amlnn-model-playground/examples/retinaface

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:

# 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

# 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:

pip install numpy opencv-python amlnnlite-1.0.0-cp310-cp310-linux_aarch64.whl

Run on device:

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:

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:

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.