首先，我参考https://blog.csdn.net/Chen_yingpeng/article/details/80692018提供的yolov3_darknet2caffe.py脚本实现了darknet-yolov3到caffemodel的转化，得到了prototxt和caffemodel。
但是，我在编译Chen提供的caffe-yolov3时，由于server上并没有sudo权限，也无法安装opencv3，故无法编译成功，为此根据之前使用caffe下的MobileNet-YOLOv3，我使用https://github.com/eric612/MobileNet-YOLO中提供的ssd_detect.cpp/yolo_detect.cpp（博主之前已经对MobileNet-YOLO已经使用cmake完全编译通了）进行测试（当然，其中根据博主的需求已经更改了一些相关code）
需要注意的是使用MobileNet-YOLO提供的ssd_detect.cpp需要将Chen提供转化过来的yolov3.prototxt做一定的修改，即要添加相应的yolo层和最后的detection_out层，具体可参考该框架中给出的prototxt文件修改。
最终可以正常测试转化过来的caffemodel，经过demo测试，结果也还正常，但是我们在专用的数据集中测试具体的指标时，发现对traffic sign的recall相对于darknet下差了很多，经过show图片发现caffemodel会把蓝色的traffic sign检出而忽略红色的，增加了训练数据之后情况稍微好转，但仍然严重。
为此，我debug了Chen提供的convert python代码，其中weights值转化的部分如下（整个code最主要的部分）：
def darknet2caffe(cfgfile, weightfile, protofile, caffemodel):
net_info = cfg2prototxt(cfgfile)
save_prototxt(net_info , protofile, region=False)
net = caffe.Net(protofile, caffe.TEST)
params = net.params
blocks = parse_cfg(cfgfile)
#Open the weights file
fp = open(weightfile, "rb")
#The first 4 values are header information
# 1. Major version number
# 2. Minor Version Number
# 3. Subversion number
# 4. IMages seen
header = np.fromfile(fp, dtype = np.int32, count = 5)
#fp = open(weightfile, 'rb')
#header = np.fromfile(fp, count=5, dtype=np.int32)
#header = np.ndarray(shape=(5,),dtype='int32',buffer=fp.read(20))
#print(header)
buf = np.fromfile(fp, dtype = np.float32)
#print(buf)
fp.close()
layers = []
layer_id = 1
start = 0
for block in blocks:
if start >= buf.size:
break
if block['type'] == 'net':
continue
elif block['type'] == 'convolutional':
batch_normalize = int(block['batch_normalize'])
if block.has_key('name'):
conv_layer_name = block['name']
bn_layer_name = '%s-bn' % block['name']
scale_layer_name = '%s-scale' % block['name']
else:
conv_layer_name = 'layer%d-conv' % layer_id
bn_layer_name = 'layer%d-bn' % layer_id
scale_layer_name = 'layer%d-scale' % layer_id
if batch_normalize:
start = load_conv_bn2caffe(buf, start, params[conv_layer_name], params[bn_layer_name], params[scale_layer_name])
else:
start = load_conv2caffe(buf, start, params[conv_layer_name])
layer_id = layer_id+1
elif block['type'] == 'depthwise_convolutional':
batch_normalize = int(block['batch_normalize'])
if block.has_key('name'):
conv_layer_name = block['name']
bn_layer_name = '%s-bn' % block['name']
scale_layer_name = '%s-scale' % block['name']
else:
conv_layer_name = 'layer%d-dwconv' % layer_id
bn_layer_name = 'layer%d-bn' % layer_id
scale_layer_name = 'layer%d-scale' % layer_id
if batch_normalize:
start = load_conv_bn2caffe(buf, start, params[conv_layer_name], params[bn_layer_name], params[scale_layer_name])
else:
start = load_conv2caffe(buf, start, params[conv_layer_name])
layer_id = layer_id+1
elif block['type'] == 'connected':
if block.has_key('name'):
fc_layer_name = block['name']
else:
fc_layer_name = 'layer%d-fc' % layer_id
start = load_fc2caffe(buf, start, params[fc_layer_name])
layer_id = layer_id+1
elif block['type'] == 'maxpool':
layer_id = layer_id+1
elif block['type'] == 'avgpool':
layer_id = layer_id+1
elif block['type'] == 'region':
layer_id = layer_id + 1
elif block['type'] == 'route':
layer_id = layer_id + 1
elif block['type'] == 'shortcut':
layer_id = layer_id + 1
elif block['type'] == 'softmax':
layer_id = layer_id + 1
elif block['type'] == 'cost':
layer_id = layer_id + 1
elif block['type'] == 'upsample':
layer_id = layer_id + 1
else:
print('unknow layer type %s ' % block['type'])
layer_id = layer_id + 1
print('save prototxt to %s' % protofile)
save_prototxt(net_info , protofile, region=True)
print('save caffemodel to %s' % caffemodel)
net.save(caffemodel)
......
def load_conv_bn2caffe(buf, start, conv_param, bn_param, scale_param):
conv_weight = conv_param[0].data
running_mean = bn_param[0].data
running_var = bn_param[1].data
scale_weight = scale_param[0].data
scale_bias = scale_param[1].data
scale_param[1].data[...] = np.reshape(buf[start:start+scale_bias.size], scale_bias.shape); start = start + scale_bias.size
#print scale_bias.size
#print scale_bias
scale_param[0].data[...] = np.reshape(buf[start:start+scale_weight.size], scale_weight.shape); start = start + scale_weight.size
#print scale_weight.size
bn_param[0].data[...] = np.reshape(buf[start:start+running_mean.size], running_mean.shape); start = start + running_mean.size
#print running_mean.size
bn_param[1].data[...] = np.reshape(buf[start:start+running_var.size], running_var.shape); start = start + running_var.size
#print running_var.size
bn_param[2].data[...] = np.array([1.0])
conv_param[0].data[...] = np.reshape(buf[start:start+conv_weight.size], conv_weight.shape); start = start + conv_weight.size
#print conv_weight.size
return start其中，buf为读取的yolov3.weights的权值，根据darknet的权重存储方式，buf为一个一维的vector，维度为61592497 x 1，这是由yolov3中所有层累加得到，而start记录每一层权值开始的位置。然而转换到caffemodel后（code中load_conv_bn2caffe()即将buf中获取的权值写入到caffemodel中）变为四维vector，例如64x32x3x3（分别表示该卷积层输入channel为64，输出channel为32，kernel大小为3x3），这一步的完成主要是由load_conv_bn2caffe中的numpy包中reshape（）函数完成。所以导致最后测试结果下降的原因是否是一维reshape为四维时顺序与darknet中不对应导致的RGB权值错位产生的影响。
然后我检查了使用的MobileNet-YOLO提供的ssd_detection.cpp，从image input网络部分入手，code示例如下：
cv::Mat img = cv::imread(fn[k]);
if (img.empty()) continue; //only proceed if sucsessful
// you probably want to do some preprocessing
CHECK(!img.empty()) << "Unable to decode image " << file;
Timer batch_timer;
batch_timer.Start();
std::vector<vector<float> > detections = detector.Detect(img);
LOG(INFO) << "Computing time: " << batch_timer.MilliSeconds() << " ms.";从上述code中发现，读取image使用的是opencv中的imread（）函数，而将读取到的img传入Detector中测试，而Detector被定义为class类，包含三个子函数，如下：
class Detector {
public:
Detector(const string& model_file,
const string& weights_file,
const string& mean_file,
const string& mean_value,
const float confidence_threshold,
const float normalize_value);
std::vector<vector<float> > Detect(const cv::Mat& img);
private:
void SetMean(const string& mean_file, const string& mean_value);
void WrapInputLayer(std::vector<cv::Mat>* input_channels);
void Preprocess(const cv::Mat& img,
std::vector<cv::Mat>* input_channels);
void Preprocess(const cv::Mat& img,
std::vector<cv::Mat>* input_channels,double normalize_value);
private:
shared_ptr<Net<float> > net_;
cv::Size input_geometry_;
int num_channels_;
cv::Mat mean_;
float nor_val = 1.0;
};这4个子函数对输入的img并未做RGB channel转换，只进行了resize操作，然后测试。
从此处我似乎得到一些启发，opencv中的imread读取的RGB图像是按照BGR顺序读取的，这是否刚好对应了traffic sign中蓝色检出而红色漏检的规律呢？然后我将读取imread读取的img channel（2）和channel（0）互换，然后input到Detector中测试，测试的结果（坐标和分数信息）使用rectangle（）函数画在未RGB channel转换的原img上，发现，检测结果正常了，与darknet下结果基本无异，这证明了Chen提供的convert code中reshape并无错误，而是我使用了cv中的imread（）直接输入检测器中导致。
之后我由考虑为什么原始darknet中和caffe中其它model（如ssd、RefineDet等）并不会出现此情况呢？我又对其进行了探究。阅读了darknet中图像输入部分的code，具体在src/image.c中函数load_image_color（），如下code：
image load_image_color(char *filename, int w, int h)
{
return load_image(filename, w, h, 3);
}
image load_image(char *filename, int w, int h, int c)
{
#ifdef OPENCV
image out = load_image_cv(filename, c);
#else
image out = load_image_stb(filename, c);
#endif
if((h && w) && (h != out.h || w != out.w)){
//按网络要求调整到(w,h)大小,前提是输入的w,h不要是0
image resized = resize_image(out, w, h);
free_image(out);
out = resized;
}
return out;
}其中的load_image（）函数调用了load_image_cv（）函数，显然darknet中仍然使用cv读取image，因为我们知道cv读取image的两种方式imread（C++）和cvLoadImage（C#）都是按照BGR格式读取，这更加使我疑惑，查看load_image_cv（）函数，code如下：
image load_image_cv(char *filename, int channels)
{
IplImage* src = 0;
int flag = -1;
if (channels == 0) flag = -1;
else if (channels == 1) flag = 0; //grayscale image
else if (channels == 3) flag = 1; //3-channel color image
else {
fprintf(stderr, "OpenCV can't force load with %d channels\n", channels);
}
//opencv api load image
if( (src = cvLoadImage(filename, flag)) == 0 )
{
fprintf(stderr, "Cannot load image \"%s\"\n", filename);
char buff[256];
sprintf(buff, "echo %s >> bad.list", filename);
system(buff);
return make_image(10,10,3);
//exit(0);
}
//将读取到的IplImage容器中的图片装入image结构中
image out = ipl_to_image(src);
cvReleaseImage(&src);
rgbgr_image(out); //convert BGR to RGB
return out;
}