Fix: total_mem → total_memory, NUM_WORKERS=0 for Windows

train.py

@@ -55,7 +55,7 @@ class Config:
     DECODER_LR = 5e-4                    # Higher LR for random decoder
     WEIGHT_DECAY = 1e-4
     USE_AMP = True                       # Mixed precision — halves VRAM usage
-    NUM_WORKERS = 4
+    NUM_WORKERS = 0                      # 0 for Windows compatibility
     GRADIENT_CLIP = 1.0
 
     # Loss
@@ -75,18 +75,7 @@ class Config:
 # CLASS MAPPING: E-SCDD 30 classes → 5 classes
 # ═══════════════════════════════════════════════════════════════
 
-# Mask pixel values in E-SCDD are integers 0-29 (Label column in CSV)
-# We remap to 5 meaningful classes:
-#   0 = background (all spacing, borders, padding, text, clamp, frame, jbox)
-#   1 = busbar (label 9)
-#   2 = crack (label 14=crack, label 10=crack_rbn_edge)
-#   3 = dark/inactive (label 11=inactive, label 17=dead_cell, label 20=edge_dark)
-#   4 = other_defect (rings, material, gridline, splice, corrosion, belt_mark, etc.)
-
-LABEL_REMAP = np.zeros(30, dtype=np.uint8)  # default: everything → 0 (background)
-
-# Background features (labels 0-8, 21-24, 29)
-# Already 0 by default
+LABEL_REMAP = np.zeros(30, dtype=np.uint8)
 
 # Busbar
 LABEL_REMAP[9] = 1    # busbars → busbar
@@ -250,7 +239,7 @@ def train():
     print(f"Device: {device}")
     if device.type == "cuda":
         print(f"GPU: {torch.cuda.get_device_name(0)}")
-        print(f"VRAM: {torch.cuda.get_device_properties(0).total_mem / 1e9:.1f} GB")
+        print(f"VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")
 
     # ── Download data ────────────────────────────────────────
     download_dataset(cfg.DATA_DIR)
@@ -395,7 +384,6 @@ def train():
 
         # Per-class dice every 10 epochs
         if (epoch + 1) % 10 == 0:
-            # Run full validation for per-class metrics
             all_per_class = {name: [] for name in cfg.CLASS_NAMES}
             with torch.no_grad():
                 for images, masks in val_loader: