File size: 9,247 Bytes

d077f18

"""
Multimodal PC Fault Detection - Configuration v2
==================================================
All hyperparameters, dataset configs, and experiment settings in one place.

v2 additions:
  - OGM-GE hyperparameters (ogm_alpha, ogm_noise_sigma)
  - Auxiliary loss weights (lambda_visual, lambda_audio)
  - Asymmetric LR multipliers (visual_lr_multiplier, audio_lr_multiplier)
"""

from dataclasses import dataclass, field
from typing import List, Optional, Literal


# ============================================================================
# Fault Taxonomy
# ============================================================================
FAULT_CLASSES = [
    "normal_operation",     # 0: Normal PC running
    "boot_failure",         # 1: BIOS beep codes, POST failures
    "overheating_fan",      # 2: Fan noise anomalies, thermal warnings
    "storage_failure",      # 3: HDD clicking, drive errors
    "system_crash",         # 4: BSOD, system halts
]
NUM_CLASSES = len(FAULT_CLASSES)

# ============================================================================
# ESC-50 → PC Fault Class Mapping (Audio) — kept for backward compat
# ============================================================================
ESC50_TO_FAULT = {
    "keyboard_typing": 0, "mouse_click": 0,
    "clock_alarm": 1, "siren": 1,
    "vacuum_cleaner": 2, "engine": 2, "washing_machine": 2,
    "clock_tick": 3, "door_wood_knock": 3, "hand_saw": 3,
    "glass_breaking": 4, "fireworks": 4, "chainsaw": 4,
}

ESC50_CATEGORY_TO_TARGET = {
    "keyboard_typing": 32, "mouse_click": 33, "clock_alarm": 37,
    "siren": 42, "vacuum_cleaner": 36, "engine": 44,
    "washing_machine": 35, "clock_tick": 38, "door_wood_knock": 30,
    "hand_saw": 49, "glass_breaking": 39, "fireworks": 48, "chainsaw": 41,
}

# ============================================================================
# Visual Data Synthesis Config — kept for backward compat
# ============================================================================
VISUAL_SYNTHESIS = {
    "normal_operation": {
        "description": "Clean desktop, green status indicators, normal task manager",
        "color_dominant": (0, 128, 0),
        "text_overlay": ["System OK", "All services running", "Temperature: Normal"],
    },
    "boot_failure": {
        "description": "BIOS POST screen with error codes, black/blue background",
        "color_dominant": (0, 0, 0),
        "text_overlay": ["BIOS ERROR", "POST Code: 3-3-1", "Memory Test Failed"],
    },
    "overheating_fan": {
        "description": "Temperature warning, red thermal display, CPU throttling",
        "color_dominant": (255, 0, 0),
        "text_overlay": ["CRITICAL TEMP", "CPU: 98°C", "Thermal Throttling Active"],
    },
    "storage_failure": {
        "description": "Disk error screen, SMART warning, data recovery prompt",
        "color_dominant": (255, 165, 0),
        "text_overlay": ["DISK ERROR", "S.M.A.R.T. WARNING", "Sector Read Failure"],
    },
    "system_crash": {
        "description": "Blue screen of death, kernel panic, stop code",
        "color_dominant": (0, 120, 215),
        "text_overlay": ["STOP: 0x0000007E", "SYSTEM_THREAD_EXCEPTION_NOT_HANDLED",
                         "Your PC ran into a problem"],
    },
}


@dataclass
class DataConfig:
    esc50_dataset: str = "ashraq/esc50"
    audioset_dataset: str = "agkphysics/AudioSet"
    audioset_config: str = "balanced"
    sample_rate: int = 32000
    audio_duration: float = 5.0
    n_fft: int = 1024
    hop_length: int = 320
    n_mels: int = 64
    fmin: int = 50
    fmax: int = 14000
    image_size: int = 224
    val_fold: int = 5
    num_synthetic_per_class: int = 200
    audio_noise_snr_db: float = 10.0
    time_shift_max: float = 0.2
    freq_mask_max: int = 10
    time_mask_max: int = 20


@dataclass
class ModelConfig:
    vit_model_name: str = "google/vit-base-patch16-224-in21k"
    vit_embed_dim: int = 768
    ast_model_name: str = "MIT/ast-finetuned-audioset-10-10-0.4593"
    ast_embed_dim: int = 768
    fusion_type: Literal["concat", "weighted_sum", "attention"] = "concat"
    fusion_dim: int = 512
    fusion_dropout: float = 0.3
    num_classes: int = NUM_CLASSES
    modality_dropout_p: float = 0.3


@dataclass
class LoRAConfig:
    enabled: bool = True
    r: int = 8
    lora_alpha: int = 16
    lora_dropout: float = 0.1
    bias: str = "none"
    vit_target_modules: List[str] = field(default_factory=lambda: ["query", "value"])
    vit_modules_to_save: List[str] = field(default_factory=lambda: ["classifier"])
    ast_target_modules: List[str] = field(default_factory=lambda: ["query", "value"])
    ast_modules_to_save: List[str] = field(default_factory=lambda: ["classifier"])


@dataclass
class TrainConfig:
    mode: Literal["multimodal", "visual_only", "audio_only"] = "multimodal"
    finetune_method: Literal["lora", "full", "linear_probe"] = "lora"
    learning_rate: float = 5e-4
    lora_learning_rate: float = 5e-3
    weight_decay: float = 0.01
    warmup_ratio: float = 0.1
    max_grad_norm: float = 1.0
    num_epochs: int = 15
    per_device_train_batch_size: int = 16
    per_device_eval_batch_size: int = 32
    gradient_accumulation_steps: int = 2
    fp16: bool = True
    eval_strategy: str = "epoch"
    metric_for_best_model: str = "macro_f1"
    output_dir: str = "./results"
    push_to_hub: bool = True
    hub_model_id: str = "Ellaft/multimodal-pc-fault-detector"
    save_strategy: str = "epoch"
    save_total_limit: int = 3
    load_best_model_at_end: bool = True
    logging_steps: int = 10
    logging_strategy: str = "steps"
    logging_first_step: bool = True
    disable_tqdm: bool = True
    seed: int = 42


@dataclass
class ExperimentConfig:
    data: DataConfig = field(default_factory=DataConfig)
    model: ModelConfig = field(default_factory=ModelConfig)
    lora: LoRAConfig = field(default_factory=LoRAConfig)
    train: TrainConfig = field(default_factory=TrainConfig)
    experiment_name: str = "multimodal_pc_fault_v2"
    description: str = "Two-branch audio-visual fusion for PC fault detection with OGM-GE anti-collapse"

    # ====================================================================
    # v2: Anti-modality-collapse hyperparameters
    # ====================================================================

    # OGM-GE: On-the-fly Gradient Modulation + Generalization Enhancement
    # From Peng et al., CVPR 2022 (arXiv: 2203.15332)
    ogm_alpha: float = 0.3          # Modulation strength (paper: 0.3-0.5)
    ogm_noise_sigma: float = 0.1    # GE noise std (paper: 0.1)

    # Auxiliary unimodal loss weights
    # Asymmetric: boost visual (weak modality), dampen audio (dominant)
    lambda_visual: float = 1.5      # Weight for visual auxiliary loss
    lambda_audio: float = 0.5       # Weight for audio auxiliary loss

    # Asymmetric learning rate multipliers (applied to lora_learning_rate)
    visual_lr_multiplier: float = 3.0   # Visual gets 3x base LR
    audio_lr_multiplier: float = 0.5    # Audio gets 0.5x base LR


def get_ablation_configs():
    """
    Generate ablation experiment configurations.
    
    v2 experiments:
    1. Multimodal + LoRA + OGM-GE (the full v2 pipeline)
    2. Visual Only + LoRA (unimodal baseline)
    3. Audio Only + LoRA (unimodal baseline)
    4. Multimodal + Full FT + OGM-GE
    5. Multimodal + Linear Probe + OGM-GE
    6. Multimodal + LoRA + High Dropout + OGM-GE (robustness)
    """
    configs = {}

    # 1. Multimodal + LoRA + OGM-GE (PRIMARY)
    cfg = ExperimentConfig()
    cfg.train.mode = "multimodal"
    cfg.experiment_name = "multimodal_lora_ogmge"
    configs["multimodal_lora_ogmge"] = cfg

    # 2. Visual Only + LoRA
    cfg = ExperimentConfig()
    cfg.train.mode = "visual_only"
    cfg.experiment_name = "visual_only_lora"
    cfg.train.hub_model_id = "Ellaft/pc-fault-visual-only"
    configs["visual_only_lora"] = cfg

    # 3. Audio Only + LoRA
    cfg = ExperimentConfig()
    cfg.train.mode = "audio_only"
    cfg.experiment_name = "audio_only_lora"
    cfg.train.hub_model_id = "Ellaft/pc-fault-audio-only"
    configs["audio_only_lora"] = cfg

    # 4. Multimodal + Full FT + OGM-GE
    cfg = ExperimentConfig()
    cfg.train.mode = "multimodal"
    cfg.train.finetune_method = "full"
    cfg.lora.enabled = False
    cfg.train.learning_rate = 2e-5
    cfg.experiment_name = "multimodal_full_ft_ogmge"
    cfg.train.hub_model_id = "Ellaft/pc-fault-multimodal-full-ft"
    configs["multimodal_full_ft_ogmge"] = cfg

    # 5. Multimodal + Linear Probe + OGM-GE
    cfg = ExperimentConfig()
    cfg.train.mode = "multimodal"
    cfg.train.finetune_method = "linear_probe"
    cfg.lora.enabled = False
    cfg.train.learning_rate = 1e-3
    cfg.experiment_name = "multimodal_linear_probe_ogmge"
    cfg.train.hub_model_id = "Ellaft/pc-fault-multimodal-linear-probe"
    configs["multimodal_linear_probe_ogmge"] = cfg

    # 6. Multimodal + LoRA + High Dropout + OGM-GE
    cfg = ExperimentConfig()
    cfg.train.mode = "multimodal"
    cfg.model.modality_dropout_p = 0.5
    cfg.experiment_name = "multimodal_robust_ogmge"
    cfg.train.hub_model_id = "Ellaft/pc-fault-multimodal-robust"
    configs["multimodal_robust_ogmge"] = cfg

    return configs