Struct GradientFlowAnalyzer 

pub struct GradientFlowAnalyzer<A> { /* private fields */ }

Gradient flow analyzer for monitoring and diagnosing gradient behavior

Implementations

impl<A> GradientFlowAnalyzer<A>

where A: Float + ScalarOperand + Debug + Sum,

pub fn new(config: GradientFlowConfig) -> Self

Create a new gradient flow analyzer with the given configuration

pub fn record_gradients( &mut self, layer_name: &str, gradients: &Array1<A>, ) -> Result<LayerGradientStats<A>>

Record gradients for a layer and compute statistics

Computes mean norm, max norm, min norm, variance, sparsity, and a histogram of gradient magnitudes. The results are stored in the internal history for later analysis.

pub fn detect_vanishing_gradients(&self) -> Vec<String>

Detect layers with vanishing gradients

Returns names of layers whose most recent mean gradient norm is below the configured vanishing threshold.

pub fn detect_exploding_gradients(&self) -> Vec<String>

Detect layers with exploding gradients

Returns names of layers whose most recent max gradient norm is above the configured exploding threshold.

pub fn get_health_report(&self) -> GradientHealthReport

Generate a comprehensive gradient health report

Analyzes all tracked layers and produces a report with:

• Lists of vanishing, exploding, and healthy layers
• An overall health assessment
• Actionable recommendations for fixing gradient issues

pub fn render_flow_chart(&self) -> Result<String>

Render an SVG flow chart showing gradient magnitudes across layers

Produces an SVG string with bars representing mean gradient norms for each layer, color-coded by health status.

pub fn get_layer_history( &self, layer_name: &str, ) -> Option<&Vec<LayerGradientStats<A>>>

Get the history of gradient statistics for a specific layer

pub fn clear_history(&mut self)

Clear all recorded gradient history