docs: Update README and compilation guides for clarity and consistency, including path corrections and improved formatting. Add copyright notices to source files and adjust file permissions for several scripts and directories.
This commit is contained in:
dian.yuan
parent
f960c5030d
commit
bd891a96dd
136 changed files with 14413 additions and 9399 deletions
|
|
@ -1,160 +1,160 @@
|
|||
# retinaface
|
||||
|
||||
## 1.Overview
|
||||
|
||||
|
||||
## 2.Model Download
|
||||
|
||||
- **Open Source model**
|
||||
|
||||
- **Open Source projects:**
|
||||
|
||||
- **Export Model Step:**
|
||||
|
||||
- **Install ultralytics**
|
||||
|
||||
pip install torch==2.4.1
|
||||
|
||||
pip install torchvision==0.19.1
|
||||
|
||||
pip install ultralytics==8.3.0
|
||||
|
||||
- **Download weights**
|
||||
|
||||
|
||||
|
||||
- **Export Model**
|
||||
|
||||
|
||||
|
||||
|
||||
- **Exported Model**
|
||||
|
||||
link to amlogic server( **onnx model or quantized tflite**)
|
||||
|
||||
|
||||
|
||||
## 3. Model Conversion
|
||||
|
||||
```
|
||||
cd model
|
||||
Usage: ./adla_covnert.sh model_path adla_tookkit_path target_platform
|
||||
|
||||
example
|
||||
|
||||
```
|
||||
|
||||
| Parameter | Discription |
|
||||
| ----------------- | ------------------------------------------------------------ |
|
||||
| model_path | onnx model path |
|
||||
| adla_tookkit_path | path to adla_toolkit |
|
||||
| target_platform | Specify target platform. for A311D2 : PRODUCT_PID0XA003. for S905X5: PRODUCT_PID0XA005 |
|
||||
|
||||
|
||||
|
||||
## 4. Demo Run
|
||||
|
||||
### CPP
|
||||
|
||||
#### 1. Compile
|
||||
|
||||
**Prerequisites:**
|
||||
- Android NDK (r25e recommended)
|
||||
- `ANDROID_NDK_PATH` environment variable set
|
||||
|
||||
**Build:**
|
||||
```bash
|
||||
# Build for arm64-v8a
|
||||
cd examples/retinaface/cpp
|
||||
AMLNN_HOME=/path/to/amlnn-toolkit ./build-android.sh -a arm64-v8a
|
||||
```
|
||||
|
||||
The executable will be generated at `build/android/retinaface_demo` (Note: executable name may vary, verify in build folder).
|
||||
|
||||
#### 2. Run
|
||||
|
||||
```bash
|
||||
# Push executable to device
|
||||
adb push build/android/retinaface_demo /data/local/tmp/
|
||||
adb push model/RetinaFace_int8_A311D2.adla /data/local/tmp/
|
||||
adb push imgs /data/local/tmp/
|
||||
|
||||
# Run on device
|
||||
adb shell
|
||||
cd /data/local/tmp
|
||||
chmod +x retinaface_demo
|
||||
export LD_LIBRARY_PATH=/vendor/lib64 or (/vendor/lib)
|
||||
|
||||
# Usage: ./retinaface_demo <model_path> <image_dir>
|
||||
./retinaface_demo RetinaFace_int8_A311D2.adla ./imgs
|
||||
```
|
||||
|
||||
**Note:** Replace `RetinaFace_int8_A311D2.adla` with your actual model file path.
|
||||
|
||||
### Python
|
||||
|
||||
**Prerequisites:**
|
||||
- Python 3.10
|
||||
- Required packages: `numpy`, `opencv-python`, `amlnnlite`
|
||||
|
||||
**Install dependencies:**
|
||||
```bash
|
||||
pip install numpy opencv-python amlnnlite-1.0.0-cp310-cp310-linux_aarch64.whl
|
||||
```
|
||||
|
||||
**Run on device:**
|
||||
```bash
|
||||
python RetinaFace.py \
|
||||
--model-path ./RetinaFace_int8_A311D2.adla \
|
||||
--image-dir ./imgs \
|
||||
--run-cycles 1 \
|
||||
--loglevel INFO
|
||||
```
|
||||
|
||||
Argument Descriptions:
|
||||
| Argument | Description |
|
||||
| ----------------- | ------------------------------------------------------------ |
|
||||
| --board-work-path | Work path on board, default is /data/local/tmp |
|
||||
| --model-path | path to .adla model |
|
||||
| --image-dir | Directory containing test images |
|
||||
| --run-cycles | Number of inference cycles, default is 1 |
|
||||
| --loglevel | Logging level: DEBUG / INFO / WARNING / ERROR, default is WARNING |
|
||||
|
||||
The script will automatically process all image files (`.jpg`, `.jpeg`, `.png`, `.bmp`) in the current directory and save results to a `{model_name}_result` folder.
|
||||
|
||||
## 5.Results
|
||||
|
||||
**Performance Feedback**
|
||||
|
||||
By setting the loglevel to INFO, the program provides real-time performance metrics upon completion. The console log will display essential hardware and execution details, including:
|
||||
- Hardware Information: System and ADLA library versions.
|
||||
- Model Overview: Basic input/output configurations.
|
||||
- NPU Metrics: Total inference time (latency) and total DRAM bandwidth consumption.
|
||||
|
||||
**Detection Output**
|
||||
|
||||
The program will print the detection count. The output images, featuring bounding boxes and five facial landmarks (eyes, nose, and mouth corners), will be saved to the `{model_name}_result` folder.
|
||||
|
||||
|
||||
You can pull the result folder back to view it:
|
||||
```bash
|
||||
adb pull /data/local/tmp/RetinaFace_int8_A311D2_result
|
||||
```
|
||||
|
||||
|
||||
|
||||
**Profiling Visualization**
|
||||
|
||||
After a successful run of the Python demo, a folder named after the model (e.g., `{model_name}`) will be generated in the script directory. This folder contains 5 HTML files that provide a visual and detailed breakdown of per-layer performance:
|
||||
- `hard_op_chart.html` & `soft_op_chart.html`: Hardware/Software op execution details.
|
||||
- `dram_rd_chart.html` & `dram_wr_chart.html`: Bandwidth read/write distribution.
|
||||
- `pie_charts_distribution.html`: Overall resource allocation.
|
||||
|
||||
You can pull the result folder back to view it:
|
||||
```bash
|
||||
adb pull /data/local/tmp/RetinaFace_int8_A311D2
|
||||
```
|
||||
|
||||
Taking hard_op_chart.html as an example (shown below), each layer's ADLA operator name includes parentheses containing the index of the corresponding quantized .tflite layer(s); by default, these indices are suppressed, and operators are labeled generically as "hardware" or "software" without numerical suffixes.
|
||||
|
||||
|
||||
# retinaface
|
||||
|
||||
## 1.Overview
|
||||
|
||||
|
||||
## 2.Model Download
|
||||
|
||||
- **Open Source model**
|
||||
|
||||
- **Open Source projects:**
|
||||
|
||||
- **Export Model Step:**
|
||||
|
||||
- **Install ultralytics**
|
||||
|
||||
pip install torch==2.4.1
|
||||
|
||||
pip install torchvision==0.19.1
|
||||
|
||||
pip install ultralytics==8.3.0
|
||||
|
||||
- **Download weights**
|
||||
|
||||
|
||||
|
||||
- **Export Model**
|
||||
|
||||
|
||||
|
||||
|
||||
- **Exported Model**
|
||||
|
||||
link to amlogic server( **onnx model or quantized tflite**)
|
||||
|
||||
|
||||
|
||||
## 3. Model Conversion
|
||||
|
||||
```
|
||||
cd model
|
||||
Usage: ./adla_convert.sh model_path adla_toolkit_path target_platform
|
||||
|
||||
example
|
||||
|
||||
```
|
||||
|
||||
| Parameter | Description |
|
||||
| ----------------- | ------------------------------------------------------------ |
|
||||
| model_path | onnx model path |
|
||||
| adla_toolkit_path | path to adla_toolkit |
|
||||
| target_platform | Specify target platform. for A311D2 : PRODUCT_PID0XA003. for S905X5: PRODUCT_PID0XA005 |
|
||||
|
||||
|
||||
|
||||
## 4. Demo Run
|
||||
|
||||
### CPP
|
||||
|
||||
#### 1. Compile
|
||||
|
||||
**Prerequisites:**
|
||||
- Android NDK (r25e recommended)
|
||||
- `ANDROID_NDK_PATH` environment variable set
|
||||
|
||||
**Build:**
|
||||
```bash
|
||||
# Build for arm64-v8a
|
||||
cd examples/retinaface/cpp
|
||||
AMLNN_HOME=/path/to/amlnn-toolkit ./build-android.sh -a arm64-v8a
|
||||
```
|
||||
|
||||
The executable will be generated at `build/android/retinaface_demo` (Note: executable name may vary, verify in build folder).
|
||||
|
||||
#### 2. Run
|
||||
|
||||
```bash
|
||||
# Push executable to device
|
||||
adb push build/android/retinaface_demo /data/local/tmp/
|
||||
adb push model/RetinaFace_int8_A311D2.adla /data/local/tmp/
|
||||
adb push imgs /data/local/tmp/
|
||||
|
||||
# Run on device
|
||||
adb shell
|
||||
cd /data/local/tmp
|
||||
chmod +x retinaface_demo
|
||||
export LD_LIBRARY_PATH=/vendor/lib64 or (/vendor/lib)
|
||||
|
||||
# Usage: ./retinaface_demo <model_path> <image_dir>
|
||||
./retinaface_demo RetinaFace_int8_A311D2.adla ./imgs
|
||||
```
|
||||
|
||||
**Note:** Replace `RetinaFace_int8_A311D2.adla` with your actual model file path.
|
||||
|
||||
### Python
|
||||
|
||||
**Prerequisites:**
|
||||
- Python 3.10
|
||||
- Required packages: `numpy`, `opencv-python`, `amlnnlite`
|
||||
|
||||
**Install dependencies:**
|
||||
```bash
|
||||
pip install numpy opencv-python amlnnlite-1.0.0-cp310-cp310-linux_aarch64.whl
|
||||
```
|
||||
|
||||
**Run on device:**
|
||||
```bash
|
||||
python RetinaFace.py \
|
||||
--model-path ./RetinaFace_int8_A311D2.adla \
|
||||
--image-dir ./imgs \
|
||||
--run-cycles 1 \
|
||||
--loglevel INFO
|
||||
```
|
||||
|
||||
Argument Descriptions:
|
||||
| Argument | Description |
|
||||
| ----------------- | ------------------------------------------------------------ |
|
||||
| --board-work-path | Work path on board, default is /data/local/tmp |
|
||||
| --model-path | path to .adla model |
|
||||
| --image-dir | Directory containing test images |
|
||||
| --run-cycles | Number of inference cycles, default is 1 |
|
||||
| --loglevel | Logging level: DEBUG / INFO / WARNING / ERROR, default is WARNING |
|
||||
|
||||
The script will automatically process all image files (`.jpg`, `.jpeg`, `.png`, `.bmp`) in the current directory and save results to a `{model_name}_result` folder.
|
||||
|
||||
## 5.Results
|
||||
|
||||
**Performance Feedback**
|
||||
|
||||
By setting the loglevel to INFO, the program provides real-time performance metrics upon completion. The console log will display essential hardware and execution details, including:
|
||||
- Hardware Information: System and ADLA library versions.
|
||||
- Model Overview: Basic input/output configurations.
|
||||
- NPU Metrics: Total inference time (latency) and total DRAM bandwidth consumption.
|
||||
|
||||
**Detection Output**
|
||||
|
||||
The program will print the detection count. The output images, featuring bounding boxes and five facial landmarks (eyes, nose, and mouth corners), will be saved to the `{model_name}_result` folder.
|
||||
|
||||
|
||||
You can pull the result folder back to view it:
|
||||
```bash
|
||||
adb pull /data/local/tmp/RetinaFace_int8_A311D2_result
|
||||
```
|
||||
|
||||
|
||||
|
||||
**Profiling Visualization**
|
||||
|
||||
After a successful run of the Python demo, a folder named after the model (e.g., `{model_name}`) will be generated in the script directory. This folder contains 5 HTML files that provide a visual and detailed breakdown of per-layer performance:
|
||||
- `hard_op_chart.html` & `soft_op_chart.html`: Hardware/Software op execution details.
|
||||
- `dram_rd_chart.html` & `dram_wr_chart.html`: Bandwidth read/write distribution.
|
||||
- `pie_charts_distribution.html`: Overall resource allocation.
|
||||
|
||||
You can pull the result folder back to view it:
|
||||
```bash
|
||||
adb pull /data/local/tmp/RetinaFace_int8_A311D2
|
||||
```
|
||||
|
||||
Taking hard_op_chart.html as an example (shown below), each layer's ADLA operator name includes parentheses containing the index of the corresponding quantized .tflite layer(s); by default, these indices are suppressed, and operators are labeled generically as "hardware" or "software" without numerical suffixes.
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -1,35 +1,35 @@
|
|||
cmake_minimum_required(VERSION 3.10...3.27)
|
||||
project(retinaface_demo)
|
||||
|
||||
set(CMAKE_CXX_STANDARD 17)
|
||||
|
||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/../../../../cmake")
|
||||
find_package(AMLNN REQUIRED)
|
||||
include_directories(${AMLNN_INCLUDE_DIR})
|
||||
link_directories(${AMLNN_LIBRARY_DIR})
|
||||
|
||||
include_directories(${CMAKE_SOURCE_DIR}/../../../../common)
|
||||
|
||||
# Set dependency path
|
||||
set(3RDPARTY_DIR "${CMAKE_SOURCE_DIR}/../../../../dependency")
|
||||
|
||||
if(CMAKE_SYSTEM_NAME STREQUAL "Android")
|
||||
# Android needs log
|
||||
link_libraries(log)
|
||||
endif()
|
||||
|
||||
# Find OpenCV
|
||||
message(STATUS "OpenCV_DIR: ${OpenCV_DIR}")
|
||||
find_package(OpenCV REQUIRED)
|
||||
include_directories(${OpenCV_INCLUDE_DIRS})
|
||||
|
||||
add_executable(retinaface_demo
|
||||
main.cpp
|
||||
postprocess.cpp
|
||||
${CMAKE_SOURCE_DIR}/../../../../common/model_loader.cpp
|
||||
)
|
||||
|
||||
target_link_libraries(retinaface_demo
|
||||
${OpenCV_LIBS}
|
||||
${AMLNN_LIBRARY}
|
||||
cmake_minimum_required(VERSION 3.10...3.27)
|
||||
project(retinaface_demo)
|
||||
|
||||
set(CMAKE_CXX_STANDARD 17)
|
||||
|
||||
list(APPEND CMAKE_MODULE_PATH "${CMAKE_SOURCE_DIR}/../../../../cmake")
|
||||
find_package(AMLNN REQUIRED)
|
||||
include_directories(${AMLNN_INCLUDE_DIR})
|
||||
link_directories(${AMLNN_LIBRARY_DIR})
|
||||
|
||||
include_directories(${CMAKE_SOURCE_DIR}/../../../../common)
|
||||
|
||||
# Set dependency path
|
||||
set(3RDPARTY_DIR "${CMAKE_SOURCE_DIR}/../../../../dependency")
|
||||
|
||||
if(CMAKE_SYSTEM_NAME STREQUAL "Android")
|
||||
# Android needs log
|
||||
link_libraries(log)
|
||||
endif()
|
||||
|
||||
# Find OpenCV
|
||||
message(STATUS "OpenCV_DIR: ${OpenCV_DIR}")
|
||||
find_package(OpenCV REQUIRED)
|
||||
include_directories(${OpenCV_INCLUDE_DIRS})
|
||||
|
||||
add_executable(retinaface_demo
|
||||
main.cpp
|
||||
postprocess.cpp
|
||||
${CMAKE_SOURCE_DIR}/../../../../common/model_loader.cpp
|
||||
)
|
||||
|
||||
target_link_libraries(retinaface_demo
|
||||
${OpenCV_LIBS}
|
||||
${AMLNN_LIBRARY}
|
||||
)
|
||||
|
|
@ -139,25 +139,25 @@ int main(int argc, char** argv) {
|
|||
auto& b = boxes[k];
|
||||
int x1 = (b[0] * kInputW - px) / scale, y1 = (b[1] * kInputH - py) / scale;
|
||||
int x2 = (b[2] * kInputW - px) / scale, y2 = (b[3] * kInputH - py) / scale;
|
||||
|
||||
|
||||
cv::rectangle(img, {x1, y1}, {x2, y2}, {0, 255, 0}, 2);
|
||||
|
||||
char score_text[16];
|
||||
std::snprintf(score_text, sizeof(score_text), "%.2f", scores_vec[k]);
|
||||
cv::putText(img, score_text, {x1, std::max(y1 - 5, 5)},
|
||||
cv::putText(img, score_text, {x1, std::max(y1 - 5, 5)},
|
||||
cv::FONT_HERSHEY_SIMPLEX, 0.5, {0, 255, 0}, 1, cv::LINE_AA);
|
||||
|
||||
auto& lm = lms[k];
|
||||
for (int j = 0; j < 5; j++) {
|
||||
int lx = (lm[2 * j] * kInputW - px) / scale;
|
||||
int ly = (lm[2 * j + 1] * kInputH - py) / scale;
|
||||
cv::circle(img, {lx, ly}, 2, {0, 0, 255}, -1);
|
||||
cv::circle(img, {lx, ly}, 2, {0, 0, 255}, -1);
|
||||
}
|
||||
}
|
||||
|
||||
std::string save_path = out_dir + "/" + filename;
|
||||
cv::imwrite(save_path, img);
|
||||
|
||||
|
||||
std::cout << " Detected " << keep.size() << " faces\n";
|
||||
std::cout << " Result saved to: " << save_path << "\n\n";
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue