code/src/enhancements/curriculum_learning.py

#!/usr/bin/env python3
"""#16 Curriculum Learning — easy-to-hard scheduling.

Reference: Bengio 2009; Hacohen 2019; SuperLoss (Castells 2020).
Düşük-uncertainty / yüksek-quality örneklerle başla, zorlaşır.

Kullanılan alanlar:
- quality_score (composite)
- uncertainty_score (heuristic / model-based)
- class_frequency_tier
"""
import json
from pathlib import Path

ROOT = Path("/arf/scratch/stakan/hitit-proje")

class CurriculumSampler:
    """Epoch'a göre sample pool'u genişleyen curriculum sampler."""
    
    def __init__(self, records, difficulty_fn=None):
        self.records = records
        self.difficulty_fn = difficulty_fn or self._default_difficulty
        self.sorted_indices = self._sort_by_difficulty()
    
    def _default_difficulty(self, r):
        """Low = easy, high = hard."""
        qs = r.get('quality_score', 0.5)
        us = r.get('uncertainty_score', 0.5)
        tier = r.get('class_frequency_tier', 'mid')
        tier_bonus = {'head': 0.0, 'mid': 0.2, 'tail': 0.5, 'rare': 0.8}.get(tier, 0.3)
        # Hard: low quality + high uncertainty + rare tier
        return (1 - qs) * 0.4 + us * 0.3 + tier_bonus * 0.3
    
    def _sort_by_difficulty(self):
        difficulties = [(i, self.difficulty_fn(r)) for i, r in enumerate(self.records)]
        difficulties.sort(key=lambda x: x[1])  # ascending: easy first
        return [i for i, _ in difficulties]
    
    def epoch_indices(self, epoch, total_epochs=100, warmup_ratio=0.2):
        """Curriculum: warmup_ratio kadar sadece en kolay, sonra linear genişle."""
        n = len(self.records)
        warmup_epochs = int(total_epochs * warmup_ratio)
        
        if epoch < warmup_epochs:
            # Sadece en kolay %50
            cutoff = n // 2
        else:
            # Linear expansion: warmup_epochs'tan sonra full pool
            progress = min(1.0, (epoch - warmup_epochs) / (total_epochs - warmup_epochs))
            cutoff = int(n // 2 + progress * (n - n // 2))
        
        return self.sorted_indices[:cutoff]

def main():
    # Sadece recipe dökümanı
    recipe = {
        "name": "Curriculum learning — quality+uncertainty+tier difficulty",
        "difficulty_formula": "(1 - quality_score) * 0.4 + uncertainty_score * 0.3 + tier_bonus * 0.3",
        "tier_bonus": {"head": 0.0, "mid": 0.2, "tail": 0.5, "rare": 0.8},
        "schedule": "First 20% epochs: easiest 50%. Then linear expansion to full pool by end.",
        "expected_gain": "+0.2-0.5% (literature mixed results)",
        "fields_used": ["quality_score", "uncertainty_score", "class_frequency_tier"],
    }
    with open(ROOT / "datasets/processed/curriculum_recipe.json", 'w') as f:
        json.dump(recipe, f, indent=2, ensure_ascii=False)
    print("Curriculum sampler hazır")

if __name__ == '__main__':
    main()