amd/movenet_demo_RAI_1.5

MoveNet Instructions

Overview

This demonstrates the use of the NPU for running the MoveNet model using ONNX Runtime with AMD's VAIP execution provider. CPU support is also included.

There are two models. One is a float32 model, the other is an int8 quantized model. The int8 model requires an xclbin file to be specified for the VAIP EP to run. It also needs special handling to ensure that the model is cached correctly.

Requirements

RAI 1.5 installation

Installation

1. git clone https://huggingface.co/datasets/amd/movenet_demo_RAI_1.5/
2. Unzip the two cache folders in the repo - these should be located in the same directory as movenet_demo.py
3. Copy the xclbins directory from C:\Program Files\RyzenAI\1.5.0\voe-4.0-win_amd64\ to the same directory as movenet_demo.py

• Note: your RAI installation path may vary

4. Install any requirements using the requirements.txt file: pip install -r requirements.txt
5. Activate RAI 1.5 conda env conda activate ryzen-ai-1.5.0

Usage

The movenet_demo.py script runs MoveNet pose estimation on a single image. It automatically detects whether you're using the FP32 or INT8 model and selects the appropriate pre-compiled cache.

Basic syntax:

python movenet_demo.py [options]

Available Parameters:

• --image: Input image path (default: images/input/test_image.jpg)
• --model: ONNX model path (default: ./movenet_fp32.onnx)
• --npu: Use NPU acceleration (required for INT8 model)
• --output, -o: Output image path (default: images/output/npu_webcam_image.png)
• --loops: Number of inference loops for benchmarking (default: 1)
• --threshold: Keypoint confidence threshold 0.0-1.0 (default: 0.3)

Basic Usage

Run with FP32 model on CPU:

python movenet_demo.py --model ./movenet_fp32.onnx

Run with INT8 model on NPU:

python movenet_demo.py --model ./movenet_int8.onnx --npu

Use custom input image:

python movenet_demo.py --image path/to/your/image.jpg

Save output to specific location:

python movenet_demo.py --output results/my_result.png

Advanced Options

Performance benchmarking with multiple inference loops:

python movenet_demo.py --loops 100  # Run 100 inferences for timing

Adjust keypoint confidence threshold:

python movenet_demo.py --threshold 0.5  # Only show keypoints with >50% confidence
python movenet_demo.py --threshold 0.1  # Show more keypoints (lower threshold)

Complete example with all options:

python movenet_demo.py \
  --image images/input/my_photo.jpg \
  --model ./movenet_int8.onnx \
  --npu \
  --output results/pose_output.png \
  --loops 50 \
  --threshold 0.4

Examples

Example 1: Basic FP32 CPU Inference

python movenet_demo.py --model ./movenet_fp32.onnx --loops 1

Output:

model ./movenet_fp32.onnx
Detected model type: fp32
Running on CPU
Model: fp32 | Input: int32 | Outputs: 1
Running 1 inference loops...
Latency: 11.8ms avg (11.8-11.8ms) | 84.6 FPS
Using keypoint threshold: 0.3
Keypoints detected: 5/17
Saved: images/output/npu_webcam_image.png

Example 2: Performance Benchmarking

python movenet_demo.py --model ./movenet_fp32.onnx --loops 10

Output:

model ./movenet_fp32.onnx
Detected model type: fp32
Running on CPU
Model: fp32 | Input: int32 | Outputs: 1
Running 10 inference loops...
Latency: 2.1ms avg (1.0-4.0ms) | 474.8 FPS
Using keypoint threshold: 0.3
Keypoints detected: 5/17
Saved: images/output/npu_webcam_image.png

Example 3: INT8 NPU Inference

python movenet_demo.py --model ./movenet_int8.onnx --npu --loops 1

Output:

model ./movenet_int8.onnx
Detected model type: int8
Running on NPU
Using xclbin: C:\...\xclbins\strix\AMD_AIE2P_4x4_Overlay.xclbin
[VitisAI EP initialization messages...]
Model: int8 | Input: float32 | Outputs: 4
Running 1 inference loops...
Latency: 4.0ms avg (4.0-4.0ms) | 251.1 FPS
Using keypoint threshold: 0.3
Keypoints detected: 0/17
Saved: images/output/npu_webcam_image.png

Example 4: Lower Threshold for More Keypoints

python movenet_demo.py --model ./movenet_fp32.onnx --threshold 0.1

Output:

model ./movenet_fp32.onnx
Detected model type: fp32
Running on CPU
Model: fp32 | Input: int32 | Outputs: 1
Running 1 inference loops...
Latency: 2.0ms avg (2.0-2.0ms) | 498.9 FPS
Using keypoint threshold: 0.1
Keypoints detected: 8/17
Saved: images/output/npu_webcam_image.png

Example 5: Custom Output Path

python movenet_demo.py --model ./movenet_fp32.onnx --output results/example_output.png

Output:

model ./movenet_fp32.onnx
Detected model type: fp32
Running on CPU
Model: fp32 | Input: int32 | Outputs: 1
Running 1 inference loops...
Latency: 2.5ms avg (2.5-2.5ms) | 398.7 FPS
Using keypoint threshold: 0.3
Keypoints detected: 5/17
Saved: results/example_output.png

Notes