Training Your Own Faster RCNN

Dependecies installation

pip install scipy pillow matplotlib pyyaml easydict opencv-python

---

This repo

https://github.com/smallcorgi/Faster-RCNN_TF

allows us to train our own Faster-RCNN. To train the network, follow the instructions in the ReadME file of the repo above until you are able to train the VOC dataset.

Over that process, you may encounter the following problems.

1. When you execute

    python ./tools/demo.py --model ./VGGnet_fast_rcnn_iter_70000.ckpt
   

   you might get following error codes

    tensorflow.python.framework.errors_impl.NotFoundError: /home/neno/workspace/OCR/Faster-      
    RCNN_TF/tools/../lib/roi_pooling_layer/roi_pooling.so: undefined symbol:_ZTIN10tensorflow8OpKernelE
   

   To solve this problem, replace $REPO/lib/make.sh with the following content and run

    *python ./tools/demo.py --model ./VGGnet_fast_rcnn_iter_70000.ckpt*
   

   again. (P.S. There is no need to execute make again after you have modified the make.sh)

    #!/usr/bin/env bash
    TF_LIB=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib())')
    TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include()       )')
    echo $TF_INC
   
    CUDA_PATH=/usr/local/cuda/
   
    cd roi_pooling_layer
   
    nvcc -std=c++11 -c -o roi_pooling_op.cu.o roi_pooling_op_gpu.cu.cc \
        -I $TF_INC -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC -arch=sm_52
   
    ## if you install tf using already-built binary, or gcc version 4.x,        uncomment the two lines below
    #g++ -std=c++11 -shared -D_GLIBCXX_USE_CXX11_ABI=0 -o roi_pooling.so        roi_pooling_op.cc \
    #   roi_pooling_op.cu.o -I $TF_INC -fPIC -lcudart -L $CUDA_PATH/lib64
   
    # for gcc5-built tf
    g++ -std=c++11 -shared -o roi_pooling.so roi_pooling_op.cc      -D_GLIBCXX_USE_CXX11_ABI=0 \
        roi_pooling_op.cu.o -I $TF_INC -L $TF_LIB -ltensorflow_framework -D         GOOGLE_CUDA=1 \
        -fPIC $CXXFLAGS -lcudart -L $CUDA_PATH/lib64
   
    cd ..
   
   
    # add building psroi_pooling layer
    cd psroi_pooling_layer
    nvcc -std=c++11 -c -o psroi_pooling_op.cu.o psroi_pooling_op_gpu.cu.cc \
        -I $TF_INC -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC -arch=sm_52
   
    g++ -std=c++11 -shared -o psroi_pooling.so psroi_pooling_op.cc \
        psroi_pooling_op.cu.o -I $TF_INC -fPIC -lcudart -L $CUDA_PATH/lib64
   
    ## if you install tf using already-built binary, or gcc version 4.x,        uncomment the two lines below
    #g++ -std=c++11 -shared -D_GLIBCXX_USE_CXX11_ABI=0 -o psroi_pooling.so      psroi_pooling_op.cc \
    #   psroi_pooling_op.cu.o -I $TF_INC -fPIC -lcudart -L $CUDA_PATH/lib64
   
    cd ..
   

2. When you execute

    python ./tools/demo.py --model ./VGGnet_fast_rcnn_iter_70000.ckpt
   

   you might get the following error message

    tensorflow/stream_executor/cuda/cuda_dnn.cc:378] Loaded runtime CuDNN library: 7104 
    (compatibility version 7100) but source was compiled with 7004 (compatibility version 7000).  If 
    using a binary install, upgrade your CuDNN library to match.  If building from sources, make sure 
    the library loaded at runtime matches a compatible version specified during compile configuration.
   

   To solve this problem, uninstall the cuDNN 7.1.4 and install cuDNN 7.0.5 instead.

---

Now, we are going to build our own dataset. First of all, you've got to prepare the images you want to rain and then we make them look "nice" so that our network can be trained effciently.

The author of this blog provides us three scripts to resize, change filename and generate index files. Here is a merged and improved version I wrote on the basis of the provided scripts . The script below firstly resizes the images and save them to the output directory and then uniform the file format to jpg and finally rename them into VOC style.

import cv2
import os
import sys
from PIL import Image
import re

path = sys.argv[1]
output_dir = sys.argv[2]

if not os.path.exists(output_dir):
    os.mkdir(output_dir)

ticket_width = 300

print('resizing raw images...')

list=os.listdir(path)
output_dir_image = output_dir + '/images/'
output_dir_label = output_dir + '/labels/'
output_dir_index = output_dir + '/indexs/'
if not os.path.exists(output_dir_image):
    os.mkdir(output_dir_image)
if not os.path.exists(output_dir_label):
    os.mkdir(output_dir_label)
if not os.path.exists(output_dir_index):
    os.mkdir(output_dir_index)
count=0

for pic in list:
    im = cv2.imread(path + '/' + pic)
    h = im.shape[0]
    w = im.shape[1]
    ratio = float(ticket_width) / w
    w_new = ticket_width
    h_new = int(ratio * h)
    im = cv2.resize(im, (w_new, h_new))
    new_path=output_dir_image + '/' + pic[0:-3] + 'jpg'
    cv2.imwrite(new_path, im)

print("renaming...")

filelist = os.listdir(output_dir_image)
total_num = len(filelist)
i = 10000 
n = 6
for item in filelist:
    if item.endswith('.jpg'):
        n = 6 - len(str(i))
        src = os.path.join(os.path.abspath(output_dir_image), item)
        dst = os.path.join(os.path.abspath(output_dir_image), str(0)*n + str(i) + '.jpg')
        try:
            os.rename(src, dst)
            i = i + 1
        except:
            continue

print('finished')

It takes two arguments to run.

python script.py $IMAGES_DIR $OUTPUT_DIR

Well done. Now you have completed all the operations that you need to do with the raw images. Here comes a heavier task ---- labeling the images.This tool provides utilities but it still costs much time. WHEN LABELING, PLEASE USE LOWER-CASE LETTERS FOR ALL LABELS. UPPER-CASE LABELS WOULD LEAD TO PROGRAM ERRORS.

The last step is to produce txt index files. This script could automatically generate index files for us.

# !/usr/bin/python
# -*- coding: utf-8 -*-
import os
import random  
import sys  

trainval_percent = 0.8  # tunable parameter
train_percent = 0.7  # tunable parameter
xmlfilepath = sys.argv[1]
txtsavepath = sys.argv[2]  
total_xml = os.listdir(xmlfilepath)  
  
num=len(total_xml)  
list=range(num)  
tv=int(num*trainval_percent)  
tr=int(tv*train_percent)  
trainval= random.sample(list,tv)  
train=random.sample(trainval,tr)  
  
ftrainval = open(txtsavepath + '/trainval.txt', 'a')  
ftest = open(txtsavepath + '/test.txt', 'a')  
ftrain = open(txtsavepath + '/train.txt', 'a')  
fval = open(txtsavepath + '/val.txt', 'a')  
  
for i  in list:  
    name=total_xml[i][:-4]+'\n'  
    if i in trainval:  
        ftrainval.write(name)  
        if i in train:  
            ftrain.write(name)  
        else:  
            fval.write(name)  
    else:  
        ftest.write(name)  
  
ftrainval.close()  
ftrain.close()  
fval.close()  
ftest .close()  

To use it, you will need to input the directory where labels store at and the directory where index files should be placed at.

python script.py $LABEL_DIR $OUTPUT_DIR

And you would get the following four text files

$OUTPUT_DIR/trainvel.txt
$OUTPUT_DIR/text.txt
$OUTPUT_DIR/train.txt
$OUTPUT_DIR/val.txt

These files will guide the neural network to locate the dataset.

---

If you have reached this line, you are already very close to start training your own FRCNN. Now we are going to replace VOC dataset our prepared data.

Put VOC images and labels into trash bin

rm -rf $VOC_DATA_DIR/VOCdevkit/VOC2007/JPEGImages/*
rm -rf $VOC_DATA_DIR/VOCdevkit/VOC2007/Annotations/*

and bring our data under the spotlight.

cp $IMAGE_DIR/* $VOC_DATA_DIR/VOCdevkit/VOC2007/JPEGImages/*
cp $LABEL_DIR/* $VOC_DATA_DIR/VOCdevkit/VOC2007/Annotations/*
cp $INDEX_DIR/* $VOC_DATA_DIR/VOCdevkit/VOC2007/ImageSets/Main/

Now the very last step is to modify the source code. Four changes have to be made.

1. $FRCNN_DIR\lib\datasets\pascal_voc.py

   Find variable _classes

    self._classes = ('__background__', # always index 0
                     'aeroplane', 'bicycle', 'bird', 'boat',
                     'bottle', 'bus', 'car', 'cat', 'chair',
                     'cow', 'diningtable', 'dog', 'horse',
                     'motorbike', 'person', 'pottedplant',
                     'sheep', 'sofa', 'train', 'tvmonitor')
   

   and append your own classes at the tail.

    self._classes = ('__background__', # always index 0
                     'aeroplane', 'bicycle', 'bird', 'boat',
                     'bottle', 'bus', 'car', 'cat', 'chair',
                     'cow', 'diningtable', 'dog', 'horse',
                     'motorbike', 'person', 'pottedplant',
                     'sheep', 'sofa', 'train', 'tvmonitor', 'class1', 'class2')
   

2. $FRCNN_DIR\lib\networks\VGGnet_train.py

   This line indicates the totoal num of all classes

    n_classes = 21
   

   If your own data has n classes to recognize, increase its value by n.

    # Two classes to recognize for example
    n_classes = 23
   

3. $FRCNN_DIR\lib\networks\VGGnet_test.py

   Same as 2

4. $FRCNN_DIR\tools\demo.py

   Find variable CLASSES and append your own classes at the tail the way same as modifying _classes in pascal_voc.py

---

Now, we are ready to train. Use this command to start training

./experiments/scripts/faster_rcnn_end2end.sh gpu 0 VGG16 pascal_voc

Here are the two errors I encountered :

1.  Traceback (most recent call last):
      File "./tools/train_net.py", line 83, in <module>
        roidb = get_training_roidb(imdb)
      File "/home/yinqsh/Ningyuan/Faster-RCNN_TF/tools/../lib/fast_rcnn/train.py", line 204, in                
      get_training_roidb
        imdb.append_flipped_images()
      File "/home/yinqsh/Ningyuan/Faster-RCNN_TF/tools/../lib/datasets/imdb.py", line 113, in       
        append_flipped_images
       assert (boxes[:, 2] >= boxes[:, 0]).all()
   

I googled it and it comes out that this error is probably caused by some illeagal boundboxes. The boundaries of these boundboxes exceed the image boundaries and therefore lead up to crashes. One solution is to delete all cache files avoiding models mix-up.

rm $FRCNN/output
rm $FRCNN/data/cache
rm $FRCNN/VOCdevkit2007/annotations_chache // if this directory exists

Another solution is to modify append_flipped_images() method in $FRCNN/lib/datasets/imdb.py. Find this line of code

boxes[:, 2] = widths[i] - oldx1 - 1

and add the following lines of code right below

boxes[:, 2] = widths[i] - oldx1 - 1
# ------------------TO-ADD-PART------------------
for b in range(len(boxes)):
    if boxes[b][2]< boxes[b][0]:
        boxes[b][0] = 0
# ------------------TO-ADD-PART------------------

2. KeyError: 'max_overlaps'
   

Solution : Delete caches

rm $FRCNN/output
rm $FRCNN/data/cache
rm $FRCNN/VOCdevkit2007/annotations_chache // if this directory exists

---

After training, you would get a trained model in

$FRCNN_DIR/output/aster_rcnn_end2end/voc_2007_trainval/

By default, model would be saved every 5000 iterations. We are going to use the model that was trained with most iterations. And there are three files for that model

model.ckpt.meta
model.ckpt.data-00000-xx-00000
model.ckpt.index

Make a copy of model.ckpt.data file under the same folder and remove suffix .data-00000-xx-00000 from the name of that copy.

Finally, we are ready to test the power of FRCNN. Put some test images into

$FRCNN/data/demo

and run

python $FRCNN/tools/demo.py --model $FRCNN/output/faster_rcnn_end2end/voc_2007_trainval/VGGnet_fast_rcnn_iter_10000.ckpt

BANG!BANG!BNAG!

---

Bugs encountered when moving codes from python2 to python 3

1. When you execute

    python ./tools/demo.py --model ./VGGnet_fast_rcnn_iter_70000.ckpt
   

   you might get the following error message

   ImportError: /media/neno/44B0AB27B0AB1F04/Faster RCNN/Faster-   
   RCNN_TF/tools/../lib/utils/cython_bbox.so: undefined symbol: _Py_ZeroStruct
   

   To solve this problem, go to

    $FRCNN_DIR/lib
   

   and run

    python3 setup.py build_ext --inplace
   

2. When you execute

    python ./tools/demo.py --model ./VGGnet_fast_rcnn_iter_70000.ckpt
   

   you might get the following error message

   ImportError: No module named 'cPickle'
   

   To solve this problem, change cPickle to pickle.

3. After upgrade pip3, pip3 crashes.

    Traceback (most recent call last):
      File "/usr/bin/pip3", line 9, in <module>
        from pip import main
    ImportError: cannot import name 'main'
   

To solve this problem, open /usr/bin/pip3 and change the following codes

    from pip import __main__
        if __name__ == '__main__':
        sys.exit(__main__._main())

to

    from pip import __main__
        if __name__ == '__main__':
        sys.exit(__main__._main())

4. When you tried to run $FRCNN/lib/make.sh, you get the following error

    fatal error: nsync_cv.h: No such file or directory
    #include "nsync_cv.h"
   

To solve this problem, open the file which causes this error and change the following two lines

  #include "external/nsync/public/nsync_cv.h"
  #include "external/nsync/public/nsync_mu.h"

5. When you tried to run demo.py, you get the following error

    cudaCheckError() : no kernel image is available for execution on the device
   

To solve this problem, go to $FRCNN/lib/make.sh file and check extra options for the nvcc compiler. There should be a option called arch which specifies the computation architecture of the Nvidia card. Check the compute ability of your card and change the arch option accordingly. In my case, I am using Telsa K80 which has 3.7 compute ability. I used to compile with -arch=sm_52 which caused this error. Then I changed it to -arch=sm_35 and things go really well now.

---

P.S.

By default, the network would load parameters from pre-trained VGG16 model. However, it might not perform well on the testset in practice. Instead, training from scratch gives a relatively better prediction.

References:

https://blog.csdn.net/zcy0xy/article/details/79614862

Follw My Wechat Official Account