項(xiàng)目實(shí)踐 | 從零開始邊緣部署輕量化人臉檢測(cè)模型——EAIDK310部署篇
繼續(xù)上一章的話題,前面我們主要聊到關(guān)于人臉檢測(cè)模型UltraFace的訓(xùn)練任務(wù),本文將和大家討論在開發(fā)板上如何部署UltraFace模型,并進(jìn)行實(shí)時(shí)視頻人臉檢測(cè),或者圖片流人臉檢測(cè)。
1Tengine簡(jiǎn)介
Tengine 由 OPEN AI LAB 主導(dǎo)開發(fā),該項(xiàng)目實(shí)現(xiàn)了深度學(xué)習(xí)神經(jīng)網(wǎng)絡(luò)模型在嵌入式設(shè)備上的快速、高效部署需求。為實(shí)現(xiàn)在眾多 AIoT 應(yīng)用中的跨平臺(tái)部署,本項(xiàng)目基于原有 Tengine 項(xiàng)目使用 C 語言進(jìn)行重構(gòu),針對(duì)嵌入式設(shè)備資源有限的特點(diǎn)進(jìn)行了深度框架裁剪。同時(shí)采用了完全分離的前后端設(shè)計(jì),有利于 CPU、GPU、NPU 等異構(gòu)計(jì)算單元的快速移植和部署,同時(shí)降低評(píng)估和遷移成本。
Tengine推理流程
依照順序調(diào)用Tengine核心API如下:
2模塊實(shí)現(xiàn)
1 模型轉(zhuǎn)換
第1步:轉(zhuǎn)換到onnx模型
model_path = "models/pretrained/version-RFB-320.pth"
net = create_Mb_Tiny_RFB_fd(len(class_names), is_test=True)
net.load(model_path)
net.eval()
net.to("cuda")
model_name = model_path.split("/")[-1].split(".")[0]
model_path = f"models/onnx/{model_name}.onnx"
dummy_input = torch.randn(1, 3, 240, 320).to("cuda")
torch.onnx.export(net, dummy_input, model_path, verbose=False, input_names=['input'], output_names=['scores', 'boxes'])
第2步:編譯Tengine模型轉(zhuǎn)換工具
依賴庫(kù)安裝
sudo apt install libprotobuf-dev protobuf-compiler
源碼編譯
mkdir build && cd build
cmake ..
make -j`nproc` && make install
編譯完成后,生成的可行性文件tm_convert_tool存放在 ./build/install/bin/ 目錄下。
第3步:轉(zhuǎn)換onnx模型為tmfile模型
./tm_convert_tool -m xxx.onnx -o xxx.tmfile
-
-m 為*.caffemodel, *.params, *.weight, *.pb, *.onnx, *.tflite等模型; -
-o 為output fp32 tmfile
2 NMS計(jì)算
偽代碼:
-
1 將各組box按照score降序排列;
-
2 從score最大值開始,置為當(dāng)前box,保存idex,然后依次遍歷后面的box,計(jì)算與當(dāng)前box的IOU值,若大于閾值,則抑制,不會(huì)輸出;
-
3 完成一輪遍歷后,繼續(xù)選擇下一個(gè)非抑制的box作為當(dāng)前box,重復(fù)步驟2;
-
4 返回沒有被抑制的index即符合條件的box;
python版本
def NMS(dects,threshhold):
"""
detcs:二維數(shù)組(n_samples,5)
5列:x1,y1,x2,y2,score
threshhold: IOU閾值
"""
x1=dects[:,0]
y1=dects[:,1]
x2=dects[:,2]
y2=dects[:,3]
score=dects[:,4]
ndects=dects.shape[0]#box的數(shù)量
area=(x2-x1+1)*(y2-y1+1)
order=score.argsort()[::-1] #score從大到小排列的indexs,一維數(shù)組
keep=[] #保存符合條件的index
suppressed=np.array([0]*ndects) #初始化為0,若大于threshhold,變?yōu)?,表示被抑制
for _i in range(ndects):
i=order[_i] #從得分最高的開始遍歷
if suppressed[i]==1:
continue
keep.append(i)
for _j in range(i+1,ndects):
j=order[_j]
if suppressed[j]==1: #若已經(jīng)被抑制,跳過
continue
xx1=np.max(x1[i],x1[j])#求兩個(gè)box的交集面積interface
yy1=np.max(y1[i],y1j])
xx2=np.min(x2[i],x2[j])
yy2=np.min(y2[i],y2[j])
w=np.max(0,xx2-xx1+1)
h=np.max(0,yy2-yy1+1)
interface=w*h
overlap=interface/(area[i]+area[j]-interface) #計(jì)算IOU(交/并)
if overlap>=threshhold:#IOU若大于閾值,則抑制
suppressed[j]=1
return keep
C++版本
void UltraFace::nms(std::vector<FaceInfo> &input, std::vector<FaceInfo> &output, int type) {
//根據(jù)score對(duì)候選框進(jìn)行 sort 排序操作
std::sort(input.begin(), input.end(), [](const FaceInfo &a, const FaceInfo &b) { return a.score > b.score; });
int box_num = input.size();
std::vector<int> merged(box_num, 0);
for (int i = 0; i < box_num; i++) {
if (merged[i])
continue;
std::vector<FaceInfo> buf;
buf.push_back(input[i]);
merged[i] = 1;
float h0 = input[i].y2 - input[i].y1 + 1;
float w0 = input[i].x2 - input[i].x1 + 1;
float area0 = h0 * w0;
for (int j = i + 1; j < box_num; j++) {
if (merged[j])
continue;
//確立每個(gè)候選框的坐標(biāo)以及寬高
float inner_x0 = input[i].x1 > input[j].x1 ? input[i].x1 : input[j].x1;
float inner_y0 = input[i].y1 > input[j].y1 ? input[i].y1 : input[j].y1;
float inner_x1 = input[i].x2 < input[j].x2 ? input[i].x2 : input[j].x2;
float inner_y1 = input[i].y2 < input[j].y2 ? input[i].y2 : input[j].y2;
float inner_h = inner_y1 - inner_y0 + 1;
float inner_w = inner_x1 - inner_x0 + 1;
if (inner_h <= 0 || inner_w <= 0)
continue;
float inner_area = inner_h * inner_w;
float h1 = input[j].y2 - input[j].y1 + 1;
float w1 = input[j].x2 - input[j].x1 + 1;
float area1 = h1 * w1;
float score;
//計(jì)算IOU
score = inner_area / (area0 + area1 - inner_area);
//根據(jù)閾值進(jìn)行極大值抑制的篩選
if (score > iou_threshold) {
merged[j] = 1;
buf.push_back(input[j]);
}
}
}
}
2.3 獲取候選框
//獲取候選框
void UltraFace::generateBBox(std::vector<FaceInfo> &bbox_collection, tensor_t scores, tensor_t boxes) {
float* scores_blob = ( float* )get_tensor_buffer(scores);
float* boxes_blob = ( float* )get_tensor_buffer(boxes);
for (int i = 0; i < num_anchors; i++) {
if (scores_blob[i * 2 + 1] > score_threshold) {
FaceInfo rects;
//確定坐標(biāo)中心以及box的寬高
float x_center = boxes_blob[i * 4] * center_variance * priors[i][2] + priors[i][0];
float y_center = boxes_blob[i * 4 + 1] * center_variance * priors[i][3] + priors[i][1];
float w = exp(boxes_blob[i * 4 + 2] * size_variance) * priors[i][2];
float h = exp(boxes_blob[i * 4 + 3] * size_variance) * priors[i][3];
//截取坐標(biāo)結(jié)果
rects.x1 = clip(x_center - w / 2.0, 1) * image_w;
rects.y1 = clip(y_center - h / 2.0, 1) * image_h;
rects.x2 = clip(x_center + w / 2.0, 1) * image_w;
rects.y2 = clip(y_center + h / 2.0, 1) * image_h;
rects.score = clip(scores_blob[i * 2 + 1], 1);
bbox_collection.push_back(rects);
}
}
}
2.4 模型檢測(cè)函數(shù)
//模型檢測(cè)函數(shù)
int UltraFace::detect(cv::Mat &raw_image, std::vector<FaceInfo> &face_list) {
if (raw_image.empty()) {
std::cout << "image is empty ,please check!" << std::endl;
return -1;
}
image_h = raw_image.rows;
image_w = raw_image.cols;
int img_size = in_w * in_h * 3;
float* input_data = ( float* )malloc(img_size * sizeof(float));
// 獲取來自opencv讀取的圖片或者視頻數(shù)據(jù),并返回一個(gè)適應(yīng)模型輸入的結(jié)果
get_input_data_cv(raw_image, input_data, in_w, in_h, mean_vals, norm_vals, 0);
if (set_tensor_buffer(input_tensor, input_data, (in_w * in_h * 3) * 4) < 0)
{
printf("Set input tensor buffer failed\n");
return -1;
}
//開始計(jì)時(shí)?
auto start = chrono::steady_clock::now();
// 6、Run網(wǎng)絡(luò)
if (run_graph(graph, 1) < 0)
{
printf("Run graph failed\n");
return -1;
}
// 獲取輸出結(jié)果
string scores = "scores";
string boxes = "boxes";
//7.1、獲取分類得分結(jié)果
tensor_t tensor_scores = get_graph_tensor(graph, scores.c_str());
//7.2、獲取檢測(cè)框坐標(biāo)結(jié)果
tensor_t tensor_boxes = get_graph_tensor(graph, boxes.c_str());
std::vector<FaceInfo> bbox_collection;
//結(jié)束計(jì)時(shí),然后計(jì)算推理時(shí)間
auto end = chrono::steady_clock::now();
chrono::duration<double> elapsed = end - start;
cout << "inference time:" << elapsed.count() << " s" << endl;
//后處理操作,主要是獲取BBox以及NMS操作
generateBBox(bbox_collection, tensor_scores, tensor_boxes);
nms(bbox_collection, face_list);
free(input_data);
return 0;
}
2.5 主函數(shù)
#include "UltraFace.hpp"
#include <iostream>
#include <opencv2/opencv.hpp>
#include <opencv2/highgui.hpp>
#include <opencv2/imgproc.hpp>
using namespace std;
int main() {
string tengine_path = "/home/chaucer/Tengine_Tutorial/2_FaceDetector/models/version-RFB-320_simplified.tmfile";
UltraFace ultraface(tengine_path, 320, 240, 4, 0.65); // config model input
cv::Mat frame;
//cv::VideoCapture capture(0);
cv::VideoCapture capture("/home/chaucer/face_detect/test_1.mp4");
//cv::Mat frame = cv::imread(image_file);
while(1)
{
capture >> frame;
auto start = chrono::steady_clock::now();
vector<FaceInfo> face_info;
ultraface.detect(frame, face_info);
cout << "face_info " << face_info.size() << endl;
for (auto face : face_info) {
cv::Point pt1(face.x1, face.y1);
cv::Point pt2(face.x2, face.y2);
cv::rectangle(frame, pt1, pt2, cv::Scalar(0, 255, 0), 2);
}
auto end = chrono::steady_clock::now();
chrono::duration<double> elapsed = end - start;
cout << "all time: " << elapsed.count() << " s" << endl;
cv::imshow("UltraFace", frame);
cv::waitKey(1);
string result_name = "result" + to_string(2) + ".jpg";
cv::imwrite(result_name, frame);
}
return 0;
}
3輸出結(jié)果
3.1 圖片檢測(cè)結(jié)果

3.2 視頻檢測(cè)結(jié)果
4參考
[1].https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB
[2].https://github.com/OAID/Tengine
[3].https://github.com/jiangzhongbo/Tengine_Tutorial
5推薦閱讀
Google新作 | 詳細(xì)解讀 Transformer那些有趣的特性(建議全文背誦)
極品Trick | 在ResNet與Transformer均適用的Skip Connection解讀
Transformer又一城 | Swin-Unet:首個(gè)純Transformer的醫(yī)學(xué)圖像分割模型解讀
輕量化卷積:TBC,不僅僅是參數(shù)共享組卷積,更具備跨通道建模
最快ViT | FaceBook提出LeViT,0.077ms的單圖處理速度卻擁有ResNet50的精度(文末附論文與源碼)
本文論文原文獲取方式,掃描下方二維碼
回復(fù)【UltraFaceC】即可獲取項(xiàng)目代碼
長(zhǎng)按掃描下方二維碼添加小助手。
可以一起討論遇到的問題
聲明:轉(zhuǎn)載請(qǐng)說明出處
掃描下方二維碼關(guān)注【集智書童】公眾號(hào),獲取更多實(shí)踐項(xiàng)目源碼和論文解讀,非常期待你我的相遇,讓我們以夢(mèng)為馬,砥礪前行!
