SAM之MaskDecoder总结(个人研究)

前言

        SAM模型主要由三部分组成，本文旨在结合代码的方式对其中mask decoder进行详细的描述与总结，不足之处还望多多指教

        mask decoder模块可以高效的将image embeddings与prompt embeddings映射到一个output mask中。为了结合image embeddings与prompt embeddings 这两个输入， 受Transformer模型的启发，作者修改了transformer中标准的Transformer Decoder作为本文所要介绍的Mask Decoder。

Mask Decoder定义

可以结合论文中该图看下列代码：

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:
        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)

        # ---upscaled---
        # ---四倍上采样---
        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(),
        )
        # ---upscaled end---
        
        # ---MLP---
        # ---对应mask数量的mlp---
        self.output_hypernetworks_mlps = nn.ModuleList(
            [
                MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)
                for i in range(self.num_mask_tokens)
            ]
        )
        # ---对应mask数量的mlp end---
        
        # ---对应iou的mlp--
        self.iou_prediction_head = MLP(
            transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth
        )
        # ---对应iou的mlp end--
        # ---MLP end---

先看传入的参数，如下：

transformer dim：表示transformer channel的维度。

transformer：表示传入的transformer模型。

num_multimask_outputs：表示mask decoder输出的mask个数(用于消除歧义), SAM原文中默认值为3。

activation：表示mask decoder 上采样过程中使用的激活函数。

iou_head_depth：表示用于预测mask的IoU质量指标时，所使用MLP层的深度。

iou_head_hidden_dim: 表示用于预测mask的IoU质量指标时，所使用MLP层中隐藏层的维度。

在下面代码中，由于在该阶段引入了一个额外的iou_token用于计算所预测mask的质量，因此要在该处+1。

self.num_mask_tokens = num_multimask_outputs + 1

Mask Decoder前向传播

def forward(
        self,
        image_embeddings: torch.Tensor,  # image encoder 输出的image embedding
        image_pe: torch.Tensor,  # image的position embedding
        sparse_prompt_embeddings: torch.Tensor,  # prompt encoder输出的sparse prompt
        dense_prompt_embeddings: torch.Tensor,  # prompt encoder 输出的dense prompt
        multimask_output: bool,  # 多类别输出,具有模糊识别的能力
    ) -> Tuple[torch.Tensor, torch.Tensor]:
       
        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,
        )

        # 根据multimask_output的bool值来对masks以及iou_pred进行选择性的切片
        if multimask_output:
            mask_slice = slice(1, None)  # 为真时， 切片选择从第一个元素到最后一个
        else:
            mask_slice = slice(0, 1)  # 为假时，只选择第一个切片
        masks = masks[:, mask_slice, :, :]
        iou_pred = iou_pred[:, mask_slice]

        # 返回mask 以及 iou的预测分数
        return masks, iou_pred

Mask Decoder中predict_masks

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)  # 对应 output_tokens + prompt tokens

        # Expand per-image data in batch direction to be per-mask
        # 扩展image_embeddings的B维度,因为boxes标记分割时,n个box时batchsize=batchsize*n
        if image_embeddings.shape[0] != tokens.shape[0]:
            # torch.repeat_interleave() 沿着指定的维度重复张量的元素
            # image_embeddings 相当于待重复的张量元素
            # tokens.shape[0] 相当于重复次数
            src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
        else:
            src = image_embeddings
        src = src + dense_prompt_embeddings  # 对应 image embedding + dense_prompt_embeddings(mask)
        pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
        b, c, h, w = src.shape

        # Run the transformer
        # hs代表transformer的输出隐藏状态 , 而src代表transformer的输入
        hs, src = self.transformer(src, pos_src, tokens)
        # [hs]是Transformer的输出，其中第一个维度表示batch中的不同样本，第二个维度表示token的序列，第三个维度表示token的特征维度。
        # 通过`hs[:, 0, :]`可以获取第一个token对应的输出，即`iou_token_out`；
        # 通过`hs[:, 1 : (1 + self.num_mask_tokens), :]`可以获取接下来的`num_mask_tokens`个token对应的输出，即`mask_tokens_out`。
        # 这样的切片操作可以有效地提取出不同类型的token对应的输出。
        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)
        # 用于存储每个mask token 对应的经过 MLP 处理后的输出。这些处理后的输出将被用于生成最终的预测 masks。
        hyper_in_list: List[torch.Tensor] = []
        # ---MLP---
        for i in range(self.num_mask_tokens):
            hyper_in_list.append(self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]))
        # 将列表 hyper_in_list 中的张量沿着指定维度进行堆叠，生成一个新的张量 hyper_in
        hyper_in = torch.stack(hyper_in_list, dim=1)
        # ---MLP End---
        
        b, c, h, w = upscaled_embedding.shape
        masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)

        # ---MLP---
        # Generate mask quality predictions
        iou_pred = self.iou_prediction_head(iou_token_out)
        # ---MLP End---

        return masks, iou_pred

该代码第一部分 Concatenate output tokens的实现原理即为, output_tokens + prompt_tokens的结合。这一部分的理解可以参考该博主的图片。【图像分割】【深度学习】SAM官方Pytorch代码-Mask decoder模块MaskDeco网络解析_sam decoder-CSDN博客

Transformer

  Transformer数据流如图所示：从底部向上看我们发现, output tokens与prompt tokens先流入一个self-attention， 然后在从 token到 image 以及 image 到 token都采用corss-attention机制。在第一个Cross attention机制中(token to image), toke当作q，image embedding当作k与v。在第二个Cross attention中(image to token)， image当作q, token充当k与v。【流程见TwoWayAttentionBlock代码】

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,  # 下采样
    ) -> None:
        """
        A transformer decoder 尝试对一个输入图片使用带有位置embedding的查询
        由多个transformer block组成, 每个block包含两个attention模块.
        输入是图像的embedding、图像的position embedding和 点的embedding,
        输出是处理后的点的embedding和处理后的图像的embedding。
        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),  # 在第一个循环中 i=0， 说明在TwoWayAttentionBlock前向传播过程中第一次进self attn
                )
            )

        
        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,  # 传入的是token = output_tokens + prompt_tokens
    ) -> Tuple[Tensor, Tensor]:
        """
        前向传播过程:
        (1) 将图像的embedding和position embedding 分别经过一个线性层，
            得到image_embedding 和 image_pe。
        (2) 将点嵌入的embedding经过一个线性层,得到point_embedding。
        (3) 对 image_embedding 和 point_embedding 进行 transformer block处理,
            得到经过处理的 image_embedding 和 point_embedding。
        (4) 对经过处理的 image_embedding 和 point_embedding 进行交叉注意力，
            得到经过处理的 point_embedding 和 image_embedding。
        
        Args:
            image_embedding (torch.Tensor): 图像嵌入张量，形状为 B x embedding_dim x h x w。
            image_pe (torch.Tensor): 图像的位置编码张量，与 image_embedding 具有相同的形状。
            point_embedding (torch.Tensor): 查询点的嵌入张量，形状为 B x N_points x embedding_dim。
        
        Returns:
            Tuple[torch.Tensor, torch.Tensor]: 经过处理的 point_embedding 和 image_embedding。
        
        """
        # Flatten image embedding to B x N_image_tokens x C
        # 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)  # image embedding 对应的 position embedding

        # 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,  # 第一次添加时, queries与query_pe相同
                key_pe=image_pe,
            )

        # Apply the final attention layer from the points to the image
        q = queries + point_embedding
        k = keys + image_pe
        # # 最后一个cross attn Final attention layer from the points to the image
        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

TwoWayAttentionBlock

        

class TwoWayAttentionBlock(nn.Module):
    #  TwoWayAttentionBlock = LayerNorm + Multi-Head Attention + MLP
    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, 
        (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]:
        
        # 第一个Self attention 模块。
        # 第一轮本身queries==query_pe
        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)

        # 第一个 Cross attention block。 tokens attending to image embedding
        # q, k, v不再是来源于同一个序列,而是多个序列. queries + query_pe充当q, k与v都由 keys提供
        # tokens to image embedding意味着，将token作为q, image_embedding 作为 k与v
        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)

        # 第二个 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

Attention

class Attention(nn.Module):
    """
    一个允许下采样embedding size的attention层
    """

    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:
        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  C_per_head表示一个head中有多少个channel

    def _recombine_heads(self, x: Tensor) -> 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:
        # 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