update

ultralytics/models/sam/modules/__init__.py

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

ultralytics/models/sam/modules/decoders.py

@@ -0,0 +1,190 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license
+
+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
+
+
+class MaskDecoder(nn.Module):
+    """
+    Decoder module for generating masks and their associated quality scores, using a transformer architecture to predict
+    masks given image and prompt embeddings.
+
+    Attributes:
+        transformer_dim (int): Channel dimension for the transformer module.
+        transformer (nn.Module): The transformer module used for mask prediction.
+        num_multimask_outputs (int): Number of masks to predict for disambiguating masks.
+        iou_token (nn.Embedding): Embedding for the IoU token.
+        num_mask_tokens (int): Number of mask tokens.
+        mask_tokens (nn.Embedding): Embedding for the mask tokens.
+        output_upscaling (nn.Sequential): Neural network sequence for upscaling the output.
+        output_hypernetworks_mlps (nn.ModuleList): Hypernetwork MLPs for generating masks.
+        iou_prediction_head (nn.Module): MLP for predicting mask quality.
+    """
+
+    def __init__(
+        self,
+        *,
+        transformer_dim: int,
+        transformer: nn.Module,
+        num_multimask_outputs: int = 3,
+        activation: Type[nn.Module] = nn.GELU,
+        iou_head_depth: int = 3,
+        iou_head_hidden_dim: int = 256,
+    ) -> None:
+        """
+        Predicts masks given an image and prompt embeddings, using a transformer architecture.
+
+        Args:
+            transformer_dim (int): the channel dimension of the transformer module
+            transformer (nn.Module): the transformer used to predict masks
+            num_multimask_outputs (int): the number of masks to predict when disambiguating masks
+            activation (nn.Module): the type of activation to use when upscaling masks
+            iou_head_depth (int): the depth of the MLP used to predict mask quality
+            iou_head_hidden_dim (int): the hidden dimension of the MLP used to predict mask quality
+        """
+        super().__init__()
+        self.transformer_dim = transformer_dim
+        self.transformer = transformer
+
+        self.num_multimask_outputs = num_multimask_outputs
+
+        self.iou_token = nn.Embedding(1, transformer_dim)
+        self.num_mask_tokens = num_multimask_outputs + 1
+        self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
+
+        self.output_upscaling = nn.Sequential(
+            nn.ConvTranspose2d(transformer_dim, transformer_dim // 4, kernel_size=2, stride=2),
+            LayerNorm2d(transformer_dim // 4),
+            activation(),
+            nn.ConvTranspose2d(transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2),
+            activation(),
+        )
+        self.output_hypernetworks_mlps = nn.ModuleList(
+            [MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3) for _ in range(self.num_mask_tokens)]
+        )
+
+        self.iou_prediction_head = MLP(transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth)
+
+    def forward(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        multimask_output: bool,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Predict masks given image and prompt embeddings.
+
+        Args:
+            image_embeddings (torch.Tensor): the embeddings from the image encoder
+            image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
+            sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
+            dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
+            multimask_output (bool): Whether to return multiple masks or a single mask.
+
+        Returns:
+            torch.Tensor: batched predicted masks
+            torch.Tensor: batched predictions of mask quality
+        """
+        masks, iou_pred = self.predict_masks(
+            image_embeddings=image_embeddings,
+            image_pe=image_pe,
+            sparse_prompt_embeddings=sparse_prompt_embeddings,
+            dense_prompt_embeddings=dense_prompt_embeddings,
+        )
+
+        # Select the correct mask or masks for output
+        mask_slice = slice(1, None) if multimask_output else slice(0, 1)
+        masks = masks[:, mask_slice, :, :]
+        iou_pred = iou_pred[:, mask_slice]
+
+        # Prepare output
+        return masks, iou_pred
+
+    def predict_masks(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Predicts masks.
+
+        See 'forward' for more details.
+        """
+        # Concatenate output tokens
+        output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)
+        output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.shape[0], -1, -1)
+        tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
+
+        # Expand per-image data in batch direction to be per-mask
+        src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
+        src = src + dense_prompt_embeddings
+        pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
+        b, c, h, w = src.shape
+
+        # Run the transformer
+        hs, src = self.transformer(src, pos_src, tokens)
+        iou_token_out = hs[:, 0, :]
+        mask_tokens_out = hs[:, 1 : (1 + self.num_mask_tokens), :]
+
+        # Upscale mask embeddings and predict masks using the mask tokens
+        src = src.transpose(1, 2).view(b, c, h, w)
+        upscaled_embedding = self.output_upscaling(src)
+        hyper_in_list: List[torch.Tensor] = [
+            self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]) for i in range(self.num_mask_tokens)
+        ]
+        hyper_in = torch.stack(hyper_in_list, dim=1)
+        b, c, h, w = upscaled_embedding.shape
+        masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)
+
+        # Generate mask quality predictions
+        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

@@ -0,0 +1,603 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license
+
+from typing import Any, Optional, Tuple, Type
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ultralytics.nn.modules import LayerNorm2d, MLPBlock
+
+
+class ImageEncoderViT(nn.Module):
+    """
+    An image encoder using Vision Transformer (ViT) architecture for encoding an image into a compact latent space. The
+    encoder takes an image, splits it into patches, and processes these patches through a series of transformer blocks.
+    The encoded patches are then processed through a neck to generate the final encoded representation.
+
+    This class and its supporting functions below lightly adapted from the ViTDet backbone available at
+    https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py.
+
+    Attributes:
+        img_size (int): Dimension of input images, assumed to be square.
+        patch_embed (PatchEmbed): Module for patch embedding.
+        pos_embed (nn.Parameter, optional): Absolute positional embedding for patches.
+        blocks (nn.ModuleList): List of transformer blocks for processing patch embeddings.
+        neck (nn.Sequential): Neck module to further process the output.
+    """
+
+    def __init__(
+        self,
+        img_size: int = 1024,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        out_chans: int = 256,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_abs_pos: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        global_attn_indexes: Tuple[int, ...] = (),
+    ) -> None:
+        """
+        Args:
+            img_size (int): Input image size.
+            patch_size (int): Patch size.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+            depth (int): Depth of ViT.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_abs_pos (bool): If True, use absolute positional embeddings.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks.
+            global_attn_indexes (list): Indexes for blocks using global attention.
+        """
+        super().__init__()
+        self.img_size = img_size
+
+        self.patch_embed = PatchEmbed(
+            kernel_size=(patch_size, patch_size),
+            stride=(patch_size, patch_size),
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+        )
+
+        self.pos_embed: Optional[nn.Parameter] = None
+        if use_abs_pos:
+            # Initialize absolute positional embedding with pretrain image size.
+            self.pos_embed = nn.Parameter(torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim))
+
+        self.blocks = nn.ModuleList()
+        for i in range(depth):
+            block = Block(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                use_rel_pos=use_rel_pos,
+                rel_pos_zero_init=rel_pos_zero_init,
+                window_size=window_size if i not in global_attn_indexes else 0,
+                input_size=(img_size // patch_size, img_size // patch_size),
+            )
+            self.blocks.append(block)
+
+        self.neck = nn.Sequential(
+            nn.Conv2d(
+                embed_dim,
+                out_chans,
+                kernel_size=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+            nn.Conv2d(
+                out_chans,
+                out_chans,
+                kernel_size=3,
+                padding=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Processes input through patch embedding, applies positional embedding if present, and passes through blocks
+        and neck.
+        """
+        x = self.patch_embed(x)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+        for blk in self.blocks:
+            x = blk(x)
+        return self.neck(x.permute(0, 3, 1, 2))
+
+
+class PromptEncoder(nn.Module):
+    """
+    Encodes different types of prompts, including points, boxes, and masks, for input to SAM's mask decoder. The encoder
+    produces both sparse and dense embeddings for the input prompts.
+
+    Attributes:
+        embed_dim (int): Dimension of the embeddings.
+        input_image_size (Tuple[int, int]): Size of the input image as (H, W).
+        image_embedding_size (Tuple[int, int]): Spatial size of the image embedding as (H, W).
+        pe_layer (PositionEmbeddingRandom): Module for random position embedding.
+        num_point_embeddings (int): Number of point embeddings for different types of points.
+        point_embeddings (nn.ModuleList): List of point embeddings.
+        not_a_point_embed (nn.Embedding): Embedding for points that are not a part of any label.
+        mask_input_size (Tuple[int, int]): Size of the input mask.
+        mask_downscaling (nn.Sequential): Neural network for downscaling the mask.
+        no_mask_embed (nn.Embedding): Embedding for cases where no mask is provided.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        image_embedding_size: Tuple[int, int],
+        input_image_size: Tuple[int, int],
+        mask_in_chans: int,
+        activation: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        """
+        Encodes prompts for input to SAM's mask decoder.
+
+        Args:
+          embed_dim (int): The prompts' embedding dimension
+          image_embedding_size (tuple(int, int)): The spatial size of the
+            image embedding, as (H, W).
+          input_image_size (int): The padded size of the image as input
+            to the image encoder, as (H, W).
+          mask_in_chans (int): The number of hidden channels used for
+            encoding input masks.
+          activation (nn.Module): The activation to use when encoding
+            input masks.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.input_image_size = input_image_size
+        self.image_embedding_size = image_embedding_size
+        self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
+
+        self.num_point_embeddings: int = 4  # pos/neg point + 2 box corners
+        point_embeddings = [nn.Embedding(1, embed_dim) for _ in range(self.num_point_embeddings)]
+        self.point_embeddings = nn.ModuleList(point_embeddings)
+        self.not_a_point_embed = nn.Embedding(1, embed_dim)
+
+        self.mask_input_size = (4 * image_embedding_size[0], 4 * image_embedding_size[1])
+        self.mask_downscaling = nn.Sequential(
+            nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans // 4),
+            activation(),
+            nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans),
+            activation(),
+            nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
+        )
+        self.no_mask_embed = nn.Embedding(1, embed_dim)
+
+    def get_dense_pe(self) -> torch.Tensor:
+        """
+        Returns the positional encoding used to encode point prompts, applied to a dense set of points the shape of the
+        image encoding.
+
+        Returns:
+          torch.Tensor: Positional encoding with shape 1x(embed_dim)x(embedding_h)x(embedding_w)
+        """
+        return self.pe_layer(self.image_embedding_size).unsqueeze(0)
+
+    def _embed_points(self, points: torch.Tensor, labels: torch.Tensor, pad: bool) -> torch.Tensor:
+        """Embeds point prompts."""
+        points = points + 0.5  # Shift to center of pixel
+        if pad:
+            padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
+            padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
+            points = torch.cat([points, padding_point], dim=1)
+            labels = torch.cat([labels, padding_label], dim=1)
+        point_embedding = self.pe_layer.forward_with_coords(points, self.input_image_size)
+        point_embedding[labels == -1] = 0.0
+        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
+        return point_embedding
+
+    def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
+        """Embeds box prompts."""
+        boxes = boxes + 0.5  # Shift to center of pixel
+        coords = boxes.reshape(-1, 2, 2)
+        corner_embedding = self.pe_layer.forward_with_coords(coords, self.input_image_size)
+        corner_embedding[:, 0, :] += self.point_embeddings[2].weight
+        corner_embedding[:, 1, :] += self.point_embeddings[3].weight
+        return corner_embedding
+
+    def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
+        """Embeds mask inputs."""
+        return self.mask_downscaling(masks)
+
+    def _get_batch_size(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> int:
+        """Gets the batch size of the output given the batch size of the input prompts."""
+        if points is not None:
+            return points[0].shape[0]
+        elif boxes is not None:
+            return boxes.shape[0]
+        elif masks is not None:
+            return masks.shape[0]
+        else:
+            return 1
+
+    def _get_device(self) -> torch.device:
+        """Returns the device of the first point embedding's weight tensor."""
+        return self.point_embeddings[0].weight.device
+
+    def forward(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Embeds different types of prompts, returning both sparse and dense embeddings.
+
+        Args:
+          points (tuple(torch.Tensor, torch.Tensor), None): point coordinates and labels to embed.
+          boxes (torch.Tensor, None): boxes to embed
+          masks (torch.Tensor, None): masks to embed
+
+        Returns:
+          torch.Tensor: sparse embeddings for the points and boxes, with shape BxNx(embed_dim), where N is determined
+            by the number of input points and boxes.
+          torch.Tensor: dense embeddings for the masks, in the shape Bx(embed_dim)x(embed_H)x(embed_W)
+        """
+        bs = self._get_batch_size(points, boxes, masks)
+        sparse_embeddings = torch.empty((bs, 0, self.embed_dim), device=self._get_device())
+        if points is not None:
+            coords, labels = points
+            point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
+            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
+        if boxes is not None:
+            box_embeddings = self._embed_boxes(boxes)
+            sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
+
+        if masks is not None:
+            dense_embeddings = self._embed_masks(masks)
+        else:
+            dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
+                bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
+            )
+
+        return sparse_embeddings, dense_embeddings
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """Positional encoding using random spatial frequencies."""
+
+    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
+        """Initializes a position embedding using random spatial frequencies."""
+        super().__init__()
+        if scale is None or scale <= 0.0:
+            scale = 1.0
+        self.register_buffer("positional_encoding_gaussian_matrix", scale * torch.randn((2, num_pos_feats)))
+
+        # Set non-deterministic for forward() error 'cumsum_cuda_kernel does not have a deterministic implementation'
+        torch.use_deterministic_algorithms(False)
+        torch.backends.cudnn.deterministic = False
+
+    def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
+        """Positionally encode points that are normalized to [0,1]."""
+        # Assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
+        coords = 2 * coords - 1
+        coords = coords @ self.positional_encoding_gaussian_matrix
+        coords = 2 * np.pi * coords
+        # Outputs d_1 x ... x d_n x C shape
+        return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
+
+    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
+        """Generate positional encoding for a grid of the specified size."""
+        h, w = size
+        device: Any = self.positional_encoding_gaussian_matrix.device
+        grid = torch.ones((h, w), device=device, dtype=torch.float32)
+        y_embed = grid.cumsum(dim=0) - 0.5
+        x_embed = grid.cumsum(dim=1) - 0.5
+        y_embed = y_embed / h
+        x_embed = x_embed / w
+
+        pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
+        return pe.permute(2, 0, 1)  # C x H x W
+
+    def forward_with_coords(self, coords_input: torch.Tensor, image_size: Tuple[int, int]) -> torch.Tensor:
+        """Positionally encode points that are not normalized to [0,1]."""
+        coords = coords_input.clone()
+        coords[:, :, 0] = coords[:, :, 0] / image_size[1]
+        coords[:, :, 1] = coords[:, :, 1] / image_size[0]
+        return self._pe_encoding(coords.to(torch.float))  # B x N x C
+
+
+class Block(nn.Module):
+    """Transformer blocks with support of window attention and residual propagation blocks."""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks. If it equals 0, then
+                use global attention.
+            input_size (tuple(int, int), None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            input_size=input_size if window_size == 0 else (window_size, window_size),
+        )
+
+        self.norm2 = norm_layer(dim)
+        self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)
+
+        self.window_size = window_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Executes a forward pass through the transformer block with window attention and non-overlapping windows."""
+        shortcut = x
+        x = self.norm1(x)
+        # Window partition
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+
+        x = self.attn(x)
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+
+        x = shortcut + x
+        return x + self.mlp(self.norm2(x))
+
+
+class Attention(nn.Module):
+    """Multi-head Attention block with relative position embeddings."""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Initialize Attention module.
+
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads.
+            qkv_bias (bool):  If True, add a learnable bias to query, key, value.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            input_size (tuple(int, int), None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        self.use_rel_pos = use_rel_pos
+        if self.use_rel_pos:
+            assert input_size is not None, "Input size must be provided if using relative positional encoding."
+            # Initialize relative positional embeddings
+            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
+            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Applies the forward operation including attention, normalization, MLP, and indexing within window limits."""
+        B, H, W, _ = x.shape
+        # qkv with shape (3, B, nHead, H * W, C)
+        qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        # q, k, v with shape (B * nHead, H * W, C)
+        q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
+
+        attn = (q * self.scale) @ k.transpose(-2, -1)
+
+        if self.use_rel_pos:
+            attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))
+
+        attn = attn.softmax(dim=-1)
+        x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)
+        return self.proj(x)
+
+
+def window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(
+    windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
+) -> torch.Tensor:
+    """
+    Window unpartition into original sequences and removing padding.
+
+    Args:
+        windows (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
+    """
+    Get relative positional embeddings according to the relative positions of query and key sizes.
+
+    Args:
+        q_size (int): size of query q.
+        k_size (int): size of key k.
+        rel_pos (Tensor): relative position embeddings (L, C).
+
+    Returns:
+        Extracted positional embeddings according to relative positions.
+    """
+    max_rel_dist = int(2 * max(q_size, k_size) - 1)
+    # Interpolate rel pos if needed.
+    if rel_pos.shape[0] != max_rel_dist:
+        # Interpolate rel pos.
+        rel_pos_resized = F.interpolate(
+            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
+            size=max_rel_dist,
+            mode="linear",
+        )
+        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
+    else:
+        rel_pos_resized = rel_pos
+
+    # Scale the coords with short length if shapes for q and k are different.
+    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
+    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
+    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
+
+    return rel_pos_resized[relative_coords.long()]
+
+
+def add_decomposed_rel_pos(
+    attn: torch.Tensor,
+    q: torch.Tensor,
+    rel_pos_h: torch.Tensor,
+    rel_pos_w: torch.Tensor,
+    q_size: Tuple[int, int],
+    k_size: Tuple[int, int],
+) -> torch.Tensor:
+    """
+    Calculate decomposed Relative Positional Embeddings from mvitv2 paper at
+    https://github.com/facebookresearch/mvit/blob/main/mvit/models/attention.py.
+
+    Args:
+        attn (Tensor): attention map.
+        q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
+        rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
+        rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
+        q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
+        k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
+
+    Returns:
+        attn (Tensor): attention map with added relative positional embeddings.
+    """
+    q_h, q_w = q_size
+    k_h, k_w = k_size
+    Rh = get_rel_pos(q_h, k_h, rel_pos_h)
+    Rw = get_rel_pos(q_w, k_w, rel_pos_w)
+
+    B, _, dim = q.shape
+    r_q = q.reshape(B, q_h, q_w, dim)
+    rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
+    rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
+
+    attn = (attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]).view(
+        B, q_h * q_w, k_h * k_w
+    )
+
+    return attn
+
+
+class PatchEmbed(nn.Module):
+    """Image to Patch Embedding."""
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, int] = (16, 16),
+        stride: Tuple[int, int] = (16, 16),
+        padding: Tuple[int, int] = (0, 0),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ) -> None:
+        """
+        Initialize PatchEmbed module.
+
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Computes patch embedding by applying convolution and transposing resulting tensor."""
+        return self.proj(x).permute(0, 2, 3, 1)  # B C H W -> B H W C

ultralytics/models/sam/modules/sam.py

@@ -0,0 +1,65 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List
+
+import torch
+from torch import nn
+
+from .decoders import MaskDecoder
+from .encoders import ImageEncoderViT, PromptEncoder
+
+
+class Sam(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.
+
+    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.
+        pixel_mean (List[float]): Mean pixel values for image normalization.
+        pixel_std (List[float]): Standard deviation values for image normalization.
+    """
+
+    mask_threshold: float = 0.0
+    image_format: str = "RGB"
+
+    def __init__(
+        self,
+        image_encoder: ImageEncoderViT,
+        prompt_encoder: PromptEncoder,
+        mask_decoder: MaskDecoder,
+        pixel_mean: List[float] = (123.675, 116.28, 103.53),
+        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.
+
+        Note:
+            All forward() operations moved to SAMPredictor.
+
+        Args:
+            image_encoder (ImageEncoderViT): The backbone used to encode the image into image embeddings.
+            prompt_encoder (PromptEncoder): Encodes various types of input prompts.
+            mask_decoder (MaskDecoder): Predicts masks from the image embeddings and encoded prompts.
+            pixel_mean (List[float], optional): Mean values for normalizing pixels in the input image. Defaults to
+                (123.675, 116.28, 103.53).
+            pixel_std (List[float], optional): Std values for normalizing pixels in the input image. Defaults to
+                (58.395, 57.12, 57.375).
+        """
+        super().__init__()
+        self.image_encoder = image_encoder
+        self.prompt_encoder = prompt_encoder
+        self.mask_decoder = mask_decoder
+        self.register_buffer("pixel_mean", torch.Tensor(pixel_mean).view(-1, 1, 1), False)
+        self.register_buffer("pixel_std", torch.Tensor(pixel_std).view(-1, 1, 1), False)

ultralytics/models/sam/modules/tiny_encoder.py

@@ -0,0 +1,742 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license
+
+# --------------------------------------------------------
+# TinyViT Model Architecture
+# Copyright (c) 2022 Microsoft
+# Adapted from LeViT and Swin Transformer
+#   LeViT: (https://github.com/facebookresearch/levit)
+#   Swin: (https://github.com/microsoft/swin-transformer)
+# Build the TinyViT Model
+# --------------------------------------------------------
+
+import itertools
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+from ultralytics.utils.instance import to_2tuple
+
+
+class Conv2d_BN(torch.nn.Sequential):
+    """A sequential container that performs 2D convolution followed by batch normalization."""
+
+    def __init__(self, a, b, ks=1, stride=1, pad=0, dilation=1, groups=1, bn_weight_init=1):
+        """Initializes the MBConv model with given input channels, output channels, expansion ratio, activation, and
+        drop path.
+        """
+        super().__init__()
+        self.add_module("c", torch.nn.Conv2d(a, b, ks, stride, pad, dilation, groups, bias=False))
+        bn = torch.nn.BatchNorm2d(b)
+        torch.nn.init.constant_(bn.weight, bn_weight_init)
+        torch.nn.init.constant_(bn.bias, 0)
+        self.add_module("bn", bn)
+
+
+class PatchEmbed(nn.Module):
+    """Embeds images into patches and projects them into a specified embedding dimension."""
+
+    def __init__(self, in_chans, embed_dim, resolution, activation):
+        """Initialize the PatchMerging class with specified input, output dimensions, resolution and activation
+        function.
+        """
+        super().__init__()
+        img_size: Tuple[int, int] = to_2tuple(resolution)
+        self.patches_resolution = (img_size[0] // 4, img_size[1] // 4)
+        self.num_patches = self.patches_resolution[0] * self.patches_resolution[1]
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+        n = embed_dim
+        self.seq = nn.Sequential(
+            Conv2d_BN(in_chans, n // 2, 3, 2, 1),
+            activation(),
+            Conv2d_BN(n // 2, n, 3, 2, 1),
+        )
+
+    def forward(self, x):
+        """Runs input tensor 'x' through the PatchMerging model's sequence of operations."""
+        return self.seq(x)
+
+
+class MBConv(nn.Module):
+    """Mobile Inverted Bottleneck Conv (MBConv) layer, part of the EfficientNet architecture."""
+
+    def __init__(self, in_chans, out_chans, expand_ratio, activation, drop_path):
+        """Initializes a convolutional layer with specified dimensions, input resolution, depth, and activation
+        function.
+        """
+        super().__init__()
+        self.in_chans = in_chans
+        self.hidden_chans = int(in_chans * expand_ratio)
+        self.out_chans = out_chans
+
+        self.conv1 = Conv2d_BN(in_chans, self.hidden_chans, ks=1)
+        self.act1 = activation()
+
+        self.conv2 = Conv2d_BN(self.hidden_chans, self.hidden_chans, ks=3, stride=1, pad=1, groups=self.hidden_chans)
+        self.act2 = activation()
+
+        self.conv3 = Conv2d_BN(self.hidden_chans, out_chans, ks=1, bn_weight_init=0.0)
+        self.act3 = activation()
+
+        # NOTE: `DropPath` is needed only for training.
+        # self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.drop_path = nn.Identity()
+
+    def forward(self, x):
+        """Implements the forward pass for the model architecture."""
+        shortcut = x
+        x = self.conv1(x)
+        x = self.act1(x)
+        x = self.conv2(x)
+        x = self.act2(x)
+        x = self.conv3(x)
+        x = self.drop_path(x)
+        x += shortcut
+        return self.act3(x)
+
+
+class PatchMerging(nn.Module):
+    """Merges neighboring patches in the feature map and projects to a new dimension."""
+
+    def __init__(self, input_resolution, dim, out_dim, activation):
+        """Initializes the ConvLayer with specific dimension, input resolution, depth, activation, drop path, and other
+        optional parameters.
+        """
+        super().__init__()
+
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.out_dim = out_dim
+        self.act = activation()
+        self.conv1 = Conv2d_BN(dim, out_dim, 1, 1, 0)
+        stride_c = 1 if out_dim in [320, 448, 576] else 2
+        self.conv2 = Conv2d_BN(out_dim, out_dim, 3, stride_c, 1, groups=out_dim)
+        self.conv3 = Conv2d_BN(out_dim, out_dim, 1, 1, 0)
+
+    def forward(self, x):
+        """Applies forward pass on the input utilizing convolution and activation layers, and returns the result."""
+        if x.ndim == 3:
+            H, W = self.input_resolution
+            B = len(x)
+            # (B, C, H, W)
+            x = x.view(B, H, W, -1).permute(0, 3, 1, 2)
+
+        x = self.conv1(x)
+        x = self.act(x)
+
+        x = self.conv2(x)
+        x = self.act(x)
+        x = self.conv3(x)
+        return x.flatten(2).transpose(1, 2)
+
+
+class ConvLayer(nn.Module):
+    """
+    Convolutional Layer featuring multiple MobileNetV3-style inverted bottleneck convolutions (MBConv).
+
+    Optionally applies downsample operations to the output, and provides support for gradient checkpointing.
+    """
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        depth,
+        activation,
+        drop_path=0.0,
+        downsample=None,
+        use_checkpoint=False,
+        out_dim=None,
+        conv_expand_ratio=4.0,
+    ):
+        """
+        Initializes the ConvLayer with the given dimensions and settings.
+
+        Args:
+            dim (int): The dimensionality of the input and output.
+            input_resolution (Tuple[int, int]): The resolution of the input image.
+            depth (int): The number of MBConv layers in the block.
+            activation (Callable): Activation function applied after each convolution.
+            drop_path (Union[float, List[float]]): Drop path rate. Single float or a list of floats for each MBConv.
+            downsample (Optional[Callable]): Function for downsampling the output. None to skip downsampling.
+            use_checkpoint (bool): Whether to use gradient checkpointing to save memory.
+            out_dim (Optional[int]): The dimensionality of the output. None means it will be the same as `dim`.
+            conv_expand_ratio (float): Expansion ratio for the MBConv layers.
+        """
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # Build blocks
+        self.blocks = nn.ModuleList(
+            [
+                MBConv(
+                    dim,
+                    dim,
+                    conv_expand_ratio,
+                    activation,
+                    drop_path[i] if isinstance(drop_path, list) else drop_path,
+                )
+                for i in range(depth)
+            ]
+        )
+
+        # Patch merging layer
+        self.downsample = (
+            None
+            if downsample is None
+            else downsample(input_resolution, dim=dim, out_dim=out_dim, activation=activation)
+        )
+
+    def forward(self, x):
+        """Processes the input through a series of convolutional layers and returns the activated output."""
+        for blk in self.blocks:
+            x = checkpoint.checkpoint(blk, x) if self.use_checkpoint else blk(x)
+        return x if self.downsample is None else self.downsample(x)
+
+
+class Mlp(nn.Module):
+    """
+    Multi-layer Perceptron (MLP) for transformer architectures.
+
+    This layer takes an input with in_features, applies layer normalization and two fully-connected layers.
+    """
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0):
+        """Initializes Attention module with the given parameters including dimension, key_dim, number of heads, etc."""
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.norm = nn.LayerNorm(in_features)
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.act = act_layer()
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        """Applies operations on input x and returns modified x, runs downsample if not None."""
+        x = self.norm(x)
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        return self.drop(x)
+
+
+class Attention(torch.nn.Module):
+    """
+    Multi-head attention module with support for spatial awareness, applying attention biases based on spatial
+    resolution. Implements trainable attention biases for each unique offset between spatial positions in the resolution
+    grid.
+
+    Attributes:
+        ab (Tensor, optional): Cached attention biases for inference, deleted during training.
+    """
+
+    def __init__(
+        self,
+        dim,
+        key_dim,
+        num_heads=8,
+        attn_ratio=4,
+        resolution=(14, 14),
+    ):
+        """
+        Initializes the Attention module.
+
+        Args:
+            dim (int): The dimensionality of the input and output.
+            key_dim (int): The dimensionality of the keys and queries.
+            num_heads (int, optional): Number of attention heads. Default is 8.
+            attn_ratio (float, optional): Attention ratio, affecting the dimensions of the value vectors. Default is 4.
+            resolution (Tuple[int, int], optional): Spatial resolution of the input feature map. Default is (14, 14).
+
+        Raises:
+            AssertionError: If `resolution` is not a tuple of length 2.
+        """
+        super().__init__()
+
+        assert isinstance(resolution, tuple) and len(resolution) == 2
+        self.num_heads = num_heads
+        self.scale = key_dim**-0.5
+        self.key_dim = key_dim
+        self.nh_kd = nh_kd = key_dim * num_heads
+        self.d = int(attn_ratio * key_dim)
+        self.dh = int(attn_ratio * key_dim) * num_heads
+        self.attn_ratio = attn_ratio
+        h = self.dh + nh_kd * 2
+
+        self.norm = nn.LayerNorm(dim)
+        self.qkv = nn.Linear(dim, h)
+        self.proj = nn.Linear(self.dh, dim)
+
+        points = list(itertools.product(range(resolution[0]), range(resolution[1])))
+        N = len(points)
+        attention_offsets = {}
+        idxs = []
+        for p1 in points:
+            for p2 in points:
+                offset = (abs(p1[0] - p2[0]), abs(p1[1] - p2[1]))
+                if offset not in attention_offsets:
+                    attention_offsets[offset] = len(attention_offsets)
+                idxs.append(attention_offsets[offset])
+        self.attention_biases = torch.nn.Parameter(torch.zeros(num_heads, len(attention_offsets)))
+        self.register_buffer("attention_bias_idxs", torch.LongTensor(idxs).view(N, N), persistent=False)
+
+    @torch.no_grad()
+    def train(self, mode=True):
+        """Sets the module in training mode and handles attribute 'ab' based on the mode."""
+        super().train(mode)
+        if mode and hasattr(self, "ab"):
+            del self.ab
+        else:
+            self.ab = self.attention_biases[:, self.attention_bias_idxs]
+
+    def forward(self, x):  # x
+        """Performs forward pass over the input tensor 'x' by applying normalization and querying keys/values."""
+        B, N, _ = x.shape  # B, N, C
+
+        # Normalization
+        x = self.norm(x)
+
+        qkv = self.qkv(x)
+        # (B, N, num_heads, d)
+        q, k, v = qkv.view(B, N, self.num_heads, -1).split([self.key_dim, self.key_dim, self.d], dim=3)
+        # (B, num_heads, N, d)
+        q = q.permute(0, 2, 1, 3)
+        k = k.permute(0, 2, 1, 3)
+        v = v.permute(0, 2, 1, 3)
+        self.ab = self.ab.to(self.attention_biases.device)
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale + (
+            self.attention_biases[:, self.attention_bias_idxs] if self.training else self.ab
+        )
+        attn = attn.softmax(dim=-1)
+        x = (attn @ v).transpose(1, 2).reshape(B, N, self.dh)
+        return self.proj(x)
+
+
+class TinyViTBlock(nn.Module):
+    """TinyViT Block that applies self-attention and a local convolution to the input."""
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        num_heads,
+        window_size=7,
+        mlp_ratio=4.0,
+        drop=0.0,
+        drop_path=0.0,
+        local_conv_size=3,
+        activation=nn.GELU,
+    ):
+        """
+        Initializes the TinyViTBlock.
+
+        Args:
+            dim (int): The dimensionality of the input and output.
+            input_resolution (Tuple[int, int]): Spatial resolution of the input feature map.
+            num_heads (int): Number of attention heads.
+            window_size (int, optional): Window size for attention. Default is 7.
+            mlp_ratio (float, optional): Ratio of mlp hidden dim to embedding dim. Default is 4.
+            drop (float, optional): Dropout rate. Default is 0.
+            drop_path (float, optional): Stochastic depth rate. Default is 0.
+            local_conv_size (int, optional): The kernel size of the local convolution. Default is 3.
+            activation (torch.nn, optional): Activation function for MLP. Default is nn.GELU.
+
+        Raises:
+            AssertionError: If `window_size` is not greater than 0.
+            AssertionError: If `dim` is not divisible by `num_heads`.
+        """
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        assert window_size > 0, "window_size must be greater than 0"
+        self.window_size = window_size
+        self.mlp_ratio = mlp_ratio
+
+        # NOTE: `DropPath` is needed only for training.
+        # self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.drop_path = nn.Identity()
+
+        assert dim % num_heads == 0, "dim must be divisible by num_heads"
+        head_dim = dim // num_heads
+
+        window_resolution = (window_size, window_size)
+        self.attn = Attention(dim, head_dim, num_heads, attn_ratio=1, resolution=window_resolution)
+
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        mlp_activation = activation
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=mlp_activation, drop=drop)
+
+        pad = local_conv_size // 2
+        self.local_conv = Conv2d_BN(dim, dim, ks=local_conv_size, stride=1, pad=pad, groups=dim)
+
+    def forward(self, x):
+        """Applies attention-based transformation or padding to input 'x' before passing it through a local
+        convolution.
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        res_x = x
+        if H == self.window_size and W == self.window_size:
+            x = self.attn(x)
+        else:
+            x = x.view(B, H, W, C)
+            pad_b = (self.window_size - H % self.window_size) % self.window_size
+            pad_r = (self.window_size - W % self.window_size) % self.window_size
+            padding = pad_b > 0 or pad_r > 0
+
+            if padding:
+                x = F.pad(x, (0, 0, 0, pad_r, 0, pad_b))
+
+            pH, pW = H + pad_b, W + pad_r
+            nH = pH // self.window_size
+            nW = pW // self.window_size
+            # Window partition
+            x = (
+                x.view(B, nH, self.window_size, nW, self.window_size, C)
+                .transpose(2, 3)
+                .reshape(B * nH * nW, self.window_size * self.window_size, C)
+            )
+            x = self.attn(x)
+            # Window reverse
+            x = x.view(B, nH, nW, self.window_size, self.window_size, C).transpose(2, 3).reshape(B, pH, pW, C)
+
+            if padding:
+                x = x[:, :H, :W].contiguous()
+
+            x = x.view(B, L, C)
+
+        x = res_x + self.drop_path(x)
+
+        x = x.transpose(1, 2).reshape(B, C, H, W)
+        x = self.local_conv(x)
+        x = x.view(B, C, L).transpose(1, 2)
+
+        return x + self.drop_path(self.mlp(x))
+
+    def extra_repr(self) -> str:
+        """Returns a formatted string representing the TinyViTBlock's parameters: dimension, input resolution, number of
+        attentions heads, window size, and MLP ratio.
+        """
+        return (
+            f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, "
+            f"window_size={self.window_size}, mlp_ratio={self.mlp_ratio}"
+        )
+
+
+class BasicLayer(nn.Module):
+    """A basic TinyViT layer for one stage in a TinyViT architecture."""
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        depth,
+        num_heads,
+        window_size,
+        mlp_ratio=4.0,
+        drop=0.0,
+        drop_path=0.0,
+        downsample=None,
+        use_checkpoint=False,
+        local_conv_size=3,
+        activation=nn.GELU,
+        out_dim=None,
+    ):
+        """
+        Initializes the BasicLayer.
+
+        Args:
+            dim (int): The dimensionality of the input and output.
+            input_resolution (Tuple[int, int]): Spatial resolution of the input feature map.
+            depth (int): Number of TinyViT blocks.
+            num_heads (int): Number of attention heads.
+            window_size (int): Local window size.
+            mlp_ratio (float, optional): Ratio of mlp hidden dim to embedding dim. Default is 4.
+            drop (float, optional): Dropout rate. Default is 0.
+            drop_path (float | tuple[float], optional): Stochastic depth rate. Default is 0.
+            downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default is None.
+            use_checkpoint (bool, optional): Whether to use checkpointing to save memory. Default is False.
+            local_conv_size (int, optional): Kernel size of the local convolution. Default is 3.
+            activation (torch.nn, optional): Activation function for MLP. Default is nn.GELU.
+            out_dim (int | None, optional): The output dimension of the layer. Default is None.
+
+        Raises:
+            ValueError: If `drop_path` is a list of float but its length doesn't match `depth`.
+        """
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # Build blocks
+        self.blocks = nn.ModuleList(
+            [
+                TinyViTBlock(
+                    dim=dim,
+                    input_resolution=input_resolution,
+                    num_heads=num_heads,
+                    window_size=window_size,
+                    mlp_ratio=mlp_ratio,
+                    drop=drop,
+                    drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                    local_conv_size=local_conv_size,
+                    activation=activation,
+                )
+                for i in range(depth)
+            ]
+        )
+
+        # Patch merging layer
+        self.downsample = (
+            None
+            if downsample is None
+            else downsample(input_resolution, dim=dim, out_dim=out_dim, activation=activation)
+        )
+
+    def forward(self, x):
+        """Performs forward propagation on the input tensor and returns a normalized tensor."""
+        for blk in self.blocks:
+            x = checkpoint.checkpoint(blk, x) if self.use_checkpoint else blk(x)
+        return x if self.downsample is None else self.downsample(x)
+
+    def extra_repr(self) -> str:
+        """Returns a string representation of the extra_repr function with the layer's parameters."""
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+
+class LayerNorm2d(nn.Module):
+    """A PyTorch implementation of Layer Normalization in 2D."""
+
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        """Initialize LayerNorm2d with the number of channels and an optional epsilon."""
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Perform a forward pass, normalizing the input tensor."""
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        return self.weight[:, None, None] * x + self.bias[:, None, None]
+
+
+class TinyViT(nn.Module):
+    """
+    The TinyViT architecture for vision tasks.
+
+    Attributes:
+        img_size (int): Input image size.
+        in_chans (int): Number of input channels.
+        num_classes (int): Number of classification classes.
+        embed_dims (List[int]): List of embedding dimensions for each layer.
+        depths (List[int]): List of depths for each layer.
+        num_heads (List[int]): List of number of attention heads for each layer.
+        window_sizes (List[int]): List of window sizes for each layer.
+        mlp_ratio (float): Ratio of MLP hidden dimension to embedding dimension.
+        drop_rate (float): Dropout rate for drop layers.
+        drop_path_rate (float): Drop path rate for stochastic depth.
+        use_checkpoint (bool): Use checkpointing for efficient memory usage.
+        mbconv_expand_ratio (float): Expansion ratio for MBConv layer.
+        local_conv_size (int): Local convolution kernel size.
+        layer_lr_decay (float): Layer-wise learning rate decay.
+
+    Note:
+        This implementation is generalized to accept a list of depths, attention heads,
+        embedding dimensions and window sizes, which allows you to create a
+        "stack" of TinyViT models of varying configurations.
+    """
+
+    def __init__(
+        self,
+        img_size=224,
+        in_chans=3,
+        num_classes=1000,
+        embed_dims=[96, 192, 384, 768],
+        depths=[2, 2, 6, 2],
+        num_heads=[3, 6, 12, 24],
+        window_sizes=[7, 7, 14, 7],
+        mlp_ratio=4.0,
+        drop_rate=0.0,
+        drop_path_rate=0.1,
+        use_checkpoint=False,
+        mbconv_expand_ratio=4.0,
+        local_conv_size=3,
+        layer_lr_decay=1.0,
+    ):
+        """
+        Initializes the TinyViT model.
+
+        Args:
+            img_size (int, optional): The input image size. Defaults to 224.
+            in_chans (int, optional): Number of input channels. Defaults to 3.
+            num_classes (int, optional): Number of classification classes. Defaults to 1000.
+            embed_dims (List[int], optional): List of embedding dimensions for each layer. Defaults to [96, 192, 384, 768].
+            depths (List[int], optional): List of depths for each layer. Defaults to [2, 2, 6, 2].
+            num_heads (List[int], optional): List of number of attention heads for each layer. Defaults to [3, 6, 12, 24].
+            window_sizes (List[int], optional): List of window sizes for each layer. Defaults to [7, 7, 14, 7].
+            mlp_ratio (float, optional): Ratio of MLP hidden dimension to embedding dimension. Defaults to 4.
+            drop_rate (float, optional): Dropout rate. Defaults to 0.
+            drop_path_rate (float, optional): Drop path rate for stochastic depth. Defaults to 0.1.
+            use_checkpoint (bool, optional): Whether to use checkpointing for efficient memory usage. Defaults to False.
+            mbconv_expand_ratio (float, optional): Expansion ratio for MBConv layer. Defaults to 4.0.
+            local_conv_size (int, optional): Local convolution kernel size. Defaults to 3.
+            layer_lr_decay (float, optional): Layer-wise learning rate decay. Defaults to 1.0.
+        """
+        super().__init__()
+        self.img_size = img_size
+        self.num_classes = num_classes
+        self.depths = depths
+        self.num_layers = len(depths)
+        self.mlp_ratio = mlp_ratio
+
+        activation = nn.GELU
+
+        self.patch_embed = PatchEmbed(
+            in_chans=in_chans, embed_dim=embed_dims[0], resolution=img_size, activation=activation
+        )
+
+        patches_resolution = self.patch_embed.patches_resolution
+        self.patches_resolution = patches_resolution
+
+        # Stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # Build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            kwargs = dict(
+                dim=embed_dims[i_layer],
+                input_resolution=(
+                    patches_resolution[0] // (2 ** (i_layer - 1 if i_layer == 3 else i_layer)),
+                    patches_resolution[1] // (2 ** (i_layer - 1 if i_layer == 3 else i_layer)),
+                ),
+                #   input_resolution=(patches_resolution[0] // (2 ** i_layer),
+                #                     patches_resolution[1] // (2 ** i_layer)),
+                depth=depths[i_layer],
+                drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])],
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint,
+                out_dim=embed_dims[min(i_layer + 1, len(embed_dims) - 1)],
+                activation=activation,
+            )
+            if i_layer == 0:
+                layer = ConvLayer(conv_expand_ratio=mbconv_expand_ratio, **kwargs)
+            else:
+                layer = BasicLayer(
+                    num_heads=num_heads[i_layer],
+                    window_size=window_sizes[i_layer],
+                    mlp_ratio=self.mlp_ratio,
+                    drop=drop_rate,
+                    local_conv_size=local_conv_size,
+                    **kwargs,
+                )
+            self.layers.append(layer)
+
+        # Classifier head
+        self.norm_head = nn.LayerNorm(embed_dims[-1])
+        self.head = nn.Linear(embed_dims[-1], num_classes) if num_classes > 0 else torch.nn.Identity()
+
+        # Init weights
+        self.apply(self._init_weights)
+        self.set_layer_lr_decay(layer_lr_decay)
+        self.neck = nn.Sequential(
+            nn.Conv2d(
+                embed_dims[-1],
+                256,
+                kernel_size=1,
+                bias=False,
+            ),
+            LayerNorm2d(256),
+            nn.Conv2d(
+                256,
+                256,
+                kernel_size=3,
+                padding=1,
+                bias=False,
+            ),
+            LayerNorm2d(256),
+        )
+
+    def set_layer_lr_decay(self, layer_lr_decay):
+        """Sets the learning rate decay for each layer in the TinyViT model."""
+        decay_rate = layer_lr_decay
+
+        # Layers -> blocks (depth)
+        depth = sum(self.depths)
+        lr_scales = [decay_rate ** (depth - i - 1) for i in range(depth)]
+
+        def _set_lr_scale(m, scale):
+            """Sets the learning rate scale for each layer in the model based on the layer's depth."""
+            for p in m.parameters():
+                p.lr_scale = scale
+
+        self.patch_embed.apply(lambda x: _set_lr_scale(x, lr_scales[0]))
+        i = 0
+        for layer in self.layers:
+            for block in layer.blocks:
+                block.apply(lambda x: _set_lr_scale(x, lr_scales[i]))
+                i += 1
+            if layer.downsample is not None:
+                layer.downsample.apply(lambda x: _set_lr_scale(x, lr_scales[i - 1]))
+        assert i == depth
+        for m in [self.norm_head, self.head]:
+            m.apply(lambda x: _set_lr_scale(x, lr_scales[-1]))
+
+        for k, p in self.named_parameters():
+            p.param_name = k
+
+        def _check_lr_scale(m):
+            """Checks if the learning rate scale attribute is present in module's parameters."""
+            for p in m.parameters():
+                assert hasattr(p, "lr_scale"), p.param_name
+
+        self.apply(_check_lr_scale)
+
+    def _init_weights(self, m):
+        """Initializes weights for linear layers and layer normalization in the given module."""
+        if isinstance(m, nn.Linear):
+            # NOTE: This initialization is needed only for training.
+            # trunc_normal_(m.weight, std=.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        """Returns a dictionary of parameter names where weight decay should not be applied."""
+        return {"attention_biases"}
+
+    def forward_features(self, x):
+        """Runs the input through the model layers and returns the transformed output."""
+        x = self.patch_embed(x)  # x input is (N, C, H, W)
+
+        x = self.layers[0](x)
+        start_i = 1
+
+        for i in range(start_i, len(self.layers)):
+            layer = self.layers[i]
+            x = layer(x)
+        B, _, C = x.shape
+        x = x.view(B, 64, 64, C)
+        x = x.permute(0, 3, 1, 2)
+        return self.neck(x)
+
+    def forward(self, x):
+        """Executes a forward pass on the input tensor through the constructed model layers."""
+        return self.forward_features(x)

ultralytics/models/sam/modules/transformer.py

@@ -0,0 +1,274 @@
+# Ultralytics YOLO 🚀, AGPL-3.0 license
+
+import math
+from typing import Tuple, Type
+
+import torch
+from torch import Tensor, nn
+
+from ultralytics.nn.modules import MLPBlock
+
+
+class TwoWayTransformer(nn.Module):
+    """
+    A Two-Way Transformer module that enables the simultaneous attention to both image and query points. This class
+    serves as a specialized transformer decoder that attends to an input image using queries whose positional embedding
+    is supplied. This is particularly useful for tasks like object detection, image segmentation, and point cloud
+    processing.
+
+    Attributes:
+        depth (int): The number of layers in the transformer.
+        embedding_dim (int): The channel dimension for the input embeddings.
+        num_heads (int): The number of heads for multihead attention.
+        mlp_dim (int): The internal channel dimension for the MLP block.
+        layers (nn.ModuleList): The list of TwoWayAttentionBlock layers that make up the transformer.
+        final_attn_token_to_image (Attention): The final attention layer applied from the queries to the image.
+        norm_final_attn (nn.LayerNorm): The layer normalization applied to the final queries.
+    """
+
+    def __init__(
+        self,
+        depth: int,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+    ) -> None:
+        """
+        A transformer decoder that attends to an input image using queries whose positional embedding is supplied.
+
+        Args:
+          depth (int): number of layers in the transformer
+          embedding_dim (int): the channel dimension for the input embeddings
+          num_heads (int): the number of heads for multihead attention. Must
+            divide embedding_dim
+          mlp_dim (int): the channel dimension internal to the MLP block
+          activation (nn.Module): the activation to use in the MLP block
+        """
+        super().__init__()
+        self.depth = depth
+        self.embedding_dim = embedding_dim
+        self.num_heads = num_heads
+        self.mlp_dim = mlp_dim
+        self.layers = nn.ModuleList()
+
+        for i in range(depth):
+            self.layers.append(
+                TwoWayAttentionBlock(
+                    embedding_dim=embedding_dim,
+                    num_heads=num_heads,
+                    mlp_dim=mlp_dim,
+                    activation=activation,
+                    attention_downsample_rate=attention_downsample_rate,
+                    skip_first_layer_pe=(i == 0),
+                )
+            )
+
+        self.final_attn_token_to_image = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)
+        self.norm_final_attn = nn.LayerNorm(embedding_dim)
+
+    def forward(
+        self,
+        image_embedding: Tensor,
+        image_pe: Tensor,
+        point_embedding: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+          image_embedding (torch.Tensor): image to attend to. Should be shape B x embedding_dim x h x w for any h and w.
+          image_pe (torch.Tensor): the positional encoding to add to the image. Must have same shape as image_embedding.
+          point_embedding (torch.Tensor): the embedding to add to the query points.
+            Must have shape B x N_points x embedding_dim for any N_points.
+
+        Returns:
+          (torch.Tensor): the processed point_embedding
+          (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)
+
+        # Prepare queries
+        queries = point_embedding
+        keys = image_embedding
+
+        # Apply transformer blocks and final layernorm
+        for layer in self.layers:
+            queries, keys = layer(
+                queries=queries,
+                keys=keys,
+                query_pe=point_embedding,
+                key_pe=image_pe,
+            )
+
+        # Apply the final attention layer from the points to the image
+        q = queries + point_embedding
+        k = keys + image_pe
+        attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm_final_attn(queries)
+
+        return queries, keys
+
+
+class TwoWayAttentionBlock(nn.Module):
+    """
+    An attention block that performs both self-attention and cross-attention in two directions: queries to keys and
+    keys to queries. This block consists of four main layers: (1) self-attention on sparse inputs, (2) cross-attention
+    of sparse inputs to dense inputs, (3) an MLP block on sparse inputs, and (4) cross-attention of dense inputs to
+    sparse inputs.
+
+    Attributes:
+        self_attn (Attention): The self-attention layer for the queries.
+        norm1 (nn.LayerNorm): Layer normalization following the first attention block.
+        cross_attn_token_to_image (Attention): Cross-attention layer from queries to keys.
+        norm2 (nn.LayerNorm): Layer normalization following the second attention block.
+        mlp (MLPBlock): MLP block that transforms the query embeddings.
+        norm3 (nn.LayerNorm): Layer normalization following the MLP block.
+        norm4 (nn.LayerNorm): Layer normalization following the third attention block.
+        cross_attn_image_to_token (Attention): Cross-attention layer from keys to queries.
+        skip_first_layer_pe (bool): Whether to skip the positional encoding in the first layer.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int = 2048,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+        skip_first_layer_pe: bool = False,
+    ) -> None:
+        """
+        A transformer block with four layers: (1) self-attention of sparse inputs, (2) cross attention of sparse
+        inputs to dense inputs, (3) mlp block on sparse inputs, and (4) cross attention of dense inputs to sparse
+        inputs.
+
+        Args:
+          embedding_dim (int): the channel dimension of the embeddings
+          num_heads (int): the number of heads in the attention layers
+          mlp_dim (int): the hidden dimension of the mlp block
+          activation (nn.Module): the activation of the mlp block
+          skip_first_layer_pe (bool): skip the PE on the first layer
+        """
+        super().__init__()
+        self.self_attn = Attention(embedding_dim, num_heads)
+        self.norm1 = nn.LayerNorm(embedding_dim)
+
+        self.cross_attn_token_to_image = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)
+        self.norm2 = nn.LayerNorm(embedding_dim)
+
+        self.mlp = MLPBlock(embedding_dim, mlp_dim, activation)
+        self.norm3 = nn.LayerNorm(embedding_dim)
+
+        self.norm4 = nn.LayerNorm(embedding_dim)
+        self.cross_attn_image_to_token = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)
+
+        self.skip_first_layer_pe = skip_first_layer_pe
+
+    def forward(self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor) -> Tuple[Tensor, Tensor]:
+        """Apply self-attention and cross-attention to queries and keys and return the processed embeddings."""
+
+        # Self attention block
+        if self.skip_first_layer_pe:
+            queries = self.self_attn(q=queries, k=queries, v=queries)
+        else:
+            q = queries + query_pe
+            attn_out = self.self_attn(q=q, k=q, v=queries)
+            queries = queries + attn_out
+        queries = self.norm1(queries)
+
+        # Cross attention block, tokens attending to image embedding
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm2(queries)
+
+        # MLP block
+        mlp_out = self.mlp(queries)
+        queries = queries + mlp_out
+        queries = self.norm3(queries)
+
+        # Cross attention block, image embedding attending to tokens
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
+        keys = keys + attn_out
+        keys = self.norm4(keys)
+
+        return queries, keys
+
+
+class Attention(nn.Module):
+    """An attention layer that allows for downscaling the size of the embedding after projection to queries, keys, and
+    values.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        downsample_rate: int = 1,
+    ) -> None:
+        """
+        Initializes the Attention model with the given dimensions and settings.
+
+        Args:
+            embedding_dim (int): The dimensionality of the input embeddings.
+            num_heads (int): The number of attention heads.
+            downsample_rate (int, optional): The factor by which the internal dimensions are downsampled. Defaults to 1.
+
+        Raises:
+            AssertionError: If 'num_heads' does not evenly divide the internal dimension (embedding_dim / downsample_rate).
+        """
+        super().__init__()
+        self.embedding_dim = 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.out_proj = nn.Linear(self.internal_dim, embedding_dim)
+
+    @staticmethod
+    def _separate_heads(x: Tensor, num_heads: int) -> Tensor:
+        """Separate the input tensor into the specified number of attention heads."""
+        b, n, c = x.shape
+        x = x.reshape(b, n, num_heads, c // num_heads)
+        return x.transpose(1, 2)  # B x N_heads x N_tokens x C_per_head
+
+    @staticmethod
+    def _recombine_heads(x: Tensor) -> Tensor:
+        """Recombine the separated attention heads into a single tensor."""
+        b, n_heads, n_tokens, c_per_head = x.shape
+        x = x.transpose(1, 2)
+        return x.reshape(b, n_tokens, n_heads * c_per_head)  # B x N_tokens x C
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
+        """Compute the attention output given the input query, key, and value tensors."""
+
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        # Attention
+        _, _, _, c_per_head = q.shape
+        attn = q @ k.permute(0, 1, 3, 2)  # B x N_heads x N_tokens x N_tokens
+        attn = attn / math.sqrt(c_per_head)
+        attn = torch.softmax(attn, dim=-1)
+
+        # Get output
+        out = attn @ v
+        out = self._recombine_heads(out)
+        return self.out_proj(out)