ultralytics 8.2.69__py3-none-any.whl → 8.2.70__py3-none-any.whl

ultralytics/__init__.py

@@ -1,6 +1,6 @@
 # Ultralytics YOLO 🚀, AGPL-3.0 license
 
-__version__ = "8.2.69"
+__version__ = "8.2.70"
 
 import os
 
@@ -8,7 +8,7 @@ import os
 os.environ["OMP_NUM_THREADS"] = "1"  # reduce CPU utilization during training
 
 from ultralytics.data.explorer.explorer import Explorer
-from ultralytics.models import NAS, RTDETR, SAM, YOLO, FastSAM, YOLOWorld
+from ultralytics.models import NAS, RTDETR, SAM, SAM2, YOLO, FastSAM, YOLOWorld
 from ultralytics.utils import ASSETS, SETTINGS
 from ultralytics.utils.checks import check_yolo as checks
 from ultralytics.utils.downloads import download
@@ -21,6 +21,7 @@ __all__ = (
     "YOLOWorld",
     "NAS",
     "SAM",
+    "SAM2",
     "FastSAM",
     "RTDETR",
     "checks",

ultralytics/cfg/__init__.py

@@ -793,6 +793,10 @@ def entrypoint(debug=""):
         from ultralytics import FastSAM
 
         model = FastSAM(model)
+    elif "sam2" in stem:
+        from ultralytics import SAM2
+
+        model = SAM2(model)
     elif "sam" in stem:
         from ultralytics import SAM
 

ultralytics/models/__init__.py

@@ -4,6 +4,7 @@ from .fastsam import FastSAM
 from .nas import NAS
 from .rtdetr import RTDETR
 from .sam import SAM
+from .sam2 import SAM2
 from .yolo import YOLO, YOLOWorld
 
-__all__ = "YOLO", "RTDETR", "SAM", "FastSAM", "NAS", "YOLOWorld"  # allow simpler import
+__all__ = "YOLO", "RTDETR", "SAM", "FastSAM", "NAS", "YOLOWorld", "SAM2"  # allow simpler import

ultralytics/models/fastsam/predict.py

@@ -21,6 +21,7 @@ class FastSAMPredictor(SegmentationPredictor):
     """
 
     def __init__(self, cfg=DEFAULT_CFG, overrides=None, _callbacks=None):
+        """Initializes a FastSAMPredictor for fast SAM segmentation tasks in Ultralytics YOLO framework."""
         super().__init__(cfg, overrides, _callbacks)
         self.prompts = {}
 

ultralytics/models/sam/build.py

@@ -14,7 +14,7 @@ from ultralytics.utils.downloads import attempt_download_asset
 
 from .modules.decoders import MaskDecoder
 from .modules.encoders import ImageEncoderViT, PromptEncoder
-from .modules.sam import Sam
+from .modules.sam import SAMModel
 from .modules.tiny_encoder import TinyViT
 from .modules.transformer import TwoWayTransformer
 
@@ -105,7 +105,7 @@ def _build_sam(
             out_chans=prompt_embed_dim,
         )
     )
-    sam = Sam(
+    sam = SAMModel(
         image_encoder=image_encoder,
         prompt_encoder=PromptEncoder(
             embed_dim=prompt_embed_dim,

ultralytics/models/sam/model.py

@@ -44,6 +44,7 @@ class SAM(Model):
         """
         if model and Path(model).suffix not in {".pt", ".pth"}:
             raise NotImplementedError("SAM prediction requires pre-trained *.pt or *.pth model.")
+        self.is_sam2 = "sam2" in Path(model).stem
         super().__init__(model=model, task="segment")
 
     def _load(self, weights: str, task=None):
@@ -54,7 +55,12 @@ class SAM(Model):
             weights (str): Path to the weights file.
             task (str, optional): Task name. Defaults to None.
         """
-        self.model = build_sam(weights)
+        if self.is_sam2:
+            from ..sam2.build import build_sam2
+
+            self.model = build_sam2(weights)
+        else:
+            self.model = build_sam(weights)
 
     def predict(self, source, stream=False, bboxes=None, points=None, labels=None, **kwargs):
         """
@@ -112,4 +118,6 @@ class SAM(Model):
         Returns:
             (dict): A dictionary mapping the 'segment' task to its corresponding 'Predictor'.
         """
-        return {"segment": {"predictor": Predictor}}
+        from ..sam2.predict import SAM2Predictor
+
+        return {"segment": {"predictor": SAM2Predictor if self.is_sam2 else Predictor}}

ultralytics/models/sam/modules/decoders.py

@@ -4,9 +4,8 @@ from typing import List, Tuple, Type
 
 import torch
 from torch import nn
-from torch.nn import functional as F
 
-from ultralytics.nn.modules import LayerNorm2d
+from ultralytics.nn.modules import MLP, LayerNorm2d
 
 
 class MaskDecoder(nn.Module):
@@ -28,7 +27,6 @@ class MaskDecoder(nn.Module):
 
     def __init__(
         self,
-        *,
         transformer_dim: int,
         transformer: nn.Module,
         num_multimask_outputs: int = 3,
@@ -149,42 +147,3 @@ class MaskDecoder(nn.Module):
         iou_pred = self.iou_prediction_head(iou_token_out)
 
         return masks, iou_pred
-
-
-class MLP(nn.Module):
-    """
-    MLP (Multi-Layer Perceptron) model lightly adapted from
-    https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py
-    """
-
-    def __init__(
-        self,
-        input_dim: int,
-        hidden_dim: int,
-        output_dim: int,
-        num_layers: int,
-        sigmoid_output: bool = False,
-    ) -> None:
-        """
-        Initializes the MLP (Multi-Layer Perceptron) model.
-
-        Args:
-            input_dim (int): The dimensionality of the input features.
-            hidden_dim (int): The dimensionality of the hidden layers.
-            output_dim (int): The dimensionality of the output layer.
-            num_layers (int): The number of hidden layers.
-            sigmoid_output (bool, optional): Apply a sigmoid activation to the output layer. Defaults to False.
-        """
-        super().__init__()
-        self.num_layers = num_layers
-        h = [hidden_dim] * (num_layers - 1)
-        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))
-        self.sigmoid_output = sigmoid_output
-
-    def forward(self, x):
-        """Executes feedforward within the neural network module and applies activation."""
-        for i, layer in enumerate(self.layers):
-            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
-        if self.sigmoid_output:
-            x = torch.sigmoid(x)
-        return x

ultralytics/models/sam/modules/encoders.py

@@ -211,6 +211,8 @@ class PromptEncoder(nn.Module):
         point_embedding[labels == -1] += self.not_a_point_embed.weight
         point_embedding[labels == 0] += self.point_embeddings[0].weight
         point_embedding[labels == 1] += self.point_embeddings[1].weight
+        point_embedding[labels == 2] += self.point_embeddings[2].weight
+        point_embedding[labels == 3] += self.point_embeddings[3].weight
         return point_embedding
 
     def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
@@ -226,8 +228,8 @@ class PromptEncoder(nn.Module):
         """Embeds mask inputs."""
         return self.mask_downscaling(masks)
 
+    @staticmethod
     def _get_batch_size(
-        self,
         points: Optional[Tuple[torch.Tensor, torch.Tensor]],
         boxes: Optional[torch.Tensor],
         masks: Optional[torch.Tensor],

ultralytics/models/sam/modules/sam.py

@@ -15,15 +15,14 @@ from .decoders import MaskDecoder
 from .encoders import ImageEncoderViT, PromptEncoder
 
 
-class Sam(nn.Module):
+class SAMModel(nn.Module):
     """
-    Sam (Segment Anything Model) is designed for object segmentation tasks. It uses image encoders to generate image
-    embeddings, and prompt encoders to encode various types of input prompts. These embeddings are then used by the mask
-    decoder to predict object masks.
+    SAMModel (Segment Anything Model) is designed for object segmentation tasks. It uses image encoders to generate
+    image embeddings, and prompt encoders to encode various types of input prompts. These embeddings are then used by
+    the mask decoder to predict object masks.
 
     Attributes:
         mask_threshold (float): Threshold value for mask prediction.
-        image_format (str): Format of the input image, default is 'RGB'.
         image_encoder (ImageEncoderViT): The backbone used to encode the image into embeddings.
         prompt_encoder (PromptEncoder): Encodes various types of input prompts.
         mask_decoder (MaskDecoder): Predicts object masks from the image and prompt embeddings.
@@ -32,7 +31,6 @@ class Sam(nn.Module):
     """
 
     mask_threshold: float = 0.0
-    image_format: str = "RGB"
 
     def __init__(
         self,
@@ -43,7 +41,7 @@ class Sam(nn.Module):
         pixel_std: List[float] = (58.395, 57.12, 57.375),
     ) -> None:
         """
-        Initialize the Sam class to predict object masks from an image and input prompts.
+        Initialize the SAMModel class to predict object masks from an image and input prompts.
 
         Note:
             All forward() operations moved to SAMPredictor.

ultralytics/models/sam/modules/transformer.py

@@ -86,7 +86,6 @@ class TwoWayTransformer(nn.Module):
           (torch.Tensor): the processed image_embedding
         """
         # BxCxHxW -> BxHWxC == B x N_image_tokens x C
-        bs, c, h, w = image_embedding.shape
         image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
         image_pe = image_pe.flatten(2).permute(0, 2, 1)
 
@@ -212,6 +211,7 @@ class Attention(nn.Module):
         embedding_dim: int,
         num_heads: int,
         downsample_rate: int = 1,
+        kv_in_dim: int = None,
     ) -> None:
         """
         Initializes the Attention model with the given dimensions and settings.
@@ -226,13 +226,14 @@ class Attention(nn.Module):
         """
         super().__init__()
         self.embedding_dim = embedding_dim
+        self.kv_in_dim = kv_in_dim if kv_in_dim is not None else embedding_dim
         self.internal_dim = embedding_dim // downsample_rate
         self.num_heads = num_heads
         assert self.internal_dim % num_heads == 0, "num_heads must divide embedding_dim."
 
         self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
-        self.k_proj = nn.Linear(embedding_dim, self.internal_dim)
-        self.v_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.k_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
+        self.v_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
         self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
 
     @staticmethod

ultralytics/models/sam/predict.py

@@ -168,7 +168,7 @@ class Predictor(BasePredictor):
                 - np.ndarray: An array of length C containing quality scores predicted by the model for each mask.
                 - np.ndarray: Low-resolution logits of shape CxHxW for subsequent inference, where H=W=256.
         """
-        features = self.model.image_encoder(im) if self.features is None else self.features
+        features = self.get_im_features(im) if self.features is None else self.features
 
         src_shape, dst_shape = self.batch[1][0].shape[:2], im.shape[2:]
         r = 1.0 if self.segment_all else min(dst_shape[0] / src_shape[0], dst_shape[1] / src_shape[1])
@@ -334,7 +334,7 @@ class Predictor(BasePredictor):
         """
         device = select_device(self.args.device, verbose=verbose)
         if model is None:
-            model = build_sam(self.args.model)
+            model = self.get_model()
         model.eval()
         self.model = model.to(device)
         self.device = device
@@ -348,6 +348,10 @@ class Predictor(BasePredictor):
         self.model.fp16 = False
         self.done_warmup = True
 
+    def get_model(self):
+        """Built Segment Anything Model (SAM) model."""
+        return build_sam(self.args.model)
+
     def postprocess(self, preds, img, orig_imgs):
         """
         Post-processes SAM's inference outputs to generate object detection masks and bounding boxes.
@@ -412,16 +416,18 @@ class Predictor(BasePredictor):
             AssertionError: If more than one image is set.
         """
         if self.model is None:
-            model = build_sam(self.args.model)
-            self.setup_model(model)
+            self.setup_model(model=None)
         self.setup_source(image)
         assert len(self.dataset) == 1, "`set_image` only supports setting one image!"
         for batch in self.dataset:
             im = self.preprocess(batch[1])
-            self.features = self.model.image_encoder(im)
-            self.im = im
+            self.features = self.get_im_features(im)
             break
 
+    def get_im_features(self, im):
+        """Get image features from the SAM image encoder."""
+        return self.model.image_encoder(im)
+
     def set_prompts(self, prompts):
         """Set prompts in advance."""
         self.prompts = prompts

ultralytics/models/sam2/__init__.py

@@ -0,0 +1,6 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license
+
+from .model import SAM2
+from .predict import SAM2Predictor
+
+__all__ = "SAM2", "SAM2Predictor"  # tuple or list

ultralytics/models/sam2/build.py

@@ -0,0 +1,156 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license
+
+import torch
+
+from ultralytics.utils.downloads import attempt_download_asset
+
+from .modules.encoders import FpnNeck, Hiera, ImageEncoder, MemoryEncoder
+from .modules.memory_attention import MemoryAttention, MemoryAttentionLayer
+from .modules.sam2 import SAM2Model
+
+
+def build_sam2_t(checkpoint=None):
+    """Build and return a Segment Anything Model (SAM2) tiny-size model with specified architecture parameters."""
+    return _build_sam2(
+        encoder_embed_dim=96,
+        encoder_stages=[1, 2, 7, 2],
+        encoder_num_heads=1,
+        encoder_global_att_blocks=[5, 7, 9],
+        encoder_window_spec=[8, 4, 14, 7],
+        encoder_backbone_channel_list=[768, 384, 192, 96],
+        checkpoint=checkpoint,
+    )
+
+
+def build_sam2_s(checkpoint=None):
+    """Builds and returns a small-size Segment Anything Model (SAM2) with specified architecture parameters."""
+    return _build_sam2(
+        encoder_embed_dim=96,
+        encoder_stages=[1, 2, 11, 2],
+        encoder_num_heads=1,
+        encoder_global_att_blocks=[7, 10, 13],
+        encoder_window_spec=[8, 4, 14, 7],
+        encoder_backbone_channel_list=[768, 384, 192, 96],
+        checkpoint=checkpoint,
+    )
+
+
+def build_sam2_b(checkpoint=None):
+    """Builds and returns a Segment Anything Model (SAM2) base-size model with specified architecture parameters."""
+    return _build_sam2(
+        encoder_embed_dim=112,
+        encoder_stages=[2, 3, 16, 3],
+        encoder_num_heads=2,
+        encoder_global_att_blocks=[12, 16, 20],
+        encoder_window_spec=[8, 4, 14, 7],
+        encoder_window_spatial_size=[14, 14],
+        encoder_backbone_channel_list=[896, 448, 224, 112],
+        checkpoint=checkpoint,
+    )
+
+
+def build_sam2_l(checkpoint=None):
+    """Build and return a Segment Anything Model (SAM2) large-size model with specified architecture parameters."""
+    return _build_sam2(
+        encoder_embed_dim=144,
+        encoder_stages=[2, 6, 36, 4],
+        encoder_num_heads=2,
+        encoder_global_att_blocks=[23, 33, 43],
+        encoder_window_spec=[8, 4, 16, 8],
+        encoder_backbone_channel_list=[1152, 576, 288, 144],
+        checkpoint=checkpoint,
+    )
+
+
+def _build_sam2(
+    encoder_embed_dim=1280,
+    encoder_stages=[2, 6, 36, 4],
+    encoder_num_heads=2,
+    encoder_global_att_blocks=[7, 15, 23, 31],
+    encoder_backbone_channel_list=[1152, 576, 288, 144],
+    encoder_window_spatial_size=[7, 7],
+    encoder_window_spec=[8, 4, 16, 8],
+    checkpoint=None,
+):
+    """Builds a SAM2 model with specified architecture parameters and optional checkpoint loading."""
+    image_encoder = ImageEncoder(
+        trunk=Hiera(
+            embed_dim=encoder_embed_dim,
+            num_heads=encoder_num_heads,
+            stages=encoder_stages,
+            global_att_blocks=encoder_global_att_blocks,
+            window_pos_embed_bkg_spatial_size=encoder_window_spatial_size,
+            window_spec=encoder_window_spec,
+        ),
+        neck=FpnNeck(
+            d_model=256,
+            backbone_channel_list=encoder_backbone_channel_list,
+            fpn_top_down_levels=[2, 3],
+            fpn_interp_model="nearest",
+        ),
+        scalp=1,
+    )
+    memory_attention = MemoryAttention(d_model=256, pos_enc_at_input=True, num_layers=4, layer=MemoryAttentionLayer())
+    memory_encoder = MemoryEncoder(out_dim=64)
+
+    sam2 = SAM2Model(
+        image_encoder=image_encoder,
+        memory_attention=memory_attention,
+        memory_encoder=memory_encoder,
+        num_maskmem=7,
+        image_size=1024,
+        sigmoid_scale_for_mem_enc=20.0,
+        sigmoid_bias_for_mem_enc=-10.0,
+        use_mask_input_as_output_without_sam=True,
+        directly_add_no_mem_embed=True,
+        use_high_res_features_in_sam=True,
+        multimask_output_in_sam=True,
+        iou_prediction_use_sigmoid=True,
+        use_obj_ptrs_in_encoder=True,
+        add_tpos_enc_to_obj_ptrs=True,
+        only_obj_ptrs_in_the_past_for_eval=True,
+        pred_obj_scores=True,
+        pred_obj_scores_mlp=True,
+        fixed_no_obj_ptr=True,
+        multimask_output_for_tracking=True,
+        use_multimask_token_for_obj_ptr=True,
+        multimask_min_pt_num=0,
+        multimask_max_pt_num=1,
+        use_mlp_for_obj_ptr_proj=True,
+        compile_image_encoder=False,
+        sam_mask_decoder_extra_args=dict(
+            dynamic_multimask_via_stability=True,
+            dynamic_multimask_stability_delta=0.05,
+            dynamic_multimask_stability_thresh=0.98,
+        ),
+    )
+
+    if checkpoint is not None:
+        checkpoint = attempt_download_asset(checkpoint)
+        with open(checkpoint, "rb") as f:
+            state_dict = torch.load(f)["model"]
+        sam2.load_state_dict(state_dict)
+    sam2.eval()
+    return sam2
+
+
+sam_model_map = {
+    "sam2_t.pt": build_sam2_t,
+    "sam2_s.pt": build_sam2_s,
+    "sam2_b.pt": build_sam2_b,
+    "sam2_l.pt": build_sam2_l,
+}
+
+
+def build_sam2(ckpt="sam_b.pt"):
+    """Constructs a Segment Anything Model (SAM2) based on the specified checkpoint, with various size options."""
+    model_builder = None
+    ckpt = str(ckpt)  # to allow Path ckpt types
+    for k in sam_model_map.keys():
+        if ckpt.endswith(k):
+            model_builder = sam_model_map.get(k)
+
+    if not model_builder:
+        raise FileNotFoundError(f"{ckpt} is not a supported SAM model. Available models are: \n {sam_model_map.keys()}")
+
+    return model_builder(ckpt)

ultralytics/models/sam2/model.py

@@ -0,0 +1,97 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license
+"""
+SAM2 model interface.
+
+This module provides an interface to the Segment Anything Model (SAM2) from Ultralytics, designed for real-time image
+segmentation tasks. The SAM2 model allows for promptable segmentation with unparalleled versatility in image analysis,
+and has been trained on the SA-1B dataset. It features zero-shot performance capabilities, enabling it to adapt to new
+image distributions and tasks without prior knowledge.
+
+Key Features:
+    - Promptable segmentation
+    - Real-time performance
+    - Zero-shot transfer capabilities
+    - Trained on SA-1B dataset
+"""
+
+from ultralytics.models.sam import SAM
+
+from .build import build_sam2
+from .predict import SAM2Predictor
+
+
+class SAM2(SAM):
+    """
+    SAM2 class for real-time image segmentation using the Segment Anything Model (SAM2).
+
+    This class extends the SAM base class, providing an interface to the SAM2 model for promptable segmentation
+    tasks. It supports loading pre-trained weights and offers zero-shot performance capabilities.
+
+    Attributes:
+        model (torch.nn.Module): The loaded SAM2 model.
+        task_map (Dict[str, Type[SAM2Predictor]]): Mapping of 'segment' task to SAM2Predictor.
+
+    Methods:
+        __init__: Initializes the SAM2 model with pre-trained weights.
+        _load: Loads specified weights into the SAM2 model.
+
+    Examples:
+        >>> sam2 = SAM2("sam2_b.pt")
+        >>> sam2._load('path/to/sam2_weights.pt')
+        >>> task_map = sam2.task_map
+        >>> print(task_map)
+        {'segment': SAM2Predictor}
+
+    Notes:
+        - Supports .pt and .pth file extensions for model weights.
+        - Offers zero-shot transfer capabilities for new image distributions and tasks.
+    """
+
+    def __init__(self, model="sam2_b.pt") -> None:
+        """
+        Initializes the SAM2 model with a pre-trained model file.
+
+        Args:
+            model (str): Path to the pre-trained SAM2 model file. File should have a .pt or .pth extension.
+
+        Raises:
+            NotImplementedError: If the model file extension is not .pt or .pth.
+
+        Examples:
+            >>> sam2 = SAM2("sam2_b.pt")
+        """
+        super().__init__(model=model)
+
+    def _load(self, weights: str, task=None):
+        """
+        Loads the specified weights into the SAM2 model.
+
+        This method is responsible for loading pre-trained weights into the SAM2 model. It supports loading
+        weights from files with .pt or .pth extensions.
+
+        Args:
+            weights (str): Path to the weights file. Should be a file with .pt or .pth extension.
+            task (str | None): Task name. If provided, it may be used to configure model-specific settings.
+
+        Examples:
+            >>> sam2_model = SAM2()
+            >>> sam2_model._load('path/to/sam2_weights.pt')
+        """
+        self.model = build_sam2(weights)
+
+    @property
+    def task_map(self):
+        """
+        Provides a mapping from the 'segment' task to its corresponding 'Predictor'.
+
+        Returns:
+            (Dict[str, Type[SAM2Predictor]]): A dictionary mapping the 'segment' task to its corresponding
+                SAM2Predictor class.
+
+        Examples:
+            >>> sam2 = SAM2()
+            >>> task_map = sam2.task_map
+            >>> print(task_map)
+            {'segment': SAM2Predictor}
+        """
+        return {"segment": {"predictor": SAM2Predictor}}

ultralytics/models/sam2/modules/__init__.py

@@ -0,0 +1 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license