【技术追踪】SAM（Segment Anything Model）代码解析与结构绘制之Mask Decoder

论文：Segment Anything
代码：https://github.com/facebookresearch/segment-anything

系列篇：
（1）【技术追踪】SAM（Segment Anything Model）代码解析与结构绘制之Image Encoder
（2）【技术追踪】SAM（Segment Anything Model）代码解析与结构绘制之Prompt Encoder

本篇示例依然采用系列篇中的狗狗图像运行代码，预测部分代码如下：

input_point = np.array([[1300, 800]])   # 输入point的坐标
input_label = np.array([1])   # label=1表示前景, label=0表示背景
# 输入box的坐标,(700,400)为左上角坐标, (1900,1100)为右下角坐标
input_box = np.array([[700, 400, 1900, 1100]])   
# 调用预测函数
masks, scores, logits = predictor.predict(point_coords=input_point,point_labels=input_label,box=input_box,multimask_output=True,
)

1. Mask Decoder代码解析

（1）输入参数

在【segment_anything/predictor.py --> SamPredictor类 -->predict_torch函数】中调用了mask_decoder实现mask预测，如下所示：

low_res_masks, iou_predictions = self.model.mask_decoder(image_embeddings=self.features,image_pe=self.model.prompt_encoder.get_dense_pe(),sparse_prompt_embeddings=sparse_embeddings,dense_prompt_embeddings=dense_embeddings,multimask_output=multimask_output,)

①参数self.features为input_image经过image_encoder嵌入后的向量，本例中大小为 ${[1, 256, 64, 64]}$ ；

②参数sparse_embeddings为prompt point和prompt box经过prompt_encoder得到的嵌入向量，本例中其大小为 ${[1, 3, 256]}$ ；

③参数dense_embeddings在本例中为无prompt mask输入时采用 nn.Embedding 的预定义嵌入向量，其大小为 ${[1, 256, 64, 64]}$ ；

④参数multimask_output是bool型参数，默认为True，支持多mask输出；

⑤参数self.model.prompt_encoder.get_dense_pe()调用PositionEmbeddingRandom实现位置编码，其大小为 ${[1, 256, 64, 64]}$ ；

  def get_dense_pe(self) -> torch.Tensor:return self.pe_layer(self.image_embedding_size).unsqueeze(0)

（2）MaskDecoder类

位置：【segment_anything/modeling/mask_decoder.py -->MaskDecoder类】
作用： 初始化网络结构，并调用predict_masks函数实现mask和iou预测

先看MaskDecoder的 $_ _ {\_\_init\_\_}$ 初始化函数和 ${forward}$ 函数：

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:super().__init__()self.transformer_dim = transformer_dim   # transformer的通道维度 = 256self.transformer = transformer  # 用于mask预测的transformer = TwoWayTransformerself.num_multimask_outputs = num_multimask_outputs  # 消除歧义时需要的mask数量 = 3self.iou_token = nn.Embedding(1, transformer_dim)  # (1, 256)self.num_mask_tokens = num_multimask_outputs + 1   # mask数目加1 = 4self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)  # (4, 256)# 以反卷积实现4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(),)# 4个mask对应的mlpself.output_hypernetworks_mlps = nn.ModuleList([MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)for i in range(self.num_mask_tokens)])# iou预测对应的mlpself.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]:masks, iou_pred = self.predict_masks(image_embeddings=image_embeddings,  # image encoder嵌入 [1, 256, 64, 64]image_pe=image_pe,  # 图像嵌入大小对应的位置编码 [1, 256, 64, 64]sparse_prompt_embeddings=sparse_prompt_embeddings,  # prompt point和box嵌入 [1, 3, 256]dense_prompt_embeddings=dense_prompt_embeddings,  # prompt mask嵌入[1, 256, 64, 64])  # 输出mask.size()=[1,4,256,256], iou_pred.size()=[1,4]# Select the correct mask or masks for outputif multimask_output:mask_slice = slice(1, None)   # 从索引1开始取后面全部else:mask_slice = slice(0, 1)   # 从索引0开始取到1结束masks = masks[:, mask_slice, :, :]  # [1, 3, 256, 256]iou_pred = iou_pred[:, mask_slice]  # [1, 3]return masks, iou_pred

传送门：【python函数】内置函数slice()用法解析

${forward}$ 的过程中主要完成了 predict_masks 函数调用；而在 $_ _ {\_\_init\_\_}$ 函数中定义了 ${transformer}$ ， ${output\_upscaling}$ ， ${output\_hypernetworks\_mlps}$ 和 ${iou\_prediction\_head}$ 这四个玩意儿，接下来咱来瞅瞅他们是啥样的。

① transformer： 在【segment_anything/build_sam.py】中可以看到为transformer定义为TwoWayTransformer，prompt_embed_dim参数为256。

        mask_decoder=MaskDecoder(num_multimask_outputs=3,transformer=TwoWayTransformer(depth=2,embedding_dim=prompt_embed_dim,  # 256mlp_dim=2048,num_heads=8,),transformer_dim=prompt_embed_dim,iou_head_depth=3,iou_head_hidden_dim=256,),

TwoWayTransformer 结构如下：

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:super().__init__()self.depth = depth   # =2self.embedding_dim = embedding_dim  # =256self.num_heads = num_heads  # =8self.mlp_dim = mlp_dim  # =2048self.layers = nn.ModuleList()# 2个TwoWayAttentionBlock模块for i in range(depth):self.layers.append(TwoWayAttentionBlock(embedding_dim=embedding_dim,  # 256num_heads=num_heads,  # 8mlp_dim=mlp_dim,  # 2048activation=activation,  # nn.ReLUattention_downsample_rate=attention_downsample_rate,  # 降采样率=2skip_first_layer_pe=(i == 0),  # 第1个TwoWayAttentionBlock为True, 第2个TwoWayAttentionBlock为False))# 1个Attention模块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,  # 图像编码:[1,256,64,64]image_pe: Tensor,   # 图像位置编码:[1,256,64,64]point_embedding: Tensor,   # iou_token,mask_tokens和sparse_prompt_embeddings的拼接向量:[1,8,256]) -> Tuple[Tensor, Tensor]:# BxCxHxW -> BxHWxC == B x N_image_tokens x Cbs, c, h, w = image_embedding.shape  # [1, 256, 64, 64]image_embedding = image_embedding.flatten(2).permute(0, 2, 1)  # [1,4096,256]image_pe = image_pe.flatten(2).permute(0, 2, 1)   # [1,4096,256]# Prepare queriesqueries = point_embedding  # 查询Q:[1,8,256]keys = image_embedding     # 键值K:[1,4096,256]# Apply transformer blocks and final layernormfor layer in self.layers:queries, keys = layer(queries=queries,keys=keys,query_pe=point_embedding,key_pe=image_pe,)  # 经过两个TwoWayAttentionBlock后, queries:[1,8,256], keys:[1,4096,256]# Apply the final attention layer from the points to the imageq = queries + point_embedding  # [1,8,256]k = keys + image_pe  # [1,4096,256]attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)  # [1,8,256]queries = queries + attn_out  # [1,8,256]queries = self.norm_final_attn(queries)  # [1,8,256]return queries, keys

Attention 结构如下：
以TwoWayAttentionBlock的第一个Attention模块为例，即：

# embedding_dim = 256, num_heads=8
self.self_attn = Attention(embedding_dim, num_heads)

Attention模块主要实现了Transformer中基本的attention机制，若参数downsample_rate不为1，则会先对维度进行下采样映射：

class Attention(nn.Module):def __init__(self,embedding_dim: int,   # 256num_heads: int,   # 8downsample_rate: int = 1,   # 1) -> None:super().__init__()self.embedding_dim = embedding_dim   # 256self.internal_dim = embedding_dim // downsample_rate   # 256self.num_heads = num_heads   # 8assert self.internal_dim % num_heads == 0, "num_heads must divide embedding_dim."self.q_proj = nn.Linear(embedding_dim, self.internal_dim)   # (256,256)self.k_proj = nn.Linear(embedding_dim, self.internal_dim)   # (256,256)self.v_proj = nn.Linear(embedding_dim, self.internal_dim)   # (256,256)self.out_proj = nn.Linear(self.internal_dim, embedding_dim)   # (256,256)def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:b, n, c = x.shapex = x.reshape(b, n, num_heads, c // num_heads)return x.transpose(1, 2)  # B x N_heads x N_tokens x C_per_headdef _recombine_heads(self, x: Tensor) -> Tensor:b, n_heads, n_tokens, c_per_head = x.shapex = x.transpose(1, 2)return x.reshape(b, n_tokens, n_heads * c_per_head)  # B x N_tokens x Cdef forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:# Input projections# 输入q:[1,8,256];k:[1,8,256];v:[1,8,256]q = self.q_proj(q)  # [1,8,256]k = self.k_proj(k)  # [1,8,256]v = self.v_proj(v)  # [1,8,256]# Separate into headsq = self._separate_heads(q, self.num_heads)  # [1,8,8,32]k = self._separate_heads(k, self.num_heads)  # [1,8,8,32]v = self._separate_heads(v, self.num_heads)  # [1,8,8,32]_, _, _, c_per_head = q.shape   # 每个head的维度c_per_head=32# attention机制-----------------------------------------------------------------------# 每个head实现q乘k的转置: [1,8,8,32]@[1,8,32,8]->[1,8,8,8]attn = q @ k.permute(0, 1, 3, 2)  # B x N_heads x N_tokens x N_tokensattn = attn / math.sqrt(c_per_head)  # q @ k(^T) / 根号dattn = torch.softmax(attn, dim=-1)  # [1,8,8,8]# -----------------------------------------------------------------------------------# Get outputout = attn @ v   # softmax( q @ k(^T) / 根号d ) @ v ---> [1,8,8,32]out = self._recombine_heads(out)  # [1,8,256]out = self.out_proj(out)  # [1,8,256]return out

为避免代码看的太晕，把Attention可视化一下，没错，就是最基本的Multi-head Attention啦~

在这里插入图片描述

TwoWayAttentionBlock 结构如下：
以TwoWayTransformer的第一个TwoWayAttentionBlock模块为例，即：

TwoWayAttentionBlock(embedding_dim=embedding_dim,  # 256num_heads=num_heads,  # 8mlp_dim=mlp_dim,  # 2048activation=activation,  # nn.ReLUattention_downsample_rate=attention_downsample_rate,  # 降采样率=2skip_first_layer_pe=(i == 0),  # 第1个TwoWayAttentionBlock为True)

TwoWayAttentionBlock模块：

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:super().__init__()self.self_attn = Attention(embedding_dim, num_heads)   # embedding_dim=256, num_heads=8self.norm1 = nn.LayerNorm(embedding_dim)  # 256self.cross_attn_token_to_image = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)   # embedding_dim=256, num_heads=8, attention_downsample_rate=2self.norm2 = nn.LayerNorm(embedding_dim)  # 256# embedding_dim=256, mlp_dim=2048, activation=nn.ReLUself.mlp = MLPBlock(embedding_dim, mlp_dim, activation)self.norm3 = nn.LayerNorm(embedding_dim)  # 256self.norm4 = nn.LayerNorm(embedding_dim)  # 256self.cross_attn_image_to_token = Attention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate)   # embedding_dim=256, num_heads=8, attention_downsample_rate=2self.skip_first_layer_pe = skip_first_layer_pe  # Truedef forward(self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor) -> Tuple[Tensor, Tensor]:# 输入queries:[1,8,256], keys:[1,4096,256], query_pe:[1,8,256], key_pe:[1,4096,256]# Self attention blockif self.skip_first_layer_pe:queries = self.self_attn(q=queries, k=queries, v=queries)  # [1,8,256]else:q = queries + query_peattn_out = self.self_attn(q=q, k=q, v=queries)queries = queries + attn_outqueries = self.norm1(queries)  # [1,8,256]# Cross attention block, tokens attending to image embeddingq = queries + query_pe  # [1,8,256]k = keys + key_pe  # [1,4096,256]attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)  # [1,8,256]queries = queries + attn_out  # [1,8,256]queries = self.norm2(queries)  # [1,8,256]# MLP blockmlp_out = self.mlp(queries)   # [1,8,256]queries = queries + mlp_out   # [1,8,256]queries = self.norm3(queries)  # [1,8,256]# Cross attention block, image embedding attending to tokensq = queries + query_pe    # [1,8,256]k = keys + key_pe   # [1,4096,256]attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)  # [1,4096,256]keys = keys + attn_out  # [1,4096,256]keys = self.norm4(keys)  # [1,4096,256]return queries, keys

可以看到TwoWayTransformer的结构以及token维度变化并不复杂，但其交错的 ${Q}$ ， ${K}$ ， ${V}$ 确实令人眼花缭乱：
在这里插入图片描述

TwoWayTransformer中的MLP：

class MLPBlock(nn.Module):def __init__(self,embedding_dim: int,mlp_dim: int,act: Type[nn.Module] = nn.GELU,) -> None:super().__init__()# embedding_dim=256, mlp_dim=2048self.lin1 = nn.Linear(embedding_dim, mlp_dim)  self.lin2 = nn.Linear(mlp_dim, embedding_dim)self.act = act()def forward(self, x: torch.Tensor) -> torch.Tensor:return self.lin2(self.act(self.lin1(x)))

MLP为简单的线性、激活、线性结构：
在这里插入图片描述

② output_upscaling：

Sequential((0): ConvTranspose2d(256, 64, kernel_size=(2, 2), stride=(2, 2))(1): LayerNorm2d()(2): GELU(approximate='none')(3): ConvTranspose2d(64, 32, kernel_size=(2, 2), stride=(2, 2))(4): GELU(approximate='none')
)

output_upscaling模块由两个反卷积、两个GELU激活和一个LayerNorm组成，实现了特征图的四倍上采样，在 predict_masks函数 中将 ${[1,256,64,64]}$ 上采样至 ${[1,32,256,256]}$ 。

src = src.transpose(1, 2).view(b, c, h, w)   # reshape: [1,4096,256]-> [1,256,64,64]
upscaled_embedding = self.output_upscaling(src) # [1,32,256,256]

③ output_hypernetworks_mlps：

ModuleList((0-3): 4 x MLP((layers): ModuleList((0-1): 2 x Linear(in_features=256, out_features=256, bias=True)(2): Linear(in_features=256, out_features=32, bias=True)))
)

output_hypernetworks_mlps由4个MLP组成，在 predict_masks函数 中将 ${[1,256]}$ 下采样至 ${[1,32]}$ 。与TwoWayAttentionBlock中的MLP不同，其结构稍稍多一丢丢：

class MLP(nn.Module):def __init__(self,input_dim: int,   # 256hidden_dim: int,  # 256output_dim: int,  # 32num_layers: int,  # 3sigmoid_output: bool = False,  # False) -> None:super().__init__()self.num_layers = num_layers  # 3h = [hidden_dim] * (num_layers - 1)  # [256,256]self.layers = nn.ModuleList(# [input_dim] + h: [256,256,256], h + [output_dim]:[256,256,32]nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))self.sigmoid_output = sigmoid_outputdef forward(self, x):for i, layer in enumerate(self.layers):# i<2经线性层后relu激活x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)  if self.sigmoid_output:x = F.sigmoid(x)return x

在这里插入图片描述

④ iou_prediction_head：

MLP((layers): ModuleList((0-1): 2 x Linear(in_features=256, out_features=256, bias=True)(2): Linear(in_features=256, out_features=4, bias=True))
)

iou_prediction_head用以实现iou预测，由1个MLP完成，其结构与output_hypernetworks_mlps中的MLP一样，只是最终将 ${[1,256]}$ 映射至 ${[1,4]}$ ，分别代表非multimask预测时的1个mask和multimask预测时的3个mask的iou。

（3）predict_masks函数

位置：【segment_anything/modeling/mask_decoder.py --> MaskDecoder类 --> predict_masks函数】
作用： 利用上述 ${transformer}$ ， ${output\_upscaling}$ ， ${output\_hypernetworks\_mlps}$ 和 ${iou\_prediction\_head}$ 四个模块，实现mask和iou预测

此时此刻，首先来重温一下，传入predict_masks函数的参数分别是什么：

① image_embeddings：image encoder嵌入，大小为 ${[1, 256, 64, 64]}$ ；
② image_pe：图像嵌入大小对应的位置编码，大小同为 ${[1, 256, 64, 64]}$ ；
③ sparse_prompt_embeddings：prompt point和box嵌入，大小为 ${[1, 3, 256]}$ ；
④ dense_prompt_embeddings：prompt mask嵌入，大小为 ${[1, 256, 64, 64]}$ ；

def predict_masks(self,image_embeddings: torch.Tensor,  # [1, 256, 64, 64]image_pe: torch.Tensor,  # [1, 256, 64, 64]sparse_prompt_embeddings: torch.Tensor,  # [1, 3, 256]dense_prompt_embeddings: torch.Tensor,  # [1, 256, 64, 64]
) -> Tuple[torch.Tensor, torch.Tensor]:"""Predicts masks. See 'forward' for more details."""# Concatenate output tokens# 拼接iou的token和mask的token: [1,256]+[4,256]->[5,256]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)  # [1,5,256]# iou的token和mask的token + prompt point和box嵌入tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)  # [1,8,256]# Expand per-image data in batch direction to be per-masksrc = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)  # 按batch重复: [1,256,64,64]src = src + dense_prompt_embeddings  # [1,256,64,64]pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)  # 按batch重复: [1,256,64,64]b, c, h, w = src.shape  # 1,256,64,64# Run the transformer# src是image encoder嵌入和prompt mask嵌入# pos_src是图像嵌入大小对应的位置编码# tokens是iou的token和mask的token + prompt point和box嵌入hs, src = self.transformer(src, pos_src, tokens)  # hs:[1,8,256], src:[1,4096,256]iou_token_out = hs[:, 0, :]  # 第1个为iou的token输出[1,256]mask_tokens_out = hs[:, 1: (1 + self.num_mask_tokens), :]  # 随后4个为mask的token输出[4,256]# Upscale mask embeddings and predict masks using the mask tokenssrc = src.transpose(1, 2).view(b, c, h, w)   # reshape: [1,4096,256]-> [1,256,64,64]upscaled_embedding = self.output_upscaling(src)  # [1,32,256,256]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)  # [1,4,32]b, c, h, w = upscaled_embedding.shape  # 1,32,256,256masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)  # [1,4,256,256]# Generate mask quality predictionsiou_pred = self.iou_prediction_head(iou_token_out)  # [1,4]return masks, iou_pred

由此可见，经TwoWayTransformer获得了iou_token_out和mask_tokens_out，iou_token_out由iou_prediction_head（1个MLP）实现iou预测，4个mask_tokens_out分别经过1个MLP所获得的映射拼接后，与经过output_upscaling上采样后的图像嵌入（包含image encoder嵌入和prompt mask嵌入）进行矩阵相乘，得到mask预测。

2. Mask Decoder结构绘制

（1）结构打印

MaskDecoder((transformer): TwoWayTransformer((layers): ModuleList((0-1): 2 x TwoWayAttentionBlock((self_attn): Attention((q_proj): Linear(in_features=256, out_features=256, bias=True)(k_proj): Linear(in_features=256, out_features=256, bias=True)(v_proj): Linear(in_features=256, out_features=256, bias=True)(out_proj): Linear(in_features=256, out_features=256, bias=True))(norm1): LayerNorm((256,), eps=1e-05, elementwise_affine=True)(cross_attn_token_to_image): Attention((q_proj): Linear(in_features=256, out_features=128, bias=True)(k_proj): Linear(in_features=256, out_features=128, bias=True)(v_proj): Linear(in_features=256, out_features=128, bias=True)(out_proj): Linear(in_features=128, out_features=256, bias=True))(norm2): LayerNorm((256,), eps=1e-05, elementwise_affine=True)(mlp): MLPBlock((lin1): Linear(in_features=256, out_features=2048, bias=True)(lin2): Linear(in_features=2048, out_features=256, bias=True)(act): ReLU())(norm3): LayerNorm((256,), eps=1e-05, elementwise_affine=True)(norm4): LayerNorm((256,), eps=1e-05, elementwise_affine=True)(cross_attn_image_to_token): Attention((q_proj): Linear(in_features=256, out_features=128, bias=True)(k_proj): Linear(in_features=256, out_features=128, bias=True)(v_proj): Linear(in_features=256, out_features=128, bias=True)(out_proj): Linear(in_features=128, out_features=256, bias=True))))(final_attn_token_to_image): Attention((q_proj): Linear(in_features=256, out_features=128, bias=True)(k_proj): Linear(in_features=256, out_features=128, bias=True)(v_proj): Linear(in_features=256, out_features=128, bias=True)(out_proj): Linear(in_features=128, out_features=256, bias=True))(norm_final_attn): LayerNorm((256,), eps=1e-05, elementwise_affine=True))(iou_token): Embedding(1, 256)(mask_tokens): Embedding(4, 256)(output_upscaling): Sequential((0): ConvTranspose2d(256, 64, kernel_size=(2, 2), stride=(2, 2))(1): LayerNorm2d()(2): GELU(approximate='none')(3): ConvTranspose2d(64, 32, kernel_size=(2, 2), stride=(2, 2))(4): GELU(approximate='none'))(output_hypernetworks_mlps): ModuleList((0-3): 4 x MLP((layers): ModuleList((0-1): 2 x Linear(in_features=256, out_features=256, bias=True)(2): Linear(in_features=256, out_features=32, bias=True))))(iou_prediction_head): MLP((layers): ModuleList((0-1): 2 x Linear(in_features=256, out_features=256, bias=True)(2): Linear(in_features=256, out_features=4, bias=True)))
)