Image.fromarray(lbl_viz).save(osp.join(out_dir, 'label_viz.png'))
   with open(osp.join(out_dir, 'label_names.txt'), 'w') as f:
for lbl_name in label_names:
   f.write(lbl_name + '\n')
   warnings.warn('info.yaml is being replaced by label_names.txt')
   info = dict(label_names=label_names)
   with open(osp.join(out_dir, 'info.yaml'), 'w') as f:
yaml.safe_dump(info, f, default_flow_style=False)
   print('Saved to: %s' % out_dir)
if __name__ == '__main__':
   main()

使用：
(py35) byz@ubuntu:~/Mask/Mask_RCNN/net_img$ python json_to_dataset.py rgb_5.json 



（二）
##################--（ubuntu_opencv将16位灰度图片转化为8位——遇到的编译问题）--###################
大家在加载灰度图时，一定要看准图片存储格式位数 
opencv默认为8位读取，如果该图为16位，则读取为全0，导致程序出错 
以下代码只需修改路径，可以批量处理图片

#include
#include
#include
#include
using namespace std;
using namespace cv;
int main(void){
    char buff1[100];
    char buff2[100];
    for(int i=1;i<901;i++){
        sprintf(buff1,"/home/byz/app/coco/tmp/json/rgb_%d_json/label.png",i);
        sprintf(buff2,"/home/byz/app/coco/tmp/mask/%d.png",i);
        Mat src;
        //Mat dst;
        src=imread(buff1,CV_LOAD_IMAGE_UNCHANGED);
        Mat ff=Mat::zeros(src.rows,src.cols,CV_8UC1);
        for(int k=0;k             for(int kk=0;kk                 int n=src.at(k,kk);
                ff.at(k,kk)=n;
            }
        }
        //src.copyTo(dst);
        //imshow("haha",ff*100);
        //waitKey(0);
        imwrite(buff2,ff);
    }
    return 0;
}

ubuntu_opencv将16位灰度图片转化为8位——遇到的编译问题：

(py35) byz@ubuntu:~/Mask/Mask_RCNN/net_img$ sudo g++ 16_to_8.cpp -lpthread -o 16_to_8
/tmp/ccKQtPkQ.o: In function `main':
16_to_8.cpp:(.text+0xaf): undefined reference to `cv::imread(cv::String const&, int)'
16_to_8.cpp:(.text+0x106): undefined reference to `cv::Mat::zeros(int, int, int)'
16_to_8.cpp:(.text+0x222): undefined reference to `cv::imwrite(cv::String const&, cv::_InputArray const&, std::vector > const&)'
/tmp/ccKQtPkQ.o: In function `cv::String::String(char const*)':
16_to_8.cpp:(.text._ZN2cv6StringC2EPKc[_ZN2cv6StringC5EPKc]+0x54): undefined reference to `cv::String::allocate(unsigned long)'
/tmp/ccKQtPkQ.o: In function `cv::String::~String()':
16_to_8.cpp:(.text._ZN2cv6StringD2Ev[_ZN2cv6StringD5Ev]+0x14): undefined reference to `cv::String::deallocate()'
/tmp/ccKQtPkQ.o: In function `cv::String::operator=(cv::String const&)':
16_to_8.cpp:(.text._ZN2cv6StringaSERKS0_[_ZN2cv6StringaSERKS0_]+0x28): undefined reference to `cv::String::deallocate()'
/tmp/ccKQtPkQ.o: In function `cv::Mat::~Mat()':
16_to_8.cpp:(.text._ZN2cv3MatD2Ev[_ZN2cv3MatD5Ev]+0x39): undefined reference to `cv::fastFree(void*)'
/tmp/ccKQtPkQ.o: In function `cv::Mat::operator=(cv::Mat const&)':
16_to_8.cpp:(.text._ZN2cv3MataSERKS0_[_ZN2cv3MataSERKS0_]+0x115): undefined reference to `cv::Mat::copySize(cv::Mat const&)'
/tmp/ccKQtPkQ.o: In function `cv::Mat::release()':
16_to_8.cpp:(.text._ZN2cv3Mat7releaseEv[_ZN2cv3Mat7releaseEv]+0x4b): undefined reference to `cv::Mat::deallocate()'
collect2: error: ld returned 1 exit status

解决：
(py35) byz@ubuntu:~/Mask/Mask_RCNN/net_img$ sudo g++ 16_to_8.cpp -lpthread -o 16_to_8 `pkg-config --cflags --libs opencv`
(py35) byz@ubuntu:~/Mask/Mask_RCNN/net_img$ ll
total 164
drwxrwxr-x 5 byz  byz   4096 Apr 27 11:58 ./
drwxrwxr-x 8 byz  byz   4096 Apr 27 11:38 ../
-rwxr-xr-x 1 root root 59424 Apr 27 11:58 16_to_8*
-rw-rw-r-- 1 byz  byz    833 Apr 27 11:48 16_to_8.cpp

补充（如果是pthread问题）：
/tmp/ccM2tvqF.o: In function `main':
thread_c.c:(.text+0x1f): undefined reference to `pthread_create'
thread_c.c:(.text+0x52): undefined reference to `pthread_create'
thread_c.c:(.text+0x7d): undefined reference to `pthread_join'
thread_c.c:(.text+0xa9): undefined reference to `pthread_join'
collect2: ld returned 1 exit status

解决：
因为pthread库不是Linux系统默认的库，连接时需要使用库libpthread.a,所以在使用pthread_create创建线程时，在编译中要加-lpthread参数:gcc test_thread.c -lpthread -o test_thread. 


（三）
###############--（运行自己的代码:~/Mask/Mask_RCNN$ python train_rcnn_test.py报错）--###################
问题：
（RuntimeError: Invalid DISPLAY variable）

具体：
Traceback (most recent call last):
  File "train_rcnn_test.py", line 210, in
    visualize.display_top_masks(image, mask, class_ids, dataset_train.class_names)
  File "/home/byz/Mask/Mask_RCNN/mrcnn/visualize.py", line 304, in display_top_masks
    display_images(to_display, titles=titles, cols=limit + 1, cmap="Blues_r")
  File "/home/byz/Mask/Mask_RCNN/mrcnn/visualize.py", line 48, in display_images
    plt.figure(figsize=(14, 14 * rows // cols))
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/matplotlib/pyplot.py", line 548, in figure
    **kwargs)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/matplotlib/backend_bases.py", line 161, in new_figure_manager
    return cls.new_figure_manager_given_figure(num, fig)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/matplotlib/backend_bases.py", line 167, in new_figure_manager_given_figure
    canvas = cls.FigureCanvas(figure)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/matplotlib/backends/backend_qt5agg.py", line 24, in __init__
    super(FigureCanvasQTAgg, self).__init__(figure=figure)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/matplotlib/backends/backend_qt5.py", line 234, in __init__
    _create_qApp()
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/matplotlib/backends/backend_qt5.py", line 125, in _create_qApp
    raise RuntimeError('Invalid DISPLAY variable')
RuntimeError: Invalid DISPLAY variable

１．问题：在本地用matplotlib绘图可以，但是在ssh远程绘图的时候会报错　RuntimeError: Invalid DISPLAY variable

２．原因：matplotlib的默认backend是TkAgg，而FltkAgg, GTK, GTKAgg, GTKCairo, TkAgg , Wx or WxAgg这几个backend都要求有GUI图形界面的，所以在ssh操作的时候会报错．

import matplotlib.pyplot as plt
plt.get_backend()

解决：
指定不需要GUI的backend（Agg, Cairo, PS, PDF or SVG）
import matplotlib.pyplot as plt
plt.switch_backend('agg')

修改源码：
mrcnn/visualize.py -- 
def display_images():
   titles = titles if titles is not None else [""] * len(images)
   plt.switch_backend('agg')   #<-------增加
   rows = len(images) // cols + 1
   plt.figure(figsize=(14, 14 * rows // cols))

 

（四）
#############################--（训练MaskRCNN模型）--###############################

执行：
model = modellib.MaskRCNN(mode="training", config=config, model_dir=MODEL_DIR)
model.load_weights(filepath=COCO_MODEL_PATH, by_name=True,
               exclude=["mrcnn_class_logits", "mrcnn_bbox_fc",
                        "mrcnn_bbox", "mrcnn_mask"])

问题：
Traceback (most recent call last):
  File "train_rcnn_test.py", line 209, in
    model = modellib.MaskRCNN(mode="training", config=config, model_dir=MODEL_DIR)
  File "/home/byz/pro/Mask_RCNN-master/mrcnn/model.py", line 1820, in __init__
    self.keras_model = self.build(mode=mode, config=config)
  File "/home/byz/pro/Mask_RCNN-master/mrcnn/model.py", line 1976, in build
    train_bn=config.TRAIN_BN)
  File "/home/byz/pro/Mask_RCNN-master/mrcnn/model.py", line 937, in fpn_classifier_graph
    name="mrcnn_class")(mrcnn_class_logits)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/engine/topology.py", line 619, in __call__
    output = self.call(inputs, **kwargs)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/layers/wrappers.py", line 213, in call
    y = self.layer.call(inputs, **kwargs)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/layers/core.py", line 304, in call
    return self.activation(inputs)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/activations.py", line 29, in softmax
    return K.softmax(x)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/backend/tensorflow_backend.py", line 2963, in softmax
    return tf.nn.softmax(x, axis=axis)
TypeError: softmax() got an unexpected keyword argument 'axis'

解决：
Keras与tensorflow版本不相符
我的是 tensorflow-gpu (1.4.1)+Keras (2.1.6)————降到2.1.0解决

新问题，在模型训练时：

Traceback (most recent call last):
  File "train_rcnn_test.py", line 232, in
    layers='heads')
  File "/home/byz/Mask/Mask_RCNN/mrcnn/model.py", line 2308, in train
    self.compile(learning_rate, self.config.LEARNING_MOMENTUM)
  File "/home/byz/Mask/Mask_RCNN/mrcnn/model.py", line 2143, in compile
    tf.reduce_mean(layer.output, keepdims=True)
TypeError: reduce_mean() got an unexpected keyword argument 'keepdims'

原因：
tensorflow版本可能不兼容 pix elnetas keep_dims: Deprecated alias for keepdims.
1.7版本已经弃用该关键字，并得到警告 
Instructions for updating:
keep_dims is deprecated, use keepdims instead

尝试升级或更改TF关键字from keepdims to keep_dims

 

解决：
修改mask-rcnn使用到该地方的代码（/Mask_RCNN/mrcnn/model.py）


（五）
###################################--（）--##########################################

问题：
Exception ignored in: >
Traceback (most recent call last):

具体：
Traceback (most recent call last):
  File "train_rcnn_test.py", line 232, in
    layers='heads')
  File "/home/byz/Mask/Mask_RCNN/mrcnn/model.py", line 2308, in train
    self.compile(learning_rate, self.config.LEARNING_MOMENTUM)
  File "/home/byz/Mask/Mask_RCNN/mrcnn/model.py", line 2143, in compile
    tf.reduce_mean(layer.output, keepdims=True)
TypeError: reduce_mean() got an unexpected keyword argument 'keepdims'
Exception ignored in: >
Traceback (most recent call last):
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 696, in __del__
TypeError: 'NoneType' object is not callable

出现该问题的原因：
这是因为在编译时，只删除了.pyc文件，留下了__pycache__文件夹。 
一般在编译python文件后，会生成__pycache__文件夹，其中包括你import的其他python文件生成的.pyc文件。 
如下图中，我运行了training.py文件，该文件中import另外两个文件model.py和input_data.py，所以生成了这两个文件的.pyc文件。 

解决：
再次编译training.py文件(你的python文件)时，应该把__pycache__文件夹整个删除。

(py35) byz@ubuntu:~/Mask/Mask_RCNN$ ll mrcnn/
total 228
drwxrwxr-x  3 byz byz   4096 Apr 27 17:50 ./
drwxrwxr-x 10 byz byz   4096 Apr 27 19:21 ../
-rw-rw-r--  1 byz byz   8374 Apr 27 11:30 config.py
-rw-rw-r--  1 byz byz      1 Apr 27 11:30 __init__.py
-rw-rw-r--  1 byz byz 123897 Apr 27 11:30 model.py
-rw-rw-r--  1 byz byz   7022 Apr 27 11:30 parallel_model.py
drwxrwxr-x  2 byz byz   4096 Apr 27 15:36 __pycache__/
-rw-rw-r--  1 byz byz  10037 Apr 27 17:37 train_rcnn_test.py
-rw-rw-r--  1 byz byz  33390 Apr 27 11:30 utils.py
-rw-rw-r--  1 byz byz  19097 Apr 27 15:36 visualize.py

(py35) byz@ubuntu:~/Mask/Mask_RCNN$ rm -rf  ./mrcnn/__pycache__/


（六）
###################################--（python多线程问题）--##########################################

警告：
UserWarning('Using a generator with `use_multiprocessing=True`'

/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/engine/training.py:2022: UserWarning: Using a generator with `use_multiprocessing=True` and multiple workers may duplicate your data. Please consider using the`keras.utils.Sequence class.
  UserWarning('Using a generator with `use_multiprocessing=True`'

请看：
https://github.com/keras-team/keras/pull/8662


（七）
#################################--（train_rcnn）--######################################

执行：
model.train(dataset_train, dataset_val,
            learning_rate=config.LEARNING_RATE,
            epochs=1,
            layers='heads')
报错：

Traceback (most recent call last):
  File "train_rcnn_test.py", line 232, in
    layers='heads')
  File "/home/byz/Mask/Mask_RCNN/mrcnn/model.py", line 2328, in train
    use_multiprocessing=True,
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/legacy/interfaces.py", line 87, in wrapper
    return func(*args, **kwargs)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/engine/training.py", line 2077, in fit_generator
    class_weight=class_weight)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/engine/training.py", line 1791, in train_on_batch
    check_batch_axis=True)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/engine/training.py", line 1409, in _standardize_user_data
    exception_prefix='input')
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/engine/training.py", line 154, in _standardize_input_data
    str(array.shape))
ValueError: Error when checking input: expected input_image_meta to have shape (None, 16) but got array with shape (1, 17)

解决：
修改(py35) byz@ubuntu:~/Mask/Mask_RCNN$ vim mrcnn/config.py 
一开始以为IMAGE_META_SIZE少了个类
self.IMAGE_META_SIZE = 1 + 3 + 3 + 4 + 1 + self.NUM_CLASSES
就改成成：
self.IMAGE_META_SIZE = 1 + 1 + 3 + 3 + 4 + 1 + self.NUM_CLASSES
程序也能跑，因为。。我代码少写了个类：
NUM_CLASSES = 1 + 3 （我有4类，应该是1+4，其中1是背景，糊涂了）


（八）
################################--（model.detect）--################################

运行：
results = model.detect([original_image], verbose=1)

报错：
+++++++=========模型输出：
original_image           shape: (512, 512, 3)         min:    0.00000  max:  255.00000  uint8
image_meta               shape: (17,)                 min:    0.00000  max:  512.00000  int64  <_________________(这)
gt_class_id              shape: (2, 4)                min:   92.00000  max:  424.00000  int32
gt_bbox                  shape: (2, 4)                min:   92.00000  max:  424.00000  int32
gt_mask                  shape: (512, 512, 2)         min:    0.00000  max:    1.00000  uint8
Processing 1 images
image                    shape: (512, 512, 3)         min:    0.00000  max:  255.00000  uint8
molded_images            shape: (1, 512, 512, 3)      min: -123.70000  max:  151.10000  float64
image_metas              shape: (1, 16)               min:    0.00000  max:  512.00000  int64   <_________________(这)
anchors                  shape: (1, 65472, 4)         min:   -0.53137  max:    1.40612  float32

Traceback (most recent call last):
  File "train_rcnn_test.py", line 280, in
    results = model.detect([original_image], verbose=1)
  File "/home/byz/Mask/Mask_RCNN/mrcnn/model.py", line 2478, in detect
    self.keras_model.predict([molded_images, image_metas, anchors], verbose=0)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/engine/training.py", line 1730, in predict
    check_batch_axis=False)
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/keras/engine/training.py", line 154, in _standardize_input_data
    str(array.shape))
ValueError: Error when checking : expected input_image_meta to have shape (None, 17) but got array with shape (1, 16)
Exception ignored in: >
Traceback (most recent call last):
  File "/home/byz/anaconda3/envs/py35/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 696, in __del__
TypeError: 'NoneType' object is not callable


解决：
检查分类数目是否正确
 

博主w~x：lovebyz99

(交流请备注，对创业有想法欢迎联系)