Hansei Reports

Toyota Way Hansei (反省) principle: Reflection and continuous improvement through systematic post-training analysis.

Usage

#![allow(unused)]
fn main() {
use entrenar::monitor::{HanseiAnalyzer, MetricsCollector, Metric};

let mut collector = MetricsCollector::new();

// During training
for epoch in 0..100 {
    collector.record(Metric::Loss, loss);
    collector.record(Metric::Accuracy, accuracy);
    collector.record(Metric::GradientNorm, grad_norm);
}

// Generate report
let analyzer = HanseiAnalyzer::new();
let report = analyzer.analyze("my-training", &collector, duration_secs);
println!("{}", analyzer.format_report(&report));
}

Report Output

═══════════════════════════════════════════════════════════════
                    HANSEI POST-TRAINING REPORT
═══════════════════════════════════════════════════════════════

Training ID: my-training
Duration: 3600.00s
Total Steps: 10000

─── Metric Summaries ───────────────────────────────────────────

Loss:
  Mean: 0.123456  Std: 0.045678
  Min: 0.089012   Max: 0.567890
  Trend: ↑ Improving

Accuracy:
  Mean: 0.945678  Std: 0.023456
  Min: 0.800000   Max: 0.980000
  Trend: ↑ Improving

─── Issues Detected ────────────────────────────────────────────

[WARNING] Gradient Health
  Possible vanishing gradients: mean norm = 1.23e-08
  → Consider using residual connections or different activation functions

─── Recommendations ────────────────────────────────────────────
1. Training completed without critical issues.
2. Consider hyperparameter search for learning rate and batch size.

═══════════════════════════════════════════════════════════════

Issue Detection

The analyzer automatically detects:

Trend Analysis

Trends are determined by comparing mean to midpoint:

• Improving: Mean closer to optimal end (low for loss, high for accuracy)
• Degrading: Mean closer to suboptimal end
• Stable: Small range relative to std
• Oscillating: High coefficient of variation (> 0.5)

Custom Thresholds

#![allow(unused)]
fn main() {
let mut analyzer = HanseiAnalyzer::new();
analyzer.gradient_explosion_threshold = 50.0;  // Default: 100.0
analyzer.gradient_vanishing_threshold = 1e-8;  // Default: 1e-7
analyzer.min_accuracy_improvement = 0.02;      // Default: 0.01
}