（简单易学）使用 CVPR 2024 的 ViTamin作为backbone进行自己的分割任务

已于 2024-08-14 15:00:12 修改
阅读量295 收藏 2
点赞数 11
分类专栏： 深度学习 文章标签： python 深度学习 算法
于 2024-08-14 11:58:45 首次发布
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/qq_37643054/article/details/141187381
版权

ViTamin的官方信息

论文链接：
https://arxiv.org/pdf/2404.02132
项目主页:
https://beckschen.github.io/vitamin

ViTamin-Seg项目

下面是本文基于vitamin作为backebone，并添加decode层实现的分割模型

项目地址：https://github.com/Human9000/vitamin-seg-torch

PS：如果你觉得项目还行的话，可以点点赞，点点星星，让更多人看到它！

模型结构说明

提取了vitamin中的patch_embed输出，vitaminblocks的第1,3,6,9,12,14层的输出。 分别作为en1,en2,...,en6 这些进行通道拼接，然后经过解码器输出分割结果 解码器如下所示(采用了：组反卷积，通道重组，批标准化，Relu激活，进行组合得到)

  (up): Sequential(
    (0): ConvTranspose2d(2304, 1152, kernel_size=(2, 2), stride=(2, 2), groups=6)
    (1): ChannelShuffle()
    (2): BatchNorm2d(1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (3): ReLU()
    (4): ConvTranspose2d(1152, 576, kernel_size=(2, 2), stride=(2, 2), groups=6)
    (5): ChannelShuffle()
    (6): Conv2d(576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=6)
    (7): BatchNorm2d(576, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (8): ReLU()
    (9): ConvTranspose2d(576, 288, kernel_size=(2, 2), stride=(2, 2), groups=6)
    (10): ChannelShuffle()
    (11): BatchNorm2d(288, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (12): ReLU()
    (13): ConvTranspose2d(288, 144, kernel_size=(2, 2), stride=(2, 2), groups=6)
    (14): BatchNorm2d(144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (15): ReLU()
    (16): Conv2d(144, 52, kernel_size=(1, 1), stride=(1, 1))
  )

计算量和参数量说明

采用SegVitamin(size=(512, 512), fact=(2, 2), in_chans=1, out_channel=13)测试模型的计算量和参数量为('15.68 GMac', '24.76 M')

下面是模型的详细参数及计算量分布情况

SegVitamin(
  24.65 M, 99.553% Params, 15.64 GMac, 99.771% MACs, 
  (vitamin): VisionTransformer(
    21.79 M, 87.995% Params, 5.79 GMac, 36.917% MACs, 
    (patch_embed): HybridEmbed(
      1.08 M, 4.346% Params, 1.89 GMac, 12.073% MACs, 
      (backbone): MbConvStages(
        1.08 M, 4.346% Params, 1.89 GMac, 12.073% MACs, 
        (stem): Stem(
          39.3 k, 0.159% Params, 643.83 MMac, 4.107% MACs, 
          (conv1): Conv2d(2.37 k, 0.010% Params, 38.8 MMac, 0.247% MACs, 4, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
          (norm1): LayerNormAct2d(
            0, 0.000% Params, 0.0 Mac, 0.000% MACs, (64,), eps=1e-06, elementwise_affine=True
            (drop): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
            (act): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          )
          (conv2): Conv2d(36.93 k, 0.149% Params, 605.03 MMac, 3.859% MACs, 64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
        )
        (stages): ModuleList(
          (0): Sequential(
            71.3 k, 0.288% Params, 497.55 MMac, 3.174% MACs, 
            (0): MbConvLNBlock(
              35.65 k, 0.144% Params, 351.54 MMac, 2.242% MACs, 
              (shortcut): Downsample2d(
                0, 0.000% Params, 1.05 MMac, 0.007% MACs, 
                (pool): AvgPool2d(0, 0.000% Params, 1.05 MMac, 0.007% MACs, kernel_size=3, stride=2, padding=1)
                (expand): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              )
              (pre_norm): LayerNormAct2d(
                0, 0.000% Params, 0.0 Mac, 0.000% MACs, (64,), eps=1e-06, elementwise_affine=True
                (drop): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
                (act): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              )
              (down): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv1_1x1): Conv2d(16.64 k, 0.067% Params, 272.63 MMac, 1.739% MACs, 64, 256, kernel_size=(1, 1), stride=(1, 1))
              (act1): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (act2): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv2_kxk): Conv2d(2.56 k, 0.010% Params, 10.49 MMac, 0.067% MACs, 256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=256)
              (conv3_1x1): Conv2d(16.45 k, 0.066% Params, 67.37 MMac, 0.430% MACs, 256, 64, kernel_size=(1, 1), stride=(1, 1))
              (drop_path): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
            )
            (1): MbConvLNBlock(
              35.65 k, 0.144% Params, 146.01 MMac, 0.931% MACs, 
              (shortcut): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (pre_norm): LayerNormAct2d(
                0, 0.000% Params, 0.0 Mac, 0.000% MACs, (64,), eps=1e-06, elementwise_affine=True
                (drop): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
                (act): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              )
              (down): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv1_1x1): Conv2d(16.64 k, 0.067% Params, 68.16 MMac, 0.435% MACs, 64, 256, kernel_size=(1, 1), stride=(1, 1))
              (act1): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (act2): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv2_kxk): Conv2d(2.56 k, 0.010% Params, 10.49 MMac, 0.067% MACs, 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=256)
              (conv3_1x1): Conv2d(16.45 k, 0.066% Params, 67.37 MMac, 0.430% MACs, 256, 64, kernel_size=(1, 1), stride=(1, 1))
              (drop_path): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
            )
          )
          (1): Sequential(
            522.88 k, 2.111% Params, 637.93 MMac, 4.069% MACs, 
            (0): MbConvLNBlock(
              112.38 k, 0.454% Params, 217.58 MMac, 1.388% MACs, 
              (shortcut): Downsample2d(
                8.32 k, 0.034% Params, 8.78 MMac, 0.056% MACs, 
                (pool): AvgPool2d(0, 0.000% Params, 262.14 KMac, 0.002% MACs, kernel_size=3, stride=2, padding=1)
                (expand): Conv2d(8.32 k, 0.034% Params, 8.52 MMac, 0.054% MACs, 64, 128, kernel_size=(1, 1), stride=(1, 1))
              )
              (pre_norm): LayerNormAct2d(
                0, 0.000% Params, 0.0 Mac, 0.000% MACs, (64,), eps=1e-06, elementwise_affine=True
                (drop): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
                (act): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              )
              (down): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv1_1x1): Conv2d(33.28 k, 0.134% Params, 136.31 MMac, 0.870% MACs, 64, 512, kernel_size=(1, 1), stride=(1, 1))
              (act1): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (act2): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv2_kxk): Conv2d(5.12 k, 0.021% Params, 5.24 MMac, 0.033% MACs, 512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=512)
              (conv3_1x1): Conv2d(65.66 k, 0.265% Params, 67.24 MMac, 0.429% MACs, 512, 128, kernel_size=(1, 1), stride=(1, 1))
              (drop_path): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
            )
            (1): MbConvLNBlock(
              136.83 k, 0.553% Params, 140.12 MMac, 0.894% MACs, 
              (shortcut): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (pre_norm): LayerNormAct2d(
                0, 0.000% Params, 0.0 Mac, 0.000% MACs, (128,), eps=1e-06, elementwise_affine=True
                (drop): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
                (act): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              )
              (down): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv1_1x1): Conv2d(66.05 k, 0.267% Params, 67.63 MMac, 0.431% MACs, 128, 512, kernel_size=(1, 1), stride=(1, 1))
              (act1): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (act2): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv2_kxk): Conv2d(5.12 k, 0.021% Params, 5.24 MMac, 0.033% MACs, 512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=512)
              (conv3_1x1): Conv2d(65.66 k, 0.265% Params, 67.24 MMac, 0.429% MACs, 512, 128, kernel_size=(1, 1), stride=(1, 1))
              (drop_path): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
            )
            (2): MbConvLNBlock(
              136.83 k, 0.553% Params, 140.12 MMac, 0.894% MACs, 
              (shortcut): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (pre_norm): LayerNormAct2d(
                0, 0.000% Params, 0.0 Mac, 0.000% MACs, (128,), eps=1e-06, elementwise_affine=True
                (drop): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
                (act): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              )
              (down): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv1_1x1): Conv2d(66.05 k, 0.267% Params, 67.63 MMac, 0.431% MACs, 128, 512, kernel_size=(1, 1), stride=(1, 1))
              (act1): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (act2): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv2_kxk): Conv2d(5.12 k, 0.021% Params, 5.24 MMac, 0.033% MACs, 512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=512)
              (conv3_1x1): Conv2d(65.66 k, 0.265% Params, 67.24 MMac, 0.429% MACs, 512, 128, kernel_size=(1, 1), stride=(1, 1))
              (drop_path): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
            )
            (3): MbConvLNBlock(
              136.83 k, 0.553% Params, 140.12 MMac, 0.894% MACs, 
              (shortcut): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (pre_norm): LayerNormAct2d(
                0, 0.000% Params, 0.0 Mac, 0.000% MACs, (128,), eps=1e-06, elementwise_affine=True
                (drop): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
                (act): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              )
              (down): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv1_1x1): Conv2d(66.05 k, 0.267% Params, 67.63 MMac, 0.431% MACs, 128, 512, kernel_size=(1, 1), stride=(1, 1))
              (act1): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (act2): GELU(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
              (conv2_kxk): Conv2d(5.12 k, 0.021% Params, 5.24 MMac, 0.033% MACs, 512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=512)
              (conv3_1x1): Conv2d(65.66 k, 0.265% Params, 67.24 MMac, 0.429% MACs, 512, 128, kernel_size=(1, 1), stride=(1, 1))
              (drop_path): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
            )
          )
        )
        (pool): StridedConv(
          442.75 k, 1.788% Params, 113.34 MMac, 0.723% MACs, 
          (proj): Conv2d(442.75 k, 1.788% Params, 113.34 MMac, 0.723% MACs, 128, 384, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
          (norm): LayerNorm2d(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (128,), eps=1e-06, elementwise_affine=True)
        )
      )
      (proj): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
    )
    (pos_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
    (patch_drop): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
    (norm_pre): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
    (blocks): Sequential(
      20.71 M, 83.644% Params, 3.89 GMac, 24.844% MACs, 
      (0): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (1): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (2): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (3): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (4): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (5): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (6): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (7): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (8): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (9): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (10): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (11): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (12): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
      (13): Block(
        1.48 M, 5.975% Params, 278.2 MMac, 1.775% MACs, 
        (norm1): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (attn): Attention(
          591.36 k, 2.388% Params, 50.82 MMac, 0.324% MACs, 
          (qkv): Linear(443.52 k, 1.791% Params, 113.54 MMac, 0.724% MACs, in_features=384, out_features=1152, bias=True)
          (q_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (k_norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
          (attn_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
          (proj): Linear(147.84 k, 0.597% Params, 37.85 MMac, 0.241% MACs, in_features=384, out_features=384, bias=True)
          (proj_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
        )
        (ls1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path1): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (norm2): LayerNorm(768, 0.003% Params, 98.3 KMac, 0.001% MACs, (384,), eps=1e-06, elementwise_affine=True)
        (mlp): GeGluMlp(
          886.66 k, 3.580% Params, 227.18 MMac, 1.449% MACs, 
          (norm): LayerNorm(0, 0.000% Params, 0.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
          (act): GELU(0, 0.000% Params, 196.61 KMac, 0.001% MACs, approximate='none')
          (w0): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w1): Linear(295.68 k, 1.194% Params, 75.69 MMac, 0.483% MACs, in_features=384, out_features=768, bias=True)
          (w2): Linear(295.3 k, 1.192% Params, 75.6 MMac, 0.482% MACs, in_features=768, out_features=384, bias=True)
        )
        (ls2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
        (drop_path2): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
      )
    )
    (norm): Identity(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
    (fc_norm): LayerNorm(768, 0.003% Params, 384.0 Mac, 0.000% MACs, (384,), eps=1e-06, elementwise_affine=True)
    (head_drop): Dropout(0, 0.000% Params, 0.0 Mac, 0.000% MACs, p=0.0, inplace=False)
    (head): Linear(385, 0.002% Params, 385.0 Mac, 0.000% MACs, in_features=384, out_features=1, bias=True)
  )
  (up): Sequential(
    2.86 M, 11.558% Params, 9.85 GMac, 62.854% MACs, 
    (0): ConvTranspose2d(1.77 M, 7.150% Params, 1.81 GMac, 11.566% MACs, 2304, 1152, kernel_size=(2, 2), stride=(2, 2), groups=6)
    (1): ChannelShuffle(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
    (2): BatchNorm2d(2.3 k, 0.009% Params, 2.36 MMac, 0.015% MACs, 1152, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (3): ReLU(0, 0.000% Params, 1.18 MMac, 0.008% MACs, )
    (4): ConvTranspose2d(442.94 k, 1.789% Params, 1.81 GMac, 11.573% MACs, 1152, 576, kernel_size=(2, 2), stride=(2, 2), groups=6)
    (5): ChannelShuffle(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
    (6): Conv2d(498.24 k, 2.012% Params, 2.04 GMac, 13.018% MACs, 576, 576, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), groups=6)
    (7): BatchNorm2d(1.15 k, 0.005% Params, 4.72 MMac, 0.030% MACs, 576, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (8): ReLU(0, 0.000% Params, 2.36 MMac, 0.015% MACs, )
    (9): ConvTranspose2d(110.88 k, 0.448% Params, 1.82 GMac, 11.588% MACs, 576, 288, kernel_size=(2, 2), stride=(2, 2), groups=6)
    (10): ChannelShuffle(0, 0.000% Params, 0.0 Mac, 0.000% MACs, )
    (11): BatchNorm2d(576, 0.002% Params, 9.44 MMac, 0.060% MACs, 288, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (12): ReLU(0, 0.000% Params, 4.72 MMac, 0.030% MACs, )
    (13): ConvTranspose2d(27.79 k, 0.112% Params, 1.82 GMac, 11.618% MACs, 288, 144, kernel_size=(2, 2), stride=(2, 2), groups=6)
    (14): BatchNorm2d(288, 0.001% Params, 18.87 MMac, 0.120% MACs, 144, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
    (15): ReLU(0, 0.000% Params, 9.44 MMac, 0.060% MACs, )
    (16): Conv2d(7.54 k, 0.030% Params, 494.14 MMac, 3.152% MACs, 144, 52, kernel_size=(1, 1), stride=(1, 1))
  )
)

瞌睡的代码仓库
关注
专栏目录
YOLOv9如何训练自己的数据集（NEU-DET为案列）
03-13 2832
教会你如何使用YOLOv9训练自己的数据集
YOLOv8独家改进：backbone改进 | TransXNet：聚合全局和局部信息的全新CNN-Transformer视觉主干| CVPR2024
03-13 1119
CVPR2024 TransXNet助力检测，代替YOLOv8 Backbone
2021-03-16
weixin_40688204的博客
03-16 188
CVPR 2021 | 全新Backbone！ReXNet：助力CV任务全面涨点 CVer今天 以下文章来源于AI人工智能初学者，作者ChaucerG AI人工智能初学者 机器学习知识点总结、深度学习知识点总结以及相关垂直领域的跟进，比如CV，NLP等方面的知识。 点击下方卡片，关注“CVer”公众号 AI/CV重磅干货，第一时间送达 CVer 一个专注侃侃计算机视觉方向的公众号。计算机视觉、图像处理、机器学习、深度学习、C/C++、Python、诗和远方等。 193篇原创内容..
CVPR 2024 | 微软提出LaViT：新的视觉backbone！只需更少的注意力！
最新发布
阿木寺的博客
08-26 220
点击下方卡片，关注“CVer”公众号AI/CV重磅干货，第一时间送达点击进入—>【Mamba/多模态/扩散】交流群添加微信号：CVer111，小助手会拉你进群！扫描下方二维码，加入CVer学术星球！可以获得最新顶会/顶刊上的论文idea和CV从入门到精通资料，及最前沿应用！发论文/搞科研/涨薪，强烈推荐！转载自：晓飞的算法工程笔记Less-Attention Vision Transform...
CVPR 2024 | 百度提出视觉新骨干ViT-CoMer，刷新密集预测任务SOTA
我爱计算机视觉
03-20 753
关注公众号，发现CV技术之美本文分享 CVPR 2024 论文ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions，由百度提出视觉新骨干 ViT-CoMer，刷新密集预测任务 SOTA。论文链接：https://arxiv.org/pdf/2403...
WateRF：用于保护版权的辐射场中的鲁棒水印（CVPR 2024） - Pytorch 代码.zip
06-28
该技术被应用于CVPR 2024（Computer Vision and Pattern Recognition）这一国际顶级计算机视觉会议上，展示了其在保护版权方面的先进性和实用性。本文将深入探讨WateRF技术的核心原理及其Pytorch实现。 WateRF技术...
CVPR2024医学图像相关文章整理
08-11
CVPR2024医学图像相关文章整理，包含了医学图像的超分、配准、分割以及生成
CVPR2024:用于野外逼真图像恢复的实用算法
03-11
(CVPR2024) Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild
Python_CVPR 2024 MagicAnimate时间一致的人类图像动画使用扩散模型.zip
05-24
【标题】"Python_CVPR 2024 MagicAnimate时间一致的人类图像动画使用扩散模型" 提供了一个关于在计算机视觉领域（CVPR）2024年会议上发表的技术，该技术着重于通过扩散模型实现时间一致的人类图像动画。扩散模型是一...
YOLOv8改进：利用CVPR2024 DynamicConv替换下采样
走向CTO的路上...
07-17 725
YOLOv8作为目前最先进的目标检测算法之一，在性能和精度方面都取得了显著进步。然而，YOLOv8模型中的下采样操作会造成信息损失，并增加模型的计算量。为了解决这些问题，本文提出了一种利用CVPR2024最新DynamicConv替换下采样方法，以提升模型精度和轻量化。])return xreturn x# ... 其他 YOLOv8 模型部分的代码 ...# ... 其他层 ...# ... 其他 YOLOv8 模型部分的代码 ...
CVPR2021|PVT--无卷积密集预测的多功能backbone
CV技术指南（微信公众号）
07-05 445
前言： 卷积神经网络(CNN)在计算机视觉领域取得了显著的成功，几乎成为所有计算机视觉任务中的一种通用和主导方法。 受transformer在自然语言处理(NLP)的成功，许多研究人员正试图探索transformer的应用，一些工作模型视觉任务作为字典查找问题可学习查询，并使用transformer decoder作为一个特定任务的头的CNN主干，如VGG和ResNet。而一些现有技术已经将注意力模块纳入到CNNs中。 论文：https://arxiv.org/pdf/2102.12122.pdf
论文阅读【backbone】：CVPR2018|Deep Layer Aggregation
qq_17614495的博客
08-25 355
文章目录前言一、Introduction二、Deep Layer Aggregation总结 论文地址：Deep Layer Aggregation 前言 视觉识别需要丰富的表示，跨度从低到高的水平，从小到大小的尺度，并且从良好到粗糙的分辨率。即使在卷积网络中的特征深度，孤立的层也不够：复合和聚合这些epresentations 改善了网络在推断what和where。网络结构正在努力探索网络骨干网的许多维度，设计更深层次或更广泛的架构，但如何在网络中最佳的聚合层和块值得进一步关注。虽然跳跃连接已被合并为组
faster rcnn论文_2020 CVPR SCNet自校准卷积网络（目标检测backbone性能提升）论文阅读...
weixin_35059431的博客
12-24 491
背景《Improving Convolutional Networks with Self-calibrated Convolutions》是2020年CVPR的论文，作者来自于南开大学程明明团队。最近各种卷积注意力组合的模块工作层出不穷，性能涨点明显，包括之前的Res2Net、李沐团队的ResNeSt，应该是近期的热点方向。论文地址：http://mftp.mmcheng.net/Papers/...
CVPR 2021 更好的Backbone，伯克利&谷歌提出BoTNet，精度达84.7%
jacke121的专栏
10-09 2090
本篇分享论文『Bottleneck Transformers for Visual Recognition』，由 UC Berkeley & Google Research(Transformer一作团队)联合提出BoTNet，ImageNet上精度达84.7%。 详细信息如下： 论文链接：https://arxiv.org/abs/2101.11605 项目链接：https://github.com/lucidrains/bottleneck-transfor..
[cvpr 2024 目标检测 前沿研究 热点] cpvr 2024中与目标检测主题有关的论文
A_Student10000的博客
08-13 1160
实现包括室内和室外场景的统一单目 3D 对象检测在机器人导航等应用中具有重要意义。然而，涉及数据的各种场景来训练模型会带来挑战，因为它们具有显著不同的特性，例如不同的几何特性和异构的域分布。为了解决这些挑战，我们构建了一种基于鸟瞰图（BEV）检测范式的检测器。
ECCV 2024 | 北大提出图像检索新方法！支持输入草图/艺术画/低分辨率！
阿木寺的博客
08-22 121
点击下方卡片，关注“CVer”公众号AI/CV重磅干货，第一时间送达点击进入—>【Mamba/多模态/扩散】交流群添加微信号：CVer111，小助手会拉你进群！扫描下方二维码，加入CVer学术星球！可以获得最新顶会/顶刊上的论文idea和CV从入门到精通资料，及最前沿应用！发论文/搞科研/涨薪，强烈推荐！李昊 投稿转载自：量子位（QbitAI）从一大堆图片中精准找图，有新招了！论文已经中了E...
CVPR2019目标检测方法进展综述
求真、务实
03-15 3894
CVPR2019目标检测方法进展综述文章目录：一、二维目标检测的优化方向1、基于目标检测的backbone和特征提取       目标检测是很多计算机视觉应用的基础，比如实例分割、人体关键点提取、人脸识别等，它结合了目标分类和定位两个任务。现代大多数目标检测器的框架是 two-stage，其中目标检测被定义为一个多任务学习问题： 1）...
CNN网络架构演进：从LeNet到DenseNet
weixin_30268071的博客
02-17 284
卷积神经网络可谓是现在深度学习领域中大红大紫的网络框架，尤其在计算机视觉领域更是一枝独秀。CNN从90年代的LeNet开始，21世纪初沉寂了10年，直到12年AlexNet开始又再焕发第二春，从ZF Net到VGG，GoogLeNet再到ResNet和最近的DenseNet，网络越来越深，架构越来越复杂，解决反向传播时梯度消失的方法也越来越巧妙。新年有假期，就好好总结一波CNN的各种经典架构吧，领...
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值