STViT-R 代码阅读记录

目录

一、SwinTransformer

1、原理

 2、代码

二、STViT-R

1、中心思想

2、代码与原文

---

本次不做具体的训练。只是看代码。所以只需搭建它的网络，执行一次前向传播即可。

一、SwinTransformer

1、原理

主要思想，将token按区域划分成窗口，只需每个窗口内的token单独进行 self-attention。

但是不同之间的窗口没有进行交互，为了解决这个问题。提出了

 2、代码

1、均匀的划分窗口

x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C  window_size 7  # 划分窗口  (64,7,7,96)
x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C  (64,49,96)

二、STViT-R

1、中心思想

在浅层的 transformer保持不变，去提取低层 特征， 保证image token 中包含丰富的空间信息。在深层时，本文提出了 STGM 去生成 语义token， 通过聚类，整个图像由一些具有高级语义信息的标记来表示。。 在第一个STGM过程中，语义token 由 intra and inter-window spatial pooling初始化。 由于这种空间初始化，语义token主要包含局部语义信息，并在空间中实现离散和均匀分布。 在接下来的注意层中，除了进一步的聚类外，语义标记还配备了全局聚类中心，网络可以自适应地选择部分语义标记，以聚焦于全局语义信息。

2、代码与原文

对应

xx = x.reshape(B, H // self.window_size, self.window_size, W // self.window_size, self.window_size, C)  # (1,2,7,2,7,384)
windows = xx.permute(0, 1, 3, 2, 4, 5).contiguous().reshape(-1, self.window_size, self.window_size, C).permute(0, 3, 1, 2)  # (4,384,7,7)
shortcut = self.multi_scale(windows)  # B*nW, W*W, C  multi_scale.py --13  (4,9,384)
if self.use_conv_pos:  # False
    shortcut = self.conv_pos(shortcut)
pool_x = self.norm1(shortcut.reshape(B, -1, C)).reshape(-1, self.multi_scale.num_samples, C)  # (4,9,384)

# 
class multi_scale_semantic_token1(nn.Module):
    def __init__(self, sample_window_size):
        super().__init__()
        self.sample_window_size = sample_window_size  # 3
        self.num_samples = sample_window_size * sample_window_size

    def forward(self, x):  # (4,384,7,7)
        B, C, _, _ = x.size()
        pool_x = F.adaptive_max_pool2d(x, (self.sample_window_size, self.sample_window_size)).view(B, C, self.num_samples).transpose(2, 1)  # (4,9,384)
        return pool_x

注意，这个是按照每个窗口内进行 pooling的。代码中，窗口size为7，分成了4个窗口，故pooling前的 x（4，384，7，7），pooling后，按窗口池化，每个窗口池化后的 size为3，故池化后的输出 （4，9，384）。 至于参数的设置，由于采用的是local，所以文中所述

而且

  

所以 有了如下的操作，将原来窗口的size扩大了，

k_windows = F.unfold(x.permute(0, 3, 1, 2), kernel_size=10, stride=4).view(B, C, 10, 10, -1).permute(0, 4, 2, 3, 1)  # (1,4,10,10,384)
k_windows = k_windows.reshape(-1, 100, C)  # (4,100,384)
k_windows = torch.cat([shortcut, k_windows], dim=1)  # (4,109,384)
k_windows = self.norm1(k_windows.reshape(B, -1, C)).reshape(-1, 100+self.multi_scale.num_samples, C)  # (4,109,384)

---

 公式1

前边的对应

# P
shortcut = self.multi_scale(windows)  

# MHA(P, X, X)

pool_x = self.norm1(shortcut.reshape(B, -1, C)).reshape(-1, self.multi_scale.num_samples, C)

if self.shortcut:
    x = shortcut + self.drop_path(self.layer_scale_1 * self.attn(pool_x, k_windows))

中间省略了Norm层，所以括号里的 P是 有Norm的，外面的P是 shortcut

后边的对应

x = x + self.drop_path(self.layer_scale_2 * self.mlp(self.norm2(x)))  # (1,36,384)

---

对应

 elif i == 2:
                if self.use_global:
                    semantic_token = blk(semantic_token+self.semantic_token2, torch.cat([semantic_token, x], dim=1))
                else:  # True
                    semantic_token = blk(semantic_token, torch.cat([semantic_token, x], dim=1))

 文中的

定义为（当只有 use_global时才使用）

        if self.use_global:
            self.semantic_token2 = nn.Parameter(torch.zeros(1, self.num_samples, embed_dim))
            trunc_normal_(self.semantic_token2, std=.02)

最终的对应

x = shortcut + self.drop_path(self.layer_scale_1 * attn)
x = x + self.drop_path(self.layer_scale_2 * self.mlp(self.norm2(x)))

---

 注意，在 i=1 到 i=5之间的层是 STGM，当i=5时，开始了哑铃的另一侧

对应代码

elif i == 5:
    x = blk(x, semantic_token)  # to layers.py--132

如图中的蓝线，原始的 image token作为Q，然后STGM的语义令牌作为KV，

---

上述过程循环往复，就组成了多个的哑铃结构 

            if i == 0:
                x = blk(x)  # (1,196,384)  to swin_transformer -- 242
            elif i == 1:
                semantic_token = blk(x)  # to layers.py --179
            elif i == 2:
                if self.use_global:  # True
                    semantic_token = blk(semantic_token+self.semantic_token2, torch.cat([semantic_token, x], dim=1))  # to layers.py--132
                else:  # True
                    semantic_token = blk(semantic_token, torch.cat([semantic_token, x], dim=1))  # to layers.py--132
            elif i > 2 and i < 5:
                semantic_token = blk(semantic_token)  # to layers.py--132
            elif i == 5:
                x = blk(x, semantic_token)  # to layers.py--132
            elif i == 6:
                x = blk(x)
            elif i == 7:
                semantic_token = blk(x)
            elif i == 8:
                semantic_token = blk(semantic_token, torch.cat([semantic_token, x], dim=1))
            elif i > 8 and i < 11:
                semantic_token = blk(semantic_token)
            elif i == 11:
                x = blk(x, semantic_token)
            elif i == 12:
                x = blk(x)
            elif i == 13:
                semantic_token = blk(x)
            elif i == 14:
                semantic_token = blk(semantic_token, torch.cat([semantic_token, x], dim=1))
            elif i > 14 and i < 17:
                semantic_token = blk(semantic_token)
            else:
                x = blk(x, semantic_token)

网络结构

SwinTransformer(
  (patch_embed): PatchEmbed(
    (proj): Sequential(
      (0): Conv2d_BN(
        (c): Conv2d(3, 48, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (bn): BatchNorm2d(48, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (1): Hardswish()
      (2): Conv2d_BN(
        (c): Conv2d(48, 96, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        (bn): BatchNorm2d(96, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
      (3): Hardswish()
    )
  )
  (pos_drop): Dropout(p=0.0, inplace=False)
  (layers): ModuleList(
    (0): BasicLayer(
      dim=96, input_resolution=(56, 56), depth=2
      (blocks): ModuleList(
        (0): SwinTransformerBlock(
          dim=96, input_resolution=(56, 56), num_heads=3, window_size=7, shift_size=0, mlp_ratio=4.0
          (norm1): LayerNorm((96,), eps=1e-05, elementwise_affine=True)
          (attn): WindowAttention(
            dim=96, window_size=(7, 7), num_heads=3
            (qkv): Linear(in_features=96, out_features=288, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=96, out_features=96, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
          )
          (drop_path): Identity()
          (norm2): LayerNorm((96,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=96, out_features=384, bias=True)
            (act): GELU()
            (fc2): Linear(in_features=384, out_features=96, bias=True)
            (drop): Dropout(p=0.0, inplace=False)
          )
        )
        (1): SwinTransformerBlock(
          dim=96, input_resolution=(56, 56), num_heads=3, window_size=7, shift_size=3, mlp_ratio=4.0
          (norm1): LayerNorm((96,), eps=1e-05, elementwise_affine=True)
          (attn): WindowAttention(
            dim=96, window_size=(7, 7), num_heads=3
            (qkv): Linear(in_features=96, out_features=288, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=96, out_features=96, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
          )
          (drop_path): DropPath(drop_prob=0.013)
          (norm2): LayerNorm((96,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=96, out_features=384, bias=True)
            (act): GELU()
            (fc2): Linear(in_features=384, out_features=96, bias=True)
            (drop): Dropout(p=0.0, inplace=False)
          )
        )
      )
      (downsample): PatchMerging(
        input_resolution=(56, 56), dim=96
        (reduction): Linear(in_features=384, out_features=192, bias=False)
        (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
      )
    )
    (1): BasicLayer(
      dim=192, input_resolution=(28, 28), depth=2
      (blocks): ModuleList(
        (0): SwinTransformerBlock(
          dim=192, input_resolution=(28, 28), num_heads=6, window_size=7, shift_size=0, mlp_ratio=4.0
          (norm1): LayerNorm((192,), eps=1e-05, elementwise_affine=True)
          (attn): WindowAttention(
            dim=192, window_size=(7, 7), num_heads=6
            (qkv): Linear(in_features=192, out_features=576, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=192, out_features=192, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
          )
          (drop_path): DropPath(drop_prob=0.026)
          (norm2): LayerNorm((192,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=192, out_features=768, bias=True)
            (act): GELU()
            (fc2): Linear(in_features=768, out_features=192, bias=True)
            (drop): Dropout(p=0.0, inplace=False)
          )
        )
        (1): SwinTransformerBlock(
          dim=192, input_resolution=(28, 28), num_heads=6, window_size=7, shift_size=3, mlp_ratio=4.0
          (norm1): LayerNorm((192,), eps=1e-05, elementwise_affine=True)
          (attn): WindowAttention(
            dim=192, window_size=(7, 7), num_heads=6
            (qkv): Linear(in_features=192, out_features=576, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=192, out_features=192, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
          )
          (drop_path): DropPath(drop_prob=0.039)
          (norm2): LayerNorm((192,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=192, out_features=768, bias=True)
            (act): GELU()
            (fc2): Linear(in_features=768, out_features=192, bias=True)
            (drop): Dropout(p=0.0, inplace=False)
          )
        )
      )
      (downsample): PatchMerging(
        input_resolution=(28, 28), dim=192
        (reduction): Linear(in_features=768, out_features=384, bias=False)
        (norm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
      )
    )
    (2): Deit(
      (blocks): ModuleList(
        (0): SwinTransformerBlock(
          dim=384, input_resolution=(14, 14), num_heads=12, window_size=7, shift_size=0, mlp_ratio=4.0
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): WindowAttention(
            dim=384, window_size=(7, 7), num_heads=12
            (qkv): Linear(in_features=384, out_features=1152, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
          )
          (drop_path): DropPath(drop_prob=0.052)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop): Dropout(p=0.0, inplace=False)
          )
        )
        (1): SemanticAttentionBlock(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (multi_scale): multi_scale_semantic_token1()
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.065)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (2): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.078)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (3): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.091)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (4): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.104)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (5): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.117)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (6): SwinTransformerBlock(
          dim=384, input_resolution=(14, 14), num_heads=12, window_size=7, shift_size=0, mlp_ratio=4.0
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): WindowAttention(
            dim=384, window_size=(7, 7), num_heads=12
            (qkv): Linear(in_features=384, out_features=1152, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
          )
          (drop_path): DropPath(drop_prob=0.130)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop): Dropout(p=0.0, inplace=False)
          )
        )
        (7): SemanticAttentionBlock(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (multi_scale): multi_scale_semantic_token1()
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.143)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (8): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.157)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (9): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.170)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (10): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.183)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (11): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.196)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (12): SwinTransformerBlock(
          dim=384, input_resolution=(14, 14), num_heads=12, window_size=7, shift_size=0, mlp_ratio=4.0
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): WindowAttention(
            dim=384, window_size=(7, 7), num_heads=12
            (qkv): Linear(in_features=384, out_features=1152, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
          )
          (drop_path): DropPath(drop_prob=0.209)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop): Dropout(p=0.0, inplace=False)
          )
        )
        (13): SemanticAttentionBlock(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (multi_scale): multi_scale_semantic_token1()
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.222)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (14): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.235)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (15): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.248)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (16): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.261)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
        (17): Block(
          (norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (attn): Attention(
            (q): Linear(in_features=384, out_features=384, bias=True)
            (kv): Linear(in_features=384, out_features=768, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=384, out_features=384, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
          )
          (drop_path): DropPath(drop_prob=0.274)
          (norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=384, out_features=1536, bias=True)
            (act): GELU()
            (drop1): Dropout(p=0.0, inplace=False)
            (fc2): Linear(in_features=1536, out_features=384, bias=True)
            (drop2): Dropout(p=0.0, inplace=False)
          )
        )
      )
      (downsample): PatchMerging(
        input_resolution=(14, 14), dim=384
        (reduction): Linear(in_features=1536, out_features=768, bias=False)
        (norm): LayerNorm((1536,), eps=1e-05, elementwise_affine=True)
      )
    )
    (3): BasicLayer(
      dim=768, input_resolution=(7, 7), depth=2
      (blocks): ModuleList(
        (0): SwinTransformerBlock(
          dim=768, input_resolution=(7, 7), num_heads=24, window_size=7, shift_size=0, mlp_ratio=4.0
          (norm1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
          (attn): WindowAttention(
            dim=768, window_size=(7, 7), num_heads=24
            (qkv): Linear(in_features=768, out_features=2304, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=768, out_features=768, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
          )
          (drop_path): DropPath(drop_prob=0.287)
          (norm2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=768, out_features=3072, bias=True)
            (act): GELU()
            (fc2): Linear(in_features=3072, out_features=768, bias=True)
            (drop): Dropout(p=0.0, inplace=False)
          )
        )
        (1): SwinTransformerBlock(
          dim=768, input_resolution=(7, 7), num_heads=24, window_size=7, shift_size=0, mlp_ratio=4.0
          (norm1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
          (attn): WindowAttention(
            dim=768, window_size=(7, 7), num_heads=24
            (qkv): Linear(in_features=768, out_features=2304, bias=True)
            (attn_drop): Dropout(p=0.0, inplace=False)
            (proj): Linear(in_features=768, out_features=768, bias=True)
            (proj_drop): Dropout(p=0.0, inplace=False)
            (softmax): Softmax(dim=-1)
          )
          (drop_path): DropPath(drop_prob=0.300)
          (norm2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
          (mlp): Mlp(
            (fc1): Linear(in_features=768, out_features=3072, bias=True)
            (act): GELU()
            (fc2): Linear(in_features=3072, out_features=768, bias=True)
            (drop): Dropout(p=0.0, inplace=False)
          )
        )
      )
    )
  )
  (norm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
  (avgpool): AdaptiveAvgPool1d(output_size=1)
  (head): Linear(in_features=768, out_features=100, bias=True)
)