File size: 3,869 Bytes
1ea7ba6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 | from pathlib import Path
import os
def _resolve_base() -> Path:
wsl = Path("/mnt/d/SpiceNet")
win = Path("D:/SpiceNet")
return wsl if wsl.exists() else win
_BASE = _resolve_base()
DATA_DIR = _BASE / "Spice_Spectrum"
DATA_DIR_SAM = _BASE / "Spice_Spectrum_SAM" # SAM-preprocessed (optional)
OUTPUT_DIR = _BASE / "outputs"
CHECKPOINT_DIR = OUTPUT_DIR / "checkpoints"
LOG_DIR = OUTPUT_DIR / "logs"
CLASSES = [
"black pepper", "cardamom", "cinnamon", "cloves", "coriander",
"cumin", "ginger", "nutmeg", "paprika", "saffron", "turmeric",
]
NUM_CLASSES = len(CLASSES)
# Hard-negative pairs (class index pairs, visually confusable)
HARD_NEG_PAIRS = [
(4, 5), # coriander ↔ cumin
(8, 10), # paprika ↔ turmeric
(0, 3), # black pepper ↔ cloves
(2, 7), # cinnamon ↔ nutmeg
]
# Splits — paper spec: 70 / 10 / 20
TRAIN_RATIO = 0.70
VAL_RATIO = 0.10
TEST_RATIO = 0.20
RANDOM_SEED = 42
# Image
IMG_SIZE = 224 # resize to this before model (efficient on GPU)
IMG_FULL = 512 # original dataset resolution
IMG_MEAN = (0.485, 0.456, 0.406)
IMG_STD = (0.229, 0.224, 0.225)
# Model
BACKBONE = "efficientnet_b4"
PRETRAINED = True
CNN_DIM = 1792 # EfficientNet-B4 num_features
TEX_DIM = 256 # texture branch output
COL_DIM = 128 # color branch output
PROJ_DIM = 128 # SupCon projection head output
DROP_RATE = 0.4
# Texture features (LBP + GLCM)
LBP_P = 8
LBP_R = 1
GLCM_DISTANCES = [1, 2]
GLCM_ANGLES_DEG = [0, 45, 90, 135]
GLCM_LEVELS = 64
TEX_INPUT_DIM = (LBP_P + 2) + (6 * len(GLCM_DISTANCES) * len(GLCM_ANGLES_DEG)) # 10 + 48 = 58
# Color features (HSV histogram)
HSV_H_BINS = 36
HSV_S_BINS = 32
HSV_V_BINS = 32
COL_INPUT_DIM = HSV_H_BINS + HSV_S_BINS + HSV_V_BINS # 100
# Phase 1 — backbone pre-training
P1_EPOCHS = 30
P1_LR = 1e-4
P1_WARMUP = 3
P1_WEIGHT_DECAY = 1e-4
P1_LABEL_SMOOTH = 0.1
P1_MIN_LR = 1e-6
# Phase 2 — contrastive fine-tuning
P2_EPOCHS = 10
P2_LR = 5e-5
P2_TEMPERATURE = 0.07
# ── ARC-V (Paper 2 — acquisition-robust DG method + baselines) ───────────────
# Defaults from ARCV_METHOD_DESIGN.md §4.3 (MixStyle/FACT papers); tune by smoke.
ARCV_MIXSTYLE_P = 0.5 # prob. of MixStyle mixing per forward
ARCV_MIXSTYLE_ALPHA = 0.1 # Beta(alpha, alpha) for the style-mix coefficient
ARCV_MIXSTYLE_STAGES = 3 # apply MixStyle hooks after the first N backbone stages
ARCV_FOURIER_P = 0.5 # prob. of Fourier amplitude-mix per batch
ARCV_FOURIER_ETA = 1.0 # strong amplitude-mix upper bound
ARCV_FOURIER_ETA_WEAK = 0.3 # weak copy (phase-consistency co-teacher)
ARCV_GAMMA_PC = 1.0 # phase-consistency loss weight (warm up in trainer)
ARCV_RSC_DROP = 0.3333 # RSC top-|grad| feature drop fraction (~1/3)
ARCV_GRL_LAMBDA = 1.0 # DANN gradient-reversal strength (baseline)
ARCV_CORAL_WEIGHT = 1.0 # Deep CORAL loss weight (baseline)
ARCV_IRM_WEIGHT = 1.0 # IRM penalty weight (baseline)
# Phase 3 — full fusion end-to-end
P3_EPOCHS = 10
P3_LR = 1e-5
P3_WEIGHT_DECAY = 1e-4
P3_LABEL_SMOOTH = 0.1
P3_ALPHA = 0.5 # CE loss weight (alpha*CE + (1-alpha)*SupCon)
# Shared training
BATCH_SIZE = 32
NUM_WORKERS = 4
GRAD_CLIP = 1.0
PATIENCE = 8
# Augmentation
AUG_ROTATION = 30
AUG_SCALE = (0.7, 1.0)
AUG_BRIGHTNESS = 0.3
AUG_CONTRAST = 0.3
AUG_SATURATION = 0.3
AUG_HUE = 0.1
# Experiment tracking (optional)
USE_WANDB = False
WANDB_PROJECT = "spicenet"
|