Case Study: APR Model Inspection

This example demonstrates the inspection tooling for .apr model files, following the Toyota Way principle of Genchi Genbutsu (go and see).

Overview

The inspection module provides comprehensive tooling to analyze .apr model files:

• Header inspection (magic, version, flags, compression)
• Metadata extraction (hyperparameters, training info, license)
• Weight statistics with health assessment
• Model diff for version comparison

Toyota Way Alignment

Running the Example

cargo run --example apr_inspection

Header Inspection

Inspect the binary header of .apr files:

let mut header = HeaderInspection::new();
header.version = (1, 2);
header.model_type = 3;  // RandomForest
header.compressed_size = 5 * 1024 * 1024;
header.uncompressed_size = 12 * 1024 * 1024;

println!("Compression Ratio: {:.2}x", header.compression_ratio());
println!("Header Valid: {}", header.is_valid());

Header Flags

Metadata Inspection

Extract model metadata including hyperparameters and provenance:

let mut meta = MetadataInspection::new("RandomForestClassifier");
meta.n_parameters = 50_000;
meta.n_features = 13;
meta.n_outputs = 3;

meta.hyperparameters.insert("n_estimators".to_string(), "100".to_string());
meta.hyperparameters.insert("max_depth".to_string(), "10".to_string());

Training Info

Track training provenance for reproducibility:

meta.training_info = Some(TrainingInfo {
    trained_at: Some("2024-12-08T10:30:00Z".to_string()),
    duration: Some(Duration::from_secs(120)),
    dataset_name: Some("iris_extended".to_string()),
    n_samples: Some(10000),
    final_loss: Some(0.0234),
    framework: Some("aprender".to_string()),
    framework_version: Some("0.15.0".to_string()),
});

Weight Statistics

Analyze model weights for health issues:

let stats = WeightStats::from_slice(&weights);

println!("Count: {}", stats.count);
println!("Min: {:.4}", stats.min);
println!("Max: {:.4}", stats.max);
println!("Mean: {:.4}", stats.mean);
println!("Std: {:.4}", stats.std);
println!("NaN Count: {}", stats.nan_count);  // CRITICAL if > 0
println!("Inf Count: {}", stats.inf_count);  // CRITICAL if > 0
println!("Sparsity: {:.2}%", stats.sparsity * 100.0);
println!("Health: {:?}", stats.health_status());

Health Status Levels

Model Diff

Compare two model versions:

let mut diff = DiffResult::new("model_v1.apr", "model_v2.apr");

diff.header_diff.push(DiffItem::new("version", "1.0", "1.1"));
diff.metadata_diff.push(DiffItem::new("n_estimators", "100", "150"));

let weight_diff = WeightDiff::from_slices(&weights_a, &weights_b);
println!("Changed Count: {}", weight_diff.changed_count);
println!("Max Diff: {:.6}", weight_diff.max_diff);
println!("Cosine Similarity: {:.4}", weight_diff.cosine_similarity);

Inspection Options

Configure inspection behavior:

// Quick inspection (no weights, no quality)
let quick = InspectOptions::quick();

// Full inspection (all checks, verbose output)
let full = InspectOptions::full();

// Default (balanced)
let default = InspectOptions::default();

Source Code

• Example: examples/apr_inspection.rs
• Module: src/inspect/mod.rs