Update: Medical Image Segmentation (2026-01-27 15:29)

app.py

@@ -36,7 +36,8 @@ class Configs:
     IMAGE_SIZE: Tuple[int, int] = (288, 288)  # W, H
     MEAN: Tuple[float, ...] = (0.485, 0.456, 0.406)
     STD: Tuple[float, ...] = (0.229, 0.224, 0.225)
-    MODEL_PATH: str = os.path.join(os.getcwd(), "segformer_trained_weights")
+    # Sử dụng model mới huấn luyện, fallback sang model cũ nếu không tìm thấy
+    MODEL_PATH: str = os.path.join(os.getcwd(), "models", "best_model") if os.path.exists(os.path.join(os.getcwd(), "models", "best_model")) else os.path.join(os.getcwd(), "segformer_trained_weights")
 
 
 def get_model(*, model_path, num_classes):

download_dataset.py

@@ -22,25 +22,42 @@ def setup_kaggle_api():
         return False
 
 def check_kaggle_credentials():
-    """Kiểm tra Kaggle credentials"""
+    """Kiểm tra Kaggle credentials - hỗ trợ cả file JSON và environment variable"""
+    
+    # Kiểm tra KAGGLE_API_TOKEN environment variable (phương thức mới)
+    kaggle_token = os.environ.get("KAGGLE_API_TOKEN")
+    if kaggle_token:
+        print("✓ Kaggle API Token được tìm thấy từ environment variable")
+        return True
+    
+    # Kiểm tra KAGGLE_USERNAME và KAGGLE_KEY (phương thức cũ với env vars)
+    kaggle_username = os.environ.get("KAGGLE_USERNAME")
+    kaggle_key = os.environ.get("KAGGLE_KEY")
+    if kaggle_username and kaggle_key:
+        print("✓ Kaggle credentials được tìm thấy từ environment variables")
+        return True
+    
+    # Kiểm tra file kaggle.json
     kaggle_dir = Path.home() / ".kaggle"
     kaggle_json = kaggle_dir / "kaggle.json"
     
     if not kaggle_json.exists():
         print("\n⚠️  Kaggle credentials không được tìm thấy!")
-        print("Hướng dẫn:")
-        print("1. Truy cập: https://www.kaggle.com/account")
-        print("2. Scroll xuống, click 'Create New API Token'")
-        print("3. File kaggle.json sẽ tải xuống")
-        print(f"4. Di chuyển file vào: {kaggle_dir}")
-        print("5. Chạy: chmod 600 ~/.kaggle/kaggle.json  (trên Linux/Mac)")
+        print("\nCách 1: Sử dụng Environment Variable (khuyến nghị)")
+        print("  PowerShell: $env:KAGGLE_API_TOKEN = 'your_token'")
+        print("  Hoặc tạo token tại: https://www.kaggle.com/settings")
+        print("\nCách 2: Sử dụng file kaggle.json")
+        print("  1. Truy cập: https://www.kaggle.com/account")
+        print("  2. Scroll xuống, click 'Create New API Token'")
+        print("  3. File kaggle.json sẽ tải xuống")
+        print(f"  4. Di chuyển file vào: {kaggle_dir}")
         return False
     
     # Đặt permissions (Linux/Mac)
     if not os.name == 'nt':  # Không phải Windows
         os.chmod(str(kaggle_json), 0o600)
     
-    print("✓ Kaggle credentials được tìm thấy")
+    print("✓ Kaggle credentials được tìm thấy từ kaggle.json")
     return True
 
 def download_dataset(competition_name="uw-madison-gi-tract-image-segmentation", 

segformer_trained_weights/config.json

@@ -1,16 +1,15 @@
 {
-  "_name_or_path": "nvidia/segformer-b4-finetuned-ade-512-512",
   "architectures": [
     "SegformerForSemanticSegmentation"
   ],
   "attention_probs_dropout_prob": 0.0,
   "classifier_dropout_prob": 0.1,
-  "decoder_hidden_size": 768,
+  "decoder_hidden_size": 256,
   "depths": [
-    3,
-    8,
-    27,
-    3
+    2,
+    2,
+    2,
+    2
   ],
   "downsampling_rates": [
     1,
@@ -19,27 +18,28 @@
     16
   ],
   "drop_path_rate": 0.1,
+  "dtype": "float32",
   "hidden_act": "gelu",
   "hidden_dropout_prob": 0.0,
   "hidden_sizes": [
+    32,
     64,
-    128,
-    320,
-    512
+    160,
+    256
   ],
   "id2label": {
-    "0": "LABEL_0",
-    "1": "LABEL_1",
-    "2": "LABEL_2",
-    "3": "LABEL_3"
+    "0": "background",
+    "1": "large_bowel",
+    "2": "small_bowel",
+    "3": "stomach"
   },
   "image_size": 224,
   "initializer_range": 0.02,
   "label2id": {
-    "LABEL_0": 0,
-    "LABEL_1": 1,
-    "LABEL_2": 2,
-    "LABEL_3": 3
+    "background": 0,
+    "large_bowel": 1,
+    "small_bowel": 2,
+    "stomach": 3
   },
   "layer_norm_eps": 1e-06,
   "mlp_ratios": [
@@ -77,6 +77,5 @@
     2,
     2
   ],
-  "torch_dtype": "float32",
-  "transformers_version": "4.30.2"
+  "transformers_version": "4.57.6"
 }

segformer_trained_weights/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f7160c06dd44ce9b8684ccf2b758dd06d3a708100a6d7cf9a8b67d49996019b
+size 14886832

train.py

@@ -116,10 +116,11 @@ class MedicalImageSegmentationTrainer:
         print("\n🧠 Loading SegFormer model...")
         
         model = SegformerForSemanticSegmentation.from_pretrained(
-            "nvidia/segformer-b0-finetuned-cityscapes-1024-1024",
+            "nvidia/mit-b0",
             num_labels=4,  # background + 3 organs
             id2label={0: "background", 1: "large_bowel", 2: "small_bowel", 3: "stomach"},
-            label2id={"background": 0, "large_bowel": 1, "small_bowel": 2, "stomach": 3}
+            label2id={"background": 0, "large_bowel": 1, "small_bowel": 2, "stomach": 3},
+            ignore_mismatched_sizes=True
         )
         
         model.to(self.device)
@@ -148,7 +149,7 @@ class MedicalImageSegmentationTrainer:
             optimizer.step()
             
             total_loss += loss.item()
-            pbar.set_postfix({'loss': loss.item():.4f})
+            pbar.set_postfix({'loss': f'{loss.item():.4f}'})
         
         return total_loss / len(train_loader)