APLUX Online AI Toolchain

Model Farm

Model Farm is APLUX's official model resource platform, providing 400+ AI models optimized for Qualcomm platforms, including measured performance benchmarks and ready-to-run sample code.

Core Functions

• Query inference performance of a specified model on A8625MY1 (latency, accuracy, memory usage)
• Download optimized model files adapted for QNN backend (supports INT4/INT8/FP16 precision)
• Obtain Python/C++ inference sample code for corresponding models
• Follow model adaptation workflows to migrate custom fine-tuned models

Standard Workflow (YOLOv5s Example)

1. Visit Model Farm: https://aiot.aidlux.com/en/models
2. Filter chip platform: QCS8625, model type: object detection
3. Find YOLOv5s-INT8-QNN236 model and review benchmark data
4. Log in with a developer account and download model files and code package
5. Upload to A8625MY1 and extract
6. Install dependencies:

pip install opencv-python numpy

7. Run test script:

cd model_farm_yolov5s/python
LD_PRELOAD=/usr/lib/libstdc++.so.6.0.30 python3 run_test.py \
  --model ../models/yolov5s_qcs8625_w8a8.qnn236.bin \
  --imgs bus.jpg \
  --invoke_nums 10

AIMO Model Optimization Platform

AIMO (AI Model Optimizer) is a web-based zero-code model conversion and optimization platform that converts mainstream framework models to Qualcomm QNN format and performs automatic quantization and performance optimization.

Core Features

• Supports mainstream frameworks including ONNX, PyTorch, TensorFlow, and PaddlePaddle
• Built-in Qualcomm-platform operator library greatly improves conversion success rate
• Supports automatic and custom quantization to balance accuracy and performance
• Generates optimized models and sample code that can run directly on A8625MY1

Standard Conversion Workflow (YOLOv8s Example)

1. Visit AIMO platform: https://aimo.aidlux.com
2. Upload source model file (for example yolov8s.onnx)
3. Select target platform: QCS8625, target framework: QNN2.36
4. Use Netron to inspect model structure and fill in input/output node names
5. Set quantization parameters: choose INT8 quantization and optionally upload calibration dataset
6. Submit conversion task and download optimized model package after completion
7. Upload to A8625MY1 and run with AidLite SDK

AI Creator Visual Training Platform

AI Creator is an all-in-one AI development platform integrating data labeling, model training, and remote deployment, optimized for edge applications.

Core Functions

• Supports pixel-level annotation: object detection, semantic segmentation, keypoint detection, OCR, classification
• Includes six built-in algorithm modules and supports serial combination of modules (for example localization + segmentation + classification)
• Automatically tunes training parameters so models can be trained without deep AI expertise
• Automatically converts trained models to QNN models adapted for A8625MY1
• Supports one-click remote deployment to connected A8625 devices