ultralytics 8.0.136 refactor and simplify package (#3748)

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

ultralytics/models/sam/modules/__init__.py

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

ultralytics/models/sam/modules/decoders.py

@@ -0,0 +1,159 @@
+# 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):
+
+    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.
+
+        Arguments:
+            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.
+
+        Arguments:
+            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.size(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):
+    """
+    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:
+        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 +1,583 @@
+# 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
+
+
+# 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 # noqa
+class ImageEncoderViT(nn.Module):
+
+    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:
+        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)
+
+        x = self.neck(x.permute(0, 3, 1, 2))
+
+        return x
+
+
+class PromptEncoder(nn.Module):
+
+    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.
+
+        Arguments:
+          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:
+        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.
+
+        Arguments:
+          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:
+        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)),
+        )
+
+    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:
+        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
+        x = x + self.mlp(self.norm2(x))
+
+        return 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:
+        """
+        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:
+        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)
+        x = self.proj(x)
+
+        return 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 :paper:`mvitv2`.
+    https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py   # noqa B950
+    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:
+        """
+        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:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x

ultralytics/models/sam/modules/sam.py

@@ -0,0 +1,173 @@
+# 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 Any, Dict, List, Tuple
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from .decoders import MaskDecoder
+from .encoders import ImageEncoderViT, PromptEncoder
+
+
+class Sam(nn.Module):
+    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] = None,
+                 pixel_std: List[float] = None) -> None:
+        """
+        SAM predicts object masks from an image and input prompts.
+
+        Arguments:
+          image_encoder (ImageEncoderViT): The backbone used to encode the
+            image into image embeddings that allow for efficient mask prediction.
+          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)): Mean values for normalizing pixels in the input image.
+          pixel_std (list(float)): Std values for normalizing pixels in the input image.
+        """
+        if pixel_mean is None:
+            pixel_mean = [123.675, 116.28, 103.53]
+        if pixel_std is None:
+            pixel_std = [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)
+
+    @property
+    def device(self) -> Any:
+        return self.pixel_mean.device
+
+    @torch.no_grad()
+    def forward(
+        self,
+        batched_input: List[Dict[str, Any]],
+        multimask_output: bool,
+    ) -> List[Dict[str, torch.Tensor]]:
+        """
+        Predicts masks end-to-end from provided images and prompts.
+        If prompts are not known in advance, using SamPredictor is
+        recommended over calling the model directly.
+
+        Arguments:
+          batched_input (list(dict)): A list over input images, each a
+            dictionary with the following keys. A prompt key can be
+            excluded if it is not present.
+              'image': The image as a torch tensor in 3xHxW format,
+                already transformed for input to the model.
+              'original_size': (tuple(int, int)) The original size of
+                the image before transformation, as (H, W).
+              'point_coords': (torch.Tensor) Batched point prompts for
+                this image, with shape BxNx2. Already transformed to the
+                input frame of the model.
+              'point_labels': (torch.Tensor) Batched labels for point prompts,
+                with shape BxN.
+              'boxes': (torch.Tensor) Batched box inputs, with shape Bx4.
+                Already transformed to the input frame of the model.
+              'mask_inputs': (torch.Tensor) Batched mask inputs to the model,
+                in the form Bx1xHxW.
+          multimask_output (bool): Whether the model should predict multiple
+            disambiguating masks, or return a single mask.
+
+        Returns:
+          (list(dict)): A list over input images, where each element is
+            as dictionary with the following keys.
+              'masks': (torch.Tensor) Batched binary mask predictions,
+                with shape BxCxHxW, where B is the number of input prompts,
+                C is determined by multimask_output, and (H, W) is the
+                original size of the image.
+              'iou_predictions': (torch.Tensor) The model's predictions
+                of mask quality, in shape BxC.
+              'low_res_logits': (torch.Tensor) Low resolution logits with
+                shape BxCxHxW, where H=W=256. Can be passed as mask input
+                to subsequent iterations of prediction.
+        """
+        input_images = torch.stack([self.preprocess(x['image']) for x in batched_input], dim=0)
+        image_embeddings = self.image_encoder(input_images)
+
+        outputs = []
+        for image_record, curr_embedding in zip(batched_input, image_embeddings):
+            if 'point_coords' in image_record:
+                points = (image_record['point_coords'], image_record['point_labels'])
+            else:
+                points = None
+            sparse_embeddings, dense_embeddings = self.prompt_encoder(
+                points=points,
+                boxes=image_record.get('boxes', None),
+                masks=image_record.get('mask_inputs', None),
+            )
+            low_res_masks, iou_predictions = self.mask_decoder(
+                image_embeddings=curr_embedding.unsqueeze(0),
+                image_pe=self.prompt_encoder.get_dense_pe(),
+                sparse_prompt_embeddings=sparse_embeddings,
+                dense_prompt_embeddings=dense_embeddings,
+                multimask_output=multimask_output,
+            )
+            masks = self.postprocess_masks(
+                low_res_masks,
+                input_size=image_record['image'].shape[-2:],
+                original_size=image_record['original_size'],
+            )
+            masks = masks > self.mask_threshold
+            outputs.append({
+                'masks': masks,
+                'iou_predictions': iou_predictions,
+                'low_res_logits': low_res_masks, })
+        return outputs
+
+    def postprocess_masks(
+        self,
+        masks: torch.Tensor,
+        input_size: Tuple[int, ...],
+        original_size: Tuple[int, ...],
+    ) -> torch.Tensor:
+        """
+        Remove padding and upscale masks to the original image size.
+
+        Arguments:
+          masks (torch.Tensor): Batched masks from the mask_decoder,
+            in BxCxHxW format.
+          input_size (tuple(int, int)): The size of the image input to the
+            model, in (H, W) format. Used to remove padding.
+          original_size (tuple(int, int)): The original size of the image
+            before resizing for input to the model, in (H, W) format.
+
+        Returns:
+          (torch.Tensor): Batched masks in BxCxHxW format, where (H, W)
+            is given by original_size.
+        """
+        masks = F.interpolate(
+            masks,
+            (self.image_encoder.img_size, self.image_encoder.img_size),
+            mode='bilinear',
+            align_corners=False,
+        )
+        masks = masks[..., :input_size[0], :input_size[1]]
+        masks = F.interpolate(masks, original_size, mode='bilinear', align_corners=False)
+        return masks
+
+    def preprocess(self, x: torch.Tensor) -> torch.Tensor:
+        """Normalize pixel values and pad to a square input."""
+        # Normalize colors
+        x = (x - self.pixel_mean) / self.pixel_std
+
+        # Pad
+        h, w = x.shape[-2:]
+        padh = self.image_encoder.img_size - h
+        padw = self.image_encoder.img_size - w
+        return F.pad(x, (0, padw, 0, padh))

ultralytics/models/sam/modules/tiny_encoder.py

@@ -0,0 +1,653 @@
+# 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):
+
+    def __init__(self, a, b, ks=1, stride=1, pad=0, dilation=1, groups=1, bn_weight_init=1):
+        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)
+
+    @torch.no_grad()
+    def fuse(self):
+        c, bn = self._modules.values()
+        w = bn.weight / (bn.running_var + bn.eps) ** 0.5
+        w = c.weight * w[:, None, None, None]
+        b = bn.bias - bn.running_mean * bn.weight / \
+            (bn.running_var + bn.eps)**0.5
+        m = torch.nn.Conv2d(w.size(1) * self.c.groups,
+                            w.size(0),
+                            w.shape[2:],
+                            stride=self.c.stride,
+                            padding=self.c.padding,
+                            dilation=self.c.dilation,
+                            groups=self.c.groups)
+        m.weight.data.copy_(w)
+        m.bias.data.copy_(b)
+        return m
+
+
+# NOTE: This module and timm package is needed only for training.
+# from ultralytics.utils.checks import check_requirements
+# check_requirements('timm')
+# from timm.models.layers import DropPath as TimmDropPath
+# from timm.models.layers import trunc_normal_
+# class DropPath(TimmDropPath):
+#
+#     def __init__(self, drop_prob=None):
+#         super().__init__(drop_prob=drop_prob)
+#         self.drop_prob = drop_prob
+#
+#     def __repr__(self):
+#         msg = super().__repr__()
+#         msg += f'(drop_prob={self.drop_prob})'
+#         return msg
+
+
+class PatchEmbed(nn.Module):
+
+    def __init__(self, in_chans, embed_dim, resolution, activation):
+        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):
+        return self.seq(x)
+
+
+class MBConv(nn.Module):
+
+    def __init__(self, in_chans, out_chans, expand_ratio, activation, drop_path):
+        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):
+        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
+        x = self.act3(x)
+
+        return x
+
+
+class PatchMerging(nn.Module):
+
+    def __init__(self, input_resolution, dim, out_dim, activation):
+        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 = 2
+        if (out_dim == 320 or out_dim == 448 or out_dim == 576):
+            stride_c = 1
+        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):
+        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)
+        x = x.flatten(2).transpose(1, 2)
+        return x
+
+
+class ConvLayer(nn.Module):
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        depth,
+        activation,
+        drop_path=0.,
+        downsample=None,
+        use_checkpoint=False,
+        out_dim=None,
+        conv_expand_ratio=4.,
+    ):
+
+        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
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, out_dim=out_dim, activation=activation)
+        else:
+            self.downsample = None
+
+    def forward(self, x):
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x
+
+
+class Mlp(nn.Module):
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        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):
+        x = self.norm(x)
+
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(torch.nn.Module):
+
+    def __init__(
+            self,
+            dim,
+            key_dim,
+            num_heads=8,
+            attn_ratio=4,
+            resolution=(14, 14),
+    ):
+        super().__init__()
+        # (h, w)
+        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):
+        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 (B,N,C)
+        B, N, _ = x.shape
+
+        # 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)
+        x = self.proj(x)
+        return x
+
+
+class TinyViTBlock(nn.Module):
+    r""" TinyViT Block.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int, int]): Input resolution.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        drop (float, optional): Dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        local_conv_size (int): the kernel size of the convolution between
+                               Attention and MLP. Default: 3
+        activation (torch.nn): the activation function. Default: nn.GELU
+    """
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        num_heads,
+        window_size=7,
+        mlp_ratio=4.,
+        drop=0.,
+        drop_path=0.,
+        local_conv_size=3,
+        activation=nn.GELU,
+    ):
+        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):
+        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)
+
+        x = x + self.drop_path(self.mlp(x))
+        return x
+
+    def extra_repr(self) -> str:
+        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.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        drop (float, optional): Dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        local_conv_size (int): the kernel size of the depthwise convolution between attention and MLP. Default: 3
+        activation (torch.nn): the activation function. Default: nn.GELU
+        out_dim (int | optional): the output dimension of the layer. Default: None
+    """
+
+    def __init__(
+        self,
+        dim,
+        input_resolution,
+        depth,
+        num_heads,
+        window_size,
+        mlp_ratio=4.,
+        drop=0.,
+        drop_path=0.,
+        downsample=None,
+        use_checkpoint=False,
+        local_conv_size=3,
+        activation=nn.GELU,
+        out_dim=None,
+    ):
+
+        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
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, out_dim=out_dim, activation=activation)
+        else:
+            self.downsample = None
+
+    def forward(self, x):
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}'
+
+
+class LayerNorm2d(nn.Module):
+
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        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:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
+
+
+class TinyViT(nn.Module):
+
+    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.,
+        drop_rate=0.,
+        drop_path_rate=0.1,
+        use_checkpoint=False,
+        mbconv_expand_ratio=4.0,
+        local_conv_size=3,
+        layer_lr_decay=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):
+        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):
+            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):
+            for p in m.parameters():
+                assert hasattr(p, 'lr_scale'), p.param_name
+
+        self.apply(_check_lr_scale)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            # NOTE: This initialization is needed only for training.
+            # trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and 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):
+        return {'attention_biases'}
+
+    def forward_features(self, x):
+        # x: (N, C, H, W)
+        x = self.patch_embed(x)
+
+        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.size()
+        x = x.view(B, 64, 64, C)
+        x = x.permute(0, 3, 1, 2)
+        x = self.neck(x)
+        return x
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        return x

ultralytics/models/sam/modules/transformer.py

@@ -0,0 +1,235 @@
+# 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):
+
+    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 the 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):
+
+    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.
+
+        Arguments:
+          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:
+        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)
+
+    def _separate_heads(self, 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
+
+    def _recombine_heads(self, 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)
+        out = self.out_proj(out)
+
+        return out