Fast RCNN 训练自己的数据集（3训练和检测）

转载请注明出处，楼燚(yì)航的blog，http://www.cnblogs.com/louyihang-loves-baiyan/

https://github.com/YihangLou/fast-rcnn-train-another-dataset 这是我在github上修改的几个文件的链接，求星星啊，求星星啊（原谅我那么不要脸~~）

在之前两篇文章中我介绍了怎么编译Fast RCNN，和怎么修改Fast RCNN的读取数据接口，接下来我来说明一下怎么来训练网络和之后的检测过程
先给看一下极好的检测效果

https://github.com/YihangLou/fast-rcnn-train-another-dataset

1.预训练模型介绍

首先在data目录下，有两个目录就是之前在1中解压好

• fast_rcnn_models/
• imagenet_models/

fast_rcnn_model文件夹下面是作者用fast rcnn训练好的三个网络,分别对应着小、中、大型网络，大家可以试用一下这几个网络，看一些检测效果，他们训练都迭代了40000次，数据集都是pascal_voc的数据集。

1. caffenet_fast_rcnn_iter_40000.caffemodel
2. vgg_cnn_m_1024_fast_rcnn_iter_40000.caffemodel
3. vgg16_fast_rcnn_iter_40000.caffemodel

imagenet_model文件夹下面是在Imagenet上训练好的通用模型，在这里用来初始化网络的参数

1. CaffeNet.v2.caffemodel
2. VGG_CNN_M_1024.v2.caffemodel
3. VGG16.v2.caffemodel

在这里我比较推荐先用中型网络训练，中型网络训练和检测的速度都比较快，效果也都比较理想，大型网络的话训练速度比较慢，我当时是5000多个标注信息，网络配置默认，中型网络训练大概两三个小时，大型网络的话用十几个小时，需要注意的是网络训练最好用GPU，CPU的话太慢了，我当时用的实验室的服务器，有16块Tesla K80，用起来真的是灰常爽！

2. 修改模型文件配置

模型文件在models下面对应的网络文件夹下，在这里我用中型网络的配置文件修改为例子
比如：我的检测目标物是car ，那么我的类别就有两个类别即 background 和 car
因此，首先打开网络的模型文件夹，打开train.prototxt
修改的地方重要有三个
分别是个地方

1. 首先在data层把num_classes 从原来的21类 20类+背景 ，改成 2类 车+背景
2. 接在在cls_score层把num_output 从原来的21 改成 2
3. 在bbox_pred层把num_output 从原来的84 改成8， 为检测类别个数乘以4，比如这里是2类那就是2*4=8

OK，如果你要进一步修改网络训练中的学习速率，步长，gamma值，以及输出模型的名字，需要在同目录下的solver.prototxt中修改。
如下图：

train_net: "models/VGG_CNN_M_1024/train.prototxt"base_lr: 0.001lr_policy: "step"gamma: 0.1stepsize: 30000display: 20average_loss: 100momentum: 0.9weight_decay: 0.0005# We disable standard caffe solver snapshotting and implement our own snapshot# functionsnapshot: 0# We still use the snapshot prefix, thoughsnapshot_prefix: "vgg_cnn_m_1024_fast_rcnn"#debug_info: true

3.启动Fast RCNN网络训练

启动训练：
./tools/train_net.py --gpu 11 --solver models/VGG_CNN_M_1024_LOUYIHANG/solver.prototxt --weights data/imagenet_models/VGG_CNN_M_1024.v2.caffemodel --imdb KakouTrain

参数讲解：

• 这里的--是两个-，markdown写的，大家不要输错
• train_net.py是网络的训练文件，之后的参数都是附带的输入参数
• --gpu 代表机器上的GPU编号，如果是nvidia系列的tesla显卡，可以在终端中输入nvidia-smi来查看当前的显卡负荷，选择合适的显卡
• --solver 代表模型的配置文件，train.prototxt的文件路径已经包含在这个文件之中
• --weights 代表初始化的权重文件，这里用的是Imagenet上预训练好的模型，中型的网络我们选择用VGG_CNN_M_1024.v2.caffemodel
• --imdb 这里给出的训练的数据库名字需要在factory.py的__sets中，我在文件里面有__sets['KakouTrain']，train_net.py这个文件会调用factory.py再生成kakou这个类，来读取数据

4.启动Fast RCNN网络检测

我修改了tools下面的demo.py这个文件，用来做检测，并且将检测的坐标结果输出到相应的txt文件中
可以看到原始的demo.py 是用网络测试了两张图像，并做可视化输出，有具体的检测效果，但是我是在Linux服务器的终端下，没有display device，因此部分代码要少做修改

下面是原始的demo.py：

#!/usr/bin/env python# --------------------------------------------------------# Fast 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_pathsfrom fast_rcnn.config import cfgfrom fast_rcnn.test import im_detectfrom utils.cython_nms import nmsfrom utils.timer import Timerimport matplotlib.pyplot as pltimport numpy as npimport scipy.io as sioimport caffe, os, sys, cv2import argparseCLASSES = ('__background__',           'aeroplane', 'bicycle', 'bird', 'boat',           'bottle', 'bus', 'car', 'cat', 'chair',           'cow', 'diningtable', 'dog', 'horse',           'motorbike', 'person', 'pottedplant',           'sheep', 'sofa', 'train', 'tvmonitor')NETS = {'vgg16': ('VGG16',                  'vgg16_fast_rcnn_iter_40000.caffemodel'),        'vgg_cnn_m_1024': ('VGG_CNN_M_1024',                           'vgg_cnn_m_1024_fast_rcnn_iter_40000.caffemodel'),        'caffenet': ('CaffeNet',                     'caffenet_fast_rcnn_iter_40000.caffemodel')}def vis_detections(im, class_name, dets, thresh=0.5):    """Draw detected bounding boxes."""    inds = np.where(dets[:, -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 i in inds:        bbox = dets[i, :4]        score = dets[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, score),                bbox=dict(facecolor='blue', alpha=0.5),                fontsize=14, color='white')    ax.set_title(('{} detections with '                  'p({} | box) >= {:.1f}').format(class_name, class_name,                                                  thresh),                  fontsize=14)    plt.axis('off')    plt.tight_layout()    plt.draw()def demo(net, image_name, classes):    """Detect object classes in an image using pre-computed object proposals."""    # Load pre-computed Selected Search object proposals    box_file = os.path.join(cfg.ROOT_DIR, 'data', 'demo',                            image_name + '_boxes.mat')    obj_proposals = sio.loadmat(box_file)['boxes']    # Load the demo image    im_file = os.path.join(cfg.ROOT_DIR, 'data', 'demo', 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, obj_proposals)    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    for cls in classes:        cls_ind = CLASSES.index(cls)        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, :]        print 'All {} detections with p({} | box) >= {:.1f}'.format(cls, cls,                                                                    CONF_THRESH)        vis_detections(im, cls, dets, thresh=CONF_THRESH)def parse_args():    """Parse input arguments."""    parser = argparse.ArgumentParser(description='Train a Fast R-CNN network')    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')    parser.add_argument('--net', dest='demo_net', help='Network to use [vgg16]',                        choices=NETS.keys(), default='vgg16')    args = parser.parse_args()    return argsif __name__ == '__main__':    args = parse_args()    prototxt = os.path.join(cfg.ROOT_DIR, 'models', NETS[args.demo_net][0],                            'test.prototxt')    caffemodel = os.path.join(cfg.ROOT_DIR, 'data', 'fast_rcnn_models',                              NETS[args.demo_net][1])    if not os.path.isfile(caffemodel):        raise IOError(('{:s} not found.\nDid you run ./data/script/'                       'fetch_fast_rcnn_models.sh?').format(caffemodel))    if args.cpu_mode:        caffe.set_mode_cpu()    else:        caffe.set_mode_gpu()        caffe.set_device(args.gpu_id)    net = caffe.Net(prototxt, caffemodel, caffe.TEST)    print '\n\nLoaded network {:s}'.format(caffemodel)    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'    print 'Demo for data/demo/000004.jpg'    demo(net, '000004', ('car',))    print '~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~'    print 'Demo for data/demo/001551.jpg'    demo(net, '001551', ('sofa', 'tvmonitor'))    plt.show()

复制这个demo.py 修改成CarFaceTest.py，下面是修改后的文件
修改后的文件主要是添加了outputDetectionResult和runDetection两个函数， 添加了部分注释

#!/usr/bin/env python# --------------------------------------------------------# Fast 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_pathsfrom fast_rcnn.config import cfgfrom fast_rcnn.test import im_detectfrom utils.cython_nms import nmsfrom utils.timer import Timerimport matplotlib.pyplot as pltimport numpy as npimport scipy.io as sioimport caffe, os, sys, cv2import argparse#CLASSES = ('__background__','aeroplane','bicycle','bird','boat',#       'bottle','bus','car','cat','chair','cow','diningtable','dog','horse'#       'motorbike','person','pottedplant','sheep','sofa','train','tvmonitor')CLASSES = ('__background__','car') #需要跟自己训练的数据集中的类别一致，原来是21类的voc数据集，自己的数据集就是car和backgroundNETS = {'vgg16': ('VGG16',                  'vgg16_fast_rcnn_iter_40000.caffemodel'),        'vgg_cnn_m_1024': ('VGG_CNN_M_1024',                           'vgg_cnn_m_1024_fast_rcnn_iter_40000.caffemodel'),    'vgg_cnn_m_1024_louyihang': ('VGG_CNN_M_1024_LOUYIHANG',               'vgg_cnn_m_1024_fast_rcnn_louyihang_iter_40000.caffemodel'),        'caffenet': ('CaffeNet',                     'caffenet_fast_rcnn_iter_40000.caffemodel'),    'caffenet_louyihang':('CaffeNet_LOUYIHANG',             'caffenet_fast_rcnn_louyihang_iter_40000.caffemodel'),    'vgg16_louyihang':('VGG16_LOUYIHANG',               'vgg16_fast_rcnn_louyihang_iter_40000.caffemodel')}#映射到对应的模型文件def outputDetectionResult(im, class_name, dets, thresh=0.5): #打开相应的输出文件    outputFile = open('CarDetectionResult.txt')    inds = np.where(dets[:,-1] >= thresh)[0]    if len(inds) == 0:        returndef runDetection (net, basePath, testFileName,classes):#这个函数是自己后加的，取代了demo函数，给定测试数据列表    ftest = open(testFileName,'r')    imageFileName = basePath+'/' + ftest.readline().strip()    num = 1    outputFile = open('CarDetectionResult.txt','w')    while imageFileName:    print imageFileName    print 'now is ', num    num +=1    imageFileBaseName = os.path.basename(imageFileName)    imageFileDir = os.path.dirname(imageFileName)    boxFileName = imageFileDir +'/'+imageFileBaseName.replace('.jpg','_boxes.mat')    print boxFileName    obj_proposals = sio.loadmat(boxFileName)['boxes']    #obj_proposals[:,2] = obj_proposals[:, 2] + obj_proposals[:, 0]#这里也需要注意，OP里面的坐标数据是否为x1y1x2y2还是x1y1wh    #obj_proposals[:,3] = obj_proposals[:, 3] + obj_proposals[:, 1]    im = cv2.imread(imageFileName)    timer = Timer()    timer.tic()    scores, boxes = im_detect(net, im, obj_proposals)#检测函数    timer.toc()    print ('Detection took {:.3f} for '               '{:d} object proposals').format(timer.total_time, boxes.shape[0])    CONF_THRESH = 0.8    NMS_THRESH = 0.3#NMS参数用来控制非极大值抑制        for cls in classes:            cls_ind = CLASSES.index(cls)            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, :]            print 'All {} detections with p({} | box) >= {:.1f}'.format(cls, cls,                                                                    CONF_THRESH)        inds = np.where(dets[:, -1] >= CONF_THRESH)[0]        print 'inds.size', inds.size        if len(inds) != 0:            outputFile.write(imageFileName+' ')        outputFile.write(str(inds.size)+' ')将检测的结果写出相应的文件里            for i in inds:            bbox = dets[i, :4]            outputFile.write(str(int(bbox[0]))+' '+ str(int(bbox[1]))+' '+ str(int(bbox[2]))+' '+ str(int(bbox[3]))+' ')            outputFile.write('\n')        else:            outputFile.write(imageFileName +' 0' '\n')    temp = ftest.readline().strip()    if temp:        imageFileName = basePath+'/' + temp    else:        breakdef vis_detections(im, class_name, dets, thresh=0.5):#这个函数需要加以说明，这个函数虽然没有用，但是我的服务器上没有输出设备    """Draw detected bounding boxes."""#因此要将部分用到显示的函数给注释掉，否则运行会报错    inds = np.where(dets[:, -1] >= thresh)[0]    print 'inds.shape', inds.shape    print inds    print 'inds.size', inds.size    if len(inds) == 0:        return        #im = im[:, :, (2, 1, 0)]    #fig, ax = plt.subplots(figsize=(12, 12))    #ax.imshow(im, aspect='equal')    #for i in inds:    #    bbox = dets[i, :4]    #    score = dets[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, score),    #            bbox=dict(facecolor='blue', alpha=0.5),    #            fontsize=14, color='white')    #ax.set_title(('{} detections with '    #              'p({} | box) >= {:.1f}').format(class_name, class_name,    #                                              thresh),    #              fontsize=14)    #plt.axis('off')    #plt.tight_layout()    #plt.draw()def demo(net, image_name, classes):#原来的demo函数，没有修改    """Detect object classes in an image using pre-computed object proposals."""    # Load pre-computed Selected Search object proposals    #box_file = os.path.join(cfg.ROOT_DIR, 'data', 'demo',image_name + '_boxes.mat')    basePath='/home/chenjie/DataSet/500CarTestDataSet2'    box_file = os.path.join(basePath,image_name + '_boxes.mat')    obj_proposals = sio.loadmat(box_file)['boxes']    # Load the demo image    #im_file = os.path.join(cfg.ROOT_DIR, 'data', 'demo', image_name + '.jpg')    im_file = os.path.join(basePath, 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, obj_proposals)    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    for cls in classes:        cls_ind = CLASSES.index(cls)        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, :]        print 'All {} detections with p({} | box) >= {:.1f}'.format(cls, cls,                                                                    CONF_THRESH)        vis_detections(im, cls, dets, thresh=CONF_THRESH)def parse_args():    """Parse input arguments."""    parser = argparse.ArgumentParser(description='Train a Fast R-CNN network')    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')    parser.add_argument('--net', dest='demo_net', help='Network to use [vgg16]',                        choices=NETS.keys(), default='vgg16')    args = parser.parse_args()    return argsif __name__ == '__main__':    args = parse_args()    prototxt = os.path.join(cfg.ROOT_DIR, 'models', NETS[args.demo_net][0],                            'test.prototxt')    #caffemodel = os.path.join(cfg.ROOT_DIR, 'data', 'fast_rcnn_models',    #                          NETS[args.demo_net][1])    #caffemodel = '/home/chenjie/fast-rcnn/output/default/KakouTrain/vgg16_fast_rcnn_louyihang_iter_40000.caffemodel'    #caffemodel = '/home/chenjie/louyihang/fast-rcnn/output/default/KakouTrain/caffenet_fast_rcnn_louyihang_iter_40000.caffemodel'    caffemodel = '/home/chenjie/fast-rcnn/output/default/KakouTrain/vgg_cnn_m_1024_fast_rcnn_louyihang_iter_40000.caffemodel'#我在这里直接指定了训练好的模型文件，训练好的模型文件是在工程根目录下的，output/default/对应的数据库名字下面    if not os.path.isfile(caffemodel):        raise IOError(('{:s} not found.\nDid you run ./data/script/'                       'fetch_fast_rcnn_models.sh?').format(caffemodel))    if args.cpu_mode:        caffe.set_mode_cpu()    else:        caffe.set_mode_gpu()        caffe.set_device(args.gpu_id)    net = caffe.Net(prototxt, caffemodel, caffe.TEST)    print '\n\nLoaded network {:s}'.format(caffemodel)    #demo(net, 'Target0/000001', ('car',))    #输入对应的测试图像列表，需要在同级目录下摆放同名的_boxes.mat文件，它会自动的替换后缀名！    #runDetection(net, '/home/chenjie/DataSet/temptest','/home/chenjie/DataSet/temptest/Imagelist.txt',('car',))    runDetection(net, '/home/chenjie/DataSet/500CarTestDataSet2','/home/chenjie/DataSet/500CarTestDataSet2/Imagelist.txt',('car',))    #runDetection(net, '/home/chenjie/DataSet/Kakou_Test_Scale0.25/','/home/chenjie/DataSet/Kakou_Test_Scale0.25/imagelist.txt',('car',))    #runDetection(net, '/home/chenjie/DataSet/Images_Version1_Test_Boxes','/home/chenjie/DataSet/Images_Version1_Test_Boxes/ImageList_Version1_List.txt',('car',))    #plt.show()

5.检测结果

训练数据集

首先给出我的训练数据集，其实我的训练数据集并不是太复杂的

测试数据集

输出检测结果到txt文件中，

测试效果

**在复杂场景下的测试效果非常好，速度也非常快，中型网络监测平均每张在K80显卡下时0.1~0.2S左右，图像的尺寸是480*640，6000张测试数据集下达到的准确率是98%!!!**