src/training/config.py

"""
config.py — Single source of truth for all hyperparameters, paths, and API keys.

All configurable values live here. Import this module anywhere you need
a path, a training parameter, or an API endpoint.
"""

import os
from pathlib import Path
from dotenv import load_dotenv

# ---------------------------------------------------------------------------
# Load .env file (if present) for API keys
# ---------------------------------------------------------------------------
load_dotenv()

# ---------------------------------------------------------------------------
# Project Root — resolved relative to this file's location
# ---------------------------------------------------------------------------
PROJECT_ROOT = Path(__file__).resolve().parents[2]  # multi-hazard-warning-system/

# ---------------------------------------------------------------------------
# Directory Structure
# ---------------------------------------------------------------------------
DATA_DIR          = PROJECT_ROOT / "data"
RAW_DIR           = DATA_DIR / "raw"
FIRMS_RAW_DIR     = RAW_DIR / "firms"
WEATHER_RAW_DIR   = RAW_DIR / "weather"
AQI_RAW_DIR       = RAW_DIR / "aqi"

PROCESSED_DIR     = DATA_DIR / "processed"
IMAGE_PATCHES_DIR = PROCESSED_DIR / "image_patches"
TIMESERIES_DIR    = PROCESSED_DIR / "timeseries"
SPLITS_DIR        = DATA_DIR / "splits"

CHECKPOINTS_DIR   = PROJECT_ROOT / "checkpoints"
OUTPUTS_DIR       = PROJECT_ROOT / "outputs"
NOTEBOOKS_DIR     = PROJECT_ROOT / "notebooks"

# Auto-create directories
for _dir in [
    FIRMS_RAW_DIR, WEATHER_RAW_DIR, AQI_RAW_DIR,
    IMAGE_PATCHES_DIR, TIMESERIES_DIR, SPLITS_DIR,
    CHECKPOINTS_DIR, OUTPUTS_DIR, NOTEBOOKS_DIR,
]:
    _dir.mkdir(parents=True, exist_ok=True)

# ---------------------------------------------------------------------------
# API Keys & Endpoints
# ---------------------------------------------------------------------------
NASA_FIRMS_API_KEY = os.getenv("NASA_FIRMS_API_KEY", "")
OPENAQ_API_KEY     = os.getenv("OPENAQ_API_KEY", "")

FIRMS_BASE_URL     = "https://firms.modaps.eosdis.nasa.gov/api/area/csv"
OPEN_METEO_URL     = "https://archive-api.open-meteo.com/v1/archive"
OPENAQ_BASE_URL    = "https://api.openaq.org/v2"

# ---------------------------------------------------------------------------
# Geographic Defaults (California wildfire region)
# ---------------------------------------------------------------------------
DEFAULT_LATITUDE   = 37.5
DEFAULT_LONGITUDE  = -120.3
DEFAULT_BBOX       = "-122.0,36.0,-118.0,39.0"  # west, south, east, north
PATCH_SIZE         = 128   # spatial resolution of image patches

# ---------------------------------------------------------------------------
# Data Pipeline
# ---------------------------------------------------------------------------
TIMESERIES_WINDOW  = 7     # days of weather/AQI history per sample
TIMESERIES_FEATURES = 6    # temp, humidity, wind_speed, wind_dir, precip, PM2.5
IMAGE_CHANNELS     = 4     # RGB + Near-Infrared
FIRE_CONFIDENCE_THRESHOLD = 80  # minimum FIRMS confidence score

NUM_SYNTHETIC_SAMPLES = 500  # fallback sample count when APIs are unavailable

# ---------------------------------------------------------------------------
# Model Architecture
# ---------------------------------------------------------------------------
CNN_FEATURE_DIM    = 2048   # ResNet-50 output features
LSTM_HIDDEN_SIZE   = 128    # per-direction hidden size
LSTM_NUM_LAYERS    = 2
LSTM_BIDIRECTIONAL = True
LSTM_FEATURE_DIM   = LSTM_HIDDEN_SIZE * (2 if LSTM_BIDIRECTIONAL else 1)  # 256

FUSION_DIM         = CNN_FEATURE_DIM + LSTM_FEATURE_DIM  # 2304
SHARED_FC_DIMS    = [512, 256]
DROPOUT_RATE       = 0.3

# Task 1: Wildfire Risk Heatmap
HEATMAP_SIZE       = (PATCH_SIZE, PATCH_SIZE)  # 128×128

# Task 2: AQI Forecast
AQI_FORECAST_HOURS = 72    # predict 72 hourly values (24–72 hrs)

# ---------------------------------------------------------------------------
# Training Hyperparameters
# ---------------------------------------------------------------------------
BATCH_SIZE         = 16
LEARNING_RATE      = 1e-4
WEIGHT_DECAY       = 1e-5
NUM_EPOCHS         = 50
EARLY_STOP_PATIENCE = 10

LAMBDA_AQI         = 0.5   # weighting factor for AQI loss in combined loss
TRAIN_RATIO        = 0.70
VAL_RATIO          = 0.15
TEST_RATIO         = 0.15

# ---------------------------------------------------------------------------
# Device
# ---------------------------------------------------------------------------
import torch
DEVICE = "cpu"  # HF Spaces free tier: CPU only

# ---------------------------------------------------------------------------
# Weights & Biases
# ---------------------------------------------------------------------------
WANDB_PROJECT      = "multi-hazard-mtl"
WANDB_ENTITY       = os.getenv("WANDB_ENTITY", None)
USE_WANDB          = False  # Disabled for deployment

# ---------------------------------------------------------------------------
# Checkpoint naming
# ---------------------------------------------------------------------------
BEST_MODEL_NAME    = "best_mtl_model.pth"
BEST_MODEL_PATH    = CHECKPOINTS_DIR / BEST_MODEL_NAME

# ---------------------------------------------------------------------------
# Inference
# ---------------------------------------------------------------------------
RISK_THRESHOLDS = {
    "Low":      (0.0, 0.25),
    "Medium":   (0.25, 0.50),
    "High":     (0.50, 0.75),
    "Extreme":  (0.75, 1.0),
}

# AQI categories (WHO / US EPA aligned)
AQI_CATEGORIES = {
    "Good":                 (0, 50),
    "Moderate":             (51, 100),
    "Unhealthy (Sensitive)": (101, 150),
    "Unhealthy":            (151, 200),
    "Very Unhealthy":       (201, 300),
    "Hazardous":            (301, 500),
}