Deep Neural Networks (DNN)

Deep learning modules in Iris are built natively on the Burn framework. This allows you to construct model structures, load weights, and execute forward passes across multiple backends (WGPU, CUDA, LibTorch, Ndarray).

Loading Pretrained Model Weights

Safetensors format

Loads weights from standard .safetensors files using the safetensors crate.

let device = Default::default();
let weights = WeightLoader::load_safetensors::<Backend>("weights/face_detector.safetensors", &device)?;

PyTorch Bin format

Loads weights from raw flat binary streams.

let weight_tensor = WeightLoader::load_bin::<Backend>("weights/model.bin", &device, [100, 100])?;

Running ONNX Models

Iris supports importing ONNX model pipelines.

// Load an ONNX model from file
let model = OnnxModel::load("weights/object_detector.onnx", &device)?;

// Preprocess the input image to get a [1, C, H, W] tensor
let input_tensor = model.preprocess(&image)?;

// Evaluate the neural network
let output_tensor: Tensor<Backend, 4> = model.predict_raw(input_tensor)?;

Direct Builders

let model = read_net::<Backend>("weights/model.onnx", &device)?;
let model = read_net_from_onnx::<Backend>("weights/model.onnx", &device)?;

Input Preprocessing (Blobs)

The blob_from_image helper handles resizing, scaling, and mean subtraction to prepare images for neural network entry.

let size = Size::new(224, 224);
let mean = Scalar::new(0.485, 0.456, 0.406, 0.0);
let scale = 1.0 / 255.0;

let blob = blob_from_image(&image, scale, size, mean, true)?; // [1, C, 224, 224]

Non-Maximum Suppression (NMS)

Filters overlapping bounding boxes based on class confidence scores and Intersection over Union (IoU) values.

let kept_indices = nms_boxes(&bboxes, &scores, 0.5, 0.4);