首先我们来生成一些样本用来训练和测试 首先上c#代码
string header = "<annotation><size><width>"+300+"</width><height>"+300+"</height></size>";
Bitmap bit = new Bitmap(300, 300);
Graphics g = Graphics.FromImage(bit);
g.FillRectangle(new SolidBrush(Color.Aquamarine), new Rectangle(0, 0, 1000, 1000));
Font font = new Font("华为宋体", 18, FontStyle.Bold);
//SizeF a=g.MeasureString("华", font);27
string name = Str(10,false);
for (int i = 0; i < 5; i++)
{
int x = rd.Next(30, 270);
int y = rd.Next(30, 270);
string text = f[rd.Next(0, f.Length - 1)];
PointF point = new PointF(x, y);
g.DrawString(text, font, Brushes.Coral, point);
int acs = (int)char.Parse(text);
header += string.Format(@"<object>
<name>a{0}</name>
<bndbox>
<xmin>{1}</xmin>
<ymin>{2}</ymin>
<xmax>{3}</xmax>
<ymax>{4}</ymax>
</bndbox>
</object>", acs.ToString(), x,y,x+30,y+30);
}
header += "</annotation>";
File.AppendAllText(name+".xml", header);
bit.Save(name+".png",System.Drawing.Imaging.ImageFormat.Jpeg);
贴部分代码 生成一批样本 然后看一下样本长啥样
然后贴一些层
class MTCNNDataLayer : public DataLayer{
public:
SETUP_LAYERFUNC(MTCNNDataLayer);
virtual int getBatchCacheSize(){
return 3;
}
virtual void loadBatch(Blob** top, int numTop){
int batch_size = top[0]->num();
//正负样本比例要是1：3
vector<Mat> ims;
vector<Rect_<float>> roi;
vector<int> label;
int numSel = 0;
int dw = top[0]->width();
int dh = top[0]->height();
while (numSel < batch_size){
string path = vfs_[cursor_];
Mat im = imread(path, top[0]->channel() == 3 ? 1 : 0);
vector<Rect_<float>> bbox = getBBoxFromPath(path);
int numPositive = (rand() % (bbox.size() - 1)) + 1;
int numNegitive = numPositive * 3;
vector<int> inds(bbox.size());
for (int i = 0; i < inds.size(); ++i)
inds[i] = i;
std::random_shuffle(inds.begin(), inds.end());
float accp = 0.5;
int numSelPos = 0;
int i = 0;
int numLab1 = 0, numLab_1 = 0;
while (numSelPos < numPositive && numSel < batch_size){
Rect crop = bbox[inds[i]];
Rect raw = crop;
float offx = randr(-crop.width * accp, crop.width * accp);
float offy = randr(-crop.height * accp, crop.height * accp);
crop.x += offx;
crop.y += offy;
crop.width += randr(-crop.width * accp, crop.width * accp);
crop.height = crop.width;
crop = crop & Rect(0, 0, im.cols, im.rows);
crop.width = min(crop.width, crop.height);
crop.height = crop.width;
float iou = IoU(crop, raw);
bool sel = true;
int lab = 0;
if (iou > 0.65){
lab = 1;
numLab1++;
}
else if (iou > 0.4 && (numLab_1+1) / (float)numPositive < 0.5){
lab = -1;
numLab_1++;
}
else{
sel = false;
}
if (sel){
numSelPos++;
Rect_<float> gt(
(raw.x - crop.x) / (float)crop.width,
(raw.y - crop.y) / (float)crop.height,
(raw.x + raw.width-1 - (crop.x + crop.width - 1)) / (float)crop.width,
(raw.y + raw.height-1 - (crop.y + crop.height - 1)) / (float)crop.height
);
Mat uim;
resize(im(crop), uim, Size(dw, dh));
ims.push_back(uim);
roi.push_back(gt);
label.push_back(lab);
numSel++;
}
i++;
if (i == numPositive)
i = 0;
}
int numSelNeg = 0;
i = 0;
while (i < numNegitive && numSel < batch_size){
int w = randr(12, min(im.rows, im.cols));
int x = randr(0, im.cols - w);
int y = randr(0, im.rows - w);
Rect crop(x, y, w, w);
crop = crop & Rect(0, 0, im.cols, im.rows);
float iou = maxIoU(bbox, crop);
if (iou < 0.3){
Rect_<float> gt;
Mat uim;
resize(im(crop), uim, Size(dw, dh));
ims.push_back(uim);
roi.push_back(gt);
label.push_back(0);
numSel++;
}
i++;
if (i == numNegitive)
i = 0;
}
cursor_++;
if (cursor_ == vfs_.size()){
cursor_ = 0;
std::random_shuffle(vfs_.begin(), vfs_.end());
}
}
CV_Assert(numSel == batch_size);
vector<int> rinds;
for (int i = 0; i < batch_size; ++i)
rinds.push_back(i);
std::random_shuffle(rinds.begin(), rinds.end());
float* data_im = top[0]->mutable_cpu_data();
float* data_label = top[1]->mutable_cpu_data();
float* data_roi = top[2]->mutable_cpu_data();
for (int i = 0; i < batch_size; ++i){
vector<Mat> ms;
for (int c = 0; c < top[0]->channel(); ++c)
ms.push_back(Mat(dh, dw, CV_32F, data_im + c * dw * dh));
int ind = rinds[i];
ims[ind].convertTo(ims[ind], CV_32F, 1 / 127.5, -1);
split(ims[ind], ms);
data_im += top[0]->channel() * dw * dh;
*data_label++ = label[ind];
*data_roi++ = roi[ind].x;
*data_roi++ = roi[ind].y;
*data_roi++ = roi[ind].width;
*data_roi++ = roi[ind].height;
}
}
void preperData(){
if (this->phase_ == PhaseTest)
paFindFiles("data-test", vfs_, "*.bmp");
else
paFindFiles("data-train", vfs_, "*.bmp");
std::random_shuffle(vfs_.begin(), vfs_.end());
this->cursor_ = 0;
if (vfs_.size() == 0){
printf("没有数据.\n");
exit(-1);
}
}
virtual void setup(const char* name, const char* type, const char* param_str, int phase, Blob** bottom, int numBottom, Blob** top, int numTop){
map<string, string> param = parseParamStr(param_str);
const int batch_size = getParamInt(param, "batch_size", phase == PhaseTest ? 10: 64);
this->phase_ = phase;
top[0]->Reshape(batch_size, 3, getParamInt(param, "width"), getParamInt(param, "height"));
top[1]->Reshape(batch_size, 1, 1, 1);
top[2]->Reshape(batch_size, 4, 1, 1);
preperData();
__super::setup(name, type, param_str, phase, bottom, numBottom, top, numTop);
}
private:
PaVfiles vfs_;
int cursor_;
int phase_;
};
class MTCNNLoss : public AbstractCustomLayer{
public:
SETUP_LAYERFUNC(MTCNNLoss);
virtual void setup(const char* name, const char* type, const char* param_str, int phase, Blob** bottom, int numBottom, Blob** top, int numTop){
diff_ = newBlob();
diff_->ReshapeLike(*bottom[0]);
top[0]->Reshape(1, 1, 1, 1);
//保证loss层的权重为1
top[0]->mutable_cpu_diff()[0] = 1;
}
virtual void forward(Blob** bottom, int numBottom, Blob** top, int numTop){
const float* label = bottom[2]->cpu_data();
int countLabel = bottom[2]->num();
int ignore_label = 0;
//label
float* diff = diff_->mutable_gpu_data();
int channel = bottom[0]->channel();
//memset(diff, 0, sizeof(float)*diff_->count());
caffe_gpu_set(diff_->count(), float(0), diff);
const float* b0 = bottom[0]->gpu_data();
const float* b1 = bottom[1]->gpu_data();
float loss = 0;
for (int i = 0; i < countLabel; ++i){
if (label[i] != ignore_label){
caffe_gpu_sub(
channel,
b0 + i * channel,
b1 + i * channel,
diff + i * channel);
//float dot = caffe_dot(channel, diff + i * channel, diff + i * channel);
float dot = 0;
caffe_gpu_dot(channel, diff + i * channel, diff + i * channel, &dot);
loss += dot / float(2);
}
}
top[0]->mutable_cpu_data()[0] = loss;
}
virtual void backward(Blob** bottom, int numBottom, Blob** top, int numTop, const bool* propagate_down){
const float* label = bottom[2]->cpu_data();
int countLabel = bottom[2]->num();
int channels = bottom[0]->channel();
int ignore_label = 0;
for (int i = 0; i < 2; ++i) {
if (propagate_down[i]) {
caffe_gpu_set(bottom[i]->count(), float(0), bottom[i]->mutable_gpu_diff());
const float sign = (i == 0) ? 1 : -1;
const float alpha = sign * top[0]->cpu_diff()[0] / bottom[i]->num();
for (int j = 0; j < countLabel; ++j){
if (label[j] != ignore_label){
caffe_gpu_axpby(
channels, // count
alpha, // alpha
diff_->gpu_data() + channels * j, // a
float(0), // beta
bottom[i]->mutable_gpu_diff() + channels * j); // b
}
}
}
}
};
virtual void reshape(Blob** bottom, int numBottom, Blob** top, int numTop){
diff_->ReshapeLike(*bottom[0]);
}
private:
WPtr<Blob> diff_;
};
训练过程不多言 贴一些结果图
可以看到在少样本取得了很好的效果；