apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

yolov5 servitization optimization (#262)

Browse Source 
* yolov5 split pre and post process

* yolov5 postprocess

* yolov5 postprocess
This commit is contained in:
heliqi
2022-09-21 05:22:39 -05:00
committed by GitHub
parent 463ee0a088
commit a8e447f225
9 changed files with 194 additions and 77 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
fastdeploy/vision/detection/contrib/yolov5.cc
View File

@@ -13,6 +13,7 @@
// limitations under the License.
#include "fastdeploy/vision/detection/contrib/yolov5.h"
#include "fastdeploy/utils/perf.h"
#include "fastdeploy/vision/utils/utils.h"
@@ -74,14 +75,14 @@ YOLOv5::YOLOv5(const std::string& model_file, const std::string& params_file,
bool YOLOv5::Initialize() {
// parameters for preprocess
size = {640, 640};
padding_value = {114.0, 114.0, 114.0};
is_mini_pad = false;
is_no_pad = false;
is_scale_up = false;
stride = 32;
max_wh = 7680.0;
multi_label = true;
size_ = {640, 640};
padding_value_ = {114.0, 114.0, 114.0};
is_mini_pad_ = false;
is_no_pad_ = false;
is_scale_up_ = false;
stride_ = 32;
max_wh_ = 7680.0;
multi_label_ = true;
if (!InitRuntime()) {
FDERROR << "Failed to initialize fastdeploy backend." << std::endl;
@@ -90,23 +91,34 @@ bool YOLOv5::Initialize() {
// Check if the input shape is dynamic after Runtime already initialized,
// Note that, We need to force is_mini_pad 'false' to keep static
// shape after padding (LetterBox) when the is_dynamic_shape is 'false'.
is_dynamic_input_ = false;
auto shape = InputInfoOfRuntime(0).shape;
for (int i = 0; i < shape.size(); ++i) {
// if height or width is dynamic
if (i >= 2 && shape[i] <= 0) {
is_dynamic_input_ = true;
break;
}
}
if (!is_dynamic_input_) {
is_mini_pad = false;
}
// TODO(qiuyanjun): remove
// is_dynamic_input_ = false;
// auto shape = InputInfoOfRuntime(0).shape;
// for (int i = 0; i < shape.size(); ++i) {
// // if height or width is dynamic
// if (i >= 2 && shape[i] <= 0) {
// is_dynamic_input_ = true;
// break;
// }
// }
// if (!is_dynamic_input_) {
// is_mini_pad_ = false;
// }
return true;
}
bool YOLOv5::Preprocess(Mat* mat, FDTensor* output,
std::map<std::string, std::array<float, 2>>* im_info) {
std::map<std::string, std::array<float, 2>>* im_info,
const std::vector<int>& size,
const std::vector<float> padding_value,
bool is_mini_pad, bool is_no_pad, bool is_scale_up,
int stride, float max_wh, bool multi_label) {
// Record the shape of image and the shape of preprocessed image
(*im_info)["input_shape"] = {static_cast<float>(mat->Height()),
static_cast<float>(mat->Width())};
(*im_info)["output_shape"] = {static_cast<float>(mat->Height()),
static_cast<float>(mat->Width())};
// process after image load
double ratio = (size[0] * 1.0) / std::max(static_cast<float>(mat->Height()),
static_cast<float>(mat->Width()));
@@ -145,9 +157,11 @@ bool YOLOv5::Preprocess(Mat* mat, FDTensor* output,
}
bool YOLOv5::Postprocess(
FDTensor& infer_result, DetectionResult* result,
std::vector<FDTensor>& infer_results, DetectionResult* result,
const std::map<std::string, std::array<float, 2>>& im_info,
float conf_threshold, float nms_iou_threshold, bool multi_label) {
float conf_threshold, float nms_iou_threshold, bool multi_label,
float max_wh) {
auto& infer_result = infer_results[0];
FDASSERT(infer_result.shape[0] == 1, "Only support batch =1 now.");
result->Clear();
if (multi_label) {
@@ -251,13 +265,9 @@ bool YOLOv5::Predict(cv::Mat* im, DetectionResult* result, float conf_threshold,
std::map<std::string, std::array<float, 2>> im_info;
// Record the shape of image and the shape of preprocessed image
im_info["input_shape"] = {static_cast<float>(mat.Height()),
static_cast<float>(mat.Width())};
im_info["output_shape"] = {static_cast<float>(mat.Height()),
static_cast<float>(mat.Width())};
if (!Preprocess(&mat, &input_tensors[0], &im_info)) {
if (!Preprocess(&mat, &input_tensors[0], &im_info, size_, padding_value_,
is_mini_pad_, is_no_pad_, is_scale_up_, stride_, max_wh_,
multi_label_)) {
FDERROR << "Failed to preprocess input image." << std::endl;
return false;
}
@@ -278,8 +288,8 @@ bool YOLOv5::Predict(cv::Mat* im, DetectionResult* result, float conf_threshold,
TIMERECORD_START(2)
#endif
if (!Postprocess(output_tensors[0], result, im_info, conf_threshold,
nms_iou_threshold, multi_label)) {
if (!Postprocess(output_tensors, result, im_info, conf_threshold,
nms_iou_threshold, multi_label_)) {
FDERROR << "Failed to post process." << std::endl;
return false;
}

78
fastdeploy/vision/detection/contrib/yolov5.h
View File

@@ -41,38 +41,18 @@ class FASTDEPLOY_DECL YOLOv5 : public FastDeployModel {
float conf_threshold = 0.25,
float nms_iou_threshold = 0.5);
// 以下为模型在预测时的一些参数，基本是前后处理所需
// 用户在创建模型后，可根据模型的要求，以及自己的需求
// 对参数进行修改
// tuple of (width, height)
std::vector<int> size;
// padding value, size should be same with Channels
std::vector<float> padding_value;
// only pad to the minimum rectange which height and width is times of stride
bool is_mini_pad;
// while is_mini_pad = false and is_no_pad = true, will resize the image to
// the set size
bool is_no_pad;
// if is_scale_up is false, the input image only can be zoom out, the maximum
// resize scale cannot exceed 1.0
bool is_scale_up;
// padding stride, for is_mini_pad
int stride;
// for offseting the boxes by classes when using NMS
float max_wh;
// for different strategies to get boxes when postprocessing
bool multi_label;
private:
// 初始化函数，包括初始化后端，以及其它模型推理需要涉及的操作
bool Initialize();
// 输入图像预处理操作
// Mat为FastDeploy定义的数据结构
// FDTensor为预处理后的Tensor数据，传给后端进行推理
// im_info为预处理过程保存的数据，在后处理中需要用到
bool Preprocess(Mat* mat, FDTensor* outputs,
std::map<std::string, std::array<float, 2>>* im_info);
static bool Preprocess(Mat* mat, FDTensor* output,
std::map<std::string, std::array<float, 2>>* im_info,
const std::vector<int>& size = {640, 640},
const std::vector<float> padding_value = {114.0, 114.0,
114.0},
bool is_mini_pad = false, bool is_no_pad = false,
bool is_scale_up = false, int stride = 32,
float max_wh = 7680.0, bool multi_label = true);
// 后端推理结果后处理，输出给用户
// infer_result 为后端推理后的输出Tensor
@@ -81,17 +61,45 @@ class FASTDEPLOY_DECL YOLOv5 : public FastDeployModel {
// conf_threshold 后处理时过滤box的置信度阈值
// nms_iou_threshold 后处理时NMS设定的iou阈值
// multi_label 后处理时box选取是否采用多标签方式
bool Postprocess(FDTensor& infer_result, DetectionResult* result,
const std::map<std::string, std::array<float, 2>>& im_info,
float conf_threshold, float nms_iou_threshold,
bool multi_label);
static bool Postprocess(
std::vector<FDTensor>& infer_results, DetectionResult* result,
const std::map<std::string, std::array<float, 2>>& im_info,
float conf_threshold, float nms_iou_threshold, bool multi_label,
float max_wh = 7680.0);
// 以下为模型在预测时的一些参数，基本是前后处理所需
// 用户在创建模型后，可根据模型的要求，以及自己的需求
// 对参数进行修改
// tuple of (width, height)
std::vector<int> size_;
// padding value, size should be same with Channels
std::vector<float> padding_value_;
// only pad to the minimum rectange which height and width is times of stride
bool is_mini_pad_;
// while is_mini_pad = false and is_no_pad = true, will resize the image to
// the set size
bool is_no_pad_;
// if is_scale_up is false, the input image only can be zoom out, the maximum
// resize scale cannot exceed 1.0
bool is_scale_up_;
// padding stride, for is_mini_pad
int stride_;
// for offseting the boxes by classes when using NMS
float max_wh_;
// for different strategies to get boxes when postprocessing
bool multi_label_;
private:
// 初始化函数，包括初始化后端，以及其它模型推理需要涉及的操作
bool Initialize();
// 查看输入是否为动态维度的 不建议直接使用 不同模型的逻辑可能不一致
bool IsDynamicInput() const { return is_dynamic_input_; }
void LetterBox(Mat* mat, std::vector<int> size, std::vector<float> color,
bool _auto, bool scale_fill = false, bool scale_up = true,
int stride = 32);
static void LetterBox(Mat* mat, std::vector<int> size,
std::vector<float> color, bool _auto,
bool scale_fill = false, bool scale_up = true,
int stride = 32);
// whether to inference with dynamic shape (e.g ONNX export with dynamic shape
// or not.)

46
fastdeploy/vision/detection/contrib/yolov5_pybind.cc
View File

@@ -26,13 +26,43 @@ void BindYOLOv5(pybind11::module& m) {
self.Predict(&mat, &res, conf_threshold, nms_iou_threshold);
return res;
})
.def_readwrite("size", &vision::detection::YOLOv5::size)
.def_readwrite("padding_value", &vision::detection::YOLOv5::padding_value)
.def_readwrite("is_mini_pad", &vision::detection::YOLOv5::is_mini_pad)
.def_readwrite("is_no_pad", &vision::detection::YOLOv5::is_no_pad)
.def_readwrite("is_scale_up", &vision::detection::YOLOv5::is_scale_up)
.def_readwrite("stride", &vision::detection::YOLOv5::stride)
.def_readwrite("max_wh", &vision::detection::YOLOv5::max_wh)
.def_readwrite("multi_label", &vision::detection::YOLOv5::multi_label);
.def_static("preprocess",
[](pybind11::array& data, const std::vector<int>& size,
const std::vector<float> padding_value, bool is_mini_pad,
bool is_no_pad, bool is_scale_up, int stride, float max_wh,
bool multi_label) {
auto mat = PyArrayToCvMat(data);
fastdeploy::vision::Mat fd_mat(mat);
FDTensor output;
std::map<std::string, std::array<float, 2>> im_info;
vision::detection::YOLOv5::Preprocess(
&fd_mat, &output, &im_info, size, padding_value,
is_mini_pad, is_no_pad, is_scale_up, stride, max_wh,
multi_label);
return make_pair(TensorToPyArray(output), im_info);
})
.def_static(
"postprocess",
[](std::vector<pybind11::array> infer_results,
const std::map<std::string, std::array<float, 2>>& im_info,
float conf_threshold, float nms_iou_threshold, bool multi_label,
float max_wh) {
std::vector<FDTensor> fd_infer_results(infer_results.size());
PyArrayToTensorList(infer_results, &fd_infer_results, true);
vision::DetectionResult result;
vision::detection::YOLOv5::Postprocess(
fd_infer_results, &result, im_info, conf_threshold,
nms_iou_threshold, multi_label, max_wh);
return result;
})
.def_readwrite("size", &vision::detection::YOLOv5::size_)
.def_readwrite("padding_value",
&vision::detection::YOLOv5::padding_value_)
.def_readwrite("is_mini_pad", &vision::detection::YOLOv5::is_mini_pad_)
.def_readwrite("is_no_pad", &vision::detection::YOLOv5::is_no_pad_)
.def_readwrite("is_scale_up", &vision::detection::YOLOv5::is_scale_up_)
.def_readwrite("stride", &vision::detection::YOLOv5::stride_)
.def_readwrite("max_wh", &vision::detection::YOLOv5::max_wh_)
.def_readwrite("multi_label", &vision::detection::YOLOv5::multi_label_);
}
} // namespace fastdeploy