0xhalfmoonkid/ritual-models

Ritual ONNX Models

This repository contains ONNX models optimized for use with the Ritual platform's ONNX precompile.

Models

1. MobileNet v2 (mobilenetv2-12.onnx)

• Size: 13.3 MB
• Purpose: Lightweight image classification
• Input: 224x224 RGB images
• Output: 1000 ImageNet classes
• ONNX Opset: 12

2. ResNet50 (resnet50-v2-7.onnx)

• Size: 97.4 MB
• Purpose: High-accuracy image classification
• Input: 224x224 RGB images
• Output: 1000 ImageNet classes
• ONNX Opset: 12

3. EfficientNet B0 (efficientnet-b0.onnx)

• Size: 20.2 MB
• Purpose: Balanced accuracy and efficiency
• Input: 224x224 RGB images
• Output: 1000 ImageNet classes
• ONNX Opset: 12

Usage with Ritual

These models can be used in Ritual smart contracts via the ONNX precompile:

// Example model ID format
string memory modelId = "huggingface/wallychang/ritual-models/mobilenetv2-12.onnx@main";

// Call ONNX precompile
(bool success, bytes memory result) = ONNX_PRECOMPILE.staticcall(
    abi.encode(
        1, // precompile version
        modelId,
        inputTensor,
        inputNumberFormat,
        outputNumberFormat,
        roundingMode
    )
);

Model Details

All models are pre-trained on ImageNet and converted to ONNX format using PyTorch. They expect:

• Input shape: [1, 3, 224, 224] (batch_size, channels, height, width)
• Input format: Float32 normalized to [0, 1]
• Output shape: [1, 1000] (batch_size, num_classes)

License

These models are based on pre-trained PyTorch models and are subject to their respective licenses.