trainer_v4_sgd_warm_restarts.py

# train_cantor_fusion_hf.py - PRODUCTION WITH ADAMW + WARM RESTARTS

"""
Cantor Fusion Classifier with AdamW + Cosine Warm Restarts
-----------------------------------------------------------
Features:
    - AdamW optimizer (best for ViTs)
    - CosineAnnealingWarmRestarts (automatic drop + restart cycles)
    - HuggingFace Hub uploads (ONE shared repo, organized by run)
    - TensorBoard logging (loss, accuracy, fusion metrics, LR tracking)
    - Easy CIFAR-10/100 switching
    - Automatic checkpoint management
    - SafeTensors format (ClamAV safe)

Author: AbstractPhil
License: MIT
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from torchvision import datasets, transforms
from torch.cuda.amp import autocast, GradScaler
from safetensors.torch import save_file, load_file

import math
import os
import json
from typing import Optional, Dict, List, Tuple, Union
from dataclasses import dataclass, asdict
import time
from pathlib import Path
from tqdm import tqdm

# HuggingFace
from huggingface_hub import HfApi, create_repo, upload_folder, upload_file
import yaml

# Import from your repo
from geovocab2.train.model.layers.attention.cantor_multiheaded_fusion import (
    CantorMultiheadFusion,
    CantorFusionConfig
)
from geovocab2.shapes.factory.cantor_route_factory import (
    CantorRouteFactory,
    RouteMode,
    SimplexConfig
)


# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# Configuration
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

@dataclass
class CantorTrainingConfig:
    """Complete configuration for Cantor fusion training with AdamW + Warm Restarts."""
    
    # Dataset
    dataset: str = "cifar10"  # "cifar10" or "cifar100"
    num_classes: int = 10
    
    # Architecture
    image_size: int = 32
    patch_size: int = 4
    embed_dim: int = 384
    num_fusion_blocks: int = 6
    num_heads: int = 8
    fusion_window: int = 32
    fusion_mode: str = "weighted"  # "weighted" or "consciousness"
    k_simplex: int = 4
    use_beatrix: bool = False
    beatrix_tau: float = 0.25
    
    # Optimization
    precompute_geometric: bool = True
    use_torch_compile: bool = True
    use_mixed_precision: bool = False
    
    # Regularization
    dropout: float = 0.1
    drop_path_rate: float = 0.1
    label_smoothing: float = 0.1
    
    # Training - Optimizer (AdamW)
    optimizer_type: str = "adamw"  # "sgd" or "adamw"
    batch_size: int = 128
    num_epochs: int = 300
    learning_rate: float = 3e-4  # AdamW default
    weight_decay: float = 0.05
    grad_clip: float = 1.0
    
    # SGD-specific (if needed)
    sgd_momentum: float = 0.9
    sgd_nesterov: bool = True
    
    # AdamW-specific  
    adamw_betas: Tuple[float, float] = (0.9, 0.999)
    adamw_eps: float = 1e-8
    
    # Learning rate schedule - WARM RESTARTS
    scheduler_type: str = "cosine_restarts"  # "multistep", "cosine", "cosine_restarts"
    
    # CosineAnnealingWarmRestarts parameters
    restart_period: int = 50  # T_0: epochs until first restart
    restart_mult: int = 2     # T_mult: multiply period after each restart
    min_lr: float = 1e-7      # eta_min: minimum learning rate
    
    # MultiStepLR (for SGD fallback)
    lr_milestones: List[int] = None
    lr_gamma: float = 0.2
    
    # Cosine annealing (regular, no restarts)
    warmup_epochs: int = 10
    
    # Data augmentation
    use_augmentation: bool = True
    use_autoaugment: bool = True
    use_cutout: bool = False
    cutout_length: int = 16
    
    # System
    device: str = "cuda" if torch.cuda.is_available() else "cpu"
    num_workers: int = 4
    seed: int = 42
    
    # Paths
    weights_dir: str = "weights"
    model_name: str = "vit-beans-v3"
    run_name: Optional[str] = None  # Auto-generated if None
    
    # HuggingFace - ONE SHARED REPO
    hf_username: str = "AbstractPhil"
    hf_repo_name: Optional[str] = None
    upload_to_hf: bool = True
    hf_token: Optional[str] = None
    
    # Logging
    log_interval: int = 50
    save_interval: int = 10
    checkpoint_upload_interval: int = 20
    
    def __post_init__(self):
        # Auto-set num_classes based on dataset
        if self.dataset == "cifar10":
            self.num_classes = 10
        elif self.dataset == "cifar100":
            self.num_classes = 100
        else:
            raise ValueError(f"Unknown dataset: {self.dataset}")
        
        # Set default milestones if None (for multistep fallback)
        if self.lr_milestones is None:
            if self.num_epochs >= 200:
                self.lr_milestones = [60, 120, 160]
            elif self.num_epochs >= 100:
                self.lr_milestones = [30, 60, 80]
            else:
                self.lr_milestones = [
                    int(self.num_epochs * 0.5),
                    int(self.num_epochs * 0.75)
                ]
        
        # Auto-generate run name
        if self.run_name is None:
            timestamp = time.strftime("%Y%m%d_%H%M%S")
            opt_name = self.optimizer_type.upper()
            sched_name = "WarmRestart" if self.scheduler_type == "cosine_restarts" else self.scheduler_type
            self.run_name = f"{self.dataset}_{self.fusion_mode}_{opt_name}_{sched_name}_{timestamp}"
        
        # ONE SHARED REPO for all runs
        if self.hf_repo_name is None:
            self.hf_repo_name = self.model_name
        
        # Set HF token from environment if not provided
        if self.hf_token is None:
            self.hf_token = os.environ.get("HF_TOKEN")
        
        # Calculate derived values
        assert self.image_size % self.patch_size == 0
        self.num_patches = (self.image_size // self.patch_size) ** 2
        self.patch_dim = self.patch_size * self.patch_size * 3
        
        # Create paths
        self.output_dir = Path(self.weights_dir) / self.model_name / self.run_name
        self.checkpoint_dir = self.output_dir / "checkpoints"
        self.tensorboard_dir = self.output_dir / "tensorboard"
        
        # Create directories
        self.output_dir.mkdir(parents=True, exist_ok=True)
        self.checkpoint_dir.mkdir(parents=True, exist_ok=True)
        self.tensorboard_dir.mkdir(parents=True, exist_ok=True)
    
    def save(self, path: Union[str, Path]):
        """Save config to YAML file."""
        path = Path(path)
        config_dict = asdict(self)
        # Convert tuples to lists for YAML
        if 'adamw_betas' in config_dict:
            config_dict['adamw_betas'] = list(config_dict['adamw_betas'])
        with open(path, 'w') as f:
            yaml.dump(config_dict, f, default_flow_style=False)
    
    @classmethod
    def load(cls, path: Union[str, Path]):
        """Load config from YAML file."""
        path = Path(path)
        with open(path, 'r') as f:
            config_dict = yaml.safe_load(f)
        # Convert lists back to tuples
        if 'adamw_betas' in config_dict:
            config_dict['adamw_betas'] = tuple(config_dict['adamw_betas'])
        return cls(**config_dict)


# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# Model Components (unchanged from previous version)
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

class PatchEmbedding(nn.Module):
    """Patch embedding layer."""
    def __init__(self, config: CantorTrainingConfig):
        super().__init__()
        self.config = config
        self.proj = nn.Conv2d(3, config.embed_dim, kernel_size=config.patch_size, stride=config.patch_size)
        self.pos_embed = nn.Parameter(torch.randn(1, config.num_patches, config.embed_dim) * 0.02)
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.proj(x)
        x = x.flatten(2).transpose(1, 2)
        x = x + self.pos_embed
        return x


class DropPath(nn.Module):
    """Stochastic depth."""
    def __init__(self, drop_prob: float = 0.0):
        super().__init__()
        self.drop_prob = drop_prob

    def forward(self, x):
        if self.drop_prob == 0. or not self.training:
            return x
        keep_prob = 1 - self.drop_prob
        shape = (x.shape[0],) + (1,) * (x.ndim - 1)
        random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
        random_tensor.floor_()
        return x.div(keep_prob) * random_tensor


class CantorFusionBlock(nn.Module):
    """Cantor fusion block."""
    def __init__(self, config: CantorTrainingConfig, drop_path: float = 0.0):
        super().__init__()
        self.norm1 = nn.LayerNorm(config.embed_dim)
        
        fusion_config = CantorFusionConfig(
            dim=config.embed_dim,
            num_heads=config.num_heads,
            fusion_window=config.fusion_window,
            fusion_mode=config.fusion_mode,
            k_simplex=config.k_simplex,
            use_beatrix_routing=config.use_beatrix,
            use_consciousness_weighting=(config.fusion_mode == "consciousness"),
            beatrix_tau=config.beatrix_tau,
            use_gating=True,
            dropout=config.dropout,
            residual=False,
            precompute_staircase=config.precompute_geometric,
            precompute_routes=config.precompute_geometric,
            precompute_distances=config.precompute_geometric,
            use_optimized_gather=True,
            staircase_cache_sizes=[config.num_patches],
            use_torch_compile=config.use_torch_compile
        )
        self.fusion = CantorMultiheadFusion(fusion_config)
        
        self.norm2 = nn.LayerNorm(config.embed_dim)
        mlp_hidden = config.embed_dim * 4
        self.mlp = nn.Sequential(
            nn.Linear(config.embed_dim, mlp_hidden),
            nn.GELU(),
            nn.Dropout(config.dropout),
            nn.Linear(mlp_hidden, config.embed_dim),
            nn.Dropout(config.dropout)
        )
        self.drop_path = DropPath(drop_path) if drop_path > 0 else nn.Identity()
    
    def forward(self, x: torch.Tensor, return_fusion_info: bool = False) -> Union[torch.Tensor, Tuple[torch.Tensor, Dict]]:
        fusion_result = self.fusion(self.norm1(x))
        x = x + self.drop_path(fusion_result['output'])
        x = x + self.drop_path(self.mlp(self.norm2(x)))
        
        if return_fusion_info:
            fusion_info = {
                'consciousness': fusion_result.get('consciousness'),
                'cantor_measure': fusion_result.get('cantor_measure')
            }
            return x, fusion_info
        return x


class CantorClassifier(nn.Module):
    """Cantor fusion classifier."""
    def __init__(self, config: CantorTrainingConfig):
        super().__init__()
        self.config = config
        
        self.patch_embed = PatchEmbedding(config)
        
        dpr = [x.item() for x in torch.linspace(0, config.drop_path_rate, config.num_fusion_blocks)]
        self.blocks = nn.ModuleList([
            CantorFusionBlock(config, drop_path=dpr[i])
            for i in range(config.num_fusion_blocks)
        ])
        
        self.norm = nn.LayerNorm(config.embed_dim)
        self.head = nn.Linear(config.embed_dim, config.num_classes)
        
        self.apply(self._init_weights)
    
    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            nn.init.trunc_normal_(m.weight, std=0.02)
            if m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)
        elif isinstance(m, nn.Conv2d):
            nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
    
    def forward(self, x: torch.Tensor, return_fusion_info: bool = False) -> Union[torch.Tensor, Tuple[torch.Tensor, List[Dict]]]:
        x = self.patch_embed(x)
        
        fusion_infos = []
        for i, block in enumerate(self.blocks):
            if return_fusion_info and i == len(self.blocks) - 1:
                x, fusion_info = block(x, return_fusion_info=True)
                fusion_infos.append(fusion_info)
            else:
                x = block(x)
        
        x = self.norm(x)
        x = x.mean(dim=1)
        logits = self.head(x)
        
        if return_fusion_info:
            return logits, fusion_infos
        return logits


# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# HuggingFace Integration
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

class HuggingFaceUploader:
    """Manages HuggingFace Hub uploads to ONE shared repo."""
    
    def __init__(self, config: CantorTrainingConfig):
        self.config = config
        self.api = HfApi(token=config.hf_token) if config.upload_to_hf else None
        self.repo_id = f"{config.hf_username}/{config.hf_repo_name}"
        self.run_prefix = f"runs/{config.run_name}"
        
        if config.upload_to_hf:
            self._create_repo()
            self._update_main_readme()
    
    def _create_repo(self):
        """Create HuggingFace repo if it doesn't exist."""
        try:
            create_repo(
                repo_id=self.repo_id,
                token=self.config.hf_token,
                exist_ok=True,
                private=False
            )
            print(f"[HF] Repository: https://huggingface.co/{self.repo_id}")
            print(f"[HF] Run folder: {self.run_prefix}")
        except Exception as e:
            print(f"[HF] Warning: Could not create repo: {e}")
    
    def _update_main_readme(self):
        """Create or update the main shared README at repo root."""
        if not self.config.upload_to_hf or self.api is None:
            return
        
        main_readme = f"""---
tags:
- image-classification
- cantor-fusion
- geometric-deep-learning
- safetensors
- vision-transformer
- warm-restarts
library_name: pytorch
datasets:
- cifar10
- cifar100
metrics:
- accuracy
---

# {self.config.hf_repo_name}

**Geometric Deep Learning with Cantor Multihead Fusion + AdamW Warm Restarts**

This repository contains multiple training runs using Cantor fusion architecture with pentachoron structures, geometric routing, and **CosineAnnealingWarmRestarts** for automatic exploration cycles.

## Training Strategy: AdamW + Warm Restarts

This model uses **AdamW with Cosine Annealing Warm Restarts** (SGDR):
- **Drop phase**: LR decays from {self.config.learning_rate} → {self.config.min_lr} over {self.config.restart_period} epochs
- **Restart phase**: LR jumps back to {self.config.learning_rate} to explore new regions
- **Cycle multiplier**: Each cycle is {self.config.restart_mult}x longer than previous
- **Benefits**: Automatic exploration + exploitation, finds better minima, robust training

### Restart Schedule
```
Epochs 0-{self.config.restart_period}:   LR: {self.config.learning_rate} → {self.config.min_lr} (first cycle)
Epoch {self.config.restart_period}:      LR: RESTART to {self.config.learning_rate} 🔄
Epochs {self.config.restart_period}-{self.config.restart_period * (1 + self.config.restart_mult)}: LR: {self.config.learning_rate} → {self.config.min_lr} (longer cycle)
...
```

## Current Run

**Latest**: `{self.config.run_name}`
- **Dataset**: {self.config.dataset.upper()}
- **Fusion Mode**: {self.config.fusion_mode}
- **Optimizer**: AdamW (adaptive moments)
- **Scheduler**: CosineAnnealingWarmRestarts
- **Architecture**: {self.config.num_fusion_blocks} blocks, {self.config.num_heads} heads
- **Simplex**: {self.config.k_simplex}-simplex ({self.config.k_simplex + 1} vertices)

## Architecture

The Cantor Fusion architecture uses:
- **Geometric Routing**: Pentachoron (5-simplex) structures for token routing
- **Cantor Multihead Fusion**: Multiple fusion heads with geometric attention
- **Beatrix Consciousness Routing**: Optional consciousness-aware token fusion
- **SafeTensors Format**: All model weights use SafeTensors (not pickle)

## Usage
```python
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

model_path = hf_hub_download(
    repo_id="{self.repo_id}",
    filename="runs/YOUR_RUN_NAME/checkpoints/best_model.safetensors"
)

state_dict = load_file(model_path)
model.load_state_dict(state_dict)
```

## Citation
```bibtex
@misc{{{self.config.hf_repo_name.replace('-', '_')},
  author = {{AbstractPhil}},
  title = {{{self.config.hf_repo_name}: Geometric Deep Learning with Warm Restarts}},
  year = {{2025}},
  publisher = {{HuggingFace}},
  url = {{https://huggingface.co/{self.repo_id}}}
}}
```

---

**Repository maintained by**: [@{self.config.hf_username}](https://huggingface.co/{self.config.hf_username})

**Latest update**: {time.strftime("%Y-%m-%d %H:%M:%S")}
"""
        
        main_readme_path = Path(self.config.weights_dir) / self.config.model_name / "MAIN_README.md"
        main_readme_path.parent.mkdir(parents=True, exist_ok=True)
        with open(main_readme_path, 'w') as f:
            f.write(main_readme)
        
        try:
            upload_file(
                path_or_fileobj=str(main_readme_path),
                path_in_repo="README.md",
                repo_id=self.repo_id,
                token=self.config.hf_token
            )
            print(f"[HF] Updated main README")
        except Exception as e:
            print(f"[HF] Main README upload failed: {e}")
    
    def upload_file(self, file_path: Path, repo_path: str):
        """Upload single file to HuggingFace."""
        if not self.config.upload_to_hf or self.api is None:
            return
        
        try:
            if not repo_path.startswith(self.run_prefix) and not repo_path.startswith("runs/"):
                full_path = f"{self.run_prefix}/{repo_path}"
            else:
                full_path = repo_path
            
            upload_file(
                path_or_fileobj=str(file_path),
                path_in_repo=full_path,
                repo_id=self.repo_id,
                token=self.config.hf_token
            )
            print(f"[HF] ✓ Uploaded: {full_path}")
        except Exception as e:
            print(f"[HF] ✗ Upload failed ({full_path}): {e}")
    
    def upload_folder_contents(self, folder_path: Path, repo_folder: str):
        """Upload entire folder to HuggingFace."""
        if not self.config.upload_to_hf or self.api is None:
            return
        
        try:
            full_path = f"{self.run_prefix}/{repo_folder}"
            upload_folder(
                folder_path=str(folder_path),
                repo_id=self.repo_id,
                path_in_repo=full_path,
                token=self.config.hf_token,
                ignore_patterns=["*.pyc", "__pycache__"]
            )
            print(f"[HF] Uploaded folder: {full_path}")
        except Exception as e:
            print(f"[HF] Folder upload failed: {e}")
    
    def create_model_card(self, trainer_stats: Dict):
        """Create and upload run-specific model card."""
        if not self.config.upload_to_hf:
            return
        
        run_card = f"""# Run: {self.config.run_name}

## Configuration
- **Dataset**: {self.config.dataset.upper()}
- **Fusion Mode**: {self.config.fusion_mode}
- **Parameters**: {trainer_stats['total_params']:,}
- **Simplex**: {self.config.k_simplex}-simplex ({self.config.k_simplex + 1} vertices)

## Performance
- **Best Validation Accuracy**: {trainer_stats['best_acc']:.2f}%
- **Training Time**: {trainer_stats['training_time']:.1f} hours
- **Final Epoch**: {trainer_stats['final_epoch']}

## Training Setup: AdamW + Warm Restarts
- **Optimizer**: AdamW (lr={self.config.learning_rate}, wd={self.config.weight_decay})
- **Scheduler**: CosineAnnealingWarmRestarts
- **Restart Period (T_0)**: {self.config.restart_period} epochs
- **Cycle Multiplier (T_mult)**: {self.config.restart_mult}x
- **Min LR**: {self.config.min_lr}
- **Batch Size**: {self.config.batch_size}
- **Mixed Precision**: {trainer_stats.get('mixed_precision', False)}

### Learning Rate Schedule
```
Cycle 1: Epochs 0-{self.config.restart_period}
  LR: {self.config.learning_rate} → {self.config.min_lr} (drop)
  Expected: Convergence to local minimum
  
Epoch {self.config.restart_period}: RESTART 🔄
  LR: {self.config.min_lr} → {self.config.learning_rate} (jump!)
  Expected: Escape local minimum, explore new regions
  
Cycle 2: Epochs {self.config.restart_period}-{self.config.restart_period * (1 + self.config.restart_mult)}
  LR: {self.config.learning_rate} → {self.config.min_lr} (longer cycle)
  Expected: Deeper convergence
  
... and so on
```

## Files
- `{self.run_prefix}/checkpoints/best_model.safetensors` - Model weights
- `{self.run_prefix}/checkpoints/best_training_state.pt` - Optimizer state
- `{self.run_prefix}/config.yaml` - Full configuration
- `{self.run_prefix}/tensorboard/` - TensorBoard logs (LR tracking!)

## Usage
```python
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="{self.repo_id}",
    filename="{self.run_prefix}/checkpoints/best_model.safetensors"
)

state_dict = load_file(model_path)
model.load_state_dict(state_dict)
```

## Training Notes

**Warm Restarts Benefits:**
- 🔄 **Exploration**: Periodic LR jumps escape local minima
- 📉 **Exploitation**: Long drop phases converge deeply
- 🎯 **Robustness**: Multiple restarts find better solutions
- 📊 **Monitoring**: Watch TensorBoard for restart effects!

**Expected Behavior:**
- Accuracy improves during each drop phase
- Brief accuracy dips after restarts (exploration)
- Overall upward trend across cycles
- Best models often found late in long cycles

---

Built with geometric consciousness-aware routing using the Devil's Staircase (Beatrix) and pentachoron parameterization.

**Training completed**: {time.strftime("%Y-%m-%d %H:%M:%S")}

[← Back to main repository](https://huggingface.co/{self.repo_id})
"""
        
        readme_path = self.config.output_dir / "RUN_README.md"
        with open(readme_path, 'w') as f:
            f.write(run_card)
        
        try:
            upload_file(
                path_or_fileobj=str(readme_path),
                path_in_repo=f"{self.run_prefix}/README.md",
                repo_id=self.repo_id,
                token=self.config.hf_token
            )
            print(f"[HF] Uploaded run README")
        except Exception as e:
            print(f"[HF] Run README upload failed: {e}")


# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# Trainer with AdamW + CosineAnnealingWarmRestarts
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

class Trainer:
    """Training manager with AdamW + Warm Restarts."""
    
    def __init__(self, config: CantorTrainingConfig):
        self.config = config
        self.device = torch.device(config.device)
        
        # Set seed
        torch.manual_seed(config.seed)
        if torch.cuda.is_available():
            torch.cuda.manual_seed(config.seed)
        
        # Model
        print("\n" + "=" * 70)
        print(f"Initializing Cantor Classifier - {config.dataset.upper()}")
        print("=" * 70)
        
        init_start = time.time()
        self.model = CantorClassifier(config).to(self.device)
        init_time = time.time() - init_start
        
        print(f"\n[Model] Initialization time: {init_time:.2f}s")
        self.print_model_info()
        
        # Track restart epochs for logging
        self.restart_epochs = self._calculate_restart_epochs()

        # Optimizer
        self.optimizer = self.create_optimizer()
        
        # Scheduler
        self.scheduler = self.create_scheduler()
        
        # Loss
        self.criterion = nn.CrossEntropyLoss(label_smoothing=config.label_smoothing)
        
        # Mixed precision
        self.use_amp = config.use_mixed_precision and config.device == "cuda"
        self.scaler = GradScaler() if self.use_amp else None
        
        if self.use_amp:
            print(f"[Training] Mixed precision enabled")
        
        # TensorBoard
        self.writer = SummaryWriter(log_dir=str(config.tensorboard_dir))
        print(f"[TensorBoard] Logging to: {config.tensorboard_dir}")
        print(f"[Checkpoints] Format: SafeTensors (ClamAV safe)")
        
        # HuggingFace
        self.hf_uploader = HuggingFaceUploader(config) if config.upload_to_hf else None
        
        # Save config
        config.save(config.output_dir / "config.yaml")
        
        # Metrics
        self.best_acc = 0.0
        self.global_step = 0
        self.start_time = time.time()
        self.upload_count = 0
        
    
    def _calculate_restart_epochs(self) -> List[int]:
        """Calculate when restarts will occur."""
        if self.config.scheduler_type != "cosine_restarts":
            return []
        
        restarts = []
        current = self.config.restart_period
        period = self.config.restart_period
        
        while current < self.config.num_epochs:
            restarts.append(current)
            period *= self.config.restart_mult
            current += period
        
        return restarts
    
    def create_optimizer(self):
        """Create optimizer based on config."""
        if self.config.optimizer_type == "sgd":
            print(f"\n[Optimizer] SGD")
            print(f"  LR: {self.config.learning_rate}")
            print(f"  Momentum: {self.config.sgd_momentum}")
            print(f"  Nesterov: {self.config.sgd_nesterov}")
            print(f"  Weight decay: {self.config.weight_decay}")
            
            return torch.optim.SGD(
                self.model.parameters(),
                lr=self.config.learning_rate,
                momentum=self.config.sgd_momentum,
                weight_decay=self.config.weight_decay,
                nesterov=self.config.sgd_nesterov
            )
        
        elif self.config.optimizer_type == "adamw":
            print(f"\n[Optimizer] AdamW")
            print(f"  LR: {self.config.learning_rate}")
            print(f"  Betas: {self.config.adamw_betas}")
            print(f"  Weight decay: {self.config.weight_decay}")
            
            return torch.optim.AdamW(
                self.model.parameters(),
                lr=self.config.learning_rate,
                betas=self.config.adamw_betas,
                eps=self.config.adamw_eps,
                weight_decay=self.config.weight_decay
            )
        
        else:
            raise ValueError(f"Unknown optimizer: {self.config.optimizer_type}")
    
    def create_scheduler(self):
        """Create LR scheduler based on config."""
        if self.config.scheduler_type == "cosine_restarts":
            print(f"\n[Scheduler] CosineAnnealingWarmRestarts")
            print(f"  T_0 (restart period): {self.config.restart_period} epochs")
            print(f"  T_mult (cycle multiplier): {self.config.restart_mult}x")
            print(f"  Min LR: {self.config.min_lr}")
            print(f"\n  Restart schedule:")
            for i, epoch in enumerate(self.restart_epochs[:5]):  # Show first 5
                print(f"    Restart #{i+1}: Epoch {epoch}")
            if len(self.restart_epochs) > 5:
                print(f"    ... and {len(self.restart_epochs) - 5} more")
            
            return torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(
                self.optimizer,
                T_0=self.config.restart_period,
                T_mult=self.config.restart_mult,
                eta_min=self.config.min_lr
            )
        
        elif self.config.scheduler_type == "multistep":
            print(f"\n[Scheduler] MultiStepLR")
            print(f"  Milestones: {self.config.lr_milestones}")
            print(f"  Gamma: {self.config.lr_gamma}")
            
            return torch.optim.lr_scheduler.MultiStepLR(
                self.optimizer,
                milestones=self.config.lr_milestones,
                gamma=self.config.lr_gamma
            )
        
        elif self.config.scheduler_type == "cosine":
            print(f"\n[Scheduler] Cosine annealing with warmup")
            print(f"  Warmup epochs: {self.config.warmup_epochs}")
            print(f"  Min LR: {self.config.min_lr}")
            
            def lr_lambda(epoch):
                if epoch < self.config.warmup_epochs:
                    return (epoch + 1) / self.config.warmup_epochs
                progress = (epoch - self.config.warmup_epochs) / (self.config.num_epochs - self.config.warmup_epochs)
                return 0.5 * (1 + math.cos(math.pi * progress))
            
            return torch.optim.lr_scheduler.LambdaLR(self.optimizer, lr_lambda)
        
        else:
            raise ValueError(f"Unknown scheduler: {self.config.scheduler_type}")
    
    def print_model_info(self):
        """Print model info."""
        total_params = sum(p.numel() for p in self.model.parameters())
        print(f"\nParameters: {total_params:,}")
        print(f"Dataset: {self.config.dataset.upper()}")
        print(f"Classes: {self.config.num_classes}")
        print(f"Fusion mode: {self.config.fusion_mode}")
        print(f"Optimizer: {self.config.optimizer_type.upper()}")
        print(f"Scheduler: {self.config.scheduler_type}")
        print(f"Output: {self.config.output_dir}")
    
    def train_epoch(self, train_loader: DataLoader, epoch: int) -> Tuple[float, float]:
        """Train one epoch."""
        self.model.train()
        total_loss, correct, total = 0.0, 0, 0
        
        # Check if this is a restart epoch
        is_restart = (epoch in self.restart_epochs)
        epoch_desc = f"Epoch {epoch+1}/{self.config.num_epochs}"
        if is_restart:
            epoch_desc += " 🔄 RESTART"
        
        pbar = tqdm(train_loader, desc=f"{epoch_desc} [Train]")
        
        for batch_idx, (images, labels) in enumerate(pbar):
            images, labels = images.to(self.device, non_blocking=True), labels.to(self.device, non_blocking=True)
            
            # Forward
            if self.use_amp:
                with autocast():
                    logits = self.model(images)
                    loss = self.criterion(logits, labels)
                self.optimizer.zero_grad(set_to_none=True)
                self.scaler.scale(loss).backward()
                self.scaler.unscale_(self.optimizer)
                torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.config.grad_clip)
                self.scaler.step(self.optimizer)
                self.scaler.update()
            else:
                logits = self.model(images)
                loss = self.criterion(logits, labels)
                self.optimizer.zero_grad(set_to_none=True)
                loss.backward()
                torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.config.grad_clip)
                self.optimizer.step()
            
            # Metrics
            total_loss += loss.item()
            _, predicted = logits.max(1)
            correct += predicted.eq(labels).sum().item()
            total += labels.size(0)
            
            # TensorBoard logging
            if batch_idx % self.config.log_interval == 0:
                current_lr = self.scheduler.get_last_lr()[0]
                self.writer.add_scalar('train/loss', loss.item(), self.global_step)
                self.writer.add_scalar('train/accuracy', 100. * correct / total, self.global_step)
                self.writer.add_scalar('train/learning_rate', current_lr, self.global_step)
            
            self.global_step += 1
            
            pbar.set_postfix({
                'loss': f'{loss.item():.4f}',
                'acc': f'{100. * correct / total:.2f}%',
                'lr': f'{self.scheduler.get_last_lr()[0]:.6f}'
            })
        
        return total_loss / len(train_loader), 100. * correct / total
    
    @torch.no_grad()
    def evaluate(self, val_loader: DataLoader, epoch: int) -> Tuple[float, Dict]:
        """Evaluate."""
        self.model.eval()
        total_loss, correct, total = 0.0, 0, 0
        consciousness_values = []
        
        pbar = tqdm(val_loader, desc=f"Epoch {epoch+1}/{self.config.num_epochs} [Val]  ")
        
        for batch_idx, (images, labels) in enumerate(pbar):
            images, labels = images.to(self.device, non_blocking=True), labels.to(self.device, non_blocking=True)
            
            # Forward with fusion info on last batch
            return_info = (batch_idx == len(val_loader) - 1)
            
            if self.use_amp:
                with autocast():
                    if return_info:
                        logits, fusion_infos = self.model(images, return_fusion_info=True)
                        if fusion_infos and fusion_infos[0].get('consciousness') is not None:
                            consciousness_values.append(fusion_infos[0]['consciousness'].mean().item())
                    else:
                        logits = self.model(images)
                    loss = self.criterion(logits, labels)
            else:
                if return_info:
                    logits, fusion_infos = self.model(images, return_fusion_info=True)
                    if fusion_infos and fusion_infos[0].get('consciousness') is not None:
                        consciousness_values.append(fusion_infos[0]['consciousness'].mean().item())
                else:
                    logits = self.model(images)
                loss = self.criterion(logits, labels)
            
            total_loss += loss.item()
            _, predicted = logits.max(1)
            correct += predicted.eq(labels).sum().item()
            total += labels.size(0)
            
            pbar.set_postfix({
                'loss': f'{total_loss / (batch_idx + 1):.4f}',
                'acc': f'{100. * correct / total:.2f}%'
            })
        
        avg_loss = total_loss / len(val_loader)
        accuracy = 100. * correct / total
        
        # TensorBoard logging
        self.writer.add_scalar('val/loss', avg_loss, epoch)
        self.writer.add_scalar('val/accuracy', accuracy, epoch)
        if consciousness_values:
            self.writer.add_scalar('val/consciousness', sum(consciousness_values) / len(consciousness_values), epoch)
        
        metrics = {
            'loss': avg_loss,
            'accuracy': accuracy,
            'consciousness': sum(consciousness_values) / len(consciousness_values) if consciousness_values else None
        }
        
        return accuracy, metrics
    
    def train(self, train_loader: DataLoader, val_loader: DataLoader):
        """Full training loop."""
        print("\n" + "=" * 70)
        print("Starting training with AdamW + Warm Restarts")
        print(f"Optimizer: {self.config.optimizer_type.upper()}")
        print(f"Scheduler: {self.config.scheduler_type}")
        print(f"Restart period: {self.config.restart_period} epochs (T_0)")
        print(f"Cycle multiplier: {self.config.restart_mult}x (T_mult)")
        print(f"Total restarts: {len(self.restart_epochs)}")
        print("=" * 70 + "\n")
        
        for epoch in range(self.config.num_epochs):
            # Train
            train_loss, train_acc = self.train_epoch(train_loader, epoch)
            
            # Evaluate
            val_acc, val_metrics = self.evaluate(val_loader, epoch)
            
            # Update scheduler
            self.scheduler.step()
            
            # Check if this is a restart epoch
            is_restart = (epoch in self.restart_epochs)
            next_lr = self.scheduler.get_last_lr()[0]
            
            # Print summary
            print(f"\n{'='*70}")
            print(f"Epoch [{epoch + 1}/{self.config.num_epochs}] Summary:")
            print(f"  Train: Loss={train_loss:.4f}, Acc={train_acc:.2f}%")
            print(f"  Val:   Loss={val_metrics['loss']:.4f}, Acc={val_acc:.2f}%")
            if val_metrics['consciousness']:
                print(f"  Consciousness: {val_metrics['consciousness']:.4f}")
            
            if is_restart:
                print(f"  🔄 WARM RESTART! Next LR: {next_lr:.6f}")
                print(f"     (Escaping local minimum, exploring new regions)")
            else:
                print(f"  Current LR: {next_lr:.6f}")
            
            # Checkpoint logic
            is_best = val_acc > self.best_acc
            should_save_regular = ((epoch + 1) % self.config.save_interval == 0)
            should_upload_regular = ((epoch + 1) % self.config.checkpoint_upload_interval == 0)
            
            if is_best:
                self.best_acc = val_acc
                print(f"  ✓ New best model! Accuracy: {val_acc:.2f}%")
                self.save_checkpoint(epoch, val_acc, prefix="best", upload=should_upload_regular)
            
            if should_save_regular:
                self.save_checkpoint(epoch, val_acc, prefix=f"epoch_{epoch+1}", upload=should_upload_regular)
            
            print(f"  HF Uploads: {self.upload_count}")
            print(f"{'='*70}\n")
            
            # Flush TensorBoard
            if (epoch + 1) % 10 == 0:
                self.writer.flush()
        
        # Training complete
        training_time = (time.time() - self.start_time) / 3600
        
        print("\n" + "=" * 70)
        print("Training Complete!")
        print(f"Best Validation Accuracy: {self.best_acc:.2f}%")
        print(f"Training Time: {training_time:.2f} hours")
        print(f"Total Uploads: {self.upload_count}")
        print(f"Warm Restarts: {len(self.restart_epochs)}")
        print("=" * 70)
        
        # Upload to HuggingFace
        if self.hf_uploader:
            print("\n[HF] Uploading final best model...")
            best_model_path = self.config.checkpoint_dir / "best_model.safetensors"
            best_state_path = self.config.checkpoint_dir / "best_training_state.pt"
            best_metadata_path = self.config.checkpoint_dir / "best_metadata.json"
            config_path = self.config.output_dir / "config.yaml"
            
            if best_model_path.exists():
                self.hf_uploader.upload_file(best_model_path, "checkpoints/best_model.safetensors")
            if best_state_path.exists():
                self.hf_uploader.upload_file(best_state_path, "checkpoints/best_training_state.pt")
            if best_metadata_path.exists():
                self.hf_uploader.upload_file(best_metadata_path, "checkpoints/best_metadata.json")
            if config_path.exists():
                self.hf_uploader.upload_file(config_path, "config.yaml")
            
            print("[HF] Final upload: TensorBoard logs...")
            self.hf_uploader.upload_folder_contents(self.config.tensorboard_dir, "tensorboard")
            
            trainer_stats = {
                'total_params': sum(p.numel() for p in self.model.parameters()),
                'best_acc': self.best_acc,
                'training_time': training_time,
                'final_epoch': self.config.num_epochs,
                'batch_size': self.config.batch_size,
                'mixed_precision': self.use_amp
            }
            self.hf_uploader.create_model_card(trainer_stats)
        
        self.writer.close()
    
    def save_checkpoint(self, epoch: int, accuracy: float, prefix: str = "checkpoint", upload: bool = False):
        """Save checkpoint as safetensors with selective upload."""
        checkpoint_dir = self.config.checkpoint_dir
        checkpoint_dir.mkdir(parents=True, exist_ok=True)
        
        # 1. Save model weights as safetensors
        model_path = checkpoint_dir / f"{prefix}_model.safetensors"
        save_file(self.model.state_dict(), str(model_path))
        
        # 2. Save optimizer/scheduler state
        training_state = {
            'optimizer_state_dict': self.optimizer.state_dict(),
            'scheduler_state_dict': self.scheduler.state_dict(),
        }
        if self.scaler is not None:
            training_state['scaler_state_dict'] = self.scaler.state_dict()
        
        training_state_path = checkpoint_dir / f"{prefix}_training_state.pt"
        torch.save(training_state, training_state_path)
        
        # 3. Save metadata
        metadata = {
            'epoch': epoch,
            'accuracy': accuracy,
            'best_accuracy': self.best_acc,
            'global_step': self.global_step,
            'timestamp': time.strftime("%Y-%m-%d %H:%M:%S"),
            'optimizer': self.config.optimizer_type,
            'scheduler': self.config.scheduler_type,
            'learning_rate': self.scheduler.get_last_lr()[0]
        }
        metadata_path = checkpoint_dir / f"{prefix}_metadata.json"
        with open(metadata_path, 'w') as f:
            json.dump(metadata, f, indent=2)
        
        is_best = (prefix == "best")
        
        if is_best:
            print(f"  💾 Saved best: {prefix}_model.safetensors")
        else:
            print(f"  💾 Saved: {prefix}_model.safetensors", end="")
        
        # Upload to HuggingFace
        if self.hf_uploader and upload:
            self.hf_uploader.upload_file(model_path, f"checkpoints/{prefix}_model.safetensors")
            self.hf_uploader.upload_file(training_state_path, f"checkpoints/{prefix}_training_state.pt")
            self.hf_uploader.upload_file(metadata_path, f"checkpoints/{prefix}_metadata.json")
            
            if is_best:
                config_path = self.config.output_dir / "config.yaml"
                if config_path.exists():
                    self.hf_uploader.upload_file(config_path, "config.yaml")
            
            self.upload_count += 1
            
            if not is_best:
                print(" → Uploaded to HF")
        else:
            if not is_best:
                print(" (local only)")


# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# Data Loading (with Cutout)
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

class Cutout:
    """Cutout data augmentation."""
    def __init__(self, length: int):
        self.length = length
    
    def __call__(self, img):
        h, w = img.size(1), img.size(2)
        mask = torch.ones((h, w), dtype=torch.float32)
        y = torch.randint(h, (1,)).item()
        x = torch.randint(w, (1,)).item()
        
        y1 = max(0, y - self.length // 2)
        y2 = min(h, y + self.length // 2)
        x1 = max(0, x - self.length // 2)
        x2 = min(w, x + self.length // 2)
        
        mask[y1:y2, x1:x2] = 0.
        mask = mask.expand_as(img)
        return img * mask


def get_data_loaders(config: CantorTrainingConfig) -> Tuple[DataLoader, DataLoader]:
    """Create data loaders."""
    
    # Normalization
    mean = (0.4914, 0.4822, 0.4465)
    std = (0.2470, 0.2435, 0.2616)
    
    # Augmentation
    if config.use_augmentation:
        transforms_list = []
        
        if config.use_autoaugment:
            policy = transforms.AutoAugmentPolicy.CIFAR10
            transforms_list.append(transforms.AutoAugment(policy))
        else:
            transforms_list.extend([
                transforms.RandomCrop(32, padding=4),
                transforms.RandomHorizontalFlip(),
            ])
        
        transforms_list.append(transforms.ToTensor())
        transforms_list.append(transforms.Normalize(mean, std))
        
        if config.use_cutout:
            transforms_list.append(Cutout(config.cutout_length))
        
        train_transform = transforms.Compose(transforms_list)
    else:
        train_transform = transforms.Compose([
            transforms.ToTensor(),
            transforms.Normalize(mean, std)
        ])
    
    val_transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize(mean, std)
    ])
    
    # Dataset selection
    if config.dataset == "cifar10":
        train_dataset = datasets.CIFAR10(root='./data', train=True, download=True, transform=train_transform)
        val_dataset = datasets.CIFAR10(root='./data', train=False, download=True, transform=val_transform)
    elif config.dataset == "cifar100":
        train_dataset = datasets.CIFAR100(root='./data', train=True, download=True, transform=train_transform)
        val_dataset = datasets.CIFAR100(root='./data', train=False, download=True, transform=val_transform)
    else:
        raise ValueError(f"Unknown dataset: {config.dataset}")
    
    train_loader = DataLoader(
        train_dataset,
        batch_size=config.batch_size,
        shuffle=True,
        num_workers=config.num_workers,
        pin_memory=(config.device == "cuda")
    )
    
    val_loader = DataLoader(
        val_dataset,
        batch_size=config.batch_size,
        shuffle=False,
        num_workers=config.num_workers,
        pin_memory=(config.device == "cuda")
    )
    
    return train_loader, val_loader


# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# Main - AdamW + CosineAnnealingWarmRestarts
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

def main():
    """Main training function with AdamW + Warm Restarts."""
    
    # ═══════════════════════════════════════════════════════════════════
    # Configuration - AdamW with Cosine Annealing Warm Restarts (SGDR)
    # ═══════════════════════════════════════════════════════════════════
    
    config = CantorTrainingConfig(
        # Dataset
        dataset="cifar100",  # "cifar10" or "cifar100"
        
        # Architecture
        embed_dim=512,
        num_fusion_blocks=6,
        num_heads=8,
        fusion_mode="weighted",  # "weighted" or "consciousness"
        k_simplex=4,
        use_beatrix=True,
        
        # Optimizer: AdamW (best for ViTs)
        optimizer_type="adamw",
        learning_rate=3e-4,  # Standard ViT learning rate
        weight_decay=0.05,   # Standard ViT weight decay
        adamw_betas=(0.9, 0.999),
        
        # Scheduler: Cosine Annealing with Warm Restarts (SGDR)
        scheduler_type="cosine_restarts",  # THE KEY!
        restart_period=20,   # T_0: First restart at epoch 50
        restart_mult=2,      # T_mult: Each cycle 2x longer
        min_lr=1e-7,         # Minimum LR at end of each cycle
        
        # Training
        num_epochs=300,      # Longer training with multiple cycles
        batch_size=128,
        grad_clip=1.0,
        label_smoothing=0.1,
        
        # Augmentation
        use_augmentation=True,
        use_autoaugment=True,
        use_cutout=False,
        cutout_length=16,
        
        # Regularization
        dropout=0.1,
        drop_path_rate=0.1,
        
        # System
        device="cuda",
        use_mixed_precision=False,
        
        # HuggingFace
        hf_username="AbstractPhil",
        upload_to_hf=True,
        checkpoint_upload_interval=25,  # Upload every 25 epochs
    )
    
    print("=" * 70)
    print(f"Cantor Fusion Classifier - {config.dataset.upper()}")
    print("Training Strategy: AdamW + Cosine Annealing Warm Restarts (SGDR)")
    print("=" * 70)
    print(f"\nConfiguration:")
    print(f"  Dataset: {config.dataset}")
    print(f"  Fusion mode: {config.fusion_mode}")
    print(f"  Optimizer: AdamW")
    print(f"  Scheduler: CosineAnnealingWarmRestarts")
    print(f"  Initial LR: {config.learning_rate}")
    print(f"  Min LR: {config.min_lr}")
    print(f"  Restart period (T_0): {config.restart_period} epochs")
    print(f"  Cycle multiplier (T_mult): {config.restart_mult}x")
    print(f"  Total epochs: {config.num_epochs}")
    
    # Calculate restart schedule
    restarts = []
    current = config.restart_period
    period = config.restart_period
    while current < config.num_epochs:
        restarts.append(current)
        period *= config.restart_mult
        current += period
    
    print(f"\n  Restart schedule ({len(restarts)} restarts):")
    for i, epoch in enumerate(restarts[:5]):
        print(f"    Restart #{i+1}: Epoch {epoch}")
    if len(restarts) > 5:
        print(f"    ... and {len(restarts) - 5} more")
    
    print(f"\n  Output: {config.output_dir}")
    print(f"  HuggingFace: {'Enabled' if config.upload_to_hf else 'Disabled'}")
    if config.upload_to_hf:
        print(f"  Repo: {config.hf_username}/{config.hf_repo_name}")
        print(f"  Run: {config.run_name}")
    
    print("\n" + "=" * 70)
    print("Expected Training Behavior:")
    print("=" * 70)
    print("📉 Cycle 1 (epochs 0-50):")
    print("   LR: 3e-4 → 1e-7 (smooth drop)")
    print("   Expected: Convergence to local minimum")
    print("")
    print("🔄 Epoch 50: RESTART!")
    print("   LR: 1e-7 → 3e-4 (jump back up)")
    print("   Expected: Escape local minimum, explore new regions")
    print("")
    print("📉 Cycle 2 (epochs 50-150):")
    print("   LR: 3e-4 → 1e-7 (longer cycle, deeper convergence)")
    print("")
    print("🔄 Epoch 150: RESTART!")
    print("   LR: 1e-7 → 3e-4")
    print("")
    print("📉 Cycle 3 (epochs 150-350+):")
    print("   LR: 3e-4 → 1e-7 (longest cycle, final polish)")
    print("")
    print("🎯 Benefits:")
    print("   - Multiple exploration cycles find better minima")
    print("   - Automatic drop + restart (no manual tuning)")
    print("   - Robust to different datasets/architectures")
    print("=" * 70)
    
    # Load data
    print("\nLoading data...")
    train_loader, val_loader = get_data_loaders(config)
    print(f"  Train: {len(train_loader.dataset)} samples")
    print(f"  Val: {len(val_loader.dataset)} samples")
    
    # Train
    trainer = Trainer(config)
    trainer.train(train_loader, val_loader)
    
    print("\n" + "=" * 70)
    print("🎯 Training complete!")
    print("   Check TensorBoard to see the warm restart cycles!")
    print(f"   tensorboard --logdir {config.tensorboard_dir}")
    print("")
    print("   Look for:")
    print("   - Smooth LR drops during each cycle")
    print("   - Sharp LR jumps at restart epochs")
    print("   - Accuracy improvements across cycles")
    print("=" * 70)


if __name__ == "__main__":
    main()