vllm.model_executor.models.deepencoder ¶

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: tuple[int, int] | None = 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) or None): Input resolution for calculating the relative
                positional parameter size.
        """  # noqa: E501
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = RelPosAttention(
            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 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, ...] = (),
        last_conv_output: int = 1024,
    ) -> 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.
        """  # noqa: E501
        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: nn.Parameter | None = 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(
            Conv2dLayer(
                embed_dim,
                out_chans,
                kernel_size=1,
                bias=False,
            ),
            LayerNorm2d(out_chans),
            Conv2dLayer(
                out_chans,
                out_chans,
                kernel_size=3,
                padding=1,
                bias=False,
            ),
            LayerNorm2d(out_chans),
        )

        self.net_2 = Conv2dLayer(
            256, 512, kernel_size=3, stride=2, padding=1, bias=False
        )
        self.net_3 = Conv2dLayer(
            512, last_conv_output, kernel_size=3, stride=2, padding=1, bias=False
        )

    def get_abs_pos(self, abs_pos: torch.Tensor, tgt_size: int):
        dtype = abs_pos.dtype

        src_size = abs_pos.size(1)

        if src_size != tgt_size:
            old_pos_embed = abs_pos.permute(0, 3, 1, 2)
            old_pos_embed = old_pos_embed.to(torch.float32)
            new_pos_embed = F.interpolate(
                old_pos_embed,
                size=(tgt_size, tgt_size),
                mode="bicubic",
                antialias=True,
                align_corners=False,
            ).to(dtype)
            new_pos_embed = new_pos_embed.permute(0, 2, 3, 1)
            return new_pos_embed
        else:
            return abs_pos

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.patch_embed(x)
        if self.pos_embed is not None:
            x = x + self.get_abs_pos(self.pos_embed, x.size(1))

        for blk in self.blocks:
            x = blk(x)

        neck_output = self.neck(x.permute(0, 3, 1, 2))
        conv2_output = self.net_2(neck_output)
        conv3_output = self.net_3(conv2_output)

        return conv3_output

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 = Conv2dLayer(
            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

class RelPosAttention(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: tuple[int, int] | None = 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) or None): Input resolution for calculating the relative
                positional parameter size.
        """  # noqa: E501
        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)

        rel_h, rel_w = None, None
        if self.use_rel_pos:
            rel_h, rel_w = add_decomposed_rel_pos(
                q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W)
            )

        q = q.view(B, self.num_heads, H * W, -1)
        k = k.view(B, self.num_heads, H * W, -1)
        v = v.view(B, self.num_heads, H * W, -1)

        if self.use_rel_pos:
            rel_h = rel_h.view(
                B, self.num_heads, rel_h.size(1), rel_h.size(2), rel_h.size(3)
            )
            rel_w = rel_w.view(
                B, self.num_heads, rel_w.size(1), rel_w.size(2), rel_w.size(3)
            )
            attn_bias = (rel_h + rel_w).view(
                B, self.num_heads, rel_h.size(2), rel_h.size(3) * rel_w.size(4)
            )
            x = torch.nn.functional.scaled_dot_product_attention(
                q, k, v, attn_mask=attn_bias
            )
        else:
            x = torch.nn.functional.scaled_dot_product_attention(q, k, v)

        x = (
            x.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