Struct TensorDiffEngine 

pub struct TensorDiffEngine {
    pub threshold: f32,
}

Engine that computes structural and numeric diffs between tensor snapshots.

Fields

threshold: f32

Element-wise absolute difference threshold below which a value is considered unchanged.

Implementations

impl TensorDiffEngine

pub fn new(threshold: f32) -> Self

Create a new engine with the given numeric threshold.

pub fn diff_tensors( &self, old: &TensorSnapshot, new: &TensorSnapshot, ) -> TensorDiff

Diff two individual tensors that share the same name.

Shape mismatch → ShapeChanged. Element-wise: count elements where |new[i] - old[i]| > threshold, accumulate max and sum. If count == 0 → Unchanged, else ValueChanged.

pub fn diff_snapshots( &self, old_set: &[TensorSnapshot], new_set: &[TensorSnapshot], ) -> Vec<TensorDiff>

Diff two full checkpoint snapshots (slices of tensors).

Results are sorted by name for deterministic output.

pub fn summarize(&self, diffs: &[TensorDiff]) -> DiffSummary

Aggregate a slice of diffs into a DiffSummary.

pub fn significant_diffs<'a>( &self, diffs: &'a [TensorDiff], ) -> Vec<&'a TensorDiff>

Filter diffs to only those that are significant at self.threshold.