SAM代码解读

Sam项目代码的初步解读，对其中的SamPredictor、ImageEncoderViT、PromptEncoder和MaskDecoder进行解读，与https://hpg123.blog.csdn.net/article/details/131194434的使用手册内容相呼应。

1、整体介绍

Sam由ImageEncoderViT,PromptEncoder,MaskDecoder三个部件组成，ImageEncoderViT负责将image输入编码为图像嵌入（描述图像的特征向量）；PromptEncoder负责将用户输入的位置提示信息（point、boxes、mask）编码为空间嵌入（描述位置的特征向量）；MaskDecoder用于对接ImageEncoderViT与PromptEncoder的输出，依据输入的图像特征向量和位置特征向量，输出对应位置的mask信息。其在交互式运行中，对于同一个图像，每一次补充位置信息都会有不同的mask输出。

Sam官方给出的是web部署策略，ImageEncoderViT在服务器端提取图像的特征向量，而PromptEncoder和MaskDecoder则在浏览器端编码用户输入和解码mask。anylabeling所嵌入的sam则是单体软件部署策略，但其编码特征预测mask的思路是一致的。注：突然发现这与百度的EISeg所提供的能力相类似，都是根据用户点击位置，生成mask

1.1 初步介绍

通过以下代码可以发现使用sam依赖sam_model_registry和SamPredictor。
代码参考自：https://hpg123.blog.csdn.net/article/details/131194434?spm=1001.2014.3001.5502

import sys
sys.path.append("..")
from segment_anything import sam_model_registry, SamPredictor

sam_checkpoint = "sam_vit_b_01ec64.pth"
model_type = "vit_b"

device = "cuda"

sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)
sam.to(device=device)

predictor = SamPredictor(sam)
#predictor.set_image(image)
masks, scores, logits = predictor.predict(
    point_coords=input_point,
    point_labels=input_label,
    multimask_output=True,
)

print(masks.shape)  # (number_of_masks) x H x W  | output (3, 600, 900)

for i, (mask, score) in enumerate(zip(masks, scores)):
    plt.figure(figsize=(10,10))
    plt.imshow(image)
    show_mask(mask, plt.gca())
    show_points(input_point, input_label, plt.gca())
    plt.title(f"Mask {i+1}, Score: {score:.3f}", fontsize=18)
    plt.axis('off')
    plt.show()

sam_model_registry的定义如下，其原本是一个dict对象，每一个value都是一个函数，函数又有一个checkpoint参数，故在使用时是sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)

def build_sam_vit_b(checkpoint=None):
    return _build_sam(
        encoder_embed_dim=768,
        encoder_depth=12,
        encoder_num_heads=12,
        encoder_global_attn_indexes=[2, 5, 8, 11],
        checkpoint=checkpoint,
    )
sam_model_registry = {
    "default": build_sam_vit_h,
    "vit_h": build_sam_vit_h,
    "vit_l": build_sam_vit_l,
    "vit_b": build_sam_vit_b,
}


def _build_sam(
    encoder_embed_dim,
    encoder_depth,
    encoder_num_heads,
    encoder_global_attn_indexes,
    checkpoint=None,
):
    prompt_embed_dim = 256
    image_size = 1024
    vit_patch_size = 16
    image_embedding_size = image_size // vit_patch_size
    sam = Sam(
        image_encoder=ImageEncoderViT(
            depth=encoder_depth,
            embed_dim=encoder_embed_dim,
            img_size=image_size,
            mlp_ratio=4,
            norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
            num_heads=encoder_num_heads,
            patch_size=vit_patch_size,
            qkv_bias=True,
            use_rel_pos=True,
            global_attn_indexes=encoder_global_attn_indexes,
            window_size=14,
            out_chans=prompt_embed_dim,
        ),
        prompt_encoder=PromptEncoder(
            embed_dim=prompt_embed_dim,
            image_embedding_size=(image_embedding_size, image_embedding_size),
            input_image_size=(image_size, image_size),
            mask_in_chans=16,
        ),
        mask_decoder=MaskDecoder(
            num_multimask_outputs=3,
            transformer=TwoWayTransformer(
                depth=2,
                embedding_dim=prompt_embed_dim,
                mlp_dim=2048,
                num_heads=8,
            ),
            transformer_dim=prompt_embed_dim,
            iou_head_depth=3,
            iou_head_hidden_dim=256,
        ),
        pixel_mean=[123.675, 116.28, 103.53],
        pixel_std=[58.395, 57.12, 57.375],
    )
    sam.eval()
    if checkpoint is not None:
        with open(checkpoint, "rb") as f:
            state_dict = torch.load(f)
        sam.load_state_dict(state_dict)
    return sam

1.2 Sam和SamPredictor介绍

Sam为实现系统级功能的集成类（torch.nn.Module）,其包含ImageEncoderViT、PromptEncoder、MaskDecoder三个组件。其并不提供预测单个输入的能力，其所要求的输入都是字典，格式是完备的（主要指位置提示【point_coords、point_labels、boxes和mask_inputs】）。有SamPredictor类对稀疏的用户输出进行封装集成。

Sam的forword

Sam的forword要求输入的是list对象，其对图像编码时是进行batch操作（一次性推理出所有图像的特征），而对提示输入和mask生成则是单独操作（通过for循环预测每一个位置提示所对应的mask）；其输出也是list对象，每个元素包含masks、iou_predictions和low_res_logits值。
其中，masks是low_res_logits的高分辨率结果，并按照mask_threshold进行二值化（也就是说模型只时预测出低分辨率的low_res_logits）。

    def forward(
        self,
        batched_input: List[Dict[str, Any]],
        multimask_output: bool,
    ) :
        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(),# 用于给整图的网格生成位置编码
                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

SamPredictor介绍

SamPredictor是对sam类的高级封装，使其能适应不同的输入数据（单点、单boxes、单mask、多点、多boxes及其间的相互组合），通过调用set_image或set_torch_image函数完成对图像的特征编码，并调用predict函数完成对缺失输入的补充（同时实现位置信息的resize，因为位置信息是针对原图size的，而sam模型要求的图像输入是1024x1024，故需要进行等比例换算）。

2、ImageEncoderViT

ImageEncoderViT是sam中的核心module，其经过了1100万个高清图像图像（图像短边为1500，比绝大部分视觉数据集要高清）的训练（其针对每个图像平均有1个手工mask【即便如此也比任何数据库多4倍的标注量】，通过自动标注后平均每个图像有100个mask）。

ImageEncoderViT是经典的Transformers（ViT）模型，其简介可以查看https://blog.csdn.net/a486259/article/details/129802747，其性能饱和慢，但随着数据增长，性能可持续增长（故用到了1100万的训练图片）。原始ViT也是在 ImageNet、ImageNet-21k和JFT- 300M进行训练，并表明JFT-300M效果更好。sam中的Vit与原始模型有细微差异，其输入shape为3x1024x1024，输出的feature map为256x64x64。
通过以下代码可以从sam中抽离ImageEncoderViT对象

import sys
sys.path.append("..")
from segment_anything import sam_model_registry, SamAutomaticMaskGenerator, SamPredictor

sam_checkpoint = "sam_vit_b_01ec64.pth"
model_type = "vit_b"

device = "cuda"

sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)
sam.to(device=device)

torch.save(sam.image_encoder,'sma_vit_b.pt')
torch.save(sam.image_encoder.state_dict(),'sma_vit_b_state_dict.pt')

2.1 主体代码

ImageEncoderViT的主要实现代码如下，其先由patch_embed对输入数据进行16倍的下采样（将patch_size设为16）;并将embed_dim[token长度]设为768（原始的vit中embed_dim是为196=img_wimg_h/(patch_wpatch_h)，img_w:224 patch_w:16）,这与原始ViT相比，是存在一定信息缺失的（64x64=4096）。

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, ...] = (),
    ) :
        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) :
        x = self.patch_embed(x)   # 实现patch_size为16x16的下采样
        if self.pos_embed is not None:
            x = x + self.pos_embed  # 进行位置编码

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

        x = self.neck(x.permute(0, 3, 1, 2))  # 维度梳理，将channel dim从embed_dim转换为out_chans

        return x

此外，ImageEncoderViT还多了一个neck层，用于将embed_dim从768转换到所需的out_chans（256）

2.2 Transformer blocks

这是与ViT存在差异的另一点，ImageEncoderViT补充了window_partition操作，对x的w和h进行填充，使其size变为window_size（window_size < patch size,然后会对wh进行补充，然后resize）

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,
    ):
        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):
        shortcut = x
        x = self.norm1(x)  # nn.LayerNorm
        # Window partition  # 对x的w和h进行填充，使其size变为window_size（window_size<patch size,然后会对wh进行补充，然后resize）
        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)   # multi head attention
        # Reverse window partition  #移除w和h方向上的填充
        if self.window_size > 0:
            x = window_unpartition(x, self.window_size, pad_hw, (H, W))

        x = shortcut + x   #shortcut
        x = x + self.mlp(self.norm2(x))  #mlp

        return x
def window_partition(x: torch.Tensor, window_size: 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:
        #https://blog.csdn.net/qq_41731861/article/details/120777887
        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h)) #三维度填充,扩展w和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]
) :
    """
    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

2.3 Attention

ImageEncoderViT中的Multi-head Attention还参考了mvitv2，在atten操作中（Q@K后）为x添加了可训练的位置信息，使模型能够对特定位置着重关注。

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,
    ) :
        """
        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 (bool): If True, add relative positional embeddings to the attention map.
            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.
        """
        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):
        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

add_decomposed_rel_pos的实现如下，其

def get_rel_pos(q_size: int, k_size: int, rel_pos: 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],
):
    """
    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

3、PromptEncoder

PromptEncoder用于对输入模型的points、boxes和masks信息进行编码，将其统一为空间特征编码的格式。
其主体代码如下所示,可见其编码器并不复杂，属于轻量化的结构。其对points、boxes和masks编码时允许有部分值空缺（空缺使用默认值），其将points和boxes组装为sparse_embeddings，将mask组装为dense_embeddings其对mask的采样由多个attention层实现，具体可见mask_downscaling。

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,
    ):
        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 i 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 forward(
        self,
        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
        boxes: Optional[torch.Tensor],
        masks: Optional[torch.Tensor],
    ) :
        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.mask_downscaling(masks)# _embed_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

3.1 _embed_points

PromptEncoder对于points附加的的labels信息操作为：-1 表示ignore，0 表示 负点（对应pe_layer输出的【0】），1 表示正点（对应pe_layer输出的【1】）
points.shape:b,n,2 n表示输入的点数（输入2个点则表示与boxes相同）

    def _embed_points(
        self,
        points: torch.Tensor,
        labels: torch.Tensor,
        pad: bool,
    ):
        """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

3.2 _embed_boxes

PromptEncoder对于points和boxes都使用PositionEmbeddingRandom（pe_layer）进行编码，可以看到其将boxes转换为2个点然后输入了模型。point输入，则对应着pe_layer输出的【2和3】

    def _embed_boxes(self, boxes: 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

3.3 PositionEmbeddingRandom

PositionEmbeddingRandom 的实现如下，其核心是_pe_encoding函数，对坐标进行标准化，然后乘上一个可训练参数，再去sin和cos做编码

class PositionEmbeddingRandom(nn.Module):
    """
    Positional encoding using random spatial frequencies.
    """

    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = 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):
        """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 #标准化到-1,1
        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]):
        """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]
    ) :
        """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

4、MaskDecoder

用于根据PromptEncoder和ImageEncoderViT的输入生成mask，其是一种transformer架构的解码头，有两个输出头：output_upscaling和iou_prediction_head，两个头输出的数量是一样的。
其forward函数仅是对输出结果进行了选择操作，核心推理是由predict_masks函数和transformer对象（TwoWayTransformer实例）完成的

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,
    ):
        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 i 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,
    ) :
        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
        if multimask_output:
            mask_slice = slice(1, None)
        else:
            mask_slice = 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,
    ):
        pass
        return masks, iou_pred

4.1 TwoWayTransformer

该Transformer对象承担了MaskDecoder在预测iou和mask时的绝大部分工作，其预测时需输入image_embedding、image_pe、point_embedding(，预测完输出queries（在后续中被作为iou_token和mask_token）, keys（在后续中被作为mask）

其基本代码如下，其输入数据为image_embedding[即：image_embeddings+dense_prompt_embeddings]、image_pe[即图像的位置编码]和point_embedding[即：cat(iou_token、mask_tokens、sparse_prompt_embeddings)]

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,
    ) :
        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,
    ) :
        # 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

4.2 predict_masks

MaskDecoder中核心函数，用于实现mask和iou的预测。

    def predict_masks(
        self,
        image_embeddings: torch.Tensor,
        image_pe: torch.Tensor,
        sparse_prompt_embeddings: torch.Tensor,
        dense_prompt_embeddings: 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] = []
        for i in range(self.num_mask_tokens):
            hyper_in_list.append(self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]))
        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