| from dotenv import load_dotenv | |
| import os | |
| load_dotenv() | |
| os.environ["HF_TOKEN"] = os.getenv("HF_TOKEN") | |
| # ── Camera & optics (single source of truth) ─────────────────────────────────── | |
| SPECS = { | |
| "s2": {"pixel_m": 7.5e-6, "focal_m": 0.600, "alt_m": 786_000.0}, | |
| "bc2": {"pixel_m": 6.0e-6, "focal_m": 0.016, "alt_m": 410_000.0, | |
| "f_number": 2.8}, | |
| } | |
| def gsd(cam): | |
| return cam["pixel_m"] * cam["alt_m"] / cam["focal_m"] | |
| GSD_S2 = gsd(SPECS["s2"]) # 9.83 m | |
| GSD_BC2 = gsd(SPECS["bc2"]) # 153.75 m | |
| SCALE_FACTOR = GSD_BC2 / GSD_S2 # ~15.65 (replaces old altitude-only ratio) | |
| # ── Camera sensor dimensions ─────────────────────────────────────────────────── | |
| CAMERA_W = 752 # physical BlueFOX sensor width (inference frame size) | |
| CAMERA_H = 480 # physical BlueFOX sensor height (inference frame size) | |
| PATCH_SIZE = 33 # output size of each training patch after GSD downsampling | |
| # = floor(512 / SCALE_FACTOR) | |
| # ── Paths ────────────────────────────────────────────────────────────────────── | |
| SAVE_DIR = "data/bluefox" | |
| CHECKPOINT = "best_model.pth" | |
| # ── Dataset ──────────────────────────────────────────────────────────────────── | |
| MAX_SAMPLES = 8490 | |
| SEED = 42 | |
| TRAIN_RATIO = 0.70 | |
| VAL_RATIO = 0.15 | |
| TEST_RATIO = 0.15 | |
| # ── Model ────────────────────────────────────────────────────────────────────── | |
| NUM_CLASSES = 4 | |
| CLASS_NAMES = ["clear", "thick_cloud", "thin_cloud", "shadow"] | |
| # ── Training ─────────────────────────────────────────────────────────────────── | |
| BATCH_SIZE = 512 # 33×33 patches are tiny -- can afford larger batches | |
| NUM_WORKERS = 0 | |
| NUM_EPOCHS = 50 | |
| PATIENCE = 8 | |
| LEARNING_RATE = 2e-4 # bumped up slightly from 2e-5 -- fine for a small CNN | |
| WEIGHT_DECAY = 0.01 |