Inference

Running trained models for prediction is straightforward in AodhML. The same model code works for both training and inference.

Basic Inference

fn predict(model: Model, input: Tensor<f32>) -> Tensor<f32> {
    // In inference mode, gradients are not computed
    with no_grad {
        let output = model.forward(input)
        return softmax(output)
    }
}

Batch Inference

fn predict_batch(model: Model, inputs: [Tensor<f32>]) -> [Tensor<f32>] {
    let batch = stack(inputs)
    with no_grad {
        let outputs = model.forward(batch)
        return unstack(softmax(outputs))
    }
}

Model Serialization

// Save model weights
model.save("model.aodh")

// Load model weights
let loaded = MLP::new()
loaded.load("model.aodh")

// Save full checkpoint (weights + optimizer state)
checkpoint.save("checkpoint.aodh", model, optimizer, epoch)

Quantization

Planned for v0.5.0:

// Quantize model to int8 for faster inference
let quantized = model.quantize(QuantConfig {
    dtype: i8,
    scheme: "per_channel",
})

// Run quantized inference
let result = quantized.forward(input)