faster rcnn训练详解
http://blog.csdn.net/zy1034092330/article/details/62044941
py-faster-rcnn训练自己的数据:流程很详细并附代码
https://huangying-zhan.github.io/2016/09/22/detection-faster-rcnn
Summary
This post records my experience with py-faster-rcnn, including how to setup py-faster-rcnn from scratch, how to perform a demo training on PASCAL VOC dataset by py-faster-rcnn, how to train your own dataset, and some errors I encountered. All the steps are based on Ubuntu 14.04 + CUDA 8.0. Faster R-CNN is an important research result for object detection with an end-to-end deep convolutional neural network architure. For the details, please refer to original paper.
The source code related to adding own dataset is provided at: (https://github.com/Huangying-Zhan/py-faster-rcnn)[https://github.com/Huangying-Zhan/py-faster-rcnn]
Contents
Part 1. Setup py-faster-rcnn
In this part, a simple instruction for install py-faster-rcnn is introduced. The instruction mainly refers to py-faster-rcnn.
Clone the Faster R-CNN repo
# Make sure to clone with --recursive
$ git clone --recursive https://github.com/rbgirshick/py-faster-rcnn.git
Lets call the directory as
$FRCN
Build the Cython modules
$ cd $FRCN/lib
$ make
Build Caffe and PyCaffe
For this part, please refer to Caffe official installation instruction or my post about Caffe installation. If you have experience with Caffe, just follow the instruction below.
$ cd $FRCN/caffe-fast-rcnn
$ cp Makefile.config.example Makefile.config # Modify Makefile.config, uncommment this line
WITH_PYTHON_LAYER := 1
# Modifiy Makefile.config according to your need, such as setup related to GPU support, cuDNN, CUDA version, Anaconda, OpenCV, etc. # After modification on Makefile.config
$ make all -j4 # -j4 is for complilation acceleration only. 4 is the number of core in your CPU, change it according to your computer CPU.
# Suppose you have installed prerequites for PyCaffe, otherwise, go back to the Caffe installation instructions.
$ make pycaffe -j4
Download pre-computed Faster R-CNN models
$ cd $FRCN
$ ./data/scripts/fetch_faster_rcnn_models.sh
Run the demo
However, in this part you might get into trouble with different errors, such as without some packages. At the end of this post, some encountered errors and solution are provided. For those unexpected error, google the error and you should be able to find a solution.
$ ./tools/demo.py
Part 2. Demo Training on PASCAL VOC
In this part, the training of py-faster-rcnn will be explained. Firstly, an original training procedure on PASCAL VOC dataset is provided. The purpose is to understand the structure of dataset and training steps.
2.1. Prepare dataset and Pre-trained model
Download VOC dataset
$ cd $FRCN/data
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar $ tar xvf VOCdevkit_08-Jun-2007.tar
$ tar xvf VOCtrainval_06-Nov-2007.tar
$ tar xvf VOCtest_06-Nov-2007.tar $ ln -s VOCdevkit VOCdevkit2007 #create a softlink
Download pre-trained models
$ cd $FRCN
$ ./data/scripts/fetch_imagenet_models.sh
$ ./data/scripts/fetch_faster_rcnn_models.sh
2.2. Training
There are 2 types of training methods provided by py-faster-rcnn. One is using the alternating optimization algrithm while another one is approximate joint training method. In this post, approximate joint training method is introduced. For the details, please refer to the paper, Faster R-CNN.
$ cd $FRCN
# ./experiments/scripts/faster_rcnn_end2end.sh [GPU_ID] [NET] [DATASET]
# Directly run this command might have an error "AssertionError: Selective search data not found at:". For the solution, please refer to Part 4.
$ ./experiments/scripts/faster_rcnn_end2end.sh 0 ZF pascal_voc
Here is a remark about the logic and idea behind the training script.
faster_rcnn_end2end.sh
This is a shell script, which is the toppest layer of the whole pipeline, it monitors the input arguments, including GPU ID, network structure(ZF-Net, VGG, or others), dataset (PASCAL VOC, COCO or others), and extra configurations.
# Part of the script
GPU_ID=$1
NET=$2
NET_lc=${NET,,}
DATASET=$3 array=( $@ )
len=${#array[@]}
EXTRA_ARGS=${array[@]:3:$len}
EXTRA_ARGS_SLUG=${EXTRA_ARGS// /_}
Then, it will call two programs,
train_net.py
and followed bytest_net.py
. As the name given,train_net.py
is to train a model whiletest_net.py
is to evaluate performance of the trained model.# Part of the script
time ./tools/train_net.py --gpu ${GPU_ID} \
--solver models/${PT_DIR}/${NET}/faster_rcnn_end2end/solver.prototxt \
--weights data/imagenet_models/${NET}.v2.caffemodel \
--imdb ${TRAIN_IMDB} \
--iters ${ITERS} \
--cfg experiments/cfgs/faster_rcnn_end2end.yml \
${EXTRA_ARGS} set +x
NET_FINAL=`grep -B 1 "done solving" ${LOG} | grep "Wrote snapshot" | awk '{print $4}'`
set -x time ./tools/test_net.py --gpu ${GPU_ID} \
--def models/${PT_DIR}/${NET}/faster_rcnn_end2end/test.prototxt \
--net ${NET_FINAL} \
--imdb ${TEST_IMDB} \
--cfg experiments/cfgs/faster_rcnn_end2end.yml \
${EXTRA_ARGS}
faster_rcnn_end2end.yml
As we can see from
faster_rcnn_end2end.sh
,cfg
comes fromfaster_rcnn_end2end.yml
, which means that this file stores many importatnt configurations. Here shows some original configurations provided.EXP_DIR: faster_rcnn_end2end
TRAIN:
HAS_RPN: True
IMS_PER_BATCH: 1
BBOX_NORMALIZE_TARGETS_PRECOMPUTED: True
RPN_POSITIVE_OVERLAP: 0.7
RPN_BATCHSIZE: 256
PROPOSAL_METHOD: gt
BG_THRESH_LO: 0.0
TEST:
HAS_RPN: True
However, if you wish to add your own configurations, such as number of iterations to take a model snapshot while training, you may refer to
$FRCN/lib/fast_rcnn/config.py
. This file contains all the configuration parameters. You don’t need to set the configuration in thisconfig.py
but just add a statement infaster_rcnn_end2end.yml
. The program can parse the arguments automatically. Of course there exists default values if you do not declare the items in the .yml file.# Example to add SNAPSHOT_ITERS into the configuration
EXP_DIR: faster_rcnn_end2end
TRAIN:
HAS_RPN: True
IMS_PER_BATCH: 1
BBOX_NORMALIZE_TARGETS_PRECOMPUTED: True
RPN_POSITIVE_OVERLAP: 0.7
RPN_BATCHSIZE: 256
PROPOSAL_METHOD: gt
BG_THRESH_LO: 0.0
SNAPSHOT_ITERS: 10000 # This line is an example to add arguments.
TEST:
HAS_RPN: True
train_net.py
Basically, there are 3 things included in the file.
# Read dataset
imdb, roidb = combined_roidb(args.imdb_name) # Pass configurations from `faster_rcnn_end2end.sh` and `faster_rcnn_end2end.yml` to lower layer programs/functions
# Call `fast_rcnn.train_net` for training
train_net(args.solver, roidb, output_dir, pretrained_model=args.pretrained_model, max_iters=args.max_iters)
combined_roidb
&pascal_voc.py
Recall that in
faster_rcnn_end2end.sh
. You have entered an argument,–imdb ${TRAIN_IMDB}
combined_roidb
do nothing but just trace back and read the datasets, such as train, val, and test using functions in$FRCN/lib/datasets/pascal_voc.py
.fast_rcnn.train_net
This function is at
$FRCN/lib/fast_rcnn/train.py
This function is the core of whole training pipeline since it callssolver.prototxt
, but in fact you don’t need to care this part in most of the time.solver.prototxt
&train.prototxt
If you are familiar with Caffe, you should know the purpose of
solver.prototxt
andtrain.prototxt
. Otherwise, you are suggested to go through Caffe’s MNIST tutorial. In here, the idea will be described briefly only.Basically,
solver.prototxt
tells the program where to find your ConvNet structure prototxt and some training setups, such as learning rate, learning policy, etc.train_net: "models/pascal_voc/ZF/faster_rcnn_end2end/train.prototxt"
base_lr: 0.001
lr_policy: "step"
gamma: 0.1
stepsize: 50000
display: 20
average_loss: 100
momentum: 0.9
weight_decay: 0.0005
snapshot_prefix: "zf_faster_rcnn"
iter_size: 2
train.prototxt
describes the network structure, including number of layer, type of layer, number of neurons in each layer, etc. Again, refer to Caffe’s MNIST tutorial in order to understandtrain.prototxt
.
Part 3. Training on new dataset
In this part, basketball detection will be used as an example to illustrate how to train a new dataset using py-faster-rcnn.
3.1. Prepare dataset
The dataset used in this part is downloaded from ImageNet.
DISCLAIMER: This dataset should be only used for non-commercial research activities. Please follow the ImageNet rules about the use of the dataset.
Download dataset
Here provides a link to download Basketball Dataset. This dataset has the following structure.
|-- basketball
|-- JPEGImages
Contains all raw .JPEG images
|-- ImageSets
.txt files state training set, validataion set. Extension is not required in these files
|-- Annotations
Bounding boxes annotation for each image. The annotation files are written in .xml format.
# Unzip the folder
$ mv basketball.tar.gz $FRCN/data/
$ cd $FRCN/data
$ tar xzf basketball.tar.gz
Add a dataset python file
Add a
basketball.py
to$FRCN/lib/datasets/
. You may check on the source code for reference. If you wish to modify this file, basically, you can just find and replace basketball by your new dataset name.Add
basketball_eval.py
Add a
basketball_eval.py
to$FRCN/lib/datasets/
. Again, check on the source code for reference. Again, find and replace basketabll.Update
/lib/datasets/factory.py
The purpose of
basketball.py
is to read a part of whole dataset, such as train_set or val_set. The purpose of function infactory.py
is to get all sets of whole dataset.Add config file
As mentioned in Part 2.2 (2), we need a
config.yml
to store configurations. In here, we can use the originalfaster_rcnn_end2end.yml
as a reference. However, there are many configurations you can set in this file. In here, we may setEXP_DIR
first and others if necessary.$ cd $FRCN/experiments/cfgs
$ cp faster_rcnn_end2end.yml config.yml
Update
imdb.py
Since new dataset may have conflicts in annotation with original PASCAL VOC dataset. For example, ImageNet images start with index 0 in row and col while PASCAL VOC dataset starts with index 1. In
imdb.py
, a part of code should be inserted inappend_flipped_images()
. Refer to source code.
3.2. Prepare network and pre-trained model
To train our own model, basically we don’t need to train the model from scratch unless you have a huge dataset which is comparable to ImageNet. Otherwise, we can train our model from fine-tuning a pre-trained Faster R-CNN model. The reason is because a pre-trained Faster R-CNN contains a lot of good lower level features, which can be used generally. Even you are using a new and self-defined architure (i.e. no existing pre-trained Faster R-CNN model), follow the training method of Faster R-CNN and train a Faster R-CNN first, followed by fine-tuning on your own dataset is suggested.
For simplicity, the network and model adopted in this part is ZF-net and a pre-trained Faster R-CNN (ZF) respectively.
$ cd $FRCN/models
# copy a well-defined network and make modification based on it
$ mkdir basketball
$ cp ./pascal_voc/ZF/faster_rcnn_end2end/* ./basketball/
$ cd basketball
Now, we should modify all files in basketball/
, including,
solver.prototxt
- train_net
- snapshot_prefix
- Others if necessary
train.prototxt
&test.prototxt
For
train.prototxt
andval.prototxt
, basically we need to update the number of output in final layers. Let’s say, in this basketball dataset, we only need 2 classes (background + basketball) and 8 output for bounding box regressor. Orignial pascal_voc have 21 classes including background and 21*4 bounding box regressor output.$ cd $FRCN/models/basketball
$ grep 21 *
$ grep 84 *
# These two commands help you to check the lines that you should modify in the files.
In this part, there are two more items we need to modify. Since we are fine-tuning a pre-trained ConvNet model on our own dataset and the number of output at last fully-connected layers (cls_score & bbox_pred) has been changed, the original weight in pre-trained ConvNet model is not suitable for our current network. The dimension is totally different. The details can be refered to Caffe’s fine-tuning tutorial. The solution is to rename the layers such that the weights for the layers will be initialized randomly instead of copying from pre-trained model (actually copying from pre-trained model will cause error).
name: "cls_score" -> name: "cls_score_basketball"
name: "bbox_pred" -> name: "bbox_pred_basketball"
However, renaming the layers may cause problems in later parts since “cls_score” and “bbox_pred” are used as keys in testing. Therefore, in the training part, we can train the model accroding to the following procedure.
- Rename the layers to cls_score_basketball and bbox_pred_basketball
- Fine-tune pre-trained Faster R-CNN (FRCN) model and snapshot at iteration 0. Let’s call the snapshot Basketball_0.caffemodel. Stop training.
- Rename the layers back to cls_score and bbox_pred.
- Fine-tune Basketball_0.caffemodel to get our final model.
The details and code will be explained in the following part.
3.3. Training and evaluation
Before training on your new dataset, you may need to check $FRCN/data/cache
to remove caches if necessary. Caches stores information of previously trained dataset. It may cause problem while training.
Rename the layers
As mentions in the previous part, rename the two layers.
Reminder: if you are using find and replace, please find the name with quotes(i.e. “cls_score”). If you just search for cls_score, without quotes, it may also replace some other layers since there is a layer named rpn_cls_score.
First fine-tuning
The purpose of first fine-tuning is to get a caffemodel which has two outputs at final fully-connected layers.
$ ./tools/train_net.py --gpu 0 --weights data/faster_rcnn_models/ZF_faster_rcnn_final.caffemodel --imdb basketball_train --cfg experiments/cfgs/config.yml --solver models/basketball/solver.prototxt --iter 0
After this fine-tuning, we should get the model we needed.
Rename the layers back
Rename the two layers back to “cls_score” and “bbox_pred”.
Second fine-tuning
This fine-tuning should train models for our final use. The pre-trained model in this stage is the model we saved in stage 2.
$ ./tools/train_net.py --gpu 0 --weights output/basketball/train/zf_faster_rcnn_basketball_iter_0.caffemodel --imdb basketball_train --cfg experiments/cfgs/config.yml --solver models/basketball/solver.prototxt --iter 10000
Evaluation / Testing
To test the performance of trained model, we can use the provided
test_net.py
for the purpose.$ ./tools/test_net.py --gpu 0 --def models/basketball/test.prototxt --net output/basketball/train/zf_faster_rcnn_basketball_iter_20000.caffemodel --imdb basketball_val --cfg experiments/cfgs/config.yml
At the end, you should be able to see something like this.
After going through such long path, training on py-faster-rcnn is completed!
Part 4. Error and solution
no easydict, cv2
# Without Anaconda
$ sudo pip install easydict
$ sudo apt-get install python-opencv # With Anaconda
$ conda install -c verydeep easydict
$ conda install opencv
# Normally, people will follow the online instruction at https://anaconda.org/auto/easydict and install auto/easydict. However, this easydict (ver.1.4) has a problem in passing the message of configuration and cause many unexpected error while verydeep/easydict (ver.1.6) won't cause these errors.
assertionError: Selective Search data is not found
Solution: install verydeep/easydict rather than auto/easydict
$ conda install -c verydeep easydict
box [:, 0] > box[:, 2]
Solution: add the following code block in imdb.py
def append_flipped_images(self):
num_images = self.num_images
widths = self._get_widths()
for i in xrange(num_images):
boxes = self.roidb[i]['boxes'].copy()
oldx1 = boxes[:, 0].copy()
oldx2 = boxes[:, 2].copy()
boxes[:, 0] = widths[i] - oldx2
boxes[:, 2] = widths[i] - oldx1
for b in range(len(boxes)):
if boxes[b][2] < boxes[b][0]:
boxes[b][0]=0
assert (boxes[:, 2] >= boxes[:, 0]).all()
For ImageNet detection dataset, no need to minus one on coordinates
# Load object bounding boxes into a data frame.
for ix, obj in enumerate(objs):
bbox = obj.find('bndbox')
# Make pixel indexes 0-based
x1 = float(bbox.find('xmin').text)
y1 = float(bbox.find('ymin').text)
x2 = float(bbox.find('xmax').text)
y2 = float(bbox.find('ymax').text)
cls = self._class_to_ind[obj.find('name').text.lower().strip()]
Reference
faster rcnn训练详解的更多相关文章
- 第三十一节,目标检测算法之 Faster R-CNN算法详解
Ren, Shaoqing, et al. “Faster R-CNN: Towards real-time object detection with region proposal network ...
- 【目标检测】Faster RCNN算法详解
Ren, Shaoqing, et al. “Faster R-CNN: Towards real-time object detection with region proposal network ...
- Faster R-CNN:详解目标检测的实现过程
本文详细解释了 Faster R-CNN 的网络架构和工作流,一步步带领读者理解目标检测的工作原理,作者本人也提供了 Luminoth 实现,供大家参考. Luminoth 实现:https:// ...
- 目标检测算法之Faster R-CNN算法详解
Fast R-CNN存在的问题:选择性搜索,非常耗时. 解决:加入一个提取边缘的神经网络,将候选框的选取交给神经网络. 在Fast R-CNN中引入Region Proposal Network(RP ...
- Faster R-CNN论文详解 - CSDN博客
废话不多说,上车吧,少年 paper链接:Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks ...
- 第三十节,目标检测算法之Fast R-CNN算法详解
Girshick, Ross. “Fast r-cnn.” Proceedings of the IEEE International Conference on Computer Vision. 2 ...
- 第二十九节,目标检测算法之R-CNN算法详解
Girshick, Ross, et al. “Rich feature hierarchies for accurate object detection and semantic segmenta ...
- 如何才能将Faster R-CNN训练起来?
如何才能将Faster R-CNN训练起来? 首先进入 Faster RCNN 的官网啦,即:https://github.com/rbgirshick/py-faster-rcnn#installa ...
- caffe学习三:使用Faster RCNN训练自己的数据
本文假设你已经完成了安装,并可以运行demo.py 不会安装且用PASCAL VOC数据集的请看另来两篇博客. caffe学习一:ubuntu16.04下跑Faster R-CNN demo (基于c ...
随机推荐
- 二分法binadySearch的用法
package com.Summer_0420.cn; import java.util.Arrays; /** * @author Summer * 二分法binadySearch的用法(测试) * ...
- Java中volatile关键字解析
一.内存模型的相关概念 大家都知道,计算机在执行程序时,每条指令都是在CPU中执行的,而执行指令过程中,势必涉及到数据的读取和写入.由于程序运行过程中的临时数据是存放在主存(物理内存)当中的,这时就存 ...
- LDAP1-安装部署LDAP服务
基于Linux部署openldap服务 参考文档: https://blog.csdn.net/computer1024/article/details/78172785 参考文档: https:/ ...
- DOM(二)
文档信息 document对象还有一些标准的Document对象所没有的属性: title属性:包含着<title>元素中的文本——显示在浏览器窗口的标题栏或标签页上,通过整个属性可以取得 ...
- MySQL常用SQL语句/函数/存储过程
一句话总结 SELECT count(*) FROM user WHERE id>0 GROUP BY name HAVING count(*)>1 ORDER BY count(*)DE ...
- Spring+Struts2+Hibernate框架整合流程
一:基本步骤 新建Maven项目,导入相关依赖(推荐) 在WEB-INF的web.xml中进行配置 ————–Hibernate配置 —————- 创建entity包,创建数据库相关实体类 根据实体类 ...
- React组件重构:嵌套+继承 与 高阶组件
前言 在最近做的一个react项目中,遇到了一个比较典型的需要重构的场景:提取两个组件中共同的部分. 最开始通过使用嵌套组件和继承的方式完成了这次重构. 但是后来又用高阶组件重新写了一遍,发现更好一点 ...
- struts2的基本使用
struts2在web中当作前端控制器,接收来自页面的请求,使用过滤器拦截模式对请求进行拦截并交给相应配置的action类处理. 所以在web中使用最重要的是struts2的核心过滤器StrutsPr ...
- vue组件化开发组件拆分原则是什么
原则:可复用.可组合: 两大类:页面组件.功能组件: 除了公共头导航.侧导航.脚部内容,还有:
- duxing201606的原味鸡树
链接 [http://murphyc.fun/problem/4011] 题意 描述 众所周知,duxing哥非常喜欢原味鸡.众所周知,原味鸡是长在原味鸡树上的. duxing哥因为是水产巨子,所以就 ...