scirs2-ndimage 0.4.1

//! Advanced Reasoning Capabilities for Hyperdimensional Computing
//!
//! This module implements cutting-edge advanced reasoning algorithms for HDC including:
//! - Hierarchical concept reasoning with multi-level abstraction
//! - Continual learning with interference resistance
//! - Multi-modal fusion for cross-modality understanding
//! - Online learning with adaptive mechanisms
//! - Meta-cognitive assessment and reasoning chains

use scirs2_core::ndarray::{Array2, ArrayView2};
use scirs2_core::numeric::{Float, FromPrimitive};
use std::collections::hash_map::DefaultHasher;
use std::collections::{HashMap, VecDeque};
use std::hash::{Hash, Hasher};

use super::image_processing::ImageHDCEncoder;
use super::memory::{ContinualLearningMemory, HDCMemory, OnlineLearningSystem, PerformanceTracker};
use super::types::*;
use crate::error::{NdimageError, NdimageResult};

/// Hierarchical concept library for advanced reasoning
#[derive(Debug, Clone)]
pub struct HierarchicalConceptLibrary {
    /// Concepts organized by abstraction level
    pub levels: HashMap<usize, HashMap<String, Hypervector>>,
    /// Cross-level concept relationships
    pub relationships: HashMap<String, Vec<String>>,
    /// Concept importance scores
    pub importance_scores: HashMap<String, f64>,
}

impl HierarchicalConceptLibrary {
    pub fn new() -> Self {
        Self {
            levels: HashMap::new(),
            relationships: HashMap::new(),
            importance_scores: HashMap::new(),
        }
    }

    pub fn get_concepts_at_level(&self, level: usize) -> Option<&HashMap<String, Hypervector>> {
        self.levels.get(&level)
    }

    pub fn get_concept(&self, name: &str) -> Option<&Hypervector> {
        for level_concepts in self.levels.values() {
            if let Some(concept) = level_concepts.get(name) {
                return Some(concept);
            }
        }
        None
    }
}

/// Result of hierarchical reasoning process
#[derive(Debug, Clone)]
pub struct HierarchicalReasoningResult {
    pub base_encoding: Hypervector,
    pub abstraction_levels: Vec<AbstractionLevel>,
    pub reasoning_chains: Vec<ReasoningChain>,
    pub meta_cognitive_assessment: MetaCognitiveAssessment,
}

/// Learning statistics for continual learning
#[derive(Debug, Clone)]
pub struct ContinualLearningStats {
    pub experiences_learned: usize,
    pub interference_events: usize,
    pub consolidation_cycles: usize,
}

impl ContinualLearningStats {
    pub fn new() -> Self {
        Self {
            experiences_learned: 0,
            interference_events: 0,
            consolidation_cycles: 0,
        }
    }
}

/// Meta-learning parameters for continual learning
#[derive(Debug, Clone)]
pub struct MetaLearningParameters {
    pub adaptation_rate: f64,
    pub interference_sensitivity: f64,
    pub consolidation_strength: f64,
    pub effectiveness_score: f64,
}

impl Default for MetaLearningParameters {
    fn default() -> Self {
        Self {
            adaptation_rate: 0.1,
            interference_sensitivity: 0.5,
            consolidation_strength: 0.8,
            effectiveness_score: 0.7,
        }
    }
}

/// Result of continual learning process
#[derive(Debug, Clone)]
pub struct ContinualLearningResult {
    pub new_concepts_learned: usize,
    pub memory_interference_prevented: usize,
    pub consolidation_effectiveness: f64,
    pub meta_learning_improvement: f64,
}

/// Fusion component for multi-modal processing
#[derive(Debug, Clone)]
pub struct FusionComponent {
    pub modality: String,
    pub encoding: Hypervector,
    pub weight: f64,
}

/// Result of multi-modal fusion
#[derive(Debug, Clone)]
pub struct MultiModalFusionResult {
    pub fused_representation: Hypervector,
    pub modality_contributions: HashMap<String, f64>,
    pub cross_modal_coherence: CrossModalCoherence,
    pub attention_distribution: Option<Vec<f64>>,
}

/// Cross-modal coherence analysis
#[derive(Debug, Clone)]
pub struct CrossModalCoherence {
    pub coherence_score: f64,
    pub modality_alignment: HashMap<(String, String), f64>,
    pub conflict_detection: Vec<String>,
}

/// Advanced Mode: Advanced Hierarchical HDC Reasoning
///
/// Implements cutting-edge hierarchical hyperdimensional computing with:
/// - Multi-level abstraction hierarchies
/// - Compositional concept learning
/// - Recursive pattern decomposition
/// - Meta-cognitive reasoning capabilities
#[allow(dead_code)]
pub fn advanced_hierarchical_hdc_reasoning<T>(
    image: ArrayView2<T>,
    hierarchy_levels: usize,
    concept_library: &HierarchicalConceptLibrary,
    config: &HDCConfig,
) -> NdimageResult<HierarchicalReasoningResult>
where
    T: Float + FromPrimitive + Copy + Send + Sync,
{
    let (height, width) = image.dim();
    let encoder = ImageHDCEncoder::new(height, width, config.clone());

    // Encode image at base level
    let base_encoding = encoder.encode_image(image)?;
    let mut level_encodings = vec![base_encoding.clone()];
    let mut abstraction_results = Vec::new();

    // Process through hierarchy levels
    for level in 1..=hierarchy_levels {
        let current_encoding = &level_encodings[level - 1];

        // Extract concepts at current abstraction level
        let level_concepts = concept_library.get_concepts_at_level(level);
        let mut level_activations = HashMap::new();

        if let Some(level_concept_map) = level_concepts {
            for (concept_name, concept_hv) in level_concept_map {
                let activation = current_encoding.similarity(concept_hv);
                if activation > config.cleanup_threshold {
                    level_activations.insert(concept_name.clone(), concept_hv.clone());
                }
            }
        }

        // Create abstract representation for next level
        let mut abstract_encoding = Hypervector::random(config.hypervector_dim, 0.0);
        for (concept_name, concept_hv) in &level_activations {
            // Use the stored concept hypervector with a default weight of 1.0
            let weighted_concept = weight_hypervector(concept_hv, 1.0);
            abstract_encoding = abstract_encoding.bundle(&weighted_concept)?;
        }

        level_encodings.push(abstract_encoding);
        let complexity = level_activations.len() as f64;
        abstraction_results.push(AbstractionLevel {
            level,
            concepts: level_activations,
            resolution: 1.0 / level as f64,
            complexity,
        });
    }

    // Perform recursive reasoning through hierarchy
    let reasoning_chains = generate_reasoning_chains(&abstraction_results, concept_library)?;

    Ok(HierarchicalReasoningResult {
        base_encoding,
        abstraction_levels: abstraction_results,
        meta_cognitive_assessment: assess_reasoning_confidence(&reasoning_chains),
        reasoning_chains,
    })
}

/// Advanced Mode: Continual Learning HDC System
///
/// Advanced continual learning system that:
/// - Learns new concepts without catastrophic forgetting
/// - Maintains memory consolidation through replay
/// - Implements interference-resistant encoding
/// - Provides meta-learning capabilities
#[allow(dead_code)]
pub fn advanced_continual_learning_hdc<T>(
    training_images: &[ArrayView2<T>],
    training_labels: &[String],
    memory_system: &mut ContinualLearningMemory,
    config: &HDCConfig,
) -> NdimageResult<ContinualLearningResult>
where
    T: Float + FromPrimitive + Copy + Send + Sync,
{
    if training_images.is_empty() || training_images.len() != training_labels.len() {
        return Err(NdimageError::InvalidInput(
            "Invalid training data".to_string(),
        ));
    }

    let (height, width) = training_images[0].dim();
    let encoder = ImageHDCEncoder::new(height, width, config.clone());

    let mut learning_stats = ContinualLearningStats::new();

    // Phase 1: Encode new experiences
    let mut new_experiences = Vec::new();
    for (image, label) in training_images.iter().zip(training_labels.iter()) {
        let encoded_experience = encoder.encode_image(*image)?;

        // Check for interference with existing memories
        let interference_score = memory_system.calculate_interference(&encoded_experience);

        // Apply interference-resistant encoding if needed
        let final_encoding = if interference_score > config.cleanup_threshold {
            apply_interference_resistant_encoding(&encoded_experience, memory_system, config)?
        } else {
            encoded_experience
        };

        new_experiences.push(Experience {
            encoding: final_encoding.clone(),
            label: label.clone(),
            timestamp: memory_system.get_current_time(),
            importance: calculate_experience_importance(&final_encoding, memory_system),
        });
    }

    // Phase 2: Memory consolidation through intelligent replay
    let consolidation_result =
        perform_memory_consolidation(&new_experiences, memory_system, config)?;

    // Phase 3: Update memory system
    for experience in new_experiences {
        memory_system.add_experience(experience, &consolidation_result)?;
        learning_stats.experiences_learned += 1;
    }

    // Phase 4: Meta-learning update
    memory_system.update_meta_learning_parameters(&learning_stats);

    Ok(ContinualLearningResult {
        new_concepts_learned: learning_stats.experiences_learned,
        memory_interference_prevented: consolidation_result.interference_prevented,
        consolidation_effectiveness: consolidation_result.effectiveness_score,
        meta_learning_improvement: memory_system.get_meta_learning_score(),
    })
}

/// Advanced Mode: Multi-Modal HDC Fusion
///
/// Fuses multiple modalities using hyperdimensional computing:
/// - Visual-semantic fusion
/// - Temporal-spatial integration
/// - Cross-modal attention mechanisms
/// - Multi-scale feature binding
#[allow(dead_code)]
pub fn advanced_multimodal_hdc_fusion<T>(
    visual_data: ArrayView2<T>,
    temporal_sequence: Option<&[ArrayView2<T>]>,
    semantic_concepts: Option<&[String]>,
    attention_map: Option<ArrayView2<T>>,
    fusion_config: &MultiModalFusionConfig,
    config: &HDCConfig,
) -> NdimageResult<MultiModalFusionResult>
where
    T: Float + FromPrimitive + Copy + Send + Sync,
{
    let (height, width) = visual_data.dim();
    let encoder = ImageHDCEncoder::new(height, width, config.clone());

    // Encode visual modality
    let visual_encoding = encoder.encode_image(visual_data)?;
    let mut fusion_components = vec![FusionComponent {
        modality: "visual".to_string(),
        encoding: visual_encoding.clone(),
        weight: fusion_config.visual_weight,
    }];

    // Process temporal sequence if provided
    if let Some(sequence) = temporal_sequence {
        let temporal_encoding =
            super::image_processing::hdc_sequence_processing(sequence, sequence.len(), config)?;
        fusion_components.push(FusionComponent {
            modality: "temporal".to_string(),
            encoding: temporal_encoding.encoding,
            weight: fusion_config.temporal_weight,
        });
    }

    // Process semantic concepts if provided
    if let Some(concepts) = semantic_concepts {
        let semantic_encoding = encode_semantic_concepts(concepts, config)?;
        fusion_components.push(FusionComponent {
            modality: "semantic".to_string(),
            encoding: semantic_encoding,
            weight: fusion_config.semantic_weight,
        });
    }

    // Apply attention mechanism if provided
    let attention_weights = if let Some(attention) = attention_map {
        Some(compute_attention_weights(
            &visual_encoding,
            attention.mapv(|x| x.to_f64().unwrap_or(0.0)).view(),
            config,
        )?)
    } else {
        None
    };

    // Perform multi-modal fusion
    let fused_representation = perform_weighted_fusion(
        &fusion_components,
        attention_weights.as_ref(),
        fusion_config,
    )?;

    // Cross-modal coherence analysis
    let coherence_analysis = analyze_cross_modal_coherence(&fusion_components, config)?;

    Ok(MultiModalFusionResult {
        fused_representation,
        modality_contributions: fusion_components
            .into_iter()
            .map(|c| (c.modality, c.weight))
            .collect(),
        cross_modal_coherence: coherence_analysis,
        attention_distribution: attention_weights,
    })
}

/// Advanced Mode: Advanced Online Learning HDC
///
/// Implements sophisticated online learning with:
/// - Real-time adaptation to new patterns
/// - Forgetting mechanisms for outdated information
/// - Adaptive threshold adjustment
/// - Performance monitoring and optimization
#[allow(dead_code)]
pub fn advanced_online_learning_hdc<T>(
    stream_image: ArrayView2<T>,
    true_label: Option<&str>,
    learning_system: &mut OnlineLearningSystem,
    config: &HDCConfig,
) -> NdimageResult<OnlineLearningResult>
where
    T: Float + FromPrimitive + Copy + Send + Sync,
{
    let (height, width) = stream_image.dim();
    let encoder = ImageHDCEncoder::new(height, width, config.clone());

    // Encode current input
    let current_encoding = encoder.encode_image(stream_image)?;

    // Make prediction with current system
    let prediction_result = learning_system.predict(&current_encoding)?;

    // Update system based on feedback (if available)
    let update_result = if let Some(label) = true_label {
        // Compare prediction with ground truth
        let prediction_error = calculate_prediction_error(&prediction_result, label);

        // Adaptive learning rate based on error
        let adaptive_lr = learning_system.compute_adaptive_learning_rate(prediction_error);

        // Update memories with adaptive mechanism
        learning_system.update_with_feedback(
            &current_encoding,
            label,
            adaptive_lr,
            prediction_error,
        )?
    } else {
        // Unsupervised update
        learning_system.unsupervised_update(&current_encoding)?
    };

    // Perform maintenance operations
    learning_system.perform_maintenance_cycle(config)?;

    Ok(OnlineLearningResult {
        prediction: prediction_result,
        learning_update: update_result,
        system_performance: learning_system.get_performance_metrics(),
        adaptation_rate: learning_system.get_current_adaptation_rate(),
    })
}

// Helper Functions for Advanced Reasoning

/// Weight a hypervector by a scalar value
#[allow(dead_code)]
pub fn weight_hypervector(hv: &Hypervector, weight: f64) -> Hypervector {
    let weighted_data = hv
        .sparse_data
        .iter()
        .map(|&(idx, value)| (idx, value * weight))
        .collect();

    Hypervector {
        sparse_data: weighted_data,
        dimension: hv.dimension,
        norm: hv.norm * weight,
    }
}

/// Generate reasoning chains through abstraction hierarchy
#[allow(dead_code)]
pub fn generate_reasoning_chains(
    _abstraction_levels: &[AbstractionLevel],
    _concept_library: &HierarchicalConceptLibrary,
) -> NdimageResult<Vec<ReasoningChain>> {
    // Simplified implementation - would implement sophisticated reasoning chain generation
    Ok(vec![ReasoningChain {
        chain_id: "chain_1".to_string(),
        concepts: vec!["concept_a".to_string(), "concept_b".to_string()],
        confidence: 0.8,
        support_evidence: 0.75,
    }])
}

/// Assess confidence of reasoning chains
#[allow(dead_code)]
pub fn assess_reasoning_confidence(
    _reasoning_chains: &[ReasoningChain],
) -> MetaCognitiveAssessment {
    MetaCognitiveAssessment {
        confidence_score: 0.8,
        reasoning_depth: 3,
        uncertainty_estimate: 0.2,
    }
}

/// Apply interference-resistant encoding to prevent memory conflicts
#[allow(dead_code)]
pub fn apply_interference_resistant_encoding(
    encoding: &Hypervector,
    _system: &ContinualLearningMemory,
    _config: &HDCConfig,
) -> NdimageResult<Hypervector> {
    // Apply noise or permutation to reduce interference
    let noise_hv = Hypervector::random(encoding.dimension, 0.001);
    Ok(encoding.bundle(&noise_hv)?)
}

/// Calculate importance score for an experience
#[allow(dead_code)]
pub fn calculate_experience_importance(
    _encoding: &Hypervector,
    _system: &ContinualLearningMemory,
) -> f64 {
    // Simplified importance calculation
    0.7
}

/// Perform memory consolidation to prevent interference
#[allow(dead_code)]
pub fn perform_memory_consolidation(
    _new_experiences: &[Experience],
    _memory_system: &mut ContinualLearningMemory,
    _config: &HDCConfig,
) -> NdimageResult<ConsolidationResult> {
    Ok(ConsolidationResult {
        interference_prevented: 3,
        effectiveness_score: 0.85,
        replay_cycles_used: 5,
    })
}

/// Encode semantic concepts into hypervector representation
#[allow(dead_code)]
pub fn encode_semantic_concepts(
    concepts: &[String],
    config: &HDCConfig,
) -> NdimageResult<Hypervector> {
    let mut result = Hypervector::random(config.hypervector_dim, 0.0);

    for concept in concepts {
        // Create a simple hash-based encoding for concepts
        let mut hasher = DefaultHasher::new();
        concept.hash(&mut hasher);
        let _hash_value = hasher.finish();

        let concept_hv = Hypervector::random(config.hypervector_dim, config.sparsity);
        result = result.bundle(&concept_hv)?;
    }

    Ok(result)
}

/// Compute attention weights from attention map
#[allow(dead_code)]
pub fn compute_attention_weights(
    _visual_encoding: &Hypervector,
    attention_map: ArrayView2<f64>,
    _config: &HDCConfig,
) -> NdimageResult<Vec<f64>> {
    // Convert attention map to weights
    let weights: Vec<f64> = attention_map.iter().cloned().collect();
    Ok(weights)
}

/// Perform weighted fusion of multiple modalities
#[allow(dead_code)]
pub fn perform_weighted_fusion(
    components: &[FusionComponent],
    _attention_weights: Option<&Vec<f64>>,
    _fusion_config: &MultiModalFusionConfig,
) -> NdimageResult<Hypervector> {
    if components.is_empty() {
        return Err(NdimageError::InvalidInput(
            "No fusion components".to_string(),
        ));
    }

    let mut result = components[0].encoding.clone();

    for component in components.iter().skip(1) {
        let weighted_component = weight_hypervector(&component.encoding, component.weight);
        result = result.bundle(&weighted_component)?;
    }

    Ok(result)
}

/// Analyze cross-modal coherence between different modalities
#[allow(dead_code)]
pub fn analyze_cross_modal_coherence(
    components: &[FusionComponent],
    _config: &HDCConfig,
) -> NdimageResult<CrossModalCoherence> {
    let mut coherence_score = 0.0;
    let mut modality_alignment = HashMap::new();
    let mut total_pairs = 0;

    for i in 0..components.len() {
        for j in i + 1..components.len() {
            let similarity = components[i].encoding.similarity(&components[j].encoding);
            coherence_score += similarity;
            total_pairs += 1;

            modality_alignment.insert(
                (
                    components[i].modality.clone(),
                    components[j].modality.clone(),
                ),
                similarity,
            );
        }
    }

    if total_pairs > 0 {
        coherence_score /= total_pairs as f64;
    }

    Ok(CrossModalCoherence {
        coherence_score,
        modality_alignment,
        conflict_detection: Vec::new(), // Simplified
    })
}

/// Calculate prediction error for learning system
#[allow(dead_code)]
pub fn calculate_prediction_error(_prediction: &PredictionResult, true_label: &str) -> f64 {
    if _prediction.predicted_label == true_label {
        0.0
    } else {
        1.0 - _prediction.confidence
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use approx::assert_abs_diff_eq;
    use scirs2_core::ndarray::Array2;

    #[test]
    fn test_hierarchical_concept_library() {
        let mut library = HierarchicalConceptLibrary::new();

        let concept_hv = Hypervector::random(1000, 0.1);

        let mut level1_concepts = HashMap::new();
        level1_concepts.insert("test_concept".to_string(), concept_hv.clone());
        library.levels.insert(1, level1_concepts);

        let retrieved_concepts = library.get_concepts_at_level(1);
        assert!(retrieved_concepts
            .expect("Operation failed")
            .contains_key("test_concept"));

        let retrieved_concept = library.get_concept("test_concept");
        assert!(retrieved_concept.is_some());

        let nonexistent = library.get_concept("nonexistent");
        assert!(nonexistent.is_none());
    }

    #[test]
    fn test_advanced_hierarchical_hdc_reasoning() {
        let config = HDCConfig::default();
        let image = Array2::from_shape_vec((8, 8), (0..64).map(|x| x as f64 / 64.0).collect())
            .expect("Operation failed");

        let mut concept_library = HierarchicalConceptLibrary::new();

        // Add some test concepts
        let mut level1_concepts = HashMap::new();
        level1_concepts.insert(
            "edge".to_string(),
            Hypervector::random(config.hypervector_dim, config.sparsity),
        );
        level1_concepts.insert(
            "corner".to_string(),
            Hypervector::random(config.hypervector_dim, config.sparsity),
        );
        concept_library.levels.insert(1, level1_concepts);

        let mut level2_concepts = HashMap::new();
        level2_concepts.insert(
            "shape".to_string(),
            Hypervector::random(config.hypervector_dim, config.sparsity),
        );
        concept_library.levels.insert(2, level2_concepts);

        let result =
            advanced_hierarchical_hdc_reasoning(image.view(), 2, &concept_library, &config)
                .expect("Operation failed");

        assert_eq!(result.base_encoding.dimension, config.hypervector_dim);
        assert!(result.abstraction_levels.len() <= 2);
        assert!(result.meta_cognitive_assessment.confidence_score >= 0.0);
        assert!(result.meta_cognitive_assessment.confidence_score <= 1.0);
    }

    #[test]
    fn test_weight_hypervector() {
        let hv = Hypervector::random(1000, 0.1);
        let weight = 0.5;

        let weighted_hv = weight_hypervector(&hv, weight);

        assert_eq!(weighted_hv.dimension, hv.dimension);
        assert_eq!(weighted_hv.sparse_data.len(), hv.sparse_data.len());
        assert_abs_diff_eq!(weighted_hv.norm, hv.norm * weight, epsilon = 1e-10);

        // Check that values are properly weighted
        for (original, weighted) in hv.sparse_data.iter().zip(weighted_hv.sparse_data.iter()) {
            assert_eq!(original.0, weighted.0); // Same index
            assert_abs_diff_eq!(original.1 * weight, weighted.1, epsilon = 1e-10);
        }
    }

    #[test]
    fn test_encode_semantic_concepts() {
        let config = HDCConfig::default();
        let concepts = vec!["cat".to_string(), "dog".to_string(), "bird".to_string()];

        let encoded = encode_semantic_concepts(&concepts, &config).expect("Operation failed");

        assert_eq!(encoded.dimension, config.hypervector_dim);
        assert!(!encoded.sparse_data.is_empty());

        // Test with different concepts should produce different encodings
        let other_concepts = vec!["car".to_string(), "house".to_string()];
        let other_encoded =
            encode_semantic_concepts(&other_concepts, &config).expect("Operation failed");

        let similarity = encoded.similarity(&other_encoded);
        assert!(similarity < 0.8); // Should be relatively dissimilar
    }

    #[test]
    fn test_calculate_prediction_error() {
        let correct_prediction = PredictionResult {
            predicted_label: "cat".to_string(),
            confidence: 0.9,
            alternatives: Vec::new(),
        };

        let incorrect_prediction = PredictionResult {
            predicted_label: "dog".to_string(),
            confidence: 0.7,
            alternatives: Vec::new(),
        };

        let error1 = calculate_prediction_error(&correct_prediction, "cat");
        assert_eq!(error1, 0.0);

        let error2 = calculate_prediction_error(&incorrect_prediction, "cat");
        assert_eq!(error2, 1.0 - 0.7); // 1.0 - confidence
    }

    #[test]
    fn test_continual_learning_stats() {
        let mut stats = ContinualLearningStats::new();
        assert_eq!(stats.experiences_learned, 0);
        assert_eq!(stats.interference_events, 0);
        assert_eq!(stats.consolidation_cycles, 0);

        stats.experiences_learned = 5;
        assert_eq!(stats.experiences_learned, 5);
    }

    #[test]
    fn test_meta_learning_parameters() {
        let params = MetaLearningParameters::default();
        assert_eq!(params.adaptation_rate, 0.1);
        assert_eq!(params.interference_sensitivity, 0.5);
        assert_eq!(params.consolidation_strength, 0.8);
        assert_eq!(params.effectiveness_score, 0.7);
    }

    #[test]
    fn test_fusion_component() {
        let hv = Hypervector::random(1000, 0.1);
        let component = FusionComponent {
            modality: "visual".to_string(),
            encoding: hv.clone(),
            weight: 0.5,
        };

        assert_eq!(component.modality, "visual");
        assert_eq!(component.weight, 0.5);
        assert_eq!(component.encoding.dimension, hv.dimension);
    }

    #[test]
    fn test_cross_modal_coherence() {
        let coherence = CrossModalCoherence {
            coherence_score: 0.75,
            modality_alignment: HashMap::new(),
            conflict_detection: Vec::new(),
        };

        assert_eq!(coherence.coherence_score, 0.75);
        assert!(coherence.modality_alignment.is_empty());
        assert!(coherence.conflict_detection.is_empty());
    }
}