Caffe-SSD-MobileNet训练

https://github.com/chuanqi305/MobileNet-SSD

https://github.com/GuoJaw/caffe-ssd-mobilenet

环境搭建

【1】去windows下载链接的代码，train和deploy模型
train和deploy模型放在MobileNet-SSD目录下
【2】将下载好的MobileNet-SSD目录放到ubuntu下的~/caffe-ssd/examples目录下
【3】测试./demo.py

训练自己的数据

（1）制作VOC格式的数据集，并转换成LMDB格式：Face_test_lmdb Face_trainval_lmdb
（2）进入~/caffe-ssd/examples/MobileNet-SSD目录下
（2-1）软链接
ln -s PATH_TO_YOUR_TRAIN_LMDB trainval_lmdb
ln -s PATH_TO_YOUR_TEST_LMDB test_lmdb

ln -s /home/gjw/caffe-ssd/examples/Face/Face_test_lmdb test_lmdb
ln -s /home/gjw/caffe-ssd/examples/Face/Face_trainval_lmdb trainval_lmdb
（2-2）将~/caffe-ssd/data/Face下的labelmap_voc.prototxt test.txt trainval.txt test_name_size.txt复制到MobileNet-SSD目录下
（2-3）gen_model.sh 5 #类别+1
注解：
在example目录下，生成文件MobileNetSSD_deploy.prototxt MobileNetSSD_test.prototxt MobileNetSSD_train.prototxt。
【修改参数】
MobileNetSSD_train.prototxt中的batch_size=24改成1
MobileNetSSD_test.prototxt中的batch_size=8改成1

（3）开始训练
sh train.sh
注解：生成的caffemodel结果保存在snapshot中。

将生成的caffemodel模型使用merge_bn.py 进行转换
修改python merge_bn.py 文件的路径

import numpy as np  
import sys,os  
caffe_root = '/home/gjw/caffe-ssd/'
sys.path.insert(0, caffe_root + 'python')  
import caffe  

train_proto = 'example/MobileNetSSD_train.prototxt'  
train_model = 'snapshot/mobilenet_iter_1200.caffemodel'  #should be your snapshot caffemodel

deploy_proto = 'example/MobileNetSSD_deploy.prototxt'  
save_model = 'example/MobileNetSSD_deploy.caffemodel'

结果：执行python merge_bn.py ，在example下生成MobileNetSSD_deploy.caffemodel

测试训练结果

1.测试精度

修改MobileNet-SSD-master/solver_test.prototxt文件内容见下：

    train_net: "example/MobileNetSSD_train.prototxt"
    test_net: "example/MobileNetSSD_test.prototxt"

执行：sh test.sh
执行结果：
I1201 16:41:51.978435 32000 solver.cpp:332] Iteration 0, loss = 4.1392
I1201 16:41:51.978489 32000 solver.cpp:433] Iteration 0, Testing net (#0)
I1201 16:41:51.985108 32000 net.cpp:693] Ignoring source layer mbox_loss
I1201 16:42:12.844115 32000 blocking_queue.cpp:50] Data layer prefetch queue empty
I1201 16:44:03.012107 32000 solver.cpp:546] Test net output #0: detection_eval = 0.211737
I1201 16:44:03.014663 32000 solver.cpp:337] Optimization Done.
I1201 16:44:03.014703 32000 caffe.cpp:254] Optimization Done.

2.测试文件夹中的图片python demo.py

caffe_root = '/home/gjw/caffe-ssd/'

net_file= 'example/MobileNetSSD_deploy.prototxt'      #必须用相对路径，不能使用绝对路径
caffe_model='example/MobileNetSSD_deploy.caffemodel'        #必须用相对路径，不能使用绝对路径
        
CLASSES = ('background',
           'open', 'close', 'yawn', 'smoke')

3.测试视频 python demo_webcam.py

import numpy as np  
import sys,os  
import cv2
caffe_root = '/home/gjw/caffe-ssd/'
sys.path.insert(0, caffe_root + 'python')  
import caffe  


net_file= 'example/MobileNetSSD_deploy.prototxt'  
caffe_model='example/MobileNetSSD_deploy.caffemodel'   
test_dir = "images"

if not os.path.exists(caffe_model):
    print("MobileNetSSD_deploy.caffemodel does not exist,")
    print("use merge_bn.py to generate it.")
    exit()
net = caffe.Net(net_file,caffe_model,caffe.TEST)  

CLASSES = ('background',
           'open', 'close', 'yawn', 'smoke')


def preprocess(src):
    img = cv2.resize(src, (300,300))
    img = img - 127.5
    img = img * 0.007843
    return img

def postprocess(img, out):   
    h = img.shape[0]
    w = img.shape[1]
    box = out['detection_out'][0,0,:,3:7] * np.array([w, h, w, h])

    cls = out['detection_out'][0,0,:,1]
    conf = out['detection_out'][0,0,:,2]
    return (box.astype(np.int32), conf, cls)

def detect(origimg):
    img = preprocess(origimg)
    img = img.astype(np.float32)
    img = img.transpose((2, 0, 1))

    net.blobs['data'].data[...] = img
    out = net.forward()  
    box, conf, cls = postprocess(origimg, out)

    for i in range(len(box)):
       p1 = (box[i][0], box[i][1])
       p2 = (box[i][2], box[i][3])
       cv2.rectangle(origimg, p1, p2, (0,255,0))
       p3 = (max(p1[0], 15), max(p1[1], 15))
       title = "%s:%.2f" % (CLASSES[int(cls[i])], conf[i])
       cv2.putText(origimg, title, p3, cv2.FONT_ITALIC, 0.6, (0, 255, 0), 1)
    cv2.imshow("SSD", origimg)
 
    k = cv2.waitKey(1) & 0xff
        #Exit if ESC pressed
    if k == 27 : return False
    return True

##############################################

cap=cv2.VideoCapture(0)

while(1):    # get a frame   
    ret, frame = cap.read()    # show a frame   
    detect(frame)  

cap.release()
cv2.destroyAllWindows()

如果数据集里面不完全是RGB三通道的，还要修改train.prototxt和test.prototxt文件：

transform_param {
    scale: 0.007843
    force_color: true  ###添加
    mean_value: 127.5
    mean_value: 127.5
    mean_value: 127.5
}