geoai-py 0.14.0__py2.py3-none-any.whl → 0.16.0__py2.py3-none-any.whl

geoai/train.py

@@ -34,6 +34,97 @@ try:
 except ImportError:
     SMP_AVAILABLE = False
 
+# Additional imports for Lightly Train
+try:
+    import lightly_train
+
+    LIGHTLY_TRAIN_AVAILABLE = True
+except ImportError:
+    LIGHTLY_TRAIN_AVAILABLE = False
+
+
+def parse_coco_annotations(
+    coco_json_path: str, images_dir: str, labels_dir: str
+) -> Tuple[List[str], List[str]]:
+    """
+    Parse COCO format annotations and return lists of image and label paths.
+
+    Args:
+        coco_json_path (str): Path to COCO annotations JSON file (instances.json).
+        images_dir (str): Directory containing image files.
+        labels_dir (str): Directory containing label mask files.
+
+    Returns:
+        Tuple[List[str], List[str]]: Lists of image paths and corresponding label paths.
+    """
+    import json
+
+    with open(coco_json_path, "r") as f:
+        coco_data = json.load(f)
+
+    # Create mapping from image_id to filename
+    image_files = []
+    label_files = []
+
+    for img_info in coco_data["images"]:
+        img_filename = img_info["file_name"]
+        img_path = os.path.join(images_dir, img_filename)
+
+        # Derive label filename (same as image filename)
+        label_path = os.path.join(labels_dir, img_filename)
+
+        if os.path.exists(img_path) and os.path.exists(label_path):
+            image_files.append(img_path)
+            label_files.append(label_path)
+
+    return image_files, label_files
+
+
+def parse_yolo_annotations(
+    data_dir: str, images_subdir: str = "images", labels_subdir: str = "labels"
+) -> Tuple[List[str], List[str]]:
+    """
+    Parse YOLO format annotations and return lists of image and label paths.
+
+    YOLO format structure:
+    - data_dir/images/: Contains image files (.tif, .png, .jpg)
+    - data_dir/labels/: Contains label masks (.tif, .png) and YOLO .txt files
+    - data_dir/classes.txt: Class names (one per line)
+
+    Args:
+        data_dir (str): Root directory containing YOLO-format data.
+        images_subdir (str): Subdirectory name for images. Defaults to 'images'.
+        labels_subdir (str): Subdirectory name for labels. Defaults to 'labels'.
+
+    Returns:
+        Tuple[List[str], List[str]]: Lists of image paths and corresponding label paths.
+    """
+    images_dir = os.path.join(data_dir, images_subdir)
+    labels_dir = os.path.join(data_dir, labels_subdir)
+
+    if not os.path.exists(images_dir):
+        raise FileNotFoundError(f"Images directory not found: {images_dir}")
+    if not os.path.exists(labels_dir):
+        raise FileNotFoundError(f"Labels directory not found: {labels_dir}")
+
+    # Get all image files
+    image_extensions = (".tif", ".tiff", ".png", ".jpg", ".jpeg")
+    image_files = []
+    label_files = []
+
+    for img_file in os.listdir(images_dir):
+        if img_file.lower().endswith(image_extensions):
+            img_path = os.path.join(images_dir, img_file)
+
+            # Find corresponding label mask (same filename)
+            label_path = os.path.join(labels_dir, img_file)
+
+            if os.path.exists(label_path):
+                image_files.append(img_path)
+                label_files.append(label_path)
+
+    return sorted(image_files), sorted(label_files)
+
 
 def get_instance_segmentation_model(
     num_classes: int = 2, num_channels: int = 3, pretrained: bool = True
@@ -617,6 +708,7 @@ def train_MaskRCNN_model(
     images_dir: str,
     labels_dir: str,
     output_dir: str,
+    input_format: str = "directory",
     num_channels: int = 3,
     model: Optional[torch.nn.Module] = None,
     pretrained: bool = True,
@@ -640,9 +732,17 @@ def train_MaskRCNN_model(
     the backbone or to continue training from a specific checkpoint.
 
     Args:
-        images_dir (str): Directory containing image GeoTIFF files.
-        labels_dir (str): Directory containing label GeoTIFF files.
+        images_dir (str): Directory containing image GeoTIFF files (for 'directory' format),
+            or root directory containing images/ subdirectory (for 'yolo' format),
+            or directory containing images (for 'coco' format).
+        labels_dir (str): Directory containing label GeoTIFF files (for 'directory' format),
+            or path to COCO annotations JSON file (for 'coco' format),
+            or not used (for 'yolo' format - labels are in images_dir/labels/).
         output_dir (str): Directory to save model checkpoints and results.
+        input_format (str): Input data format - 'directory' (default), 'coco', or 'yolo'.
+            - 'directory': Standard directory structure with separate images_dir and labels_dir
+            - 'coco': COCO JSON format (labels_dir should be path to instances.json)
+            - 'yolo': YOLO format (images_dir is root with images/ and labels/ subdirectories)
         num_channels (int, optional): Number of input channels. If None, auto-detected.
             Defaults to 3.
         model (torch.nn.Module, optional): Predefined model. If None, a new model is created.
@@ -688,45 +788,63 @@ def train_MaskRCNN_model(
         device = get_device()
     print(f"Using device: {device}")
 
-    # Get all image and label files
-    # Support multiple image formats: GeoTIFF, PNG, JPG, JPEG, TIF, TIFF
-    image_extensions = (".tif", ".tiff", ".png", ".jpg", ".jpeg")
-    label_extensions = (".tif", ".tiff", ".png", ".jpg", ".jpeg")
+    # Get all image and label files based on input format
+    if input_format.lower() == "coco":
+        # Parse COCO format annotations
+        if verbose:
+            print(f"Loading COCO format annotations from {labels_dir}")
+        # For COCO format, labels_dir is path to instances.json
+        # Labels are typically in a "labels" directory parallel to "annotations"
+        coco_root = os.path.dirname(os.path.dirname(labels_dir))  # Go up two levels
+        labels_directory = os.path.join(coco_root, "labels")
+        image_files, label_files = parse_coco_annotations(
+            labels_dir, images_dir, labels_directory
+        )
+    elif input_format.lower() == "yolo":
+        # Parse YOLO format annotations
+        if verbose:
+            print(f"Loading YOLO format data from {images_dir}")
+        image_files, label_files = parse_yolo_annotations(images_dir)
+    else:
+        # Default: directory format
+        # Support multiple image formats: GeoTIFF, PNG, JPG, JPEG, TIF, TIFF
+        image_extensions = (".tif", ".tiff", ".png", ".jpg", ".jpeg")
+        label_extensions = (".tif", ".tiff", ".png", ".jpg", ".jpeg")
+
+        image_files = sorted(
+            [
+                os.path.join(images_dir, f)
+                for f in os.listdir(images_dir)
+                if f.lower().endswith(image_extensions)
+            ]
+        )
+        label_files = sorted(
+            [
+                os.path.join(labels_dir, f)
+                for f in os.listdir(labels_dir)
+                if f.lower().endswith(label_extensions)
+            ]
+        )
 
-    image_files = sorted(
-        [
-            os.path.join(images_dir, f)
-            for f in os.listdir(images_dir)
-            if f.lower().endswith(image_extensions)
-        ]
-    )
-    label_files = sorted(
-        [
-            os.path.join(labels_dir, f)
-            for f in os.listdir(labels_dir)
-            if f.lower().endswith(label_extensions)
-        ]
-    )
+        # Ensure matching files
+        if len(image_files) != len(label_files):
+            print("Warning: Number of image files and label files don't match!")
+            # Find matching files by basename
+            basenames = [os.path.basename(f) for f in image_files]
+            label_files = [
+                os.path.join(labels_dir, os.path.basename(f))
+                for f in image_files
+                if os.path.exists(os.path.join(labels_dir, os.path.basename(f)))
+            ]
+            image_files = [
+                f
+                for f, b in zip(image_files, basenames)
+                if os.path.exists(os.path.join(labels_dir, b))
+            ]
+            print(f"Using {len(image_files)} matching files")
 
     print(f"Found {len(image_files)} image files and {len(label_files)} label files")
 
-    # Ensure matching files
-    if len(image_files) != len(label_files):
-        print("Warning: Number of image files and label files don't match!")
-        # Find matching files by basename
-        basenames = [os.path.basename(f) for f in image_files]
-        label_files = [
-            os.path.join(labels_dir, os.path.basename(f))
-            for f in image_files
-            if os.path.exists(os.path.join(labels_dir, os.path.basename(f)))
-        ]
-        image_files = [
-            f
-            for f, b in zip(image_files, basenames)
-            if os.path.exists(os.path.join(labels_dir, b))
-        ]
-        print(f"Using {len(image_files)} matching files")
-
     # Split data into train and validation sets
     train_imgs, val_imgs, train_labels, val_labels = train_test_split(
         image_files, label_files, test_size=val_split, random_state=seed
@@ -1915,14 +2033,14 @@ def get_smp_model(
         )
 
 
-def dice_coefficient(
+def f1_score(
     pred: torch.Tensor,
     target: torch.Tensor,
     smooth: float = 1e-6,
     num_classes: Optional[int] = None,
 ) -> float:
     """
-    Calculate Dice coefficient for segmentation (binary or multi-class).
+    Calculate F1 score (also known as Dice coefficient) for segmentation (binary or multi-class).
 
     Args:
         pred (torch.Tensor): Predicted mask (probabilities or logits) with shape [C, H, W] or [H, W].
@@ -1931,7 +2049,7 @@ def dice_coefficient(
         num_classes (int, optional): Number of classes. If None, auto-detected.
 
     Returns:
-        float: Mean Dice coefficient across all classes.
+        float: Mean F1 score across all classes.
     """
     # Convert predictions to class predictions
     if pred.dim() == 3:  # [C, H, W] format
@@ -1946,8 +2064,8 @@ def dice_coefficient(
     if num_classes is None:
         num_classes = max(pred_classes.max().item(), target.max().item()) + 1
 
-    # Calculate Dice for each class and average
-    dice_scores = []
+    # Calculate F1 score for each class and average
+    f1_scores = []
     for class_id in range(num_classes):
         pred_class = (pred_classes == class_id).float()
         target_class = (target == class_id).float()
@@ -1956,10 +2074,10 @@ def dice_coefficient(
         union = pred_class.sum() + target_class.sum()
 
         if union > 0:
-            dice = (2.0 * intersection + smooth) / (union + smooth)
-            dice_scores.append(dice.item())
+            f1 = (2.0 * intersection + smooth) / (union + smooth)
+            f1_scores.append(f1.item())
 
-    return sum(dice_scores) / len(dice_scores) if dice_scores else 0.0
+    return sum(f1_scores) / len(f1_scores) if f1_scores else 0.0
 
 
 def iou_coefficient(
@@ -2009,6 +2127,108 @@ def iou_coefficient(
     return sum(iou_scores) / len(iou_scores) if iou_scores else 0.0
 
 
+def precision_score(
+    pred: torch.Tensor,
+    target: torch.Tensor,
+    smooth: float = 1e-6,
+    num_classes: Optional[int] = None,
+) -> float:
+    """
+    Calculate precision score for segmentation (binary or multi-class).
+
+    Precision = TP / (TP + FP), where:
+    - TP (True Positives): Correctly predicted positive pixels
+    - FP (False Positives): Incorrectly predicted positive pixels
+
+    Args:
+        pred (torch.Tensor): Predicted mask (probabilities or logits) with shape [C, H, W] or [H, W].
+        target (torch.Tensor): Ground truth mask with shape [H, W].
+        smooth (float): Smoothing factor to avoid division by zero.
+        num_classes (int, optional): Number of classes. If None, auto-detected.
+
+    Returns:
+        float: Mean precision score across all classes.
+    """
+    # Convert predictions to class predictions
+    if pred.dim() == 3:  # [C, H, W] format
+        pred = torch.softmax(pred, dim=0)
+        pred_classes = torch.argmax(pred, dim=0)
+    elif pred.dim() == 2:  # [H, W] format
+        pred_classes = pred
+    else:
+        raise ValueError(f"Unexpected prediction dimensions: {pred.shape}")
+
+    # Auto-detect number of classes if not provided
+    if num_classes is None:
+        num_classes = max(pred_classes.max().item(), target.max().item()) + 1
+
+    # Calculate precision for each class and average
+    precision_scores = []
+    for class_id in range(num_classes):
+        pred_class = (pred_classes == class_id).float()
+        target_class = (target == class_id).float()
+
+        true_positives = (pred_class * target_class).sum()
+        predicted_positives = pred_class.sum()
+
+        if predicted_positives > 0:
+            precision = (true_positives + smooth) / (predicted_positives + smooth)
+            precision_scores.append(precision.item())
+
+    return sum(precision_scores) / len(precision_scores) if precision_scores else 0.0
+
+
+def recall_score(
+    pred: torch.Tensor,
+    target: torch.Tensor,
+    smooth: float = 1e-6,
+    num_classes: Optional[int] = None,
+) -> float:
+    """
+    Calculate recall score (also known as sensitivity) for segmentation (binary or multi-class).
+
+    Recall = TP / (TP + FN), where:
+    - TP (True Positives): Correctly predicted positive pixels
+    - FN (False Negatives): Incorrectly predicted negative pixels
+
+    Args:
+        pred (torch.Tensor): Predicted mask (probabilities or logits) with shape [C, H, W] or [H, W].
+        target (torch.Tensor): Ground truth mask with shape [H, W].
+        smooth (float): Smoothing factor to avoid division by zero.
+        num_classes (int, optional): Number of classes. If None, auto-detected.
+
+    Returns:
+        float: Mean recall score across all classes.
+    """
+    # Convert predictions to class predictions
+    if pred.dim() == 3:  # [C, H, W] format
+        pred = torch.softmax(pred, dim=0)
+        pred_classes = torch.argmax(pred, dim=0)
+    elif pred.dim() == 2:  # [H, W] format
+        pred_classes = pred
+    else:
+        raise ValueError(f"Unexpected prediction dimensions: {pred.shape}")
+
+    # Auto-detect number of classes if not provided
+    if num_classes is None:
+        num_classes = max(pred_classes.max().item(), target.max().item()) + 1
+
+    # Calculate recall for each class and average
+    recall_scores = []
+    for class_id in range(num_classes):
+        pred_class = (pred_classes == class_id).float()
+        target_class = (target == class_id).float()
+
+        true_positives = (pred_class * target_class).sum()
+        actual_positives = target_class.sum()
+
+        if actual_positives > 0:
+            recall = (true_positives + smooth) / (actual_positives + smooth)
+            recall_scores.append(recall.item())
+
+    return sum(recall_scores) / len(recall_scores) if recall_scores else 0.0
+
+
 def train_semantic_one_epoch(
     model: torch.nn.Module,
     optimizer: torch.optim.Optimizer,
@@ -2090,13 +2310,15 @@ def evaluate_semantic(
         num_classes (int): Number of classes for evaluation metrics.
 
     Returns:
-        dict: Evaluation metrics including loss, IoU, and Dice.
+        dict: Evaluation metrics including loss, IoU, F1, precision, and recall.
     """
     model.eval()
 
     total_loss = 0
-    dice_scores = []
+    f1_scores = []
     iou_scores = []
+    precision_scores = []
+    recall_scores = []
     num_batches = len(data_loader)
 
     with torch.no_grad():
@@ -2112,23 +2334,38 @@ def evaluate_semantic(
 
             # Calculate metrics for each sample in the batch
             for pred, target in zip(outputs, targets):
-                dice = dice_coefficient(pred, target, num_classes=num_classes)
+                f1 = f1_score(pred, target, num_classes=num_classes)
                 iou = iou_coefficient(pred, target, num_classes=num_classes)
-                dice_scores.append(dice)
+                precision = precision_score(pred, target, num_classes=num_classes)
+                recall = recall_score(pred, target, num_classes=num_classes)
+                f1_scores.append(f1)
                 iou_scores.append(iou)
+                precision_scores.append(precision)
+                recall_scores.append(recall)
 
     # Calculate metrics
     avg_loss = total_loss / num_batches
-    avg_dice = sum(dice_scores) / len(dice_scores) if dice_scores else 0
+    avg_f1 = sum(f1_scores) / len(f1_scores) if f1_scores else 0
     avg_iou = sum(iou_scores) / len(iou_scores) if iou_scores else 0
-
-    return {"loss": avg_loss, "Dice": avg_dice, "IoU": avg_iou}
+    avg_precision = (
+        sum(precision_scores) / len(precision_scores) if precision_scores else 0
+    )
+    avg_recall = sum(recall_scores) / len(recall_scores) if recall_scores else 0
+
+    return {
+        "loss": avg_loss,
+        "F1": avg_f1,
+        "IoU": avg_iou,
+        "Precision": avg_precision,
+        "Recall": avg_recall,
+    }
 
 
 def train_segmentation_model(
     images_dir: str,
     labels_dir: str,
     output_dir: str,
+    input_format: str = "directory",
     architecture: str = "unet",
     encoder_name: str = "resnet34",
     encoder_weights: Optional[str] = "imagenet",
@@ -2150,6 +2387,7 @@ def train_segmentation_model(
     target_size: Optional[Tuple[int, int]] = None,
     resize_mode: str = "resize",
     num_workers: Optional[int] = None,
+    early_stopping_patience: Optional[int] = None,
     **kwargs: Any,
 ) -> torch.nn.Module:
     """
@@ -2160,9 +2398,17 @@ def train_segmentation_model(
     this approach treats the task as pixel-level binary classification.
 
     Args:
-        images_dir (str): Directory containing image GeoTIFF files.
-        labels_dir (str): Directory containing label GeoTIFF files.
+        images_dir (str): Directory containing image GeoTIFF files (for 'directory' format),
+            or root directory containing images/ subdirectory (for 'yolo' format),
+            or directory containing images (for 'coco' format).
+        labels_dir (str): Directory containing label GeoTIFF files (for 'directory' format),
+            or path to COCO annotations JSON file (for 'coco' format),
+            or not used (for 'yolo' format - labels are in images_dir/labels/).
         output_dir (str): Directory to save model checkpoints and results.
+        input_format (str): Input data format - 'directory' (default), 'coco', or 'yolo'.
+            - 'directory': Standard directory structure with separate images_dir and labels_dir
+            - 'coco': COCO JSON format (labels_dir should be path to instances.json)
+            - 'yolo': YOLO format (images_dir is root with images/ and labels/ subdirectories)
         architecture (str): Model architecture ('unet', 'deeplabv3', 'deeplabv3plus', 'fpn',
             'pspnet', 'linknet', 'manet'). Defaults to 'unet'.
         encoder_name (str): Encoder backbone name (e.g., 'resnet34', 'resnet50', 'efficientnet-b0').
@@ -2194,6 +2440,8 @@ def train_segmentation_model(
             'resize' - Resize images to target_size (may change aspect ratio)
             'pad' - Pad images to target_size (preserves aspect ratio). Defaults to 'resize'.
         num_workers (int): Number of workers for data loading. If None, uses 0 on macOS and Windows, 8 otherwise.
+        early_stopping_patience (int, optional): Number of epochs with no improvement after which
+            training will be stopped. If None, early stopping is disabled. Defaults to None.
         **kwargs: Additional arguments passed to smp.create_model().
     Returns:
         None: Model weights are saved to output_dir.
@@ -2225,45 +2473,63 @@ def train_segmentation_model(
         device = get_device()
     print(f"Using device: {device}")
 
-    # Get all image and label files
-    # Support multiple image formats: GeoTIFF, PNG, JPG, JPEG, TIF, TIFF
-    image_extensions = (".tif", ".tiff", ".png", ".jpg", ".jpeg")
-    label_extensions = (".tif", ".tiff", ".png", ".jpg", ".jpeg")
+    # Get all image and label files based on input format
+    if input_format.lower() == "coco":
+        # Parse COCO format annotations
+        if verbose:
+            print(f"Loading COCO format annotations from {labels_dir}")
+        # For COCO format, labels_dir is path to instances.json
+        # Labels are typically in a "labels" directory parallel to "annotations"
+        coco_root = os.path.dirname(os.path.dirname(labels_dir))  # Go up two levels
+        labels_directory = os.path.join(coco_root, "labels")
+        image_files, label_files = parse_coco_annotations(
+            labels_dir, images_dir, labels_directory
+        )
+    elif input_format.lower() == "yolo":
+        # Parse YOLO format annotations
+        if verbose:
+            print(f"Loading YOLO format data from {images_dir}")
+        image_files, label_files = parse_yolo_annotations(images_dir)
+    else:
+        # Default: directory format
+        # Support multiple image formats: GeoTIFF, PNG, JPG, JPEG, TIF, TIFF
+        image_extensions = (".tif", ".tiff", ".png", ".jpg", ".jpeg")
+        label_extensions = (".tif", ".tiff", ".png", ".jpg", ".jpeg")
+
+        image_files = sorted(
+            [
+                os.path.join(images_dir, f)
+                for f in os.listdir(images_dir)
+                if f.lower().endswith(image_extensions)
+            ]
+        )
+        label_files = sorted(
+            [
+                os.path.join(labels_dir, f)
+                for f in os.listdir(labels_dir)
+                if f.lower().endswith(label_extensions)
+            ]
+        )
 
-    image_files = sorted(
-        [
-            os.path.join(images_dir, f)
-            for f in os.listdir(images_dir)
-            if f.lower().endswith(image_extensions)
-        ]
-    )
-    label_files = sorted(
-        [
-            os.path.join(labels_dir, f)
-            for f in os.listdir(labels_dir)
-            if f.lower().endswith(label_extensions)
-        ]
-    )
+        # Ensure matching files
+        if len(image_files) != len(label_files):
+            print("Warning: Number of image files and label files don't match!")
+            # Find matching files by basename
+            basenames = [os.path.basename(f) for f in image_files]
+            label_files = [
+                os.path.join(labels_dir, os.path.basename(f))
+                for f in image_files
+                if os.path.exists(os.path.join(labels_dir, os.path.basename(f)))
+            ]
+            image_files = [
+                f
+                for f, b in zip(image_files, basenames)
+                if os.path.exists(os.path.join(labels_dir, b))
+            ]
+            print(f"Using {len(image_files)} matching files")
 
     print(f"Found {len(image_files)} image files and {len(label_files)} label files")
 
-    # Ensure matching files
-    if len(image_files) != len(label_files):
-        print("Warning: Number of image files and label files don't match!")
-        # Find matching files by basename
-        basenames = [os.path.basename(f) for f in image_files]
-        label_files = [
-            os.path.join(labels_dir, os.path.basename(f))
-            for f in image_files
-            if os.path.exists(os.path.join(labels_dir, os.path.basename(f)))
-        ]
-        image_files = [
-            f
-            for f, b in zip(image_files, basenames)
-            if os.path.exists(os.path.join(labels_dir, b))
-        ]
-        print(f"Using {len(image_files)} matching files")
-
     if len(image_files) == 0:
         raise FileNotFoundError("No matching image and label files found")
 
@@ -2405,7 +2671,7 @@ def train_segmentation_model(
         print(f"Using {torch.cuda.device_count()} GPUs for training")
         model = torch.nn.DataParallel(model)
 
-    # Set up loss function (CrossEntropyLoss for multi-class, can also use DiceLoss)
+    # Set up loss function (CrossEntropyLoss for multi-class, can also use F1Loss)
     criterion = torch.nn.CrossEntropyLoss()
 
     # Set up optimizer
@@ -2423,8 +2689,11 @@ def train_segmentation_model(
     train_losses = []
     val_losses = []
     val_ious = []
-    val_dices = []
+    val_f1s = []
+    val_precisions = []
+    val_recalls = []
     start_epoch = 0
+    epochs_without_improvement = 0
 
     # Load checkpoint if provided
     if checkpoint_path is not None:
@@ -2459,8 +2728,15 @@ def train_segmentation_model(
                         val_losses = checkpoint["val_losses"]
                     if "val_ious" in checkpoint:
                         val_ious = checkpoint["val_ious"]
-                    if "val_dices" in checkpoint:
-                        val_dices = checkpoint["val_dices"]
+                    if "val_f1s" in checkpoint:
+                        val_f1s = checkpoint["val_f1s"]
+                    # Also check for old val_dices format for backward compatibility
+                    elif "val_dices" in checkpoint:
+                        val_f1s = checkpoint["val_dices"]
+                    if "val_precisions" in checkpoint:
+                        val_precisions = checkpoint["val_precisions"]
+                    if "val_recalls" in checkpoint:
+                        val_recalls = checkpoint["val_recalls"]
 
                     print(f"Resuming training from epoch {start_epoch}")
                     print(f"Previous best IoU: {best_iou:.4f}")
@@ -2500,7 +2776,9 @@ def train_segmentation_model(
         )
         val_losses.append(eval_metrics["loss"])
         val_ious.append(eval_metrics["IoU"])
-        val_dices.append(eval_metrics["Dice"])
+        val_f1s.append(eval_metrics["F1"])
+        val_precisions.append(eval_metrics["Precision"])
+        val_recalls.append(eval_metrics["Recall"])
 
         # Update learning rate
         lr_scheduler.step(eval_metrics["loss"])
@@ -2511,14 +2789,28 @@ def train_segmentation_model(
             f"Train Loss: {train_loss:.4f}, "
             f"Val Loss: {eval_metrics['loss']:.4f}, "
             f"Val IoU: {eval_metrics['IoU']:.4f}, "
-            f"Val Dice: {eval_metrics['Dice']:.4f}"
+            f"Val F1: {eval_metrics['F1']:.4f}, "
+            f"Val Precision: {eval_metrics['Precision']:.4f}, "
+            f"Val Recall: {eval_metrics['Recall']:.4f}"
         )
 
-        # Save best model
+        # Save best model and check for early stopping
         if eval_metrics["IoU"] > best_iou:
             best_iou = eval_metrics["IoU"]
+            epochs_without_improvement = 0
             print(f"Saving best model with IoU: {best_iou:.4f}")
             torch.save(model.state_dict(), os.path.join(output_dir, "best_model.pth"))
+        else:
+            epochs_without_improvement += 1
+            if (
+                early_stopping_patience is not None
+                and epochs_without_improvement >= early_stopping_patience
+            ):
+                print(
+                    f"\nEarly stopping triggered after {epochs_without_improvement} epochs without improvement"
+                )
+                print(f"Best validation IoU: {best_iou:.4f}")
+                break
 
         # Save checkpoint every 10 epochs (if not save_best_only)
         if not save_best_only and ((epoch + 1) % 10 == 0 or epoch == num_epochs - 1):
@@ -2536,7 +2828,9 @@ def train_segmentation_model(
                     "train_losses": train_losses,
                     "val_losses": val_losses,
                     "val_ious": val_ious,
-                    "val_dices": val_dices,
+                    "val_f1s": val_f1s,
+                    "val_precisions": val_precisions,
+                    "val_recalls": val_recalls,
                 },
                 os.path.join(output_dir, f"checkpoint_epoch_{epoch+1}.pth"),
             )
@@ -2549,7 +2843,9 @@ def train_segmentation_model(
         "train_losses": train_losses,
         "val_losses": val_losses,
         "val_ious": val_ious,
-        "val_dices": val_dices,
+        "val_f1s": val_f1s,
+        "val_precisions": val_precisions,
+        "val_recalls": val_recalls,
     }
     torch.save(history, os.path.join(output_dir, "training_history.pth"))
 
@@ -2565,7 +2861,9 @@ def train_segmentation_model(
         f.write(f"Total epochs: {num_epochs}\n")
         f.write(f"Best validation IoU: {best_iou:.4f}\n")
         f.write(f"Final validation IoU: {val_ious[-1]:.4f}\n")
-        f.write(f"Final validation Dice: {val_dices[-1]:.4f}\n")
+        f.write(f"Final validation F1: {val_f1s[-1]:.4f}\n")
+        f.write(f"Final validation Precision: {val_precisions[-1]:.4f}\n")
+        f.write(f"Final validation Recall: {val_recalls[-1]:.4f}\n")
         f.write(f"Final validation loss: {val_losses[-1]:.4f}\n")
 
     print(f"Training complete! Best IoU: {best_iou:.4f}")
@@ -2594,10 +2892,10 @@ def train_segmentation_model(
             plt.grid(True)
 
             plt.subplot(1, 3, 3)
-            plt.plot(val_dices, label="Val Dice")
-            plt.title("Dice Score")
+            plt.plot(val_f1s, label="Val F1")
+            plt.title("F1 Score")
             plt.xlabel("Epoch")
-            plt.ylabel("Dice")
+            plt.ylabel("F1")
             plt.legend()
             plt.grid(True)
 
@@ -2627,6 +2925,7 @@ def semantic_inference_on_geotiff(
     device: Optional[torch.device] = None,
     probability_path: Optional[str] = None,
     probability_threshold: Optional[float] = None,
+    save_class_probabilities: bool = False,
     quiet: bool = False,
     **kwargs: Any,
 ) -> Tuple[str, float]:
@@ -2648,6 +2947,8 @@ def semantic_inference_on_geotiff(
         probability_threshold (float, optional): Probability threshold for binary classification.
             Only used when num_classes=2. If provided, pixels with class 1 probability >= threshold
             are classified as class 1, otherwise class 0. If None (default), uses argmax.
+        save_class_probabilities (bool): If True and probability_path is provided, saves each
+            class probability as a separate single-band file. Defaults to False.
         quiet (bool): If True, suppress progress bar. Defaults to False.
         **kwargs: Additional arguments.
 
@@ -2864,7 +3165,7 @@ def semantic_inference_on_geotiff(
             prob_meta = meta.copy()
             prob_meta.update({"count": num_classes, "dtype": "float32"})
 
-            # Save normalized probabilities
+            # Save normalized probabilities as multi-band raster
             with rasterio.open(probability_path, "w", **prob_meta) as dst:
                 for class_idx in range(num_classes):
                     # Normalize probabilities
@@ -2878,6 +3179,36 @@ def semantic_inference_on_geotiff(
             if not quiet:
                 print(f"Saved probability map to {probability_path}")
 
+            # Save individual class probabilities if requested
+            if save_class_probabilities:
+                # Prepare single-band metadata
+                single_band_meta = meta.copy()
+                single_band_meta.update({"count": 1, "dtype": "float32"})
+
+                # Get base filename and extension
+                prob_base = os.path.splitext(probability_path)[0]
+                prob_ext = os.path.splitext(probability_path)[1]
+
+                for class_idx in range(num_classes):
+                    # Create filename for this class
+                    class_prob_path = f"{prob_base}_class_{class_idx}{prob_ext}"
+
+                    # Normalize probabilities
+                    prob_band = np.zeros((height, width), dtype=np.float32)
+                    prob_band[valid_pixels] = (
+                        prob_accumulator[class_idx, valid_pixels]
+                        / count_accumulator[valid_pixels]
+                    )
+
+                    # Save single-band file
+                    with rasterio.open(class_prob_path, "w", **single_band_meta) as dst:
+                        dst.write(prob_band, 1)
+
+                    if not quiet:
+                        print(
+                            f"Saved class {class_idx} probability to {class_prob_path}"
+                        )
+
         return output_path, inference_time
 
 
@@ -2894,6 +3225,7 @@ def semantic_inference_on_image(
     binary_output: bool = True,
     probability_path: Optional[str] = None,
     probability_threshold: Optional[float] = None,
+    save_class_probabilities: bool = False,
     quiet: bool = False,
     **kwargs: Any,
 ) -> Tuple[str, float]:
@@ -2916,6 +3248,8 @@ def semantic_inference_on_image(
         probability_threshold (float, optional): Probability threshold for binary classification.
             Only used when num_classes=2. If provided, pixels with class 1 probability >= threshold
             are classified as class 1, otherwise class 0. If None (default), uses argmax.
+        save_class_probabilities (bool): If True and probability_path is provided, saves each
+            class probability as a separate single-band file. Defaults to False.
         quiet (bool): If True, suppress progress bar. Defaults to False.
         **kwargs: Additional arguments.
 
@@ -3194,7 +3528,7 @@ def semantic_inference_on_image(
                 "transform": transform,
             }
 
-            # Save normalized probabilities
+            # Save normalized probabilities as multi-band raster
             with rasterio.open(probability_path, "w", **prob_meta) as dst:
                 for class_idx in range(num_classes):
                     # Normalize probabilities
@@ -3205,6 +3539,39 @@ def semantic_inference_on_image(
             if not quiet:
                 print(f"Saved probability map to {probability_path}")
 
+            # Save individual class probabilities if requested
+            if save_class_probabilities:
+                # Prepare single-band metadata
+                single_band_meta = {
+                    "driver": "GTiff",
+                    "height": height,
+                    "width": width,
+                    "count": 1,
+                    "dtype": "float32",
+                    "transform": transform,
+                }
+
+                # Get base filename and extension
+                prob_base = os.path.splitext(probability_path)[0]
+                prob_ext = os.path.splitext(probability_path)[1]
+
+                for class_idx in range(num_classes):
+                    # Create filename for this class
+                    class_prob_path = f"{prob_base}_class_{class_idx}{prob_ext}"
+
+                    # Normalize probabilities
+                    prob_band = np.zeros((height, width), dtype=np.float32)
+                    prob_band[valid_pixels] = normalized_probs[class_idx, valid_pixels]
+
+                    # Save single-band file
+                    with rasterio.open(class_prob_path, "w", **single_band_meta) as dst:
+                        dst.write(prob_band, 1)
+
+                    if not quiet:
+                        print(
+                            f"Saved class {class_idx} probability to {class_prob_path}"
+                        )
+
         return output_path, inference_time
 
 
@@ -3222,6 +3589,7 @@ def semantic_segmentation(
     device: Optional[torch.device] = None,
     probability_path: Optional[str] = None,
     probability_threshold: Optional[float] = None,
+    save_class_probabilities: bool = False,
     quiet: bool = False,
     **kwargs: Any,
 ) -> None:
@@ -3244,11 +3612,16 @@ def semantic_segmentation(
         batch_size (int): Batch size for inference.
         device (torch.device, optional): Device to run inference on.
         probability_path (str, optional): Path to save probability map. If provided,
-            the normalized class probabilities will be saved as a multi-band raster.
+            the normalized class probabilities will be saved as a multi-band raster
+            where each band contains probabilities for each class.
         probability_threshold (float, optional): Probability threshold for binary classification.
             Only used when num_classes=2. If provided, pixels with class 1 probability >= threshold
             are classified as class 1, otherwise class 0. If None (default), uses argmax.
             Must be between 0 and 1.
+        save_class_probabilities (bool): If True and probability_path is provided, saves each
+            class probability as a separate single-band file. Files will be named like
+            "probability_class_0.tif", "probability_class_1.tif", etc. in the same directory
+            as probability_path. Defaults to False.
         quiet (bool): If True, suppress progress bar. Defaults to False.
         **kwargs: Additional arguments.
 
@@ -3325,6 +3698,7 @@ def semantic_segmentation(
             device=device,
             probability_path=probability_path,
             probability_threshold=probability_threshold,
+            save_class_probabilities=save_class_probabilities,
             quiet=quiet,
             **kwargs,
         )
@@ -3345,6 +3719,7 @@ def semantic_segmentation(
             binary_output=True,  # Convert to binary output for better visualization
             probability_path=probability_path,
             probability_threshold=probability_threshold,
+            save_class_probabilities=save_class_probabilities,
             quiet=quiet,
             **kwargs,
         )
@@ -3527,6 +3902,7 @@ def train_instance_segmentation_model(
     images_dir: str,
     labels_dir: str,
     output_dir: str,
+    input_format: str = "directory",
     num_classes: int = 2,
     num_channels: int = 3,
     batch_size: int = 4,
@@ -3545,9 +3921,17 @@ def train_instance_segmentation_model(
     This is a wrapper function for train_MaskRCNN_model with clearer naming.
 
     Args:
-        images_dir (str): Directory containing image GeoTIFF files.
-        labels_dir (str): Directory containing label GeoTIFF files.
+        images_dir (str): Directory containing image GeoTIFF files (for 'directory' format),
+            or root directory containing images/ subdirectory (for 'yolo' format),
+            or directory containing images (for 'coco' format).
+        labels_dir (str): Directory containing label GeoTIFF files (for 'directory' format),
+            or path to COCO annotations JSON file (for 'coco' format),
+            or not used (for 'yolo' format - labels are in images_dir/labels/).
         output_dir (str): Directory to save model checkpoints and results.
+        input_format (str): Input data format - 'directory' (default), 'coco', or 'yolo'.
+            - 'directory': Standard directory structure with separate images_dir and labels_dir
+            - 'coco': COCO JSON format (labels_dir should be path to instances.json)
+            - 'yolo': YOLO format (images_dir is root with images/ and labels/ subdirectories)
         num_classes (int): Number of classes (including background). Defaults to 2.
         num_channels (int): Number of input channels. Defaults to 3.
         batch_size (int): Batch size for training. Defaults to 4.
@@ -3572,6 +3956,7 @@ def train_instance_segmentation_model(
         images_dir=images_dir,
         labels_dir=labels_dir,
         output_dir=output_dir,
+        input_format=input_format,
         num_channels=num_channels,
         model=model,
         batch_size=batch_size,
@@ -3756,3 +4141,307 @@ def instance_segmentation_batch(
             continue
 
     print(f"Batch processing completed. Results saved to {output_dir}")
+
+
+def lightly_train_model(
+    data_dir: str,
+    output_dir: str,
+    model: str = "torchvision/resnet50",
+    method: str = "dinov2_distillation",
+    epochs: int = 100,
+    batch_size: int = 64,
+    learning_rate: float = 1e-4,
+    **kwargs: Any,
+) -> str:
+    """
+    Train a model using Lightly Train for self-supervised pretraining.
+
+    Args:
+        data_dir (str): Directory containing unlabeled images for training.
+        output_dir (str): Directory to save training outputs and model checkpoints.
+        model (str): Model architecture to train. Supports models from torchvision,
+            timm, ultralytics, etc. Default is "torchvision/resnet50".
+        method (str): Self-supervised learning method. Options include:
+            - "simclr": Works with CNN models (ResNet, EfficientNet, etc.)
+            - "dino": Works with both CNNs and ViTs
+            - "dinov2": Requires ViT models only
+            - "dinov2_distillation": Requires ViT models only (recommended for ViTs)
+            Default is "dinov2_distillation".
+        epochs (int): Number of training epochs. Default is 100.
+        batch_size (int): Batch size for training. Default is 64.
+        learning_rate (float): Learning rate for training. Default is 1e-4.
+        **kwargs: Additional arguments passed to lightly_train.train().
+
+    Returns:
+        str: Path to the exported model file.
+
+    Raises:
+        ImportError: If lightly-train is not installed.
+        ValueError: If data_dir does not exist, is empty, or incompatible model/method.
+
+    Note:
+        Model/Method compatibility:
+        - CNN models (ResNet, EfficientNet): Use "simclr" or "dino"
+        - ViT models: Use "dinov2", "dinov2_distillation", or "dino"
+
+    Example:
+        >>> # For CNN models (ResNet, EfficientNet)
+        >>> model_path = lightly_train_model(
+        ...     data_dir="path/to/unlabeled/images",
+        ...     output_dir="path/to/output",
+        ...     model="torchvision/resnet50",
+        ...     method="simclr",  # Use simclr for CNNs
+        ...     epochs=50
+        ... )
+        >>> # For ViT models
+        >>> model_path = lightly_train_model(
+        ...     data_dir="path/to/unlabeled/images",
+        ...     output_dir="path/to/output",
+        ...     model="timm/vit_base_patch16_224",
+        ...     method="dinov2",  # dinov2 requires ViT
+        ...     epochs=50
+        ... )
+    """
+    if not LIGHTLY_TRAIN_AVAILABLE:
+        raise ImportError(
+            "lightly-train is not installed. Please install it with: "
+            "pip install lightly-train"
+        )
+
+    if not os.path.exists(data_dir):
+        raise ValueError(f"Data directory does not exist: {data_dir}")
+
+    # Check if data directory contains images
+    image_extensions = ["*.jpg", "*.jpeg", "*.png", "*.tif", "*.tiff", "*.bmp"]
+    image_files = []
+    for ext in image_extensions:
+        image_files.extend(glob.glob(os.path.join(data_dir, "**", ext), recursive=True))
+
+    if not image_files:
+        raise ValueError(f"No image files found in {data_dir}")
+
+    # Validate model/method compatibility
+    is_vit_model = "vit" in model.lower() or "vision_transformer" in model.lower()
+
+    if method in ["dinov2", "dinov2_distillation"] and not is_vit_model:
+        raise ValueError(
+            f"Method '{method}' requires a Vision Transformer (ViT) model, but got '{model}'.\n"
+            f"Solutions:\n"
+            f"  1. Use a ViT model: model='timm/vit_base_patch16_224'\n"
+            f"  2. Use a CNN-compatible method: method='simclr' or method='dino'\n"
+            f"\nFor CNN models (ResNet, EfficientNet), use 'simclr' or 'dino'.\n"
+            f"For ViT models, use 'dinov2', 'dinov2_distillation', or 'dino'."
+        )
+
+    print(f"Found {len(image_files)} images in {data_dir}")
+    print(f"Starting self-supervised pretraining with {method} method...")
+    print(f"Model: {model}")
+
+    # Create output directory
+    os.makedirs(output_dir, exist_ok=True)
+
+    # Detect if running in notebook environment and set appropriate configuration
+    def is_notebook():
+        try:
+            from IPython import get_ipython
+
+            if get_ipython() is not None:
+                return True
+        except (ImportError, NameError):
+            pass
+        return False
+
+    # Force single-device training in notebooks to avoid DDP strategy issues
+    if is_notebook():
+        # Only override if not explicitly set by user
+        if "accelerator" not in kwargs:
+            # Use CPU in notebooks to avoid DDP incompatibility
+            # Users can still override by passing accelerator='gpu'
+            kwargs["accelerator"] = "cpu"
+        if "devices" not in kwargs:
+            kwargs["devices"] = 1  # Force single device
+
+    # Train the model using Lightly Train
+    lightly_train.train(
+        out=output_dir,
+        data=data_dir,
+        model=model,
+        method=method,
+        epochs=epochs,
+        batch_size=batch_size,
+        **kwargs,
+    )
+
+    # Return path to the exported model
+    exported_model_path = os.path.join(
+        output_dir, "exported_models", "exported_last.pt"
+    )
+
+    if os.path.exists(exported_model_path):
+        print(
+            f"Model training completed. Exported model saved to: {exported_model_path}"
+        )
+        return exported_model_path
+    else:
+        # Check for alternative export paths
+        possible_paths = [
+            os.path.join(output_dir, "exported_models", "exported_best.pt"),
+            os.path.join(output_dir, "checkpoints", "last.ckpt"),
+        ]
+
+        for path in possible_paths:
+            if os.path.exists(path):
+                print(f"Model training completed. Exported model saved to: {path}")
+                return path
+
+        print(f"Model training completed. Output saved to: {output_dir}")
+        return output_dir
+
+
+def load_lightly_pretrained_model(
+    model_path: str,
+    model_architecture: str = "torchvision/resnet50",
+    device: str = None,
+) -> torch.nn.Module:
+    """
+    Load a pretrained model from Lightly Train.
+
+    Args:
+        model_path (str): Path to the pretrained model file (.pt format).
+        model_architecture (str): Architecture of the model to load.
+            Default is "torchvision/resnet50".
+        device (str): Device to load the model on. If None, uses CPU.
+
+    Returns:
+        torch.nn.Module: Loaded pretrained model ready for fine-tuning.
+
+    Raises:
+        FileNotFoundError: If model_path does not exist.
+        ImportError: If required libraries are not available.
+
+    Example:
+        >>> model = load_lightly_pretrained_model(
+        ...     model_path="path/to/pretrained_model.pt",
+        ...     model_architecture="torchvision/resnet50",
+        ...     device="cuda"
+        ... )
+        >>> # Fine-tune the model with your existing training pipeline
+    """
+    if not os.path.exists(model_path):
+        raise FileNotFoundError(f"Model file not found: {model_path}")
+
+    print(f"Loading pretrained model from: {model_path}")
+
+    # Load the model based on architecture
+    if model_architecture.startswith("torchvision/"):
+        model_name = model_architecture.replace("torchvision/", "")
+
+        # Import the model from torchvision
+        if hasattr(torchvision.models, model_name):
+            model = getattr(torchvision.models, model_name)()
+        else:
+            raise ValueError(f"Unknown torchvision model: {model_name}")
+
+    elif model_architecture.startswith("timm/"):
+        try:
+            import timm
+
+            model_name = model_architecture.replace("timm/", "")
+            model = timm.create_model(model_name)
+        except ImportError:
+            raise ImportError(
+                "timm is required for TIMM models. Install with: pip install timm"
+            )
+
+    else:
+        # For other architectures, try to import from torchvision as default
+        try:
+            model = getattr(torchvision.models, model_architecture)()
+        except AttributeError:
+            raise ValueError(f"Unsupported model architecture: {model_architecture}")
+
+    # Load the pretrained weights
+    try:
+        state_dict = torch.load(model_path, map_location=device, weights_only=True)
+    except TypeError:
+        # For backward compatibility with older PyTorch versions
+        state_dict = torch.load(model_path, map_location=device)
+    model.load_state_dict(state_dict)
+
+    print(f"Successfully loaded pretrained model: {model_architecture}")
+    return model
+
+
+def lightly_embed_images(
+    data_dir: str,
+    model_path: str,
+    output_path: str,
+    model_architecture: str = None,  # Deprecated, kept for backwards compatibility
+    batch_size: int = 64,
+    **kwargs: Any,
+) -> str:
+    """
+    Generate embeddings for images using a Lightly Train pretrained model.
+
+    Args:
+        data_dir (str): Directory containing images to embed.
+        model_path (str): Path to the pretrained model checkpoint file (.ckpt).
+        output_path (str): Path to save the embeddings (as .pt file).
+        model_architecture (str): Architecture of the pretrained model (deprecated,
+            kept for backwards compatibility but not used). The model architecture
+            is automatically loaded from the checkpoint.
+        batch_size (int): Batch size for embedding generation. Default is 64.
+        **kwargs: Additional arguments passed to lightly_train.embed().
+            Supported kwargs include: image_size, num_workers, accelerator, etc.
+
+    Returns:
+        str: Path to the saved embeddings file.
+
+    Raises:
+        ImportError: If lightly-train is not installed.
+        FileNotFoundError: If data_dir or model_path does not exist.
+
+    Note:
+        The model_path should point to a .ckpt file from the training output,
+        typically located at: output_dir/checkpoints/last.ckpt
+
+    Example:
+        >>> embeddings_path = lightly_embed_images(
+        ...     data_dir="path/to/images",
+        ...     model_path="output_dir/checkpoints/last.ckpt",
+        ...     output_path="embeddings.pt",
+        ...     batch_size=32
+        ... )
+        >>> print(f"Embeddings saved to: {embeddings_path}")
+    """
+    if not LIGHTLY_TRAIN_AVAILABLE:
+        raise ImportError(
+            "lightly-train is not installed. Please install it with: "
+            "pip install lightly-train"
+        )
+
+    if not os.path.exists(data_dir):
+        raise FileNotFoundError(f"Data directory does not exist: {data_dir}")
+
+    if not os.path.exists(model_path):
+        raise FileNotFoundError(f"Model file does not exist: {model_path}")
+
+    print(f"Generating embeddings for images in: {data_dir}")
+    print(f"Using pretrained model: {model_path}")
+
+    output_dir = os.path.dirname(output_path)
+    if output_dir:
+        os.makedirs(output_dir, exist_ok=True)
+
+    # Generate embeddings using Lightly Train
+    # Note: model_architecture is not used - it's inferred from the checkpoint
+    lightly_train.embed(
+        out=output_path,
+        data=data_dir,
+        checkpoint=model_path,
+        batch_size=batch_size,
+        **kwargs,
+    )
+
+    print(f"Embeddings saved to: {output_path}")
+    return output_path