Skip to content

P
PaddleX

PaddlePaddle / PaddleX

通知 138
Star 4
Fork 0
P
PaddleX
未验证 提交 c7b5a892 编写于 作者: M mamingjie-China 提交者： GitHub
浏览文件
操作

Merge pull request #3 from PaddlePaddle/develop 

update
上级 8c0c13e1 dd323e8d
展开全部 隐藏空白更改
内联 并排
Showing with 3482 addition and 506 deletion
README.md
浏览文件 @ c7b5a892
......@@ -14,7 +14,7 @@
![support os](https://img.shields.io/badge/os-linux%2C%20win%2C%20mac-yellow.svg)
![QQGroup](https://img.shields.io/badge/QQ_Group-1045148026-52B6EF?style=social&logo=tencent-qq&logoColor=000&logoWidth=20)
集成飞桨智能视觉领域**图像分类**、**目标检测**、**语义分割**、**实例分割**任务能力，将深度学习开发全流程从**数据准备**、**模型训练与优化**到**多端部署**端到端打通，无需分别安装不同功能模块，并提供统一任务API接口，以**低代码**的形式为开发者提供飞桨全流程开发的最佳体验。
集成飞桨智能视觉领域**图像分类**、**目标检测**、**语义分割**、**实例分割**任务能力，将深度学习开发全流程从**数据准备**、**模型训练与优化**到**多端部署**端到端打通，并提供**统一任务API接口**及**图形化开发界面Demo**。开发者无需分别安装不同套件，以**低代码**的形式即可快速完成飞桨全流程开发。
**PaddleX** 经过**质检**、**安防**、**巡检**、**遥感**、**零售**、**医疗**等十多个行业实际应用场景验证，沉淀产业实际经验，**并提供丰富的案例实践教程**，全程助力开发者产业实践落地。
......@@ -48,8 +48,6 @@ pip install paddlex -i https://mirror.baidu.com/pypi/simple
- 前往[PaddleX GUI使用教程](./docs/gui/how_to_use.md)了解PaddleX GUI使用详情。
- https://aistudio.baidu.com/aistudio/projectdetail/440197
## 产品模块说明
......
deploy/cpp/CMakeLists.txt
浏览文件 @ c7b5a892
......@@ -305,6 +305,19 @@ add_executable(segmenter demo/segmenter.cpp src/transforms.cpp src/paddlex.cpp s
ADD_DEPENDENCIES(segmenter ext-yaml-cpp)
target_link_libraries(segmenter ${DEPS})
add_executable(video_classifier demo/video_classifier.cpp src/transforms.cpp src/paddlex.cpp src/visualize.cpp)
ADD_DEPENDENCIES(video_classifier ext-yaml-cpp)
target_link_libraries(video_classifier ${DEPS})
add_executable(video_detector demo/video_detector.cpp src/transforms.cpp src/paddlex.cpp src/visualize.cpp)
ADD_DEPENDENCIES(video_detector ext-yaml-cpp)
target_link_libraries(video_detector ${DEPS})
add_executable(video_segmenter demo/video_segmenter.cpp src/transforms.cpp src/paddlex.cpp src/visualize.cpp)
ADD_DEPENDENCIES(video_segmenter ext-yaml-cpp)
target_link_libraries(video_segmenter ${DEPS})
if (WIN32 AND WITH_MKL)
add_custom_command(TARGET classifier POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./paddlex_inference/Release/mklml.dll
......@@ -326,7 +339,27 @@ if (WIN32 AND WITH_MKL)
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./paddlex_inference/Release/mkldnn.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./release/mklml.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./release/libiomp5md.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./release/mkldnn.dll
)
add_custom_command(TARGET video_classifier POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./paddlex_inference/Release/mklml.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./paddlex_inference/Release/libiomp5md.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./paddlex_inference/Release/mkldnn.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./release/mklml.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./release/libiomp5md.dll
)
add_custom_command(TARGET video_detector POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./paddlex_inference/Release/mklml.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./paddlex_inference/Release/libiomp5md.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./paddlex_inference/Release/mkldnn.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./release/mklml.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./release/libiomp5md.dll
)
add_custom_command(TARGET video_segmenter POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./paddlex_inference/Release/mklml.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./paddlex_inference/Release/libiomp5md.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./paddlex_inference/Release/mkldnn.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./release/mklml.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./release/libiomp5md.dll
)
# for encryption
if (EXISTS "${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll")
......@@ -342,6 +375,18 @@ if (WIN32 AND WITH_MKL)
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./pmodel-decrypt.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./release/pmodel-decrypt.dll
)
add_custom_command(TARGET video_classifier POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./pmodel-decrypt.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./release/pmodel-decrypt.dll
)
add_custom_command(TARGET video_detector POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./pmodel-decrypt.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./release/pmodel-decrypt.dll
)
add_custom_command(TARGET video_segmenter POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./pmodel-decrypt.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./release/pmodel-decrypt.dll
)
endif()
endif()
......
deploy/cpp/demo/classifier.cpp
浏览文件 @ c7b5a892
......@@ -37,7 +37,6 @@ DEFINE_int32(batch_size, 1, "Batch size of infering");
DEFINE_int32(thread_num,
omp_get_num_procs(),
"Number of preprocessing threads");
DEFINE_bool(use_ir_optim, true, "use ir optimization");
int main(int argc, char** argv) {
// Parsing command-line
......@@ -52,16 +51,15 @@ int main(int argc, char** argv) {
return -1;
}
// 加载模型
// Load model
PaddleX::Model model;
model.Init(FLAGS_model_dir,
FLAGS_use_gpu,
FLAGS_use_trt,
FLAGS_gpu_id,
FLAGS_key,
FLAGS_use_ir_optim);
FLAGS_key);
// 进行预测
// Predict
int imgs = 1;
if (FLAGS_image_list != "") {
std::ifstream inf(FLAGS_image_list);
......@@ -69,7 +67,7 @@ int main(int argc, char** argv) {
std::cerr << "Fail to open file " << FLAGS_image_list << std::endl;
return -1;
}
// 多batch预测
// Mini-batch predict
std::string image_path;
std::vector<std::string> image_paths;
while (getline(inf, image_path)) {
......@@ -77,7 +75,7 @@ int main(int argc, char** argv) {
}
imgs = image_paths.size();
for (int i = 0; i < image_paths.size(); i += FLAGS_batch_size) {
// 读图像
// Read image
int im_vec_size =
std::min(static_cast<int>(image_paths.size()), i + FLAGS_batch_size);
std::vector<cv::Mat> im_vec(im_vec_size - i);
......
deploy/cpp/demo/detector.cpp
浏览文件 @ c7b5a892
......@@ -43,10 +43,9 @@ DEFINE_double(threshold,
DEFINE_int32(thread_num,
omp_get_num_procs(),
"Number of preprocessing threads");
DEFINE_bool(use_ir_optim, true, "use ir optimization");
int main(int argc, char** argv) {
// 解析命令行参数
// Parsing command-line
google::ParseCommandLineFlags(&argc, &argv, true);
if (FLAGS_model_dir == "") {
......@@ -57,17 +56,16 @@ int main(int argc, char** argv) {
std::cerr << "--image or --image_list need to be defined" << std::endl;
return -1;
}
// 加载模型
// Load model
PaddleX::Model model;
model.Init(FLAGS_model_dir,
FLAGS_use_gpu,
FLAGS_use_trt,
FLAGS_gpu_id,
FLAGS_key,
FLAGS_use_ir_optim);
FLAGS_key);
int imgs = 1;
std::string save_dir = "output";
// 进行预测
// Predict
if (FLAGS_image_list != "") {
std::ifstream inf(FLAGS_image_list);
if (!inf) {
......@@ -92,7 +90,7 @@ int main(int argc, char** argv) {
im_vec[j - i] = std::move(cv::imread(image_paths[j], 1));
}
model.predict(im_vec, &results, thread_num);
// 输出结果目标框
// Output predicted bounding boxes
for (int j = 0; j < im_vec_size - i; ++j) {
for (int k = 0; k < results[j].boxes.size(); ++k) {
std::cout << "image file: " << image_paths[i + j] << ", ";
......@@ -106,7 +104,7 @@ int main(int argc, char** argv) {
<< results[j].boxes[k].coordinate[3] << ")" << std::endl;
}
}
// 可视化
// Visualize results
for (int j = 0; j < im_vec_size - i; ++j) {
cv::Mat vis_img = PaddleX::Visualize(
im_vec[j], results[j], model.labels, FLAGS_threshold);
......@@ -120,7 +118,7 @@ int main(int argc, char** argv) {
PaddleX::DetResult result;
cv::Mat im = cv::imread(FLAGS_image, 1);
model.predict(im, &result);
// 输出结果目标框
// Output predicted bounding boxes
for (int i = 0; i < result.boxes.size(); ++i) {
std::cout << "image file: " << FLAGS_image << std::endl;
std::cout << ", predict label: " << result.boxes[i].category
......@@ -132,7 +130,7 @@ int main(int argc, char** argv) {
<< result.boxes[i].coordinate[3] << ")" << std::endl;
}
// 可视化
// Visualize results
cv::Mat vis_img =
PaddleX::Visualize(im, result, model.labels, FLAGS_threshold);
std::string save_path =
......
deploy/cpp/demo/segmenter.cpp
浏览文件 @ c7b5a892
......@@ -39,10 +39,9 @@ DEFINE_int32(batch_size, 1, "Batch size of infering");
DEFINE_int32(thread_num,
omp_get_num_procs(),
"Number of preprocessing threads");
DEFINE_bool(use_ir_optim, false, "use ir optimization");
int main(int argc, char** argv) {
// 解析命令行参数
// Parsing command-line
google::ParseCommandLineFlags(&argc, &argv, true);
if (FLAGS_model_dir == "") {
......@@ -54,16 +53,15 @@ int main(int argc, char** argv) {
return -1;
}
// 加载模型
// Load model
PaddleX::Model model;
model.Init(FLAGS_model_dir,
FLAGS_use_gpu,
FLAGS_use_trt,
FLAGS_gpu_id,
FLAGS_key,
FLAGS_use_ir_optim);
FLAGS_key);
int imgs = 1;
// 进行预测
// Predict
if (FLAGS_image_list != "") {
std::ifstream inf(FLAGS_image_list);
if (!inf) {
......@@ -88,7 +86,7 @@ int main(int argc, char** argv) {
im_vec[j - i] = std::move(cv::imread(image_paths[j], 1));
}
model.predict(im_vec, &results, thread_num);
// 可视化
// Visualize results
for (int j = 0; j < im_vec_size - i; ++j) {
cv::Mat vis_img =
PaddleX::Visualize(im_vec[j], results[j], model.labels);
......@@ -102,7 +100,7 @@ int main(int argc, char** argv) {
PaddleX::SegResult result;
cv::Mat im = cv::imread(FLAGS_image, 1);
model.predict(im, &result);
// 可视化
// Visualize results
cv::Mat vis_img = PaddleX::Visualize(im, result, model.labels);
std::string save_path =
PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_image);
......
deploy/cpp/demo/video_classifier.cpp 0 → 100644
浏览文件 @ c7b5a892
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <glog/logging.h>
#include <omp.h>
#include <algorithm>
#include <chrono> // NOLINT
#include <fstream>
#include <iostream>
#include <string>
#include <vector>
#include <utility>
#include "include/paddlex/paddlex.h"
#include "include/paddlex/visualize.h"
#if defined(__arm__) || defined(__aarch64__)
#include <opencv2/videoio/legacy/constants_c.h>
#endif
using namespace std::chrono; // NOLINT
DEFINE_string(model_dir, "", "Path of inference model");
DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
DEFINE_bool(use_trt, false, "Infering with TensorRT");
DEFINE_int32(gpu_id, 0, "GPU card id");
DEFINE_string(key, "", "key of encryption");
DEFINE_bool(use_camera, false, "Infering with Camera");
DEFINE_int32(camera_id, 0, "Camera id");
DEFINE_string(video_path, "", "Path of input video");
DEFINE_bool(show_result, false, "show the result of each frame with a window");
DEFINE_bool(save_result, true, "save the result of each frame to a video");
DEFINE_string(save_dir, "output", "Path to save visualized image");
int main(int argc, char** argv) {
// Parsing command-line
google::ParseCommandLineFlags(&argc, &argv, true);
if (FLAGS_model_dir == "") {
std::cerr << "--model_dir need to be defined" << std::endl;
return -1;
}
if (FLAGS_video_path == "" & FLAGS_use_camera == false) {
std::cerr << "--video_path or --use_camera need to be defined" << std::endl;
return -1;
}
// Load model
PaddleX::Model model;
model.Init(FLAGS_model_dir,
FLAGS_use_gpu,
FLAGS_use_trt,
FLAGS_gpu_id,
FLAGS_key);
// Open video
cv::VideoCapture capture;
if (FLAGS_use_camera) {
capture.open(FLAGS_camera_id);
if (!capture.isOpened()) {
std::cout << "Can not open the camera "
<< FLAGS_camera_id << "."
<< std::endl;
return -1;
}
} else {
capture.open(FLAGS_video_path);
if (!capture.isOpened()) {
std::cout << "Can not open the video "
<< FLAGS_video_path << "."
<< std::endl;
return -1;
}
}
// Create a VideoWriter
cv::VideoWriter video_out;
std::string video_out_path;
if (FLAGS_save_result) {
// Get video information: resolution, fps
int video_width = static_cast<int>(capture.get(CV_CAP_PROP_FRAME_WIDTH));
int video_height = static_cast<int>(capture.get(CV_CAP_PROP_FRAME_HEIGHT));
int video_fps = static_cast<int>(capture.get(CV_CAP_PROP_FPS));
int video_fourcc;
if (FLAGS_use_camera) {
video_fourcc = 828601953;
} else {
video_fourcc = static_cast<int>(capture.get(CV_CAP_PROP_FOURCC));
}
if (FLAGS_use_camera) {
time_t now = time(0);
video_out_path =
PaddleX::generate_save_path(FLAGS_save_dir,
std::to_string(now) + ".mp4");
} else {
video_out_path =
PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_video_path);
}
video_out.open(video_out_path.c_str(),
video_fourcc,
video_fps,
cv::Size(video_width, video_height),
true);
if (!video_out.isOpened()) {
std::cout << "Create video writer failed!" << std::endl;
return -1;
}
}
PaddleX::ClsResult result;
cv::Mat frame;
int key;
while (capture.read(frame)) {
if (FLAGS_show_result || FLAGS_use_camera) {
key = cv::waitKey(1);
// When pressing `ESC`, then exit program and result video is saved
if (key == 27) {
break;
}
} else if (frame.empty()) {
break;
}
// Begin to predict
model.predict(frame, &result);
// Visualize results
cv::Mat vis_img = frame.clone();
auto colormap = PaddleX::GenerateColorMap(model.labels.size());
int c1 = colormap[3 * result.category_id + 0];
int c2 = colormap[3 * result.category_id + 1];
int c3 = colormap[3 * result.category_id + 2];
cv::Scalar text_color = cv::Scalar(c1, c2, c3);
std::string text = result.category;
text += std::to_string(static_cast<int>(result.score * 100)) + "%";
int font_face = cv::FONT_HERSHEY_SIMPLEX;
double font_scale = 0.5f;
float thickness = 0.5;
cv::Size text_size =
cv::getTextSize(text, font_face, font_scale, thickness, nullptr);
cv::Point origin;
origin.x = frame.cols / 2;
origin.y = frame.rows / 2;
cv::Rect text_back = cv::Rect(origin.x,
origin.y - text_size.height,
text_size.width,
text_size.height);
cv::rectangle(vis_img, text_back, text_color, -1);
cv::putText(vis_img,
text,
origin,
font_face,
font_scale,
cv::Scalar(255, 255, 255),
thickness);
if (FLAGS_show_result || FLAGS_use_camera) {
cv::imshow("video_classifier", vis_img);
}
if (FLAGS_save_result) {
video_out.write(vis_img);
}
std::cout << "Predict label: " << result.category
<< ", label_id:" << result.category_id
<< ", score: " << result.score << std::endl;
}
capture.release();
if (FLAGS_save_result) {
video_out.release();
std::cout << "Visualized output saved as " << video_out_path << std::endl;
}
if (FLAGS_show_result || FLAGS_use_camera) {
cv::destroyAllWindows();
}
return 0;
}
deploy/cpp/demo/video_detector.cpp 0 → 100644
浏览文件 @ c7b5a892
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <glog/logging.h>
#include <omp.h>
#include <algorithm>
#include <chrono> // NOLINT
#include <fstream>
#include <iostream>
#include <string>
#include <vector>
#include <utility>
#include "include/paddlex/paddlex.h"
#include "include/paddlex/visualize.h"
#if defined(__arm__) || defined(__aarch64__)
#include <opencv2/videoio/legacy/constants_c.h>
#endif
using namespace std::chrono; // NOLINT
DEFINE_string(model_dir, "", "Path of inference model");
DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
DEFINE_bool(use_trt, false, "Infering with TensorRT");
DEFINE_int32(gpu_id, 0, "GPU card id");
DEFINE_bool(use_camera, false, "Infering with Camera");
DEFINE_int32(camera_id, 0, "Camera id");
DEFINE_string(video_path, "", "Path of input video");
DEFINE_bool(show_result, false, "show the result of each frame with a window");
DEFINE_bool(save_result, true, "save the result of each frame to a video");
DEFINE_string(key, "", "key of encryption");
DEFINE_string(save_dir, "output", "Path to save visualized image");
DEFINE_double(threshold,
0.5,
"The minimum scores of target boxes which are shown");
int main(int argc, char** argv) {
// Parsing command-line
google::ParseCommandLineFlags(&argc, &argv, true);
if (FLAGS_model_dir == "") {
std::cerr << "--model_dir need to be defined" << std::endl;
return -1;
}
if (FLAGS_video_path == "" & FLAGS_use_camera == false) {
std::cerr << "--video_path or --use_camera need to be defined" << std::endl;
return -1;
}
// Load model
PaddleX::Model model;
model.Init(FLAGS_model_dir,
FLAGS_use_gpu,
FLAGS_use_trt,
FLAGS_gpu_id,
FLAGS_key);
// Open video
cv::VideoCapture capture;
if (FLAGS_use_camera) {
capture.open(FLAGS_camera_id);
if (!capture.isOpened()) {
std::cout << "Can not open the camera "
<< FLAGS_camera_id << "."
<< std::endl;
return -1;
}
} else {
capture.open(FLAGS_video_path);
if (!capture.isOpened()) {
std::cout << "Can not open the video "
<< FLAGS_video_path << "."
<< std::endl;
return -1;
}
}
// Create a VideoWriter
cv::VideoWriter video_out;
std::string video_out_path;
if (FLAGS_save_result) {
// Get video information: resolution, fps
int video_width = static_cast<int>(capture.get(CV_CAP_PROP_FRAME_WIDTH));
int video_height = static_cast<int>(capture.get(CV_CAP_PROP_FRAME_HEIGHT));
int video_fps = static_cast<int>(capture.get(CV_CAP_PROP_FPS));
int video_fourcc;
if (FLAGS_use_camera) {
video_fourcc = 828601953;
} else {
video_fourcc = static_cast<int>(capture.get(CV_CAP_PROP_FOURCC));
}
if (FLAGS_use_camera) {
time_t now = time(0);
video_out_path =
PaddleX::generate_save_path(FLAGS_save_dir,
std::to_string(now) + ".mp4");
} else {
video_out_path =
PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_video_path);
}
video_out.open(video_out_path.c_str(),
video_fourcc,
video_fps,
cv::Size(video_width, video_height),
true);
if (!video_out.isOpened()) {
std::cout << "Create video writer failed!" << std::endl;
return -1;
}
}
PaddleX::DetResult result;
cv::Mat frame;
int key;
while (capture.read(frame)) {
if (FLAGS_show_result || FLAGS_use_camera) {
key = cv::waitKey(1);
// When pressing `ESC`, then exit program and result video is saved
if (key == 27) {
break;
}
} else if (frame.empty()) {
break;
}
// Begin to predict
model.predict(frame, &result);
// Visualize results
cv::Mat vis_img =
PaddleX::Visualize(frame, result, model.labels, FLAGS_threshold);
if (FLAGS_show_result || FLAGS_use_camera) {
cv::imshow("video_detector", vis_img);
}
if (FLAGS_save_result) {
video_out.write(vis_img);
}
result.clear();
}
capture.release();
if (FLAGS_save_result) {
std::cout << "Visualized output saved as " << video_out_path << std::endl;
video_out.release();
}
if (FLAGS_show_result || FLAGS_use_camera) {
cv::destroyAllWindows();
}
return 0;
}
deploy/cpp/demo/video_segmenter.cpp 0 → 100644
浏览文件 @ c7b5a892
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <glog/logging.h>
#include <omp.h>
#include <algorithm>
#include <chrono> // NOLINT
#include <fstream>
#include <iostream>
#include <string>
#include <vector>
#include <utility>
#include <ctime>
#include "include/paddlex/paddlex.h"
#include "include/paddlex/visualize.h"
#if defined(__arm__) || defined(__aarch64__)
#include <opencv2/videoio/legacy/constants_c.h>
#endif
using namespace std::chrono; // NOLINT
DEFINE_string(model_dir, "", "Path of inference model");
DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
DEFINE_bool(use_trt, false, "Infering with TensorRT");
DEFINE_int32(gpu_id, 0, "GPU card id");
DEFINE_string(key, "", "key of encryption");
DEFINE_bool(use_camera, false, "Infering with Camera");
DEFINE_int32(camera_id, 0, "Camera id");
DEFINE_string(video_path, "", "Path of input video");
DEFINE_bool(show_result, false, "show the result of each frame with a window");
DEFINE_bool(save_result, true, "save the result of each frame to a video");
DEFINE_string(save_dir, "output", "Path to save visualized image");
int main(int argc, char** argv) {
// Parsing command-line
google::ParseCommandLineFlags(&argc, &argv, true);
if (FLAGS_model_dir == "") {
std::cerr << "--model_dir need to be defined" << std::endl;
return -1;
}
if (FLAGS_video_path == "" & FLAGS_use_camera == false) {
std::cerr << "--video_path or --use_camera need to be defined" << std::endl;
return -1;
}
// Load model
PaddleX::Model model;
model.Init(FLAGS_model_dir,
FLAGS_use_gpu,
FLAGS_use_trt,
FLAGS_gpu_id,
FLAGS_key);
// Open video
cv::VideoCapture capture;
if (FLAGS_use_camera) {
capture.open(FLAGS_camera_id);
if (!capture.isOpened()) {
std::cout << "Can not open the camera "
<< FLAGS_camera_id << "."
<< std::endl;
return -1;
}
} else {
capture.open(FLAGS_video_path);
if (!capture.isOpened()) {
std::cout << "Can not open the video "
<< FLAGS_video_path << "."
<< std::endl;
return -1;
}
}
// Create a VideoWriter
cv::VideoWriter video_out;
std::string video_out_path;
if (FLAGS_save_result) {
// Get video information: resolution, fps
int video_width = static_cast<int>(capture.get(CV_CAP_PROP_FRAME_WIDTH));
int video_height = static_cast<int>(capture.get(CV_CAP_PROP_FRAME_HEIGHT));
int video_fps = static_cast<int>(capture.get(CV_CAP_PROP_FPS));
int video_fourcc;
if (FLAGS_use_camera) {
video_fourcc = 828601953;
} else {
video_fourcc = static_cast<int>(capture.get(CV_CAP_PROP_FOURCC));
}
if (FLAGS_use_camera) {
time_t now = time(0);
video_out_path =
PaddleX::generate_save_path(FLAGS_save_dir,
std::to_string(now) + ".mp4");
} else {
video_out_path =
PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_video_path);
}
video_out.open(video_out_path.c_str(),
video_fourcc,
video_fps,
cv::Size(video_width, video_height),
true);
if (!video_out.isOpened()) {
std::cout << "Create video writer failed!" << std::endl;
return -1;
}
}
PaddleX::SegResult result;
cv::Mat frame;
int key;
while (capture.read(frame)) {
if (FLAGS_show_result || FLAGS_use_camera) {
key = cv::waitKey(1);
// When pressing `ESC`, then exit program and result video is saved
if (key == 27) {
break;
}
} else if (frame.empty()) {
break;
}
// Begin to predict
model.predict(frame, &result);
// Visualize results
cv::Mat vis_img = PaddleX::Visualize(frame, result, model.labels);
if (FLAGS_show_result || FLAGS_use_camera) {
cv::imshow("video_segmenter", vis_img);
}
if (FLAGS_save_result) {
video_out.write(vis_img);
}
result.clear();
}
capture.release();
if (FLAGS_save_result) {
video_out.release();
std::cout << "Visualized output saved as " << video_out_path << std::endl;
}
if (FLAGS_show_result || FLAGS_use_camera) {
cv::destroyAllWindows();
}
return 0;
}
deploy/cpp/include/paddlex/visualize.h
浏览文件 @ c7b5a892
......@@ -24,8 +24,8 @@
#include <dirent.h>
// #include <sys/io.h>
#if defined(__arm__) || defined(__aarch64__) // for arm
#include <aarch64-linux-gpu/sys/stat.h>
#include <aarch64-linux-gpu/sys/types.h>
#include <aarch64-linux-gnu/sys/stat.h>
#include <aarch64-linux-gnu/sys/types.h>
#else
#include <sys/stat.h>
#include <sys/types.h>
......
deploy/cpp/scripts/bootstrap.sh
浏览文件 @ c7b5a892
......@@ -7,12 +7,12 @@ if [ ! -d "./paddlex-encryption" ]; then
fi
# download pre-compiled opencv lib
OPENCV_URL=https://paddleseg.bj.bcebos.com/deploy/docker/opencv3gcc4.8.tar.bz2
if [ ! -d "./deps/opencv3gcc4.8" ]; then
OPENCV_URL=https://bj.bcebos.com/paddleseg/deploy/opencv3.4.6gcc4.8ffmpeg.tar.gz2
if [ ! -d "./deps/opencv3.4.6gcc4.8ffmpeg/" ]; then
mkdir -p deps
cd deps
wget -c ${OPENCV_URL}
tar xvfj opencv3gcc4.8.tar.bz2
rm -rf opencv3gcc4.8.tar.bz2
tar xvfj opencv3.4.6gcc4.8ffmpeg.tar.gz2
rm -rf opencv3.4.6gcc4.8ffmpeg.tar.gz2
cd ..
fi
deploy/cpp/scripts/build.sh
浏览文件 @ c7b5a892
......@@ -24,7 +24,7 @@ ENCRYPTION_DIR=$(pwd)/paddlex-encryption
# OPENCV 路径, 如果使用自带预编译版本可不修改
sh $(pwd)/scripts/bootstrap.sh # 下载预编译版本的opencv
OPENCV_DIR=$(pwd)/deps/opencv3gcc4.8/
OPENCV_DIR=$(pwd)/deps/opencv3.4.6gcc4.8ffmpeg/
# 以下无需改动
rm -rf build
......@@ -42,4 +42,4 @@ cmake .. \
-DCUDNN_LIB=${CUDNN_LIB} \
-DENCRYPTION_DIR=${ENCRYPTION_DIR} \
-DOPENCV_DIR=${OPENCV_DIR}
make
make -j16
deploy/cpp/src/paddlex.cpp
浏览文件 @ c7b5a892
......@@ -65,7 +65,11 @@ void Model::create_predictor(const std::string& model_dir,
config.SwitchUseFeedFetchOps(false);
config.SwitchSpecifyInputNames(true);
// 开启图优化
#if defined(__arm__) || defined(__aarch64__)
config.SwitchIrOptim(false);
#else
config.SwitchIrOptim(use_ir_optim);
#endif
// 开启内存优化
config.EnableMemoryOptim();
if (use_trt) {
......
docs/deploy/nvidia-jetson.md
浏览文件 @ c7b5a892
......@@ -57,13 +57,6 @@ CUDA_LIB=/usr/local/cuda/lib64
# CUDNN 的 lib 路径
CUDNN_LIB=/usr/local/cuda/lib64
# 是否加载加密后的模型
WITH_ENCRYPTION=OFF
# OPENCV 路径, 如果使用自带预编译版本可不修改
sh $(pwd)/scripts/jetson_bootstrap.sh # 下载预编译版本的opencv
OPENCV_DIR=$(pwd)/deps/opencv3/
# 以下无需改动
rm -rf build
mkdir -p build
......@@ -77,18 +70,13 @@ cmake .. \
-DPADDLE_DIR=${PADDLE_DIR} \
-DWITH_STATIC_LIB=${WITH_STATIC_LIB} \
-DCUDA_LIB=${CUDA_LIB} \
-DCUDNN_LIB=${CUDNN_LIB} \
-DENCRYPTION_DIR=${ENCRYPTION_DIR} \
-DOPENCV_DIR=${OPENCV_DIR}
-DCUDNN_LIB=${CUDNN_LIB}
make
```
**注意：** linux环境下编译会自动下载OPENCV和YAML，如果编译环境无法访问外网，可手动下载：
**注意：** linux环境下编译会自动下载YAML，如果编译环境无法访问外网，可手动下载：
- [opencv3_aarch.tgz](https://bj.bcebos.com/paddlex/deploy/tools/opencv3_aarch.tgz)
- [yaml-cpp.zip](https://bj.bcebos.com/paddlex/deploy/deps/yaml-cpp.zip)
opencv3_aarch.tgz文件下载后解压，然后在script/build.sh中指定`OPENCE_DIR`为解压后的路径。
yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://bj.bcebos.com/paddlex/deploy/deps/yaml-cpp.zip` 中的网址，改为下载文件的路径。
修改脚本设置好主要参数后，执行`build`脚本：
......@@ -100,7 +88,7 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
**在加载模型前，请检查你的模型目录中文件应该包括`model.yml`、`__model__`和`__params__`三个文件。如若不满足这个条件，请参考[模型导出为Inference文档](export_model.md)将模型导出为部署格式。**
编译成功后，预测demo的可执行程序分别为`build/demo/detector`，`build/demo/classifier`，`build/demo/segmenter`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
* 编译成功后，图片预测demo的可执行程序分别为`build/demo/detector`，`build/demo/classifier`，`build/demo/segmenter`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
| 参数 | 说明 |
| ---- | ---- |
......@@ -111,10 +99,26 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
| use_trt | 是否使用 TensorRT 预测, 支持值为0或1(默认值为0) |
| gpu_id | GPU 设备ID, 默认值为0 |
| save_dir | 保存可视化结果的路径, 默认值为"output"，**classfier无该参数** |
| key | 加密过程中产生的密钥信息，默认值为""表示加载的是未加密的模型 |
| batch_size | 预测的批量大小，默认为1 |
| thread_num | 预测的线程数，默认为cpu处理器个数 |
| use_ir_optim | 是否使用图优化策略，支持值为0或1（默认值为1，图像分割默认值为0）|
* 编译成功后，视频预测demo的可执行程序分别为`build/demo/video_detector`，`build/demo/video_classifier`，`build/demo/video_segmenter`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
| 参数 | 说明 |
| ---- | ---- |
| model_dir | 导出的预测模型所在路径 |
| use_camera | 是否使用摄像头预测，支持值为0或1(默认值为0) |
| camera_id | 摄像头设备ID，默认值为0 |
| video_path | 视频文件的路径 |
| use_gpu | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
| use_trt | 是否使用 TensorRT 预测, 支持值为0或1(默认值为0) |
| gpu_id | GPU 设备ID, 默认值为0 |
| show_result | 对视频文件做预测时，是否在屏幕上实时显示预测可视化结果(因加入了延迟处理，故显示结果不能反映真实的帧率)，支持值为0或1(默认值为0) |
| save_result | 是否将每帧的预测可视结果保存为视频文件，支持值为0或1(默认值为1) |
| save_dir | 保存可视化结果的路径, 默认值为"output" |
**注意：若系统无GUI，则不要将show_result设置为1。当使用摄像头预测时，按`ESC`键可关闭摄像头并推出预测程序。**
## 样例
......@@ -143,3 +147,21 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
./build/demo/detector --model_dir=/root/projects/inference_model --image_list=/root/projects/images_list.txt --use_gpu=1 --save_dir=output --batch_size=2 --thread_num=2
```
图片文件`可视化预测结果`会保存在`save_dir`参数设置的目录下。
**样例三：**
使用摄像头预测：
```shell
./build/demo/video_detector --model_dir=/root/projects/inference_model --use_camera=1 --use_gpu=1 --save_dir=output --save_result=1
```
当`save_result`设置为1时，`可视化预测结果`会以视频文件的格式保存在`save_dir`参数设置的目录下。
**样例四：**
对视频文件进行预测：
```shell
./build/demo/video_detector --model_dir=/root/projects/inference_model --video_path=/path/to/video_file --use_gpu=1 --save_dir=output --show_result=1 --save_result=1
```
当`save_result`设置为1时，`可视化预测结果`会以视频文件的格式保存在`save_dir`参数设置的目录下。如果系统有GUI，通过将`show_result`设置为1在屏幕上观看可视化预测结果。
docs/deploy/server/cpp/linux.md
浏览文件 @ c7b5a892
......@@ -116,7 +116,7 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
**在加载模型前，请检查你的模型目录中文件应该包括`model.yml`、`__model__`和`__params__`三个文件。如若不满足这个条件，请参考[模型导出为Inference文档](../../export_model.md)将模型导出为部署格式。**
编译成功后，预测demo的可执行程序分别为`build/demo/detector`，`build/demo/classifier`，`build/demo/segmenter`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
* 编译成功后，图片预测demo的可执行程序分别为`build/demo/detector`，`build/demo/classifier`，`build/demo/segmenter`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
| 参数 | 说明 |
| ---- | ---- |
......@@ -130,7 +130,24 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
| key | 加密过程中产生的密钥信息，默认值为""表示加载的是未加密的模型 |
| batch_size | 预测的批量大小，默认为1 |
| thread_num | 预测的线程数，默认为cpu处理器个数 |
| use_ir_optim | 是否使用图优化策略，支持值为0或1（默认值为1，图像分割默认值为0）|
* 编译成功后，视频预测demo的可执行程序分别为`build/demo/video_detector`，`build/demo/video_classifier`，`build/demo/video_segmenter`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
| 参数 | 说明 |
| ---- | ---- |
| model_dir | 导出的预测模型所在路径 |
| use_camera | 是否使用摄像头预测，支持值为0或1(默认值为0) |
| camera_id | 摄像头设备ID，默认值为0 |
| video_path | 视频文件的路径 |
| use_gpu | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
| use_trt | 是否使用 TensorRT 预测, 支持值为0或1(默认值为0) |
| gpu_id | GPU 设备ID, 默认值为0 |
| show_result | 对视频文件做预测时，是否在屏幕上实时显示预测可视化结果(因加入了延迟处理，故显示结果不能反映真实的帧率)，支持值为0或1(默认值为0) |
| save_result | 是否将每帧的预测可视结果保存为视频文件，支持值为0或1(默认值为1) |
| save_dir | 保存可视化结果的路径, 默认值为"output"|
| key | 加密过程中产生的密钥信息，默认值为""表示加载的是未加密的模型 |
**注意：若系统无GUI，则不要将show_result设置为1。当使用摄像头预测时，按`ESC`键可关闭摄像头并推出预测程序。**
## 样例
......@@ -138,7 +155,7 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
> 关于预测速度的说明：加载模型后前几张图片的预测速度会较慢，这是因为运行启动时涉及到内存显存初始化等步骤，通常在预测20-30张图片后模型的预测速度达到稳定。
`样例一`：
**样例一：**
不使用`GPU`测试图片 `/root/projects/images/xiaoduxiong.jpeg`
......@@ -148,7 +165,7 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
图片文件`可视化预测结果`会保存在`save_dir`参数设置的目录下。
`样例二`:
**样例二:**
使用`GPU`预测多个图片`/root/projects/image_list.txt`，image_list.txt内容的格式如下：
```
......@@ -161,3 +178,21 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
./build/demo/detector --model_dir=/root/projects/inference_model --image_list=/root/projects/images_list.txt --use_gpu=1 --save_dir=output --batch_size=2 --thread_num=2
```
图片文件`可视化预测结果`会保存在`save_dir`参数设置的目录下。
**样例三：**
使用摄像头预测：
```shell
./build/demo/video_detector --model_dir=/root/projects/inference_model --use_camera=1 --use_gpu=1 --save_dir=output --save_result=1
```
当`save_result`设置为1时，`可视化预测结果`会以视频文件的格式保存在`save_dir`参数设置的目录下。
**样例四：**
对视频文件进行预测：
```shell
./build/demo/video_detector --model_dir=/root/projects/inference_model --video_path=/path/to/video_file --use_gpu=1 --save_dir=output --show_result=1 --save_result=1
```
当`save_result`设置为1时，`可视化预测结果`会以视频文件的格式保存在`save_dir`参数设置的目录下。如果系统有GUI，通过将`show_result`设置为1在屏幕上观看可视化预测结果。
docs/deploy/server/cpp/windows.md
浏览文件 @ c7b5a892
......@@ -101,7 +101,7 @@ D:
cd D:\projects\PaddleX\deploy\cpp\out\build\x64-Release
```
编译成功后，预测demo的入口程序为`paddlex_inference\detector.exe`，`paddlex_inference\classifier.exe`，`paddlex_inference\segmenter.exe`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
* 编译成功后，图片预测demo的入口程序为`paddlex_inference\detector.exe`，`paddlex_inference\classifier.exe`，`paddlex_inference\segmenter.exe`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
| 参数 | 说明 |
| ---- | ---- |
......@@ -114,7 +114,24 @@ cd D:\projects\PaddleX\deploy\cpp\out\build\x64-Release
| key | 加密过程中产生的密钥信息，默认值为""表示加载的是未加密的模型 |
| batch_size | 预测的批量大小，默认为1 |
| thread_num | 预测的线程数，默认为cpu处理器个数 |
| use_ir_optim | 是否使用图优化策略，支持值为0或1（默认值为1，图像分割默认值为0）|
* 编译成功后，视频预测demo的入口程序为`paddlex_inference\video_detector.exe`，`paddlex_inference\video_classifier.exe`，`paddlex_inference\video_segmenter.exe`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
| 参数 | 说明 |
| ---- | ---- |
| model_dir | 导出的预测模型所在路径 |
| use_camera | 是否使用摄像头预测，支持值为0或1(默认值为0) |
| camera_id | 摄像头设备ID，默认值为0 |
| video_path | 视频文件的路径 |
| use_gpu | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
| gpu_id | GPU 设备ID, 默认值为0 |
| show_result | 对视频文件做预测时，是否在屏幕上实时显示预测可视化结果(因加入了延迟处理，故显示结果不能反映真实的帧率)，支持值为0或1(默认值为0) |
| save_result | 是否将每帧的预测可视结果保存为视频文件，支持值为0或1(默认值为1) |
| save_dir | 保存可视化结果的路径, 默认值为"output" |
| key | 加密过程中产生的密钥信息，默认值为""表示加载的是未加密的模型 |
**注意：若系统无GUI，则不要将show_result设置为1。当使用摄像头预测时，按`ESC`键可关闭摄像头并推出预测程序。**
## 样例
......@@ -157,3 +174,18 @@ D:\images\xiaoduxiongn.jpeg
```
`--key`传入加密工具输出的密钥，例如`kLAl1qOs5uRbFt0/RrIDTZW2+tOf5bzvUIaHGF8lJ1c=`, 图片文件可视化预测结果会保存在`save_dir`参数设置的目录下。
### 样例四：(使用未加密的模型开启摄像头预测)
```shell
.\paddlex_inference\video_detector.exe --model_dir=D:\projects\inference_model --use_camera=1 --use_gpu=1 --save_dir=output
```
当`save_result`设置为1时，`可视化预测结果`会以视频文件的格式保存在`save_dir`参数设置的目录下。
### 样例五：(使用未加密的模型对视频文件做预测)
```shell
.\paddlex_inference\video_detector.exe --model_dir=D:\projects\inference_model --video_path=D:\projects\video_test.mp4 --use_gpu=1 --show_result=1 --save_dir=output
```
当`save_result`设置为1时，`可视化预测结果`会以视频文件的格式保存在`save_dir`参数设置的目录下。如果系统有GUI，通过将`show_result`设置为1在屏幕上观看可视化预测结果。
docs/deploy/server/python.md
浏览文件 @ c7b5a892
......@@ -30,6 +30,25 @@ image_list = ['xiaoduxiong_test_image/JPEGImages/WeChatIMG110.jpeg',
result = predictor.predict(image_list=image_list)
```
* 视频流预测
```
import cv2
import paddlex as pdx
predictor = pdx.deploy.Predictor('./inference_model')
cap = cv2.VideoCapture(0)
while cap.isOpened():
ret, frame = cap.read()
if ret:
result = predictor.predict(frame)
vis_img = pdx.det.visualize(frame, result, threshold=0.6, save_dir=None)
cv2.imshow('Xiaoduxiong', vis_img)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
else:
break
cap.release()
```
> 关于预测速度的说明：加载模型后前几张图片的预测速度会较慢，这是因为运行启动时涉及到内存显存初始化等步骤，通常在预测20-30张图片后模型的预测速度达到稳定。
## 预测性能对比
......
docs/examples/human_segmentation.md
浏览文件 @ c7b5a892
# 人像分割模型
本教程基于PaddleX核心分割模型实现人像分割，开放预训练模型和测试数据、支持视频流人像分割、提供模型Fine-tune到Paddle Lite移动端部署的全流程应用指南。
本教程基于PaddleX核心分割模型实现人像分割，开放预训练模型和测试数据、支持视频流人像分割、提供模型Fine-tune到Paddle Lite移动端及Nvidia Jeston嵌入式设备部署的全流程应用指南。
## 预训练模型和测试数据
#### 预训练模型
本案例开放了两个在大规模人像数据集上训练好的模型，以满足服务器端场景和移动端场景的需求。使用这些模型可以快速体验视频流人像分割，也可以部署到移动端进行实时人像分割，也可以用于完成模型Fine-tuning。
本案例开放了两个在大规模人像数据集上训练好的模型，以满足服务器端场景和移动端场景的需求。使用这些模型可以快速体验视频流人像分割，也可以部署到移动端或嵌入式设备进行实时人像分割，也可以用于完成模型Fine-tuning。
| 模型类型 | Checkpoint Parameter | Inference Model | Quant Inference Model | 备注 |
| --- | --- | --- | ---| --- |
......@@ -243,15 +243,17 @@ python quant_offline.py --model_dir output/best_model \
* `--save_dir`: 量化模型保存路径
* `--image_shape`: 网络输入图像大小（w, h）
## Paddle Lite移动端部署
## 推理部署
### Paddle Lite移动端部署
本案例将人像分割模型在移动端进行部署，部署流程展示如下，通用的移动端部署流程参见[Paddle Lite移动端部署](../../docs/deploy/paddlelite/android.md)。
### 1. 将PaddleX模型导出为inference模型
#### 1. 将PaddleX模型导出为inference模型
本案例使用humanseg_mobile_quant预训练模型，该模型已经是inference模型，不需要再执行模型导出步骤。如果不使用预训练模型，则执行上一章节`模型训练`中的`模型导出`将自己训练的模型导出为inference格式。
### 2. 将inference模型优化为Paddle Lite模型
#### 2. 将inference模型优化为Paddle Lite模型
下载并解压 [模型优化工具opt](https://bj.bcebos.com/paddlex/deploy/lite/model_optimize_tool_11cbd50e.tar.gz)，进入模型优化工具opt所在路径后，执行以下命令：
......@@ -273,16 +275,16 @@ python quant_offline.py --model_dir output/best_model \
更详细的使用方法和参数含义请参考: [使用opt转化模型](https://paddle-lite.readthedocs.io/zh/latest/user_guides/opt/opt_bin.html)
### 3. 移动端预测
#### 3. 移动端预测
PaddleX提供了基于PaddleX Android SDK的安卓demo，可供用户体验图像分类、目标检测、实例分割和语义分割，该demo位于`PaddleX/deploy/lite/android/demo`，用户将模型、配置文件和测试图片拷贝至该demo下进行预测。
#### 3.1 前置依赖
##### 3.1 前置依赖
* Android Studio 3.4
* Android手机或开发板
#### 3.2 拷贝模型、配置文件和测试图片
##### 3.2 拷贝模型、配置文件和测试图片
* 将Lite模型（.nb文件）拷贝到`PaddleX/deploy/lite/android/demo/app/src/main/assets/model/`目录下, 根据.nb文件的名字，修改文件`PaddleX/deploy/lite/android/demo/app/src/main/res/values/strings.xml`中的`MODEL_PATH_DEFAULT`；
......@@ -290,7 +292,7 @@ PaddleX提供了基于PaddleX Android SDK的安卓demo，可供用户体验图
* 将测试图片拷贝到`PaddleX/deploy/lite/android/demo/app/src/main/assets/images/`目录下，根据图片文件的名字，修改文件`PaddleX/deploy/lite/android/demo/app/src/main/res/values/strings.xml`中的`IMAGE_PATH_DEFAULT`。
#### 3.3 导入工程并运行
##### 3.3 导入工程并运行
* 打开Android Studio，在"Welcome to Android Studio"窗口点击"Open an existing Android Studio project"，在弹出的路径选择窗口中进入`PaddleX/deploy/lite/android/demo`目录，然后点击右下角的"Open"按钮，导入工程；
......@@ -303,3 +305,58 @@ PaddleX提供了基于PaddleX Android SDK的安卓demo，可供用户体验图
测试图片及其分割结果如下所示：
![](./images/beauty.png)
### Nvidia Jetson嵌入式设备部署
#### c++部署
step 1. 下载PaddleX源码
```
git clone https://github.com/PaddlePaddle/PaddleX
```
step 2. 将`PaddleX/examples/human_segmentation/deploy/cpp`下的`human_segmenter.cpp`和`CMakeList.txt`拷贝至`PaddleX/deploy/cpp`目录下，拷贝之前可以将`PaddleX/deploy/cpp`下原本的`CMakeList.txt`做好备份。
step 3. 按照[Nvidia Jetson开发板部署](../deploy/nvidia-jetson.md)中的Step2至Step3完成C++预测代码的编译。
step 4. 编译成功后，可执行程为`build/human_segmenter`，其主要命令参数说明如下：
| 参数 | 说明 |
| ---- | ---- |
| model_dir | 人像分割模型路径 |
| use_gpu | 是否使用 GPU 预测, 支持值为0或1(默认值为0)|
| gpu_id | GPU 设备ID, 默认值为0 |
| use_camera | 是否使用摄像头采集图片，支持值为0或1(默认值为0) |
| camera_id | 摄像头设备ID，默认值为0 |
| video_path | 视频文件的路径 |
| show_result | 对视频文件做预测时，是否在屏幕上实时显示预测可视化结果，支持值为0或1(默认值为0) |
| save_result | 是否将每帧的预测可视结果保存为视频文件，支持值为0或1(默认值为1) |
| image | 待预测的图片路径 |
| save_dir | 保存可视化结果的路径, 默认值为"output"|
step 5. 推理预测
用于部署推理的模型应为inference格式，本案例使用humanseg_server_inference预训练模型，该模型已经是inference模型，不需要再执行模型导出步骤。如果不使用预训练模型，则执行第2章节`模型训练`中的`模型导出`将自己训练的模型导出为inference格式。
* 使用未加密的模型对单张图片做预测
待测试图片位于本案例提供的测试数据中，可以替换成自己的图片。
```shell
./build/human_segmenter --model_dir=/path/to/humanseg_server_inference --image=/path/to/data/mini_supervisely/Images/pexels-photo-63776.png --use_gpu=1 --save_dir=output
```
* 使用未加密的模型开启摄像头做预测
```shell
./build/human_segmenter --model_dir=/path/to/humanseg_server_inference --use_camera=1 --save_result=1 --use_gpu=1 --save_dir=output
```
* 使用未加密的模型对视频文件做预测
待测试视频文件位于本案例提供的测试数据中，可以替换成自己的视频文件。
```shell
./build/human_segmenter --model_dir=/path/to/humanseg_server_inference --video_path=/path/to/data/mini_supervisely/video_test.mp4 --save_result=1 --use_gpu=1 --save_dir=output
```
docs/examples/meter_reader.md
浏览文件 @ c7b5a892
......@@ -245,7 +245,6 @@ step 5. 推理预测：
./build/meter_reader/meter_reader --det_model_dir=/path/to/det_inference_model --seg_model_dir=/path/to/seg_inference_model --use_camera=1 --use_gpu=1 --use_erode=1 --save_dir=output
```
## 模型训练
......
examples/human_segmentation/deploy/cpp/CMakeLists.txt 0 → 100644
浏览文件 @ c7b5a892
cmake_minimum_required(VERSION 3.0)
project(PaddleX CXX C)
option(WITH_MKL "Compile human_segmenter with MKL/OpenBlas support,defaultuseMKL." ON)
option(WITH_GPU "Compile human_segmenter with GPU/CPU, default use CPU." ON)
if (NOT WIN32)
option(WITH_STATIC_LIB "Compile human_segmenter with static/shared library, default use static." OFF)
else()
option(WITH_STATIC_LIB "Compile human_segmenter with static/shared library, default use static." ON)
endif()
option(WITH_TENSORRT "Compile human_segmenter with TensorRT." OFF)
option(WITH_ENCRYPTION "Compile human_segmenter with encryption tool." OFF)
SET(TENSORRT_DIR "" CACHE PATH "Location of libraries")
SET(PADDLE_DIR "" CACHE PATH "Location of libraries")
SET(OPENCV_DIR "" CACHE PATH "Location of libraries")
SET(ENCRYPTION_DIR"" CACHE PATH "Location of libraries")
SET(CUDA_LIB "" CACHE PATH "Location of libraries")
if (NOT WIN32)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
else()
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/paddlex_inference)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/paddlex_inference)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/paddlex_inference)
endif()
if (NOT WIN32)
SET(YAML_BUILD_TYPE ON CACHE BOOL "yaml build shared library.")
else()
SET(YAML_BUILD_TYPE OFF CACHE BOOL "yaml build shared library.")
endif()
include(cmake/yaml-cpp.cmake)
include_directories("${CMAKE_SOURCE_DIR}/")
include_directories("${CMAKE_CURRENT_BINARY_DIR}/ext/yaml-cpp/src/ext-yaml-cpp/include")
link_directories("${CMAKE_CURRENT_BINARY_DIR}/ext/yaml-cpp/lib")
macro(safe_set_static_flag)
foreach(flag_var
CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
if(${flag_var} MATCHES "/MD")
string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
endif(${flag_var} MATCHES "/MD")
endforeach(flag_var)
endmacro()
if (WITH_ENCRYPTION)
add_definitions( -DWITH_ENCRYPTION=${WITH_ENCRYPTION})
endif()
if (WITH_MKL)
ADD_DEFINITIONS(-DUSE_MKL)
endif()
if (NOT DEFINED PADDLE_DIR OR ${PADDLE_DIR} STREQUAL "")
message(FATAL_ERROR "please set PADDLE_DIR with -DPADDLE_DIR=/path/paddle_influence_dir")
endif()
if (NOT (${CMAKE_SYSTEM_PROCESSOR} STREQUAL "aarch64"))
if (NOT DEFINED OPENCV_DIR OR ${OPENCV_DIR} STREQUAL "")
message(FATAL_ERROR "please set OPENCV_DIR with -DOPENCV_DIR=/path/opencv")
endif()
endif()
include_directories("${CMAKE_SOURCE_DIR}/")
include_directories("${PADDLE_DIR}/")
include_directories("${PADDLE_DIR}/third_party/install/protobuf/include")
include_directories("${PADDLE_DIR}/third_party/install/glog/include")
include_directories("${PADDLE_DIR}/third_party/install/gflags/include")
include_directories("${PADDLE_DIR}/third_party/install/xxhash/include")
if (EXISTS "${PADDLE_DIR}/third_party/install/snappy/include")
include_directories("${PADDLE_DIR}/third_party/install/snappy/include")
endif()
if(EXISTS "${PADDLE_DIR}/third_party/install/snappystream/include")
include_directories("${PADDLE_DIR}/third_party/install/snappystream/include")
endif()
# zlib does not exist in 1.8.1
if (EXISTS "${PADDLE_DIR}/third_party/install/zlib/include")
include_directories("${PADDLE_DIR}/third_party/install/zlib/include")
endif()
include_directories("${PADDLE_DIR}/third_party/boost")
include_directories("${PADDLE_DIR}/third_party/eigen3")
if (EXISTS "${PADDLE_DIR}/third_party/install/snappy/lib")
link_directories("${PADDLE_DIR}/third_party/install/snappy/lib")
endif()
if(EXISTS "${PADDLE_DIR}/third_party/install/snappystream/lib")
link_directories("${PADDLE_DIR}/third_party/install/snappystream/lib")
endif()
if (EXISTS "${PADDLE_DIR}/third_party/install/zlib/lib")
link_directories("${PADDLE_DIR}/third_party/install/zlib/lib")
endif()
link_directories("${PADDLE_DIR}/third_party/install/protobuf/lib")
link_directories("${PADDLE_DIR}/third_party/install/glog/lib")
link_directories("${PADDLE_DIR}/third_party/install/gflags/lib")
link_directories("${PADDLE_DIR}/third_party/install/xxhash/lib")
link_directories("${PADDLE_DIR}/paddle/lib/")
link_directories("${CMAKE_CURRENT_BINARY_DIR}")
if (WIN32)
include_directories("${PADDLE_DIR}/paddle/fluid/inference")
include_directories("${PADDLE_DIR}/paddle/include")
link_directories("${PADDLE_DIR}/paddle/fluid/inference")
find_package(OpenCV REQUIRED PATHS ${OPENCV_DIR}/build/ NO_DEFAULT_PATH)
unset(OpenCV_DIR CACHE)
else ()
if (${CMAKE_SYSTEM_PROCESSOR} STREQUAL "aarch64") # x86_64 aarch64
set(OpenCV_INCLUDE_DIRS "/usr/include/opencv4")
file(GLOB OpenCV_LIBS /usr/lib/aarch64-linux-gnu/libopencv_*${CMAKE_SHARED_LIBRARY_SUFFIX})
message("OpenCV libs: ${OpenCV_LIBS}")
else()
find_package(OpenCV REQUIRED PATHS ${OPENCV_DIR}/share/OpenCV NO_DEFAULT_PATH)
endif()
include_directories("${PADDLE_DIR}/paddle/include")
link_directories("${PADDLE_DIR}/paddle/lib")
endif ()
include_directories(${OpenCV_INCLUDE_DIRS})
if (WIN32)
add_definitions("/DGOOGLE_GLOG_DLL_DECL=")
find_package(OpenMP REQUIRED)
if (OPENMP_FOUND)
message("OPENMP FOUND")
set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} ${OpenMP_C_FLAGS}")
set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} ${OpenMP_C_FLAGS}")
set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} ${OpenMP_CXX_FLAGS}")
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} ${OpenMP_CXX_FLAGS}")
endif()
set(CMAKE_C_FLAGS_DEBUG "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd")
set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT")
set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd")
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
if (WITH_STATIC_LIB)
safe_set_static_flag()
add_definitions(-DSTATIC_LIB)
endif()
else()
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -o2 -fopenmp -std=c++11")
set(CMAKE_STATIC_LIBRARY_PREFIX "")
endif()
if (WITH_GPU)
if (NOT DEFINED CUDA_LIB OR ${CUDA_LIB} STREQUAL "")
message(FATAL_ERROR "please set CUDA_LIB with -DCUDA_LIB=/path/cuda/lib64")
endif()
if (NOT WIN32)
if (NOT DEFINED CUDNN_LIB)
message(FATAL_ERROR "please set CUDNN_LIB with -DCUDNN_LIB=/path/cudnn/")
endif()
endif(NOT WIN32)
endif()
if (NOT WIN32)
if (WITH_TENSORRT AND WITH_GPU)
include_directories("${TENSORRT_DIR}/include")
link_directories("${TENSORRT_DIR}/lib")
endif()
endif(NOT WIN32)
if (NOT WIN32)
set(NGRAPH_PATH "${PADDLE_DIR}/third_party/install/ngraph")
if(EXISTS ${NGRAPH_PATH})
include(GNUInstallDirs)
include_directories("${NGRAPH_PATH}/include")
link_directories("${NGRAPH_PATH}/${CMAKE_INSTALL_LIBDIR}")
set(NGRAPH_LIB ${NGRAPH_PATH}/${CMAKE_INSTALL_LIBDIR}/libngraph${CMAKE_SHARED_LIBRARY_SUFFIX})
endif()
endif()
if(WITH_MKL)
include_directories("${PADDLE_DIR}/third_party/install/mklml/include")
if (WIN32)
set(MATH_LIB ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.lib
${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.lib)
else ()
set(MATH_LIB ${PADDLE_DIR}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
execute_process(COMMAND cp -r ${PADDLE_DIR}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} /usr/lib)
endif ()
set(MKLDNN_PATH "${PADDLE_DIR}/third_party/install/mkldnn")
if(EXISTS ${MKLDNN_PATH})
include_directories("${MKLDNN_PATH}/include")
if (WIN32)
set(MKLDNN_LIB ${MKLDNN_PATH}/lib/mkldnn.lib)
else ()
set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
endif ()
endif()
else()
set(MATH_LIB ${PADDLE_DIR}/third_party/install/openblas/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX})
endif()
if (WIN32)
if(EXISTS "${PADDLE_DIR}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}")
set(DEPS
${PADDLE_DIR}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
else()
set(DEPS
${PADDLE_DIR}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
endif()
endif()
if(WITH_STATIC_LIB)
set(DEPS
${PADDLE_DIR}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
else()
if (NOT WIN32)
set(DEPS
${PADDLE_DIR}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX})
else()
set(DEPS
${PADDLE_DIR}/paddle/lib/paddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX})
endif()
endif()
if (NOT WIN32)
set(DEPS ${DEPS}
${MATH_LIB} ${MKLDNN_LIB}
glog gflags protobuf z xxhash yaml-cpp
)
if(EXISTS "${PADDLE_DIR}/third_party/install/snappystream/lib")
set(DEPS ${DEPS} snappystream)
endif()
if (EXISTS "${PADDLE_DIR}/third_party/install/snappy/lib")
set(DEPS ${DEPS} snappy)
endif()
else()
set(DEPS ${DEPS}
${MATH_LIB} ${MKLDNN_LIB}
glog gflags_static libprotobuf xxhash libyaml-cppmt)
if (EXISTS "${PADDLE_DIR}/third_party/install/zlib/lib")
set(DEPS ${DEPS} zlibstatic)
endif()
set(DEPS ${DEPS} libcmt shlwapi)
if (EXISTS "${PADDLE_DIR}/third_party/install/snappy/lib")
set(DEPS ${DEPS} snappy)
endif()
if (EXISTS "${PADDLE_DIR}/third_party/install/snappystream/lib")
set(DEPS ${DEPS} snappystream)
endif()
endif(NOT WIN32)
if(WITH_GPU)
if(NOT WIN32)
if (WITH_TENSORRT)
set(DEPS ${DEPS} ${TENSORRT_DIR}/lib/libnvinfer${CMAKE_SHARED_LIBRARY_SUFFIX})
set(DEPS ${DEPS} ${TENSORRT_DIR}/lib/libnvinfer_plugin${CMAKE_SHARED_LIBRARY_SUFFIX})
endif()
set(DEPS ${DEPS} ${CUDA_LIB}/libcudart${CMAKE_SHARED_LIBRARY_SUFFIX})
set(DEPS ${DEPS} ${CUDNN_LIB}/libcudnn${CMAKE_SHARED_LIBRARY_SUFFIX})
else()
set(DEPS ${DEPS} ${CUDA_LIB}/cudart${CMAKE_STATIC_LIBRARY_SUFFIX} )
set(DEPS ${DEPS} ${CUDA_LIB}/cublas${CMAKE_STATIC_LIBRARY_SUFFIX} )
set(DEPS ${DEPS} ${CUDA_LIB}/cudnn${CMAKE_STATIC_LIBRARY_SUFFIX})
endif()
endif()
if(WITH_ENCRYPTION)
if(NOT WIN32)
include_directories("${ENCRYPTION_DIR}/include")
link_directories("${ENCRYPTION_DIR}/lib")
set(DEPS ${DEPS} ${ENCRYPTION_DIR}/lib/libpmodel-decrypt${CMAKE_SHARED_LIBRARY_SUFFIX})
else()
include_directories("${ENCRYPTION_DIR}/include")
link_directories("${ENCRYPTION_DIR}/lib")
set(DEPS ${DEPS} ${ENCRYPTION_DIR}/lib/pmodel-decrypt${CMAKE_STATIC_LIBRARY_SUFFIX})
endif()
endif()
if (NOT WIN32)
set(EXTERNAL_LIB "-ldl -lrt -lgomp -lz -lm -lpthread")
set(DEPS ${DEPS} ${EXTERNAL_LIB})
endif()
set(DEPS ${DEPS} ${OpenCV_LIBS})
add_library(paddlex_inference SHARED src/visualize src/transforms.cpp src/paddlex.cpp)
ADD_DEPENDENCIES(paddlex_inference ext-yaml-cpp)
target_link_libraries(paddlex_inference ${DEPS})
add_executable(human_segmenter human_segmenter.cpp src/transforms.cpp src/paddlex.cpp src/visualize.cpp)
ADD_DEPENDENCIES(human_segmenter ext-yaml-cpp)
target_link_libraries(human_segmenter ${DEPS})
if (WIN32 AND WITH_MKL)
add_custom_command(TARGET human_segmenter POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./mklml.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./libiomp5md.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./mkldnn.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./release/mklml.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./release/libiomp5md.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./release/mkldnn.dll
)
# for encryption
if (EXISTS "${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll")
add_custom_command(TARGET human_segmenter POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./pmodel-decrypt.dll
COMMAND ${CMAKE_COMMAND} -E copy_if_different ${ENCRYPTION_DIR}/lib/pmodel-decrypt.dll ./release/pmodel-decrypt.dll
)
endif()
endif()
file(COPY "${CMAKE_SOURCE_DIR}/include/paddlex/visualize.h"
DESTINATION "${CMAKE_BINARY_DIR}/include/" )
file(COPY "${CMAKE_SOURCE_DIR}/include/paddlex/config_parser.h"
DESTINATION "${CMAKE_BINARY_DIR}/include/" )
file(COPY "${CMAKE_SOURCE_DIR}/include/paddlex/transforms.h"
DESTINATION "${CMAKE_BINARY_DIR}/include/" )
file(COPY "${CMAKE_SOURCE_DIR}/include/paddlex/results.h"
DESTINATION "${CMAKE_BINARY_DIR}/include/" )
file(COPY "${CMAKE_SOURCE_DIR}/include/paddlex/paddlex.h"
DESTINATION "${CMAKE_BINARY_DIR}/include/" )
examples/human_segmentation/deploy/cpp/human_segmenter.cpp 0 → 100644
浏览文件 @ c7b5a892
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <glog/logging.h>
#include <omp.h>
#include <algorithm>
#include <chrono> // NOLINT
#include <fstream>
#include <iostream>
#include <string>
#include <vector>
#include <utility>
#include <ctime>
#include "include/paddlex/paddlex.h"
#include "include/paddlex/visualize.h"
#if defined(__arm__) || defined(__aarch64__)
#include <opencv2/videoio/legacy/constants_c.h>
#endif
using namespace std::chrono; // NOLINT
DEFINE_string(model_dir, "", "Path of inference model");
DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
DEFINE_bool(use_trt, false, "Infering with TensorRT");
DEFINE_int32(gpu_id, 0, "GPU card id");
DEFINE_string(key, "", "key of encryption");
DEFINE_string(image, "", "Path of test image file");
DEFINE_bool(use_camera, false, "Infering with Camera");
DEFINE_int32(camera_id, 0, "Camera id");
DEFINE_string(video_path, "", "Path of input video");
DEFINE_bool(show_result, false, "show the result of each frame with a window");
DEFINE_bool(save_result, true, "save the result of each frame to a video");
DEFINE_string(save_dir, "output", "Path to save visualized image");
int main(int argc, char** argv) {
// Parsing command-line
google::ParseCommandLineFlags(&argc, &argv, true);
if (FLAGS_model_dir == "") {
std::cerr << "--model_dir need to be defined" << std::endl;
return -1;
}
if (FLAGS_image == "" & FLAGS_video_path == ""
& FLAGS_use_camera == false) {
std::cerr << "--image or --video_path or --use_camera need to be defined"
<< std::endl;
return -1;
}
// Load model
PaddleX::Model model;
model.Init(FLAGS_model_dir,
FLAGS_use_gpu,
FLAGS_use_trt,
FLAGS_gpu_id,
FLAGS_key);
if (FLAGS_use_camera || FLAGS_video_path != "") {
// Open video
cv::VideoCapture capture;
if (FLAGS_use_camera) {
capture.open(FLAGS_camera_id);
if (!capture.isOpened()) {
std::cout << "Can not open the camera "
<< FLAGS_camera_id << "."
<< std::endl;
return -1;
}
} else {
capture.open(FLAGS_video_path);
if (!capture.isOpened()) {
std::cout << "Can not open the video "
<< FLAGS_video_path << "."
<< std::endl;
return -1;
}
}
// Create a VideoWriter
cv::VideoWriter video_out;
std::string video_out_path;
if (FLAGS_save_result) {
// Get video information: resolution, fps
int video_width = static_cast<int>(capture.get(CV_CAP_PROP_FRAME_WIDTH));
int video_height =
static_cast<int>(capture.get(CV_CAP_PROP_FRAME_HEIGHT));
int video_fps = static_cast<int>(capture.get(CV_CAP_PROP_FPS));
int video_fourcc;
if (FLAGS_use_camera) {
video_fourcc = 828601953;
} else {
video_fourcc = static_cast<int>(capture.get(CV_CAP_PROP_FOURCC));
}
if (FLAGS_use_camera) {
time_t now = time(0);
video_out_path =
PaddleX::generate_save_path(FLAGS_save_dir,
std::to_string(now) + ".mp4");
} else {
video_out_path =
PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_video_path);
}
video_out.open(video_out_path.c_str(),
video_fourcc,
video_fps,
cv::Size(video_width, video_height),
true);
if (!video_out.isOpened()) {
std::cout << "Create video writer failed!" << std::endl;
return -1;
}
}
PaddleX::SegResult result;
cv::Mat frame;
int key;
while (capture.read(frame)) {
if (FLAGS_show_result || FLAGS_use_camera) {
key = cv::waitKey(1);
// When pressing `ESC`, then exit program and result video is saved
if (key == 27) {
break;
}
} else if (frame.empty()) {
break;
}
// Begin to predict
model.predict(frame, &result);
// Visualize results
std::vector<uint8_t> label_map(result.label_map.data.begin(),
result.label_map.data.end());
cv::Mat mask(result.label_map.shape[0],
result.label_map.shape[1],
CV_8UC1,
label_map.data());
int rows = result.label_map.shape[0];
int cols = result.label_map.shape[1];
cv::Mat vis_img = frame.clone();
for (int i = 0; i < rows; i++) {
for (int j = 0; j < cols; j++) {
int category_id = static_cast<int>(mask.at<uchar>(i, j));
if (category_id == 0) {
vis_img.at<cv::Vec3b>(i, j)[0] = 255;
vis_img.at<cv::Vec3b>(i, j)[1] = 255;
vis_img.at<cv::Vec3b>(i, j)[2] = 255;
}
}
}
if (FLAGS_show_result || FLAGS_use_camera) {
cv::imshow("human_seg", vis_img);
}
if (FLAGS_save_result) {
video_out.write(vis_img);
}
result.clear();
}
capture.release();
if (FLAGS_save_result) {
video_out.release();
std::cout << "Visualized output saved as " << video_out_path << std::endl;
}
if (FLAGS_show_result || FLAGS_use_camera) {
cv::destroyAllWindows();
}
} else {
PaddleX::SegResult result;
cv::Mat im = cv::imread(FLAGS_image, 1);
model.predict(im, &result);
// Visualize results
std::vector<uint8_t> label_map(result.label_map.data.begin(),
result.label_map.data.end());
cv::Mat mask(result.label_map.shape[0],
result.label_map.shape[1],
CV_8UC1,
label_map.data());
int rows = result.label_map.shape[0];
int cols = result.label_map.shape[1];
cv::Mat vis_img = im.clone();
for (int i = 0; i < rows; i++) {
for (int j = 0; j < cols; j++) {
int category_id = static_cast<int>(mask.at<uchar>(i, j));
if (category_id == 0) {
vis_img.at<cv::Vec3b>(i, j)[0] = 255;
vis_img.at<cv::Vec3b>(i, j)[1] = 255;
vis_img.at<cv::Vec3b>(i, j)[2] = 255;
}
}
}
std::string save_path =
PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_image);
cv::imwrite(save_path, vis_img);
result.clear();
std::cout << "Visualized output saved as " << save_path << std::endl;
}
return 0;
}
examples/meter_reader/deploy/cpp/meter_reader/meter_reader.cpp
浏览文件 @ c7b5a892
......@@ -51,7 +51,8 @@ DEFINE_string(seg_key, "", "Segmenter model key of encryption");
DEFINE_string(image, "", "Path of test image file");
DEFINE_string(image_list, "", "Path of test image list file");
DEFINE_string(save_dir, "output", "Path to save visualized image");
DEFINE_double(score_threshold, 0.5, "Detected bbox whose score is lower than this threshlod is filtered");
DEFINE_double(score_threshold, 0.5,
"Detected bbox whose score is lower than this threshlod is filtered");
void predict(const cv::Mat &input_image, PaddleX::Model *det_model,
PaddleX::Model *seg_model, const std::string save_dir,
......@@ -207,7 +208,7 @@ int main(int argc, char **argv) {
return -1;
}
// 加载模型
// Load model
PaddleX::Model det_model;
det_model.Init(FLAGS_det_model_dir, FLAGS_use_gpu, FLAGS_use_trt,
FLAGS_gpu_id, FLAGS_det_key);
......
paddlex/cv/__init__.py
浏览文件 @ c7b5a892
......@@ -26,6 +26,7 @@ ResNet50 = models.ResNet50
DarkNet53 = models.DarkNet53
# detection
YOLOv3 = models.YOLOv3
PPYOLO = models.PPYOLO
#EAST = models.EAST
FasterRCNN = models.FasterRCNN
MaskRCNN = models.MaskRCNN
......
paddlex/cv/datasets/dataset.py
浏览文件 @ c7b5a892
......@@ -115,7 +115,7 @@ def multithread_reader(mapper,
while not isinstance(sample, EndSignal):
batch_data.append(sample)
if len(batch_data) == batch_size:
batch_data = generate_minibatch(batch_data)
batch_data = generate_minibatch(batch_data, mapper=mapper)
yield batch_data
batch_data = []
sample = out_queue.get()
......@@ -127,11 +127,11 @@ def multithread_reader(mapper,
else:
batch_data.append(sample)
if len(batch_data) == batch_size:
batch_data = generate_minibatch(batch_data)
batch_data = generate_minibatch(batch_data, mapper=mapper)
yield batch_data
batch_data = []
if not drop_last and len(batch_data) != 0:
batch_data = generate_minibatch(batch_data)
batch_data = generate_minibatch(batch_data, mapper=mapper)
yield batch_data
batch_data = []
......@@ -188,18 +188,21 @@ def multiprocess_reader(mapper,
else:
batch_data.append(sample)
if len(batch_data) == batch_size:
batch_data = generate_minibatch(batch_data)
batch_data = generate_minibatch(batch_data, mapper=mapper)
yield batch_data
batch_data = []
if len(batch_data) != 0 and not drop_last:
batch_data = generate_minibatch(batch_data)
batch_data = generate_minibatch(batch_data, mapper=mapper)
yield batch_data
batch_data = []
return queue_reader
def generate_minibatch(batch_data, label_padding_value=255):
def generate_minibatch(batch_data, label_padding_value=255, mapper=None):
if mapper is not None and mapper.batch_transforms is not None:
for op in mapper.batch_transforms:
batch_data = op(batch_data)
# if batch_size is 1, do not pad the image
if len(batch_data) == 1:
return batch_data
......
paddlex/cv/models/__init__.py
浏览文件 @ c7b5a892
......@@ -38,6 +38,7 @@ from .classifier import HRNet_W18
from .classifier import AlexNet
from .base import BaseAPI
from .yolo_v3 import YOLOv3
from .ppyolo import PPYOLO
from .faster_rcnn import FasterRCNN
from .mask_rcnn import MaskRCNN
from .unet import UNet
......
paddlex/cv/models/base.py
浏览文件 @ c7b5a892
......@@ -548,6 +548,8 @@ class BaseAPI:
current_save_dir = osp.join(save_dir, "epoch_{}".format(i + 1))
if not osp.isdir(current_save_dir):
os.makedirs(current_save_dir)
if hasattr(self, 'use_ema'):
self.exe.run(self.ema.apply_program)
if eval_dataset is not None and eval_dataset.num_samples > 0:
self.eval_metrics, self.eval_details = self.evaluate(
eval_dataset=eval_dataset,
......@@ -574,6 +576,8 @@ class BaseAPI:
log_writer.add_scalar(
"Metrics/Eval(Epoch): {}".format(k), v, i + 1)
self.save_model(save_dir=current_save_dir)
if hasattr(self, 'use_ema'):
self.exe.run(self.ema.restore_program)
time_eval_one_epoch = time.time() - eval_epoch_start_time
eval_epoch_start_time = time.time()
if best_model_epoch > 0:
......
paddlex/cv/models/ppyolo.py 0 → 100644
浏览文件 @ c7b5a892
此差异已折叠。
paddlex/cv/models/yolo_v3.py
浏览文件 @ c7b5a892
......@@ -15,21 +15,11 @@
from __future__ import absolute_import
import math
import tqdm
import os.path as osp
import numpy as np
from multiprocessing.pool import ThreadPool
import paddle.fluid as fluid
import paddlex.utils.logging as logging
import paddlex
import copy
from paddlex.cv.transforms import arrange_transforms
from paddlex.cv.datasets import generate_minibatch
from .base import BaseAPI
from collections import OrderedDict
from .utils.detection_eval import eval_results, bbox2out
from .ppyolo import PPYOLO
class YOLOv3(BaseAPI):
class YOLOv3(PPYOLO):
"""构建YOLOv3，并实现其训练、评估、预测和模型导出。
Args:
......@@ -65,12 +55,12 @@ class YOLOv3(BaseAPI):
320, 352, 384, 416, 448, 480, 512, 544, 576, 608
]):
self.init_params = locals()
super(YOLOv3, self).__init__('detector')
backbones = [
'DarkNet53', 'ResNet34', 'MobileNetV1', 'MobileNetV3_large'
]
assert backbone in backbones, "backbone should be one of {}".format(
backbones)
super(YOLOv3, self).__init__('detector')
self.backbone = backbone
self.num_classes = num_classes
self.anchors = anchors
......@@ -84,6 +74,16 @@ class YOLOv3(BaseAPI):
self.sync_bn = True
self.train_random_shapes = train_random_shapes
self.fixed_input_shape = None
self.use_fine_grained_loss = False
self.use_coord_conv = False
self.use_iou_aware = False
self.use_spp = False
self.use_drop_block = False
self.use_iou_loss = False
self.scale_x_y = 1.
self.use_matrix_nms = False
self.use_ema = False
self.with_dcn_v2 = False
def _get_backbone(self, backbone_name):
if backbone_name == 'DarkNet53':
......@@ -104,59 +104,6 @@ class YOLOv3(BaseAPI):
norm_type='sync_bn', model_name=model_name)
return backbone
def build_net(self, mode='train'):
model = paddlex.cv.nets.detection.YOLOv3(
backbone=self._get_backbone(self.backbone),
num_classes=self.num_classes,
mode=mode,
anchors=self.anchors,
anchor_masks=self.anchor_masks,
ignore_threshold=self.ignore_threshold,
label_smooth=self.label_smooth,
nms_score_threshold=self.nms_score_threshold,
nms_topk=self.nms_topk,
nms_keep_topk=self.nms_keep_topk,
nms_iou_threshold=self.nms_iou_threshold,
train_random_shapes=self.train_random_shapes,
fixed_input_shape=self.fixed_input_shape)
inputs = model.generate_inputs()
model_out = model.build_net(inputs)
outputs = OrderedDict([('bbox', model_out)])
if mode == 'train':
self.optimizer.minimize(model_out)
outputs = OrderedDict([('loss', model_out)])
return inputs, outputs
def default_optimizer(self, learning_rate, warmup_steps, warmup_start_lr,
lr_decay_epochs, lr_decay_gamma,
num_steps_each_epoch):
if warmup_steps > lr_decay_epochs[0] * num_steps_each_epoch:
logging.error(
"In function train(), parameters should satisfy: warmup_steps <= lr_decay_epochs[0]*num_samples_in_train_dataset",
exit=False)
logging.error(
"See this doc for more information: https://github.com/PaddlePaddle/PaddleX/blob/develop/docs/appendix/parameters.md#notice",
exit=False)
logging.error(
"warmup_steps should less than {} or lr_decay_epochs[0] greater than {}, please modify 'lr_decay_epochs' or 'warmup_steps' in train function".
format(lr_decay_epochs[0] * num_steps_each_epoch, warmup_steps
// num_steps_each_epoch))
boundaries = [b * num_steps_each_epoch for b in lr_decay_epochs]
values = [(lr_decay_gamma**i) * learning_rate
for i in range(len(lr_decay_epochs) + 1)]
lr_decay = fluid.layers.piecewise_decay(
boundaries=boundaries, values=values)
lr_warmup = fluid.layers.linear_lr_warmup(
learning_rate=lr_decay,
warmup_steps=warmup_steps,
start_lr=warmup_start_lr,
end_lr=learning_rate)
optimizer = fluid.optimizer.Momentum(
learning_rate=lr_warmup,
momentum=0.9,
regularization=fluid.regularizer.L2DecayRegularizer(5e-04))
return optimizer
def train(self,
num_epochs,
train_dataset,
......@@ -214,259 +161,11 @@ class YOLOv3(BaseAPI):
ValueError: 评估类型不在指定列表中。
ValueError: 模型从inference model进行加载。
"""
if not self.trainable:
raise ValueError("Model is not trainable from load_model method.")
if metric is None:
if isinstance(train_dataset, paddlex.datasets.CocoDetection):
metric = 'COCO'
elif isinstance(train_dataset, paddlex.datasets.VOCDetection) or \
isinstance(train_dataset, paddlex.datasets.EasyDataDet):
metric = 'VOC'
else:
raise ValueError(
"train_dataset should be datasets.VOCDetection or datasets.COCODetection or datasets.EasyDataDet."
)
assert metric in ['COCO', 'VOC'], "Metric only support 'VOC' or 'COCO'"
self.metric = metric
self.labels = train_dataset.labels
# 构建训练网络
if optimizer is None:
# 构建默认的优化策略
num_steps_each_epoch = train_dataset.num_samples // train_batch_size
optimizer = self.default_optimizer(
learning_rate=learning_rate,
warmup_steps=warmup_steps,
warmup_start_lr=warmup_start_lr,
lr_decay_epochs=lr_decay_epochs,
lr_decay_gamma=lr_decay_gamma,
num_steps_each_epoch=num_steps_each_epoch)
self.optimizer = optimizer
# 构建训练、验证、预测网络
self.build_program()
# 初始化网络权重
self.net_initialize(
startup_prog=fluid.default_startup_program(),
pretrain_weights=pretrain_weights,
save_dir=save_dir,
sensitivities_file=sensitivities_file,
eval_metric_loss=eval_metric_loss,
resume_checkpoint=resume_checkpoint)
# 训练
self.train_loop(
num_epochs=num_epochs,
train_dataset=train_dataset,
train_batch_size=train_batch_size,
eval_dataset=eval_dataset,
save_interval_epochs=save_interval_epochs,
log_interval_steps=log_interval_steps,
save_dir=save_dir,
use_vdl=use_vdl,
early_stop=early_stop,
early_stop_patience=early_stop_patience)
def evaluate(self,
eval_dataset,
batch_size=1,
epoch_id=None,
metric=None,
return_details=False):
"""评估。
Args:
eval_dataset (paddlex.datasets): 验证数据读取器。
batch_size (int): 验证数据批大小。默认为1。
epoch_id (int): 当前评估模型所在的训练轮数。
metric (bool): 训练过程中评估的方式，取值范围为['COCO', 'VOC']。默认为None，
根据用户传入的Dataset自动选择，如为VOCDetection，则metric为'VOC';
如为COCODetection，则metric为'COCO'。
return_details (bool): 是否返回详细信息。
Returns:
tuple (metrics, eval_details) | dict (metrics): 当return_details为True时，返回(metrics, eval_details)，
当return_details为False时，返回metrics。metrics为dict，包含关键字：'bbox_mmap'或者’bbox_map‘，
分别表示平均准确率平均值在各个IoU阈值下的结果取平均值的结果（mmAP）、平均准确率平均值（mAP）。
eval_details为dict，包含关键字：'bbox'，对应元素预测结果列表，每个预测结果由图像id、
预测框类别id、预测框坐标、预测框得分；’gt‘：真实标注框相关信息。
"""
arrange_transforms(
model_type=self.model_type,
class_name=self.__class__.__name__,
transforms=eval_dataset.transforms,
mode='eval')
if metric is None:
if hasattr(self, 'metric') and self.metric is not None:
metric = self.metric
else:
if isinstance(eval_dataset, paddlex.datasets.CocoDetection):
metric = 'COCO'
elif isinstance(eval_dataset, paddlex.datasets.VOCDetection):
metric = 'VOC'
else:
raise Exception(
"eval_dataset should be datasets.VOCDetection or datasets.COCODetection."
)
assert metric in ['COCO', 'VOC'], "Metric only support 'VOC' or 'COCO'"
total_steps = math.ceil(eval_dataset.num_samples * 1.0 / batch_size)
results = list()
data_generator = eval_dataset.generator(
batch_size=batch_size, drop_last=False)
logging.info(
"Start to evaluating(total_samples={}, total_steps={})...".format(
eval_dataset.num_samples, total_steps))
for step, data in tqdm.tqdm(
enumerate(data_generator()), total=total_steps):
images = np.array([d[0] for d in data])
im_sizes = np.array([d[1] for d in data])
feed_data = {'image': images, 'im_size': im_sizes}
with fluid.scope_guard(self.scope):
outputs = self.exe.run(
self.test_prog,
feed=[feed_data],
fetch_list=list(self.test_outputs.values()),
return_numpy=False)
res = {
'bbox': (np.array(outputs[0]),
outputs[0].recursive_sequence_lengths())
}
res_id = [np.array([d[2]]) for d in data]
res['im_id'] = (res_id, [])
if metric == 'VOC':
res_gt_box = [d[3].reshape(-1, 4) for d in data]
res_gt_label = [d[4].reshape(-1, 1) for d in data]
res_is_difficult = [d[5].reshape(-1, 1) for d in data]
res_id = [np.array([d[2]]) for d in data]
res['gt_box'] = (res_gt_box, [])
res['gt_label'] = (res_gt_label, [])
res['is_difficult'] = (res_is_difficult, [])
results.append(res)
logging.debug("[EVAL] Epoch={}, Step={}/{}".format(epoch_id, step +
1, total_steps))
box_ap_stats, eval_details = eval_results(
results, metric, eval_dataset.coco_gt, with_background=False)
evaluate_metrics = OrderedDict(
zip(['bbox_mmap'
if metric == 'COCO' else 'bbox_map'], box_ap_stats))
if return_details:
return evaluate_metrics, eval_details
return evaluate_metrics
@staticmethod
def _preprocess(images, transforms, model_type, class_name, thread_num=1):
arrange_transforms(
model_type=model_type,
class_name=class_name,
transforms=transforms,
mode='test')
pool = ThreadPool(thread_num)
batch_data = pool.map(transforms, images)
pool.close()
pool.join()
padding_batch = generate_minibatch(batch_data)
im = np.array(
[data[0] for data in padding_batch],
dtype=padding_batch[0][0].dtype)
im_size = np.array([data[1] for data in padding_batch], dtype=np.int32)
return im, im_size
@staticmethod
def _postprocess(res, batch_size, num_classes, labels):
clsid2catid = dict({i: i for i in range(num_classes)})
xywh_results = bbox2out([res], clsid2catid)
preds = [[] for i in range(batch_size)]
for xywh_res in xywh_results:
image_id = xywh_res['image_id']
del xywh_res['image_id']
xywh_res['category'] = labels[xywh_res['category_id']]
preds[image_id].append(xywh_res)
return preds
def predict(self, img_file, transforms=None):
"""预测。
Args:
img_file (str|np.ndarray): 预测图像路径，或者是解码后的排列格式为（H, W, C）且类型为float32且为BGR格式的数组。
transforms (paddlex.det.transforms): 数据预处理操作。
Returns:
list: 预测结果列表，每个预测结果由预测框类别标签、
预测框类别名称、预测框坐标(坐标格式为[xmin, ymin, w, h]）、
预测框得分组成。
"""
if transforms is None and not hasattr(self, 'test_transforms'):
raise Exception("transforms need to be defined, now is None.")
if isinstance(img_file, (str, np.ndarray)):
images = [img_file]
else:
raise Exception("img_file must be str/np.ndarray")
if transforms is None:
transforms = self.test_transforms
im, im_size = YOLOv3._preprocess(images, transforms, self.model_type,
self.__class__.__name__)
with fluid.scope_guard(self.scope):
result = self.exe.run(self.test_prog,
feed={'image': im,
'im_size': im_size},
fetch_list=list(self.test_outputs.values()),
return_numpy=False,
use_program_cache=True)
res = {
k: (np.array(v), v.recursive_sequence_lengths())
for k, v in zip(list(self.test_outputs.keys()), result)
}
res['im_id'] = (np.array(
[[i] for i in range(len(images))]).astype('int32'), [[]])
preds = YOLOv3._postprocess(res,
len(images), self.num_classes, self.labels)
return preds[0]
def batch_predict(self, img_file_list, transforms=None, thread_num=2):
"""预测。
Args:
img_file_list (list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是图像路径，也可以是解码后的排列格式为（H，W，C）
且类型为float32且为BGR格式的数组。
transforms (paddlex.det.transforms): 数据预处理操作。
thread_num (int): 并发执行各图像预处理时的线程数。
Returns:
list: 每个元素都为列表，表示各图像的预测结果。在各图像的预测结果列表中，每个预测结果由预测框类别标签、
预测框类别名称、预测框坐标(坐标格式为[xmin, ymin, w, h]）、
预测框得分组成。
"""
if transforms is None and not hasattr(self, 'test_transforms'):
raise Exception("transforms need to be defined, now is None.")
if not isinstance(img_file_list, (list, tuple)):
raise Exception("im_file must be list/tuple")
if transforms is None:
transforms = self.test_transforms
im, im_size = YOLOv3._preprocess(img_file_list, transforms,
self.model_type,
self.__class__.__name__, thread_num)
with fluid.scope_guard(self.scope):
result = self.exe.run(self.test_prog,
feed={'image': im,
'im_size': im_size},
fetch_list=list(self.test_outputs.values()),
return_numpy=False,
use_program_cache=True)
res = {
k: (np.array(v), v.recursive_sequence_lengths())
for k, v in zip(list(self.test_outputs.keys()), result)
}
res['im_id'] = (np.array(
[[i] for i in range(len(img_file_list))]).astype('int32'), [[]])
preds = YOLOv3._postprocess(res,
len(img_file_list), self.num_classes,
self.labels)
return preds
return super(YOLOv3, self).train(
num_epochs, train_dataset, train_batch_size, eval_dataset,
save_interval_epochs, log_interval_steps, save_dir,
pretrain_weights, optimizer, learning_rate, warmup_steps,
warmup_start_lr, lr_decay_epochs, lr_decay_gamma, metric, use_vdl,
sensitivities_file, eval_metric_loss, early_stop,
early_stop_patience, resume_checkpoint, False)
paddlex/cv/nets/detection/iou_aware.py 0 → 100644
浏览文件 @ c7b5a892
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from paddle import fluid
def _split_ioup(output, an_num, num_classes):
"""
Split new output feature map to output, predicted iou
along channel dimension
"""
ioup = fluid.layers.slice(output, axes=[1], starts=[0], ends=[an_num])
ioup = fluid.layers.sigmoid(ioup)
oriout = fluid.layers.slice(
output, axes=[1], starts=[an_num], ends=[an_num * (num_classes + 6)])
return (ioup, oriout)
def _de_sigmoid(x, eps=1e-7):
x = fluid.layers.clip(x, eps, 1 / eps)
one = fluid.layers.fill_constant(
shape=[1, 1, 1, 1], dtype=x.dtype, value=1.)
x = fluid.layers.clip((one / x - 1.0), eps, 1 / eps)
x = -fluid.layers.log(x)
return x
def _postprocess_output(ioup, output, an_num, num_classes, iou_aware_factor):
"""
post process output objectness score
"""
tensors = []
stride = output.shape[1] // an_num
for m in range(an_num):
tensors.append(
fluid.layers.slice(
output,
axes=[1],
starts=[stride * m + 0],
ends=[stride * m + 4]))
obj = fluid.layers.slice(
output, axes=[1], starts=[stride * m + 4], ends=[stride * m + 5])
obj = fluid.layers.sigmoid(obj)
ip = fluid.layers.slice(ioup, axes=[1], starts=[m], ends=[m + 1])
new_obj = fluid.layers.pow(obj, (
1 - iou_aware_factor)) * fluid.layers.pow(ip, iou_aware_factor)
new_obj = _de_sigmoid(new_obj)
tensors.append(new_obj)
tensors.append(
fluid.layers.slice(
output,
axes=[1],
starts=[stride * m + 5],
ends=[stride * m + 5 + num_classes]))
output = fluid.layers.concat(tensors, axis=1)
return output
def get_iou_aware_score(output, an_num, num_classes, iou_aware_factor):
ioup, output = _split_ioup(output, an_num, num_classes)
output = _postprocess_output(ioup, output, an_num, num_classes,
iou_aware_factor)
return output
paddlex/cv/nets/detection/loss/iou_aware_loss.py 0 → 100644
浏览文件 @ c7b5a892
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.initializer import NumpyArrayInitializer
from paddle import fluid
from .iou_loss import IouLoss
class IouAwareLoss(IouLoss):
"""
iou aware loss, see https://arxiv.org/abs/1912.05992
Args:
loss_weight (float): iou aware loss weight, default is 1.0
max_height (int): max height of input to support random shape input
max_width (int): max width of input to support random shape input
"""
def __init__(self, loss_weight=1.0, max_height=608, max_width=608):
super(IouAwareLoss, self).__init__(
loss_weight=loss_weight,
max_height=max_height,
max_width=max_width)
def __call__(self,
ioup,
x,
y,
w,
h,
tx,
ty,
tw,
th,
anchors,
downsample_ratio,
batch_size,
scale_x_y,
eps=1.e-10):
'''
Args:
ioup ([Variables]): the predicted iou
x | y | w | h ([Variables]): the output of yolov3 for encoded x|y|w|h
tx |ty |tw |th ([Variables]): the target of yolov3 for encoded x|y|w|h
anchors ([float]): list of anchors for current output layer
downsample_ratio (float): the downsample ratio for current output layer
batch_size (int): training batch size
eps (float): the decimal to prevent the denominator eqaul zero
'''
pred = self._bbox_transform(x, y, w, h, anchors, downsample_ratio,
batch_size, False, scale_x_y, eps)
gt = self._bbox_transform(tx, ty, tw, th, anchors, downsample_ratio,
batch_size, True, scale_x_y, eps)
iouk = self._iou(pred, gt, ioup, eps)
iouk.stop_gradient = True
loss_iou_aware = fluid.layers.cross_entropy(
ioup, iouk, soft_label=True)
loss_iou_aware = loss_iou_aware * self._loss_weight
return loss_iou_aware
paddlex/cv/nets/detection/loss/iou_loss.py 0 → 100644
浏览文件 @ c7b5a892
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.initializer import NumpyArrayInitializer
from paddle import fluid
class IouLoss(object):
"""
iou loss, see https://arxiv.org/abs/1908.03851
loss = 1.0 - iou * iou
Args:
loss_weight (float): iou loss weight, default is 2.5
max_height (int): max height of input to support random shape input
max_width (int): max width of input to support random shape input
ciou_term (bool): whether to add ciou_term
loss_square (bool): whether to square the iou term
"""
def __init__(self,
loss_weight=2.5,
max_height=608,
max_width=608,
ciou_term=False,
loss_square=True):
self._loss_weight = loss_weight
self._MAX_HI = max_height
self._MAX_WI = max_width
self.ciou_term = ciou_term
self.loss_square = loss_square
def __call__(self,
x,
y,
w,
h,
tx,
ty,
tw,
th,
anchors,
downsample_ratio,
batch_size,
scale_x_y=1.,
ioup=None,
eps=1.e-10):
'''
Args:
x | y | w | h ([Variables]): the output of yolov3 for encoded x|y|w|h
tx |ty |tw |th ([Variables]): the target of yolov3 for encoded x|y|w|h
anchors ([float]): list of anchors for current output layer
downsample_ratio (float): the downsample ratio for current output layer
batch_size (int): training batch size
eps (float): the decimal to prevent the denominator eqaul zero
'''
pred = self._bbox_transform(x, y, w, h, anchors, downsample_ratio,
batch_size, False, scale_x_y, eps)
gt = self._bbox_transform(tx, ty, tw, th, anchors, downsample_ratio,
batch_size, True, scale_x_y, eps)
iouk = self._iou(pred, gt, ioup, eps)
if self.loss_square:
loss_iou = 1. - iouk * iouk
else:
loss_iou = 1. - iouk
loss_iou = loss_iou * self._loss_weight
return loss_iou
def _iou(self, pred, gt, ioup=None, eps=1.e-10):
x1, y1, x2, y2 = pred
x1g, y1g, x2g, y2g = gt
x2 = fluid.layers.elementwise_max(x1, x2)
y2 = fluid.layers.elementwise_max(y1, y2)
xkis1 = fluid.layers.elementwise_max(x1, x1g)
ykis1 = fluid.layers.elementwise_max(y1, y1g)
xkis2 = fluid.layers.elementwise_min(x2, x2g)
ykis2 = fluid.layers.elementwise_min(y2, y2g)
intsctk = (xkis2 - xkis1) * (ykis2 - ykis1)
intsctk = intsctk * fluid.layers.greater_than(
xkis2, xkis1) * fluid.layers.greater_than(ykis2, ykis1)
unionk = (x2 - x1) * (y2 - y1) + (x2g - x1g) * (y2g - y1g
) - intsctk + eps
iouk = intsctk / unionk
if self.ciou_term:
ciou = self.get_ciou_term(pred, gt, iouk, eps)
iouk = iouk - ciou
return iouk
def get_ciou_term(self, pred, gt, iouk, eps):
x1, y1, x2, y2 = pred
x1g, y1g, x2g, y2g = gt
cx = (x1 + x2) / 2
cy = (y1 + y2) / 2
w = (x2 - x1) + fluid.layers.cast((x2 - x1) == 0, 'float32')
h = (y2 - y1) + fluid.layers.cast((y2 - y1) == 0, 'float32')
cxg = (x1g + x2g) / 2
cyg = (y1g + y2g) / 2
wg = x2g - x1g
hg = y2g - y1g
# A or B
xc1 = fluid.layers.elementwise_min(x1, x1g)
yc1 = fluid.layers.elementwise_min(y1, y1g)
xc2 = fluid.layers.elementwise_max(x2, x2g)
yc2 = fluid.layers.elementwise_max(y2, y2g)
# DIOU term
dist_intersection = (cx - cxg) * (cx - cxg) + (cy - cyg) * (cy - cyg)
dist_union = (xc2 - xc1) * (xc2 - xc1) + (yc2 - yc1) * (yc2 - yc1)
diou_term = (dist_intersection + eps) / (dist_union + eps)
# CIOU term
ciou_term = 0
ar_gt = wg / hg
ar_pred = w / h
arctan = fluid.layers.atan(ar_gt) - fluid.layers.atan(ar_pred)
ar_loss = 4. / np.pi / np.pi * arctan * arctan
alpha = ar_loss / (1 - iouk + ar_loss + eps)
alpha.stop_gradient = True
ciou_term = alpha * ar_loss
return diou_term + ciou_term
def _bbox_transform(self, dcx, dcy, dw, dh, anchors, downsample_ratio,
batch_size, is_gt, scale_x_y, eps):
grid_x = int(self._MAX_WI / downsample_ratio)
grid_y = int(self._MAX_HI / downsample_ratio)
an_num = len(anchors) // 2
shape_fmp = fluid.layers.shape(dcx)
shape_fmp.stop_gradient = True
# generate the grid_w x grid_h center of feature map
idx_i = np.array([[i for i in range(grid_x)]])
idx_j = np.array([[j for j in range(grid_y)]]).transpose()
gi_np = np.repeat(idx_i, grid_y, axis=0)
gi_np = np.reshape(gi_np, newshape=[1, 1, grid_y, grid_x])
gi_np = np.tile(gi_np, reps=[batch_size, an_num, 1, 1])
gj_np = np.repeat(idx_j, grid_x, axis=1)
gj_np = np.reshape(gj_np, newshape=[1, 1, grid_y, grid_x])
gj_np = np.tile(gj_np, reps=[batch_size, an_num, 1, 1])
gi_max = self._create_tensor_from_numpy(gi_np.astype(np.float32))
gi = fluid.layers.crop(x=gi_max, shape=dcx)
gi.stop_gradient = True
gj_max = self._create_tensor_from_numpy(gj_np.astype(np.float32))
gj = fluid.layers.crop(x=gj_max, shape=dcx)
gj.stop_gradient = True
grid_x_act = fluid.layers.cast(shape_fmp[3], dtype="float32")
grid_x_act.stop_gradient = True
grid_y_act = fluid.layers.cast(shape_fmp[2], dtype="float32")
grid_y_act.stop_gradient = True
if is_gt:
cx = fluid.layers.elementwise_add(dcx, gi) / grid_x_act
cx.gradient = True
cy = fluid.layers.elementwise_add(dcy, gj) / grid_y_act
cy.gradient = True
else:
dcx_sig = fluid.layers.sigmoid(dcx)
dcy_sig = fluid.layers.sigmoid(dcy)
if (abs(scale_x_y - 1.0) > eps):
dcx_sig = scale_x_y * dcx_sig - 0.5 * (scale_x_y - 1)
dcy_sig = scale_x_y * dcy_sig - 0.5 * (scale_x_y - 1)
cx = fluid.layers.elementwise_add(dcx_sig, gi) / grid_x_act
cy = fluid.layers.elementwise_add(dcy_sig, gj) / grid_y_act
anchor_w_ = [anchors[i] for i in range(0, len(anchors)) if i % 2 == 0]
anchor_w_np = np.array(anchor_w_)
anchor_w_np = np.reshape(anchor_w_np, newshape=[1, an_num, 1, 1])
anchor_w_np = np.tile(
anchor_w_np, reps=[batch_size, 1, grid_y, grid_x])
anchor_w_max = self._create_tensor_from_numpy(
anchor_w_np.astype(np.float32))
anchor_w = fluid.layers.crop(x=anchor_w_max, shape=dcx)
anchor_w.stop_gradient = True
anchor_h_ = [anchors[i] for i in range(0, len(anchors)) if i % 2 == 1]
anchor_h_np = np.array(anchor_h_)
anchor_h_np = np.reshape(anchor_h_np, newshape=[1, an_num, 1, 1])
anchor_h_np = np.tile(
anchor_h_np, reps=[batch_size, 1, grid_y, grid_x])
anchor_h_max = self._create_tensor_from_numpy(
anchor_h_np.astype(np.float32))
anchor_h = fluid.layers.crop(x=anchor_h_max, shape=dcx)
anchor_h.stop_gradient = True
# e^tw e^th
exp_dw = fluid.layers.exp(dw)
exp_dh = fluid.layers.exp(dh)
pw = fluid.layers.elementwise_mul(exp_dw, anchor_w) / \
(grid_x_act * downsample_ratio)
ph = fluid.layers.elementwise_mul(exp_dh, anchor_h) / \
(grid_y_act * downsample_ratio)
if is_gt:
exp_dw.stop_gradient = True
exp_dh.stop_gradient = True
pw.stop_gradient = True
ph.stop_gradient = True
x1 = cx - 0.5 * pw
y1 = cy - 0.5 * ph
x2 = cx + 0.5 * pw
y2 = cy + 0.5 * ph
if is_gt:
x1.stop_gradient = True
y1.stop_gradient = True
x2.stop_gradient = True
y2.stop_gradient = True
return x1, y1, x2, y2
def _create_tensor_from_numpy(self, numpy_array):
paddle_array = fluid.layers.create_parameter(
attr=ParamAttr(),
shape=numpy_array.shape,
dtype=numpy_array.dtype,
default_initializer=NumpyArrayInitializer(numpy_array))
paddle_array.stop_gradient = True
return paddle_array
paddlex/cv/nets/detection/loss/yolo_loss.py 0 → 100644
浏览文件 @ c7b5a892
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from paddle import fluid
try:
from collections.abc import Sequence
except Exception:
from collections import Sequence
class YOLOv3Loss(object):
"""
Combined loss for YOLOv3 network
Args:
batch_size (int): training batch size
ignore_thresh (float): threshold to ignore confidence loss
label_smooth (bool): whether to use label smoothing
use_fine_grained_loss (bool): whether use fine grained YOLOv3 loss
instead of fluid.layers.yolov3_loss
"""
def __init__(self,
batch_size=8,
ignore_thresh=0.7,
label_smooth=True,
use_fine_grained_loss=False,
iou_loss=None,
iou_aware_loss=None,
downsample=[32, 16, 8],
scale_x_y=1.,
match_score=False):
self._batch_size = batch_size
self._ignore_thresh = ignore_thresh
self._label_smooth = label_smooth
self._use_fine_grained_loss = use_fine_grained_loss
self._iou_loss = iou_loss
self._iou_aware_loss = iou_aware_loss
self.downsample = downsample
self.scale_x_y = scale_x_y
self.match_score = match_score
def __call__(self, outputs, gt_box, gt_label, gt_score, targets, anchors,
anchor_masks, mask_anchors, num_classes, prefix_name):
if self._use_fine_grained_loss:
return self._get_fine_grained_loss(
outputs, targets, gt_box, self._batch_size, num_classes,
mask_anchors, self._ignore_thresh)
else:
losses = []
for i, output in enumerate(outputs):
scale_x_y = self.scale_x_y if not isinstance(
self.scale_x_y, Sequence) else self.scale_x_y[i]
anchor_mask = anchor_masks[i]
loss = fluid.layers.yolov3_loss(
x=output,
gt_box=gt_box,
gt_label=gt_label,
gt_score=gt_score,
anchors=anchors,
anchor_mask=anchor_mask,
class_num=num_classes,
ignore_thresh=self._ignore_thresh,
downsample_ratio=self.downsample[i],
use_label_smooth=self._label_smooth,
scale_x_y=scale_x_y,
name=prefix_name + "yolo_loss" + str(i))
losses.append(fluid.layers.reduce_mean(loss))
return {'loss': sum(losses)}
def _get_fine_grained_loss(self,
outputs,
targets,
gt_box,
batch_size,
num_classes,
mask_anchors,
ignore_thresh,
eps=1.e-10):
"""
Calculate fine grained YOLOv3 loss
Args:
outputs ([Variables]): List of Variables, output of backbone stages
targets ([Variables]): List of Variables, The targets for yolo
loss calculatation.
gt_box (Variable): The ground-truth boudding boxes.
batch_size (int): The training batch size
num_classes (int): class num of dataset
mask_anchors ([[float]]): list of anchors in each output layer
ignore_thresh (float): prediction bbox overlap any gt_box greater
than ignore_thresh, objectness loss will
be ignored.
Returns:
Type: dict
xy_loss (Variable): YOLOv3 (x, y) coordinates loss
wh_loss (Variable): YOLOv3 (w, h) coordinates loss
obj_loss (Variable): YOLOv3 objectness score loss
cls_loss (Variable): YOLOv3 classification loss
"""
assert len(outputs) == len(targets), \
"YOLOv3 output layer number not equal target number"
loss_xys, loss_whs, loss_objs, loss_clss = [], [], [], []
if self._iou_loss is not None:
loss_ious = []
if self._iou_aware_loss is not None:
loss_iou_awares = []
for i, (output, target,
anchors) in enumerate(zip(outputs, targets, mask_anchors)):
downsample = self.downsample[i]
an_num = len(anchors) // 2
if self._iou_aware_loss is not None:
ioup, output = self._split_ioup(output, an_num, num_classes)
x, y, w, h, obj, cls = self._split_output(output, an_num,
num_classes)
tx, ty, tw, th, tscale, tobj, tcls = self._split_target(target)
tscale_tobj = tscale * tobj
scale_x_y = self.scale_x_y if not isinstance(
self.scale_x_y, Sequence) else self.scale_x_y[i]
if (abs(scale_x_y - 1.0) < eps):
loss_x = fluid.layers.sigmoid_cross_entropy_with_logits(
x, tx) * tscale_tobj
loss_x = fluid.layers.reduce_sum(loss_x, dim=[1, 2, 3])
loss_y = fluid.layers.sigmoid_cross_entropy_with_logits(
y, ty) * tscale_tobj
loss_y = fluid.layers.reduce_sum(loss_y, dim=[1, 2, 3])
else:
dx = scale_x_y * fluid.layers.sigmoid(x) - 0.5 * (scale_x_y -
1.0)
dy = scale_x_y * fluid.layers.sigmoid(y) - 0.5 * (scale_x_y -
1.0)
loss_x = fluid.layers.abs(dx - tx) * tscale_tobj
loss_x = fluid.layers.reduce_sum(loss_x, dim=[1, 2, 3])
loss_y = fluid.layers.abs(dy - ty) * tscale_tobj
loss_y = fluid.layers.reduce_sum(loss_y, dim=[1, 2, 3])
# NOTE: we refined loss function of (w, h) as L1Loss
loss_w = fluid.layers.abs(w - tw) * tscale_tobj
loss_w = fluid.layers.reduce_sum(loss_w, dim=[1, 2, 3])
loss_h = fluid.layers.abs(h - th) * tscale_tobj
loss_h = fluid.layers.reduce_sum(loss_h, dim=[1, 2, 3])
if self._iou_loss is not None:
loss_iou = self._iou_loss(x, y, w, h, tx, ty, tw, th, anchors,
downsample, self._batch_size,
scale_x_y)
loss_iou = loss_iou * tscale_tobj
loss_iou = fluid.layers.reduce_sum(loss_iou, dim=[1, 2, 3])
loss_ious.append(fluid.layers.reduce_mean(loss_iou))
if self._iou_aware_loss is not None:
loss_iou_aware = self._iou_aware_loss(
ioup, x, y, w, h, tx, ty, tw, th, anchors, downsample,
self._batch_size, scale_x_y)
loss_iou_aware = loss_iou_aware * tobj
loss_iou_aware = fluid.layers.reduce_sum(
loss_iou_aware, dim=[1, 2, 3])
loss_iou_awares.append(
fluid.layers.reduce_mean(loss_iou_aware))
loss_obj_pos, loss_obj_neg = self._calc_obj_loss(
output, obj, tobj, gt_box, self._batch_size, anchors,
num_classes, downsample, self._ignore_thresh, scale_x_y)
loss_cls = fluid.layers.sigmoid_cross_entropy_with_logits(cls,
tcls)
loss_cls = fluid.layers.elementwise_mul(loss_cls, tobj, axis=0)
loss_cls = fluid.layers.reduce_sum(loss_cls, dim=[1, 2, 3, 4])
loss_xys.append(fluid.layers.reduce_mean(loss_x + loss_y))
loss_whs.append(fluid.layers.reduce_mean(loss_w + loss_h))
loss_objs.append(
fluid.layers.reduce_mean(loss_obj_pos + loss_obj_neg))
loss_clss.append(fluid.layers.reduce_mean(loss_cls))
losses_all = {
"loss_xy": fluid.layers.sum(loss_xys),
"loss_wh": fluid.layers.sum(loss_whs),
"loss_obj": fluid.layers.sum(loss_objs),
"loss_cls": fluid.layers.sum(loss_clss),
}
if self._iou_loss is not None:
losses_all["loss_iou"] = fluid.layers.sum(loss_ious)
if self._iou_aware_loss is not None:
losses_all["loss_iou_aware"] = fluid.layers.sum(loss_iou_awares)
return losses_all
def _split_ioup(self, output, an_num, num_classes):
"""
Split output feature map to output, predicted iou
along channel dimension
"""
ioup = fluid.layers.slice(output, axes=[1], starts=[0], ends=[an_num])
ioup = fluid.layers.sigmoid(ioup)
oriout = fluid.layers.slice(
output,
axes=[1],
starts=[an_num],
ends=[an_num * (num_classes + 6)])
return (ioup, oriout)
def _split_output(self, output, an_num, num_classes):
"""
Split output feature map to x, y, w, h, objectness, classification
along channel dimension
"""
x = fluid.layers.strided_slice(
output,
axes=[1],
starts=[0],
ends=[output.shape[1]],
strides=[5 + num_classes])
y = fluid.layers.strided_slice(
output,
axes=[1],
starts=[1],
ends=[output.shape[1]],
strides=[5 + num_classes])
w = fluid.layers.strided_slice(
output,
axes=[1],
starts=[2],
ends=[output.shape[1]],
strides=[5 + num_classes])
h = fluid.layers.strided_slice(
output,
axes=[1],
starts=[3],
ends=[output.shape[1]],
strides=[5 + num_classes])
obj = fluid.layers.strided_slice(
output,
axes=[1],
starts=[4],
ends=[output.shape[1]],
strides=[5 + num_classes])
clss = []
stride = output.shape[1] // an_num
for m in range(an_num):
clss.append(
fluid.layers.slice(
output,
axes=[1],
starts=[stride * m + 5],
ends=[stride * m + 5 + num_classes]))
cls = fluid.layers.transpose(
fluid.layers.stack(
clss, axis=1), perm=[0, 1, 3, 4, 2])
return (x, y, w, h, obj, cls)
def _split_target(self, target):
"""
split target to x, y, w, h, objectness, classification
along dimension 2
target is in shape [N, an_num, 6 + class_num, H, W]
"""
tx = target[:, :, 0, :, :]
ty = target[:, :, 1, :, :]
tw = target[:, :, 2, :, :]
th = target[:, :, 3, :, :]
tscale = target[:, :, 4, :, :]
tobj = target[:, :, 5, :, :]
tcls = fluid.layers.transpose(
target[:, :, 6:, :, :], perm=[0, 1, 3, 4, 2])
tcls.stop_gradient = True
return (tx, ty, tw, th, tscale, tobj, tcls)
def _calc_obj_loss(self, output, obj, tobj, gt_box, batch_size, anchors,
num_classes, downsample, ignore_thresh, scale_x_y):
# A prediction bbox overlap any gt_bbox over ignore_thresh,
# objectness loss will be ignored, process as follows:
# 1. get pred bbox, which is same with YOLOv3 infer mode, use yolo_box here
# NOTE: img_size is set as 1.0 to get noramlized pred bbox
bbox, prob = fluid.layers.yolo_box(
x=output,
img_size=fluid.layers.ones(
shape=[batch_size, 2], dtype="int32"),
anchors=anchors,
class_num=num_classes,
conf_thresh=0.,
downsample_ratio=downsample,
clip_bbox=False,
scale_x_y=scale_x_y)
# 2. split pred bbox and gt bbox by sample, calculate IoU between pred bbox
# and gt bbox in each sample
if batch_size > 1:
preds = fluid.layers.split(bbox, batch_size, dim=0)
gts = fluid.layers.split(gt_box, batch_size, dim=0)
else:
preds = [bbox]
gts = [gt_box]
probs = [prob]
ious = []
for pred, gt in zip(preds, gts):
def box_xywh2xyxy(box):
x = box[:, 0]
y = box[:, 1]
w = box[:, 2]
h = box[:, 3]
return fluid.layers.stack(
[
x - w / 2.,
y - h / 2.,
x + w / 2.,
y + h / 2.,
], axis=1)
pred = fluid.layers.squeeze(pred, axes=[0])
gt = box_xywh2xyxy(fluid.layers.squeeze(gt, axes=[0]))
ious.append(fluid.layers.iou_similarity(pred, gt))
iou = fluid.layers.stack(ious, axis=0)
# 3. Get iou_mask by IoU between gt bbox and prediction bbox,
# Get obj_mask by tobj(holds gt_score), calculate objectness loss
max_iou = fluid.layers.reduce_max(iou, dim=-1)
iou_mask = fluid.layers.cast(max_iou <= ignore_thresh, dtype="float32")
if self.match_score:
max_prob = fluid.layers.reduce_max(prob, dim=-1)
iou_mask = iou_mask * fluid.layers.cast(
max_prob <= 0.25, dtype="float32")
output_shape = fluid.layers.shape(output)
an_num = len(anchors) // 2
iou_mask = fluid.layers.reshape(iou_mask, (-1, an_num, output_shape[2],
output_shape[3]))
iou_mask.stop_gradient = True
# NOTE: tobj holds gt_score, obj_mask holds object existence mask
obj_mask = fluid.layers.cast(tobj > 0., dtype="float32")
obj_mask.stop_gradient = True
# For positive objectness grids, objectness loss should be calculated
# For negative objectness grids, objectness loss is calculated only iou_mask == 1.0
loss_obj = fluid.layers.sigmoid_cross_entropy_with_logits(obj,
obj_mask)
loss_obj_pos = fluid.layers.reduce_sum(loss_obj * tobj, dim=[1, 2, 3])
loss_obj_neg = fluid.layers.reduce_sum(
loss_obj * (1.0 - obj_mask) * iou_mask, dim=[1, 2, 3])
return loss_obj_pos, loss_obj_neg
paddlex/cv/nets/detection/ops.py 0 → 100644
浏览文件 @ c7b5a892
# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
from numbers import Integral
import math
import six
import paddle
from paddle import fluid
def DropBlock(input, block_size, keep_prob, is_test):
if is_test:
return input
def CalculateGamma(input, block_size, keep_prob):
input_shape = fluid.layers.shape(input)
feat_shape_tmp = fluid.layers.slice(input_shape, [0], [3], [4])
feat_shape_tmp = fluid.layers.cast(feat_shape_tmp, dtype="float32")
feat_shape_t = fluid.layers.reshape(feat_shape_tmp, [1, 1, 1, 1])
feat_area = fluid.layers.pow(feat_shape_t, factor=2)
block_shape_t = fluid.layers.fill_constant(
shape=[1, 1, 1, 1], value=block_size, dtype='float32')
block_area = fluid.layers.pow(block_shape_t, factor=2)
useful_shape_t = feat_shape_t - block_shape_t + 1
useful_area = fluid.layers.pow(useful_shape_t, factor=2)
upper_t = feat_area * (1 - keep_prob)
bottom_t = block_area * useful_area
output = upper_t / bottom_t
return output
gamma = CalculateGamma(input, block_size=block_size, keep_prob=keep_prob)
input_shape = fluid.layers.shape(input)
p = fluid.layers.expand_as(gamma, input)
input_shape_tmp = fluid.layers.cast(input_shape, dtype="int64")
random_matrix = fluid.layers.uniform_random(
input_shape_tmp, dtype='float32', min=0.0, max=1.0)
one_zero_m = fluid.layers.less_than(random_matrix, p)
one_zero_m.stop_gradient = True
one_zero_m = fluid.layers.cast(one_zero_m, dtype="float32")
mask_flag = fluid.layers.pool2d(
one_zero_m,
pool_size=block_size,
pool_type='max',
pool_stride=1,
pool_padding=block_size // 2)
mask = 1.0 - mask_flag
elem_numel = fluid.layers.reduce_prod(input_shape)
elem_numel_m = fluid.layers.cast(elem_numel, dtype="float32")
elem_numel_m.stop_gradient = True
elem_sum = fluid.layers.reduce_sum(mask)
elem_sum_m = fluid.layers.cast(elem_sum, dtype="float32")
elem_sum_m.stop_gradient = True
output = input * mask * elem_numel_m / elem_sum_m
return output
class MultiClassNMS(object):
def __init__(self,
score_threshold=.05,
nms_top_k=-1,
keep_top_k=100,
nms_threshold=.5,
normalized=False,
nms_eta=1.0,
background_label=0):
super(MultiClassNMS, self).__init__()
self.score_threshold = score_threshold
self.nms_top_k = nms_top_k
self.keep_top_k = keep_top_k
self.nms_threshold = nms_threshold
self.normalized = normalized
self.nms_eta = nms_eta
self.background_label = background_label
def __call__(self, bboxes, scores):
return fluid.layers.multiclass_nms(
bboxes=bboxes,
scores=scores,
score_threshold=self.score_threshold,
nms_top_k=self.nms_top_k,
keep_top_k=self.keep_top_k,
normalized=self.normalized,
nms_threshold=self.nms_threshold,
nms_eta=self.nms_eta,
background_label=self.background_label)
class MatrixNMS(object):
def __init__(self,
score_threshold=.05,
post_threshold=.05,
nms_top_k=-1,
keep_top_k=100,
use_gaussian=False,
gaussian_sigma=2.,
normalized=False,
background_label=0):
super(MatrixNMS, self).__init__()
self.score_threshold = score_threshold
self.post_threshold = post_threshold
self.nms_top_k = nms_top_k
self.keep_top_k = keep_top_k
self.normalized = normalized
self.use_gaussian = use_gaussian
self.gaussian_sigma = gaussian_sigma
self.background_label = background_label
def __call__(self, bboxes, scores):
return paddle.fluid.layers.matrix_nms(
bboxes=bboxes,
scores=scores,
score_threshold=self.score_threshold,
post_threshold=self.post_threshold,
nms_top_k=self.nms_top_k,
keep_top_k=self.keep_top_k,
normalized=self.normalized,
use_gaussian=self.use_gaussian,
gaussian_sigma=self.gaussian_sigma,
background_label=self.background_label)
class MultiClassSoftNMS(object):
def __init__(
self,
score_threshold=0.01,
keep_top_k=300,
softnms_sigma=0.5,
normalized=False,
background_label=0, ):
super(MultiClassSoftNMS, self).__init__()
self.score_threshold = score_threshold
self.keep_top_k = keep_top_k
self.softnms_sigma = softnms_sigma
self.normalized = normalized
self.background_label = background_label
def __call__(self, bboxes, scores):
def create_tmp_var(program, name, dtype, shape, lod_level):
return program.current_block().create_var(
name=name, dtype=dtype, shape=shape, lod_level=lod_level)
def _soft_nms_for_cls(dets, sigma, thres):
"""soft_nms_for_cls"""
dets_final = []
while len(dets) > 0:
maxpos = np.argmax(dets[:, 0])
dets_final.append(dets[maxpos].copy())
ts, tx1, ty1, tx2, ty2 = dets[maxpos]
scores = dets[:, 0]
# force remove bbox at maxpos
scores[maxpos] = -1
x1 = dets[:, 1]
y1 = dets[:, 2]
x2 = dets[:, 3]
y2 = dets[:, 4]
eta = 0 if self.normalized else 1
areas = (x2 - x1 + eta) * (y2 - y1 + eta)
xx1 = np.maximum(tx1, x1)
yy1 = np.maximum(ty1, y1)
xx2 = np.minimum(tx2, x2)
yy2 = np.minimum(ty2, y2)
w = np.maximum(0.0, xx2 - xx1 + eta)
h = np.maximum(0.0, yy2 - yy1 + eta)
inter = w * h
ovr = inter / (areas + areas[maxpos] - inter)
weight = np.exp(-(ovr * ovr) / sigma)
scores = scores * weight
idx_keep = np.where(scores >= thres)
dets[:, 0] = scores
dets = dets[idx_keep]
dets_final = np.array(dets_final).reshape(-1, 5)
return dets_final
def _soft_nms(bboxes, scores):
class_nums = scores.shape[-1]
softnms_thres = self.score_threshold
softnms_sigma = self.softnms_sigma
keep_top_k = self.keep_top_k
cls_boxes = [[] for _ in range(class_nums)]
cls_ids = [[] for _ in range(class_nums)]
start_idx = 1 if self.background_label == 0 else 0
for j in range(start_idx, class_nums):
inds = np.where(scores[:, j] >= softnms_thres)[0]
scores_j = scores[inds, j]
rois_j = bboxes[inds, j, :] if len(
bboxes.shape) > 2 else bboxes[inds, :]
dets_j = np.hstack((scores_j[:, np.newaxis], rois_j)).astype(
np.float32, copy=False)
cls_rank = np.argsort(-dets_j[:, 0])
dets_j = dets_j[cls_rank]
cls_boxes[j] = _soft_nms_for_cls(
dets_j, sigma=softnms_sigma, thres=softnms_thres)
cls_ids[j] = np.array([j] * cls_boxes[j].shape[0]).reshape(-1,
1)
cls_boxes = np.vstack(cls_boxes[start_idx:])
cls_ids = np.vstack(cls_ids[start_idx:])
pred_result = np.hstack([cls_ids, cls_boxes])
# Limit to max_per_image detections **over all classes**
image_scores = cls_boxes[:, 0]
if len(image_scores) > keep_top_k:
image_thresh = np.sort(image_scores)[-keep_top_k]
keep = np.where(cls_boxes[:, 0] >= image_thresh)[0]
pred_result = pred_result[keep, :]
return pred_result
def _batch_softnms(bboxes, scores):
batch_offsets = bboxes.lod()
bboxes = np.array(bboxes)
scores = np.array(scores)
out_offsets = [0]
pred_res = []
if len(batch_offsets) > 0:
batch_offset = batch_offsets[0]
for i in range(len(batch_offset) - 1):
s, e = batch_offset[i], batch_offset[i + 1]
pred = _soft_nms(bboxes[s:e], scores[s:e])
out_offsets.append(pred.shape[0] + out_offsets[-1])
pred_res.append(pred)
else:
assert len(bboxes.shape) == 3
assert len(scores.shape) == 3
for i in range(bboxes.shape[0]):
pred = _soft_nms(bboxes[i], scores[i])
out_offsets.append(pred.shape[0] + out_offsets[-1])
pred_res.append(pred)
res = fluid.LoDTensor()
res.set_lod([out_offsets])
if len(pred_res) == 0:
pred_res = np.array([[1]], dtype=np.float32)
res.set(np.vstack(pred_res).astype(np.float32), fluid.CPUPlace())
return res
pred_result = create_tmp_var(
fluid.default_main_program(),
name='softnms_pred_result',
dtype='float32',
shape=[-1, 6],
lod_level=1)
fluid.layers.py_func(
func=_batch_softnms, x=[bboxes, scores], out=pred_result)
return pred_result
paddlex/cv/transforms/__init__.py
浏览文件 @ c7b5a892
......@@ -91,7 +91,10 @@ def arrange_transforms(model_type, class_name, transforms, mode='train'):
elif model_type == 'segmenter':
arrange_transform = seg_transforms.ArrangeSegmenter
elif model_type == 'detector':
arrange_name = 'Arrange{}'.format(class_name)
if class_name == "PPYOLO":
arrange_name = 'ArrangeYOLOv3'
else:
arrange_name = 'Arrange{}'.format(class_name)
arrange_transform = getattr(det_transforms, arrange_name)
else:
raise Exception("Unrecognized model type: {}".format(self.model_type))
......
paddlex/cv/transforms/cls_transforms.py
浏览文件 @ c7b5a892
......@@ -46,7 +46,7 @@ class Compose(ClsTransform):
raise ValueError('The length of transforms ' + \
'must be equal or larger than 1!')
self.transforms = transforms
self.batch_transforms = None
# 检查transforms里面的操作，目前支持PaddleX定义的或者是imgaug操作
for op in self.transforms:
if not isinstance(op, ClsTransform):
......
paddlex/cv/transforms/det_transforms.py
浏览文件 @ c7b5a892
......@@ -55,6 +55,7 @@ class Compose(DetTransform):
raise ValueError('The length of transforms ' + \
'must be equal or larger than 1!')
self.transforms = transforms
self.batch_transforms = None
self.use_mixup = False
for t in self.transforms:
if type(t).__name__ == 'MixupImage':
......@@ -1385,3 +1386,187 @@ class ComposedYOLOv3Transforms(Compose):
mean=mean, std=std)
]
super(ComposedYOLOv3Transforms, self).__init__(transforms)
class BatchRandomShape(DetTransform):
"""调整图像大小（resize）。
对batch数据中的每张图像全部resize到random_shapes中任意一个大小。
注意：当插值方式为“RANDOM”时，则随机选取一种插值方式进行resize。
Args:
random_shapes (list): resize大小选择列表。
默认为[320, 352, 384, 416, 448, 480, 512, 544, 576, 608]。
interp (str): resize的插值方式，与opencv的插值方式对应，取值范围为
['NEAREST', 'LINEAR', 'CUBIC', 'AREA', 'LANCZOS4', 'RANDOM']。默认为"RANDOM"。
Raises:
ValueError: 插值方式不在['NEAREST', 'LINEAR', 'CUBIC',
'AREA', 'LANCZOS4', 'RANDOM']中。
"""
# The interpolation mode
interp_dict = {
'NEAREST': cv2.INTER_NEAREST,
'LINEAR': cv2.INTER_LINEAR,
'CUBIC': cv2.INTER_CUBIC,
'AREA': cv2.INTER_AREA,
'LANCZOS4': cv2.INTER_LANCZOS4
}
def __init__(
self,
random_shapes=[320, 352, 384, 416, 448, 480, 512, 544, 576, 608],
interp='RANDOM'):
if not (interp == "RANDOM" or interp in self.interp_dict):
raise ValueError("interp should be one of {}".format(
self.interp_dict.keys()))
self.random_shapes = random_shapes
self.interp = interp
def __call__(self, batch_data):
"""
Args:
batch_data (list): 由与图像相关的各种信息组成的batch数据。
Returns:
list: 由与图像相关的各种信息组成的batch数据。
"""
shape = np.random.choice(self.random_shapes)
if self.interp == "RANDOM":
interp = random.choice(list(self.interp_dict.keys()))
else:
interp = self.interp
for data_id, data in enumerate(batch_data):
data_list = list(data)
im = data_list[0]
im = np.swapaxes(im, 1, 0)
im = np.swapaxes(im, 1, 2)
im = resize(im, shape, self.interp_dict[interp])
im = np.swapaxes(im, 1, 2)
im = np.swapaxes(im, 1, 0)
data_list[0] = im
batch_data[data_id] = tuple(data_list)
return batch_data
class GenerateYoloTarget(object):
"""生成YOLOv3的ground truth（真实标注框）在不同特征层的位置转换信息。
该transform只在YOLOv3计算细粒度loss时使用。
Args:
anchors (list|tuple): anchor框的宽度和高度。
anchor_masks (list|tuple): 在计算损失时，使用anchor的mask索引。
num_classes (int): 类别数。默认为80。
iou_thresh (float): iou阈值，当anchor和真实标注框的iou大于该阈值时，计入target。默认为1.0。
"""
def __init__(self,
anchors,
anchor_masks,
downsample_ratios,
num_classes=80,
iou_thresh=1.):
super(GenerateYoloTarget, self).__init__()
self.anchors = anchors
self.anchor_masks = anchor_masks
self.downsample_ratios = downsample_ratios
self.num_classes = num_classes
self.iou_thresh = iou_thresh
def __call__(self, batch_data):
"""
Args:
batch_data (list): 由与图像相关的各种信息组成的batch数据。
Returns:
list: 由与图像相关的各种信息组成的batch数据。
其中，每个数据新添加的字段为：
- target0 (np.ndarray): YOLOv3的ground truth在特征层0的位置转换信息，
形状为(特征层0的anchor数量, 6+类别数, 特征层0的h, 特征层0的w)。
- target1 (np.ndarray): YOLOv3的ground truth在特征层1的位置转换信息，
形状为(特征层1的anchor数量, 6+类别数, 特征层1的h, 特征层1的w)。
- ...
-targetn (np.ndarray): YOLOv3的ground truth在特征层n的位置转换信息，
形状为(特征层n的anchor数量, 6+类别数, 特征层n的h, 特征层n的w)。
n的是大小由anchor_masks的长度决定。
"""
im = batch_data[0][0]
h = im.shape[1]
w = im.shape[2]
an_hw = np.array(self.anchors) / np.array([[w, h]])
for data_id, data in enumerate(batch_data):
gt_bbox = data[1]
gt_class = data[2]
gt_score = data[3]
im_shape = data[4]
origin_h = float(im_shape[0])
origin_w = float(im_shape[1])
data_list = list(data)
for i, (
mask, downsample_ratio
) in enumerate(zip(self.anchor_masks, self.downsample_ratios)):
grid_h = int(h / downsample_ratio)
grid_w = int(w / downsample_ratio)
target = np.zeros(
(len(mask), 6 + self.num_classes, grid_h, grid_w),
dtype=np.float32)
for b in range(gt_bbox.shape[0]):
gx = gt_bbox[b, 0] / float(origin_w)
gy = gt_bbox[b, 1] / float(origin_h)
gw = gt_bbox[b, 2] / float(origin_w)
gh = gt_bbox[b, 3] / float(origin_h)
cls = gt_class[b]
score = gt_score[b]
if gw <= 0. or gh <= 0. or score <= 0.:
continue
# find best match anchor index
best_iou = 0.
best_idx = -1
for an_idx in range(an_hw.shape[0]):
iou = jaccard_overlap(
[0., 0., gw, gh],
[0., 0., an_hw[an_idx, 0], an_hw[an_idx, 1]])
if iou > best_iou:
best_iou = iou
best_idx = an_idx
gi = int(gx * grid_w)
gj = int(gy * grid_h)
# gtbox should be regresed in this layes if best match
# anchor index in anchor mask of this layer
if best_idx in mask:
best_n = mask.index(best_idx)
# x, y, w, h, scale
target[best_n, 0, gj, gi] = gx * grid_w - gi
target[best_n, 1, gj, gi] = gy * grid_h - gj
target[best_n, 2, gj, gi] = np.log(
gw * w / self.anchors[best_idx][0])
target[best_n, 3, gj, gi] = np.log(
gh * h / self.anchors[best_idx][1])
target[best_n, 4, gj, gi] = 2.0 - gw * gh
# objectness record gt_score
target[best_n, 5, gj, gi] = score
# classification
target[best_n, 6 + cls, gj, gi] = 1.
# For non-matched anchors, calculate the target if the iou
# between anchor and gt is larger than iou_thresh
if self.iou_thresh < 1:
for idx, mask_i in enumerate(mask):
if mask_i == best_idx: continue
iou = jaccard_overlap(
[0., 0., gw, gh],
[0., 0., an_hw[mask_i, 0], an_hw[mask_i, 1]])
if iou > self.iou_thresh:
# x, y, w, h, scale
target[idx, 0, gj, gi] = gx * grid_w - gi
target[idx, 1, gj, gi] = gy * grid_h - gj
target[idx, 2, gj, gi] = np.log(
gw * w / self.anchors[mask_i][0])
target[idx, 3, gj, gi] = np.log(
gh * h / self.anchors[mask_i][1])
target[idx, 4, gj, gi] = 2.0 - gw * gh
# objectness record gt_score
target[idx, 5, gj, gi] = score
# classification
target[idx, 6 + cls, gj, gi] = 1.
data_list.append(target)
batch_data[data_id] = tuple(data_list)
return batch_data
paddlex/cv/transforms/seg_transforms.py
浏览文件 @ c7b5a892
......@@ -49,6 +49,7 @@ class Compose(SegTransform):
raise ValueError('The length of transforms ' + \
'must be equal or larger than 1!')
self.transforms = transforms
self.batch_transforms = None
self.to_rgb = False
# 检查transforms里面的操作，目前支持PaddleX定义的或者是imgaug操作
for op in self.transforms:
......
paddlex/det.py
浏览文件 @ c7b5a892
......@@ -17,6 +17,7 @@ from . import cv
FasterRCNN = cv.models.FasterRCNN
YOLOv3 = cv.models.YOLOv3
PPYOLO = cv.models.PPYOLO
MaskRCNN = cv.models.MaskRCNN
transforms = cv.transforms.det_transforms
visualize = cv.models.utils.visualize.visualize_detection
......
tutorials/train/object_detection/ppyolo.py 0 → 100644
浏览文件 @ c7b5a892
# 环境变量配置，用于控制是否使用GPU
# 说明文档：https://paddlex.readthedocs.io/zh_CN/develop/appendix/parameters.html#gpu
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
from paddlex.det import transforms
import paddlex as pdx
# 下载和解压昆虫检测数据集
insect_dataset = 'https://bj.bcebos.com/paddlex/datasets/insect_det.tar.gz'
pdx.utils.download_and_decompress(insect_dataset, path='./')
# 定义训练和验证时的transforms
# API说明 https://paddlex.readthedocs.io/zh_CN/develop/apis/transforms/det_transforms.html
train_transforms = transforms.Compose([
transforms.MixupImage(mixup_epoch=250), transforms.RandomDistort(),
transforms.RandomExpand(), transforms.RandomCrop(), transforms.Resize(
target_size=608, interp='RANDOM'), transforms.RandomHorizontalFlip(),
transforms.Normalize()
])
eval_transforms = transforms.Compose([
transforms.Resize(
target_size=608, interp='CUBIC'), transforms.Normalize()
])
# 定义训练和验证所用的数据集
# API说明：https://paddlex.readthedocs.io/zh_CN/develop/apis/datasets.html#paddlex-datasets-vocdetection
train_dataset = pdx.datasets.VOCDetection(
data_dir='insect_det',
file_list='insect_det/train_list.txt',
label_list='insect_det/labels.txt',
transforms=train_transforms,
shuffle=True)
eval_dataset = pdx.datasets.VOCDetection(
data_dir='insect_det',
file_list='insect_det/val_list.txt',
label_list='insect_det/labels.txt',
transforms=eval_transforms)
# 初始化模型，并进行训练
# 可使用VisualDL查看训练指标，参考https://paddlex.readthedocs.io/zh_CN/develop/train/visualdl.html
num_classes = len(train_dataset.labels)
# API说明: https://paddlex.readthedocs.io/zh_CN/develop/apis/models/detection.html#paddlex-det-yolov3
model = pdx.det.PPYOLO(num_classes=num_classes)
# API说明: https://paddlex.readthedocs.io/zh_CN/develop/apis/models/detection.html#train
# 各参数介绍与调整说明：https://paddlex.readthedocs.io/zh_CN/develop/appendix/parameters.html
model.train(
num_epochs=270,
train_dataset=train_dataset,
train_batch_size=8,
eval_dataset=eval_dataset,
learning_rate=0.000125,
lr_decay_epochs=[210, 240],
save_dir='output/ppyolo',
use_vdl=True)
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或