本文针对RBG的 faster rcnn 代码,做以下工作:
1)完成安装及配置
2)使用自己的数据做训练和测试
faster-rcnn安装
与 fast-rcnn的安装 非常类似:
1) 克隆源代码
git clone --recursive https://github.com/rbgirshick/py-faster-rcnn.git
2)编译Cython模块
cd caffe-fast-rcnn/lib
make
3)替换最新caffe文件,解决cudnn5的错误
include/caffe/layers/cudnn_relu_layer.hpp
src/caffe/layers/cudnn_relu_layer.cpp
src/caffe/layers/cudnn_relu_layer.cu
include/caffe/layers/cudnn_sigmoid_layer.hpp
src/caffe/layers/cudnn_sigmoid_layer.cpp
src/caffe/layers/cudnn_sigmoid_layer.cu
include/caffe/layers/cudnn_tanh_layer.hpp
src/caffe/layers/cudnn_tanh_layer.cpp
src/caffe/layers/cudnn_tanh_layer.cu
include/caffe/layers/cudnn_conv_layer.hpp
src/caffe/layers/cudnn_conv_layer.cpp
src/caffe/layers/cudnn_conv_layer.cu
include/caffe/util/cudnn.hpp
4)编译
cd caffe-fast-rcnn
make -j8 && make pycaffe
5)下载模型文件,测试
# caffe-fast-rcnn/data/fetch_faster_rcnn_models.sh
# caffe-fast-rcnn/data/fetch_imagenet_models.sh
# 从以上两个文件中拿到下载地址,国内需要翻墙下载
./tools/demo.py
图1. faster-rcnn 测试demo结果(平均耗时90ms)
使用自己的数据训练
参考这篇文章,思路是:
1)首先将自己的数据转成 pascal_voc 的格式;
2)然后修改faster-rcnn中数据读取接口;
3)然后根据自己数据的类别数修改训练模型;
4)使用/tools/train.py训练;
5)修改/tools/demo.py做测试
训练数据格式化
faster-rcnn使用 pascal_voc 数据做训练和测试,为了最少的变动代码,将我们的数据也换成pascal_voc格式
推荐使用开源工具 labelImg ,可以很方便的在图像中框选目标,并自动生成pascal_voc格式的xml文件。这里特别需要注意坐标格式,框选的时候一定要从左上到右下的顺序
完成后将图片放在 py-faster-rcnn/trainval/JPEGImages/ 目录下
将xml文件放在 py-faster-rcnn/trainval/Annotations/ 目录下
在 py-faster-rcnn/trainval/ 目录下新建一个文件 ImageList.txt,逐行记录图片的名称(不要后缀)
修改数据读取接口
1)修改 py-faster-rcnn/lib/pascal_voc.py 文件
复制pascal_voc.py为my_voc.py
a.修改初始化 函数
class my_voc(imdb):
def __init__(self, image_set, devkit_path=None):
imdb.__init__(self, image_set)
self._image_set = image_set
self._devkit_path = devkit_path
self._data_path = os.path.join(self._devkit_path)
self._classes = ('__background__', # always index 0
'target1', 'target2', 'target3')
self._class_to_ind = dict(zip(self._classes, xrange(len(self._classes))))
self._image_ext = '.jpg'
self._image_index = self._load_image_set_index('/ImageList.txt')
# Default to roidb handler
self._roidb_handler = self.selective_search_roidb
self._salt = str(uuid.uuid4())
self._comp_id = 'comp4'
...
b.修改 _load_image_set_index 函数
image_set_file = os.path.join(self._data_path + imagelist)
...
c.修改 _load_pascal_annotation 函数
...
x1 = float(bbox.find('xmin').text)
y1 = float(bbox.find('ymin').text)
x2 = float(bbox.find('xmax').text)
y2 = float(bbox.find('ymax').text)
...
d.修改 main 函数
if __name__ == '__main__':
from datasets.my_voc import my_voc
d = my_voc('trainval/')
res = d.roidb
from IPython import embed; embed()
2)修改 py-faster-rcnn/lib/factory.py 文件
a.import 换成 my_voc 模块
from datasets.my_voc import my_voc
c.注释掉接下来的 3段 set up
d.修改 name, devkit
name = 'my'
devkit = '/home/xxx/py-faster-rcnn/trainval'
__sets['my'] = (lambda name = name, devkit = devkit: my_voc(name, devkit))
在训练时可能还会有其他错误,比如这里提到的“pb2.text_format”的问题,是由于 protobuf版本的原因,直接 import google.prototxt.text_format 可解决
修改训练模型
faster-rcnn里提供了 vgg16,vgg_cnn_m_1024,zf 大中小3个模型
目录 py-faster-rcnn/data/faster_rcnn_models 是作者使用pascal_voc数据训练好的模型,可以直接拿来检测
目录 py-faster-rcnn/data/imagenet_models 是在 imagenet上训练好的通用模型,用来初始化网络(其实也就是finetunning)
目录 py-faster-rcnn/models/pascal_voc 是3个模型的porototxt文件,在这里需要修改相应参数
复制 pascal_voc 为 my_voc ,以 vgg_cnn_m_1024模型为例,修改 VGG_CNN_M_1024/faster_rcnn_end2end/ 下的3个文件:
1)solver.prototxt
修改 train_net 路径为
models/my_voc/VGG_CNN_M_1024/faster_rcnn_end2end/train.prototxt
2)train.prototxt
a.修改 input-data 层
param_str: "'num_classes': 4"
# 新的数据加上 背景 一共 4 类
b.修改 roi-data 层
param_str: "'num_classes': 4"
c.修改 cls_score 层
num_output: 4
d.修改 bbox_pred 层
num_output: 16 # 4*4
3) test.prototxt
a.修改 cls_score 层
num_output: 4
b.修改 bbox_pred 层
num_output: 16
开始训练
python ./tools/train_net.py --gpu 0 --solver models/my_voc/VGG_CNN_M_1024/faster_rcnn_end2end/solver.prototxt --weights data/imagenet_models/VGG_CNN_M_1024.v2.caffemodel --imdb my --iters 80000 --cfg experiments/cfgs/faster_rcnn_end2end.yml
注意:必须要在data/cache/目录下把数据库的缓存文件.pkl给删除掉,否则其不会重新读取相应的数据库,而是直接从之前读入然后缓存的pkl文件中读取进来,这样修改的数据库并没有进入网络,而是加载了老版本的数据。
测试
仿照 /tools/demo.py,修改一个detect.py
主要修改了
1)数据读取方式;
2)原代码中对于一张图片,如果检测到2个类别,会生成2个图像窗口分别表示;这里将所有检测结果放在一个图像窗口里
#!/usr/bin/env python
# --------------------------------------------------------
# Faster R-CNN
# Copyright (c) 2015 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Ross Girshick
# --------------------------------------------------------
"""
Demo script showing detections in sample images.
See README.md for installation instructions before running.
"""
import _init_paths
from fast_rcnn.config import cfg
from fast_rcnn.test import im_detect
from fast_rcnn.nms_wrapper import nms
from utils.timer import Timer
import matplotlib.pyplot as plt
import numpy as np
import scipy.io as sio
import caffe, os, sys, cv2
import argparse
CLASSES = ('__background__',
'target1', 'target2', 'target3')
def vis_detections(im, class_name, dets, thresh=0.5):
"""Draw detected bounding boxes."""
inds = []
for index in range(len(class_name)):
inds.append(np.where(dets[index][:, -1] >= thresh)[0])
'''
if len(inds) == 0:
return
'''
im = im[:, :, (2, 1, 0)]
fig, ax = plt.subplots(figsize=(12, 12))
ax.imshow(im, aspect='equal')
for index in range(len(inds)):
if len(inds[index]) == 0:
continue
for i in inds[index]:
bbox = dets[index][i, :4]
score = dets[index][i, -1]
ax.add_patch(
plt.Rectangle((bbox[0], bbox[1]),
bbox[2] - bbox[0],
bbox[3] - bbox[1], fill=False,
edgecolor='red', linewidth=3.5)
)
ax.text(bbox[0], bbox[1] - 2,
'{:s} {:.3f}'.format(class_name[index], score),
bbox=dict(facecolor='blue', alpha=0.5),
fontsize=14, color='white')
ax.set_title(('detections with '
'p({} | box) >= {:.1f}').format(class_name,
thresh),
fontsize=14)
plt.axis('off')
plt.tight_layout()
plt.draw()
def demo(net, image_name):
"""Detect object classes in an image using pre-computed object proposals."""
# Load the demo image
im_file = os.path.join('/home/xxx/py-faster-rcnn/trainval/JPEGImages/'+image_name+'.jpg')
im = cv2.imread(im_file)
# Detect all object classes and regress object bounds
timer = Timer()
timer.tic()
scores, boxes = im_detect(net, im)
timer.toc()
print ('Detection took {:.3f}s for '
'{:d} object proposals').format(timer.total_time, boxes.shape[0])
# Visualize detections for each class
CONF_THRESH = 0.8
NMS_THRESH = 0.3
cls_ = []
dets_ = []
for cls_ind, cls in enumerate(CLASSES[1:]):
cls_ind += 1 # because we skipped background
cls_boxes = boxes[:, 4*cls_ind:4*(cls_ind + 1)]
cls_scores = scores[:, cls_ind]
dets = np.hstack((cls_boxes,
cls_scores[:, np.newaxis])).astype(np.float32)
keep = nms(dets, NMS_THRESH)
dets = dets[keep, :]
cls_.append(cls)
dets_.append(dets)
vis_detections(im, cls_, dets_, thresh=CONF_THRESH)
def parse_args():
"""Parse input arguments."""
parser = argparse.ArgumentParser(description='Faster R-CNN demo')
parser.add_argument('--gpu', dest='gpu_id', help='GPU device id to use [0]',
default=0, type=int)
parser.add_argument('--cpu', dest='cpu_mode',
help='Use CPU mode (overrides --gpu)',
action='store_true')
args = parser.parse_args()
return args
if __name__ == '__main__':
cfg.TEST.HAS_RPN = True # Use RPN for proposals
args = parse_args()
prototxt = os.path.join('/home/xxx/py-faster-rcnn/models/my_voc/VGG_CNN_M_1024/faster_rcnn_end2end/test.prototxt')
caffemodel = os.path.join('/home/xxx/py-faster-rcnn/output/faster_rcnn_end2end/my/vgg_cnn_m_1024_faster_rcnn_iter_80000.caffemodel')
if not os.path.isfile(caffemodel):
raise IOError(('{:s} not found.').format(caffemodel))
if args.cpu_mode:
caffe.set_mode_cpu()
else:
caffe.set_mode_gpu()
caffe.set_device(args.gpu_id)
cfg.GPU_ID = args.gpu_id
net = caffe.Net(prototxt, caffemodel, caffe.TEST)
print '\n\nLoaded network {:s}'.format(caffemodel)
# Warmup on a dummy image
im = 128 * np.ones((300, 500, 3), dtype=np.uint8)
for i in xrange(2):
_, _= im_detect(net, im)
test_file_list = '/home/xxx/py-faster-rcnn/trainval/ImageList_.txt'
im_names = []
with open(test_file_list, 'r') as f:
im_names.extend(f.readlines())
key = ''
for im_name in im_names:
print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'
print 'detect for {}'.format(im_name)
demo(net, im_name.strip())
plt.show()
key = raw_input("\'q\' to quit:")
if key == 'q':
break