Struct ActivationEstimator 

pub struct ActivationEstimator { /* private fields */ }

Runs calibration samples through a model and collects activation statistics.

Usage:

let model = OnnxModel::load("model.onnx")?;
let mut estimator = ActivationEstimator::new(model, "model.onnx")?;
let dataset = CalibrationDataset::from_numpy("samples.npy")?;
estimator.calibrate(&dataset)?;
let stats = estimator.get_layer_stats();  // HashMap<layer_name, &ActivationStats>

Implementations

impl ActivationEstimator

pub fn from_path(model: OnnxModel, onnx_path: &str) -> Result<Self>

Load model and prepare for calibration.

This:

1. Reloads the ONNX file with tract (we need the filepath)
2. Exposes all layer outputs as model outputs
3. Optimizes the graph
4. Creates a runnable plan

Important: The model parameter must have been loaded from a file on disk. We re-parse that file with tract. If the model was constructed programmatically or the file no longer exists, this will fail.

pub fn new(model: OnnxModel, onnx_path: &str) -> Result<Self>

Convenience constructor when you have the model and its path.

pub fn calibrate(&mut self, dataset: &CalibrationDataset) -> Result<()>

Run calibration samples through the model and collect activation statistics.

For each sample:

• Run inference
• Capture all intermediate tensors
• Update min/max/histogram for each layer

Progress is printed every 10 batches.

pub fn get_layer_stats(&self) -> HashMap<String, &ActivationStats>

Get collected activation statistics for all layers (borrowed).

Returns a map from layer name → &ActivationStats. These stats include min/max (for range optimization) and histogram (for entropy/MSE methods).

pub fn into_layer_stats(self) -> HashMap<String, ActivationStats>

Consume and return owned activation statistics.

Use this when passing stats to Quantizer::with_calibration, which expects HashMap<String, ActivationStats> (owned, not borrowed).

pub fn get_layer_stats_mut(&mut self) -> &mut HashMap<String, ActivationStats>

Get mutable reference to stats (for advanced use cases)

pub fn into_model(self) -> OnnxModel

Consume the estimator and return the original OnnxModel.

Useful when you need the model back but have already extracted stats with get_layer_stats() (borrowed). For the typical quantization pipeline, use into_layer_stats() to get owned stats, then reload the model separately for quantization.

pub fn model(&self) -> &OnnxModel

Borrow the original model.

Trait Implementations

impl Debug for ActivationEstimator

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.