• YOLOv8+swin_transfomer

    测试环境：cuda11.3  pytorch1.11 rtx3090  wsl2 ubuntu20.04

    本科在读，中九以上老师或者课题组捞捞我，孩子想读书，求课题组师兄内推qaq

    踩了很多坑，网上很多博主的代码根本跑不通，自己去github仓库复现修改的

    网上博主的代码日常出现cpu,gpu混合，或许是人家分布式训练了，哈哈哈

    下面上干货吧，宝子们点个关注，点个赞，没有废话
    ————————————————
    首先上yaml文件，诚意满满，我是做的分割，做检测就修改最后那个检测头就好了

    1. # Ultralytics YOLO 🚀, AGPL-3.0 license
    2. # YOLOv8-seg instance segmentation model. For Usage examples see https://docs.ultralytics.com/tasks/segment
    3.  
    4. # Parameters
    5. nc: 1  # number of classes
    6. scales: # model compound scaling constants, i.e. 'model=yolov8n-seg.yaml' will call yolov8-seg.yaml with scale 'n'
    7.   # [depth, width, max_channels]
    8.   n: [0.33, 0.25, 1024]
    9.   s: [0.33, 0.50, 1024]
    10.   m: [0.67, 0.75, 768]
    11.   l: [1.00, 1.00, 512]
    12.   x: [1.00, 1.25, 512]
    13.  
    14. # YOLOv8.0n backbone
    15. backbone:
    16.   # [from, repeats, module, args]
    17.   - [-1, 1, Conv, [64, 3, 2]]  # 0-P1/2
    18.   - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4
    19.   - [-1, 3, C2f, [128, True]]
    20.   - [-1, 1, Conv, [256, 3, 2]]  # 3-P3/8
    21.   - [-1, 6, C2f, [256, True]]
    22.   - [-1, 1, Conv, [512, 3, 2]]  # 5-P4/16
    23.   - [-1, 9, C3STR, [512]]
    24.   - [-1, 1, Conv, [1024, 3, 2]]  # 7-P5/32
    25.   - [-1, 3, C3STR, [1024]]
    26.   - [-1, 1, SPPF, [1024, 5]]  # 9
    27.  
    28. # YOLOv8.0n head
    29. head:
    30.   - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
    31.   - [[-1, 6], 1, Concat, [1]]  # cat backbone P4
    32.   - [-1, 3, C2f, [512]]  # 12
    33.  
    34.   - [-1, 1, nn.Upsample, [None, 2, 'nearest']]
    35.   - [[-1, 4], 1, Concat, [1]]  # cat backbone P3
    36.   - [-1, 3, C2f, [256]]  # 15 (P3/8-small)
    37.  
    38.   - [-1, 1, Conv, [256, 3, 2]]
    39.   - [[-1, 12], 1, Concat, [1]]  # cat head P4
    40.   - [-1, 3, C2f, [512]]  # 18 (P4/16-medium)
    41.  
    42.   - [-1, 1, Conv, [512, 3, 2]]
    43.   - [[-1, 9], 1, Concat, [1]]  # cat head P5
    44.   - [-1, 3, C2f, [1024]]  # 21 (P5/32-large)
    45.  
    46.   - [[15, 18, 21], 1, Segment, [nc, 32, 256]]  # Segment(P3, P4, P5)

    然后在nn/modules/block.py最下面加入

    1.  
    2. class SwinTransformerBlock(nn.Module):
    3.     def __init__(self, c1, c2, num_heads, num_layers, window_size=8):
    4.         super().__init__()
    5.         self.conv = None
    6.         if c1 != c2:
    7.             self.conv = Conv(c1, c2)
    8.  
    9.         # remove input_resolution
    10.         self.blocks = nn.Sequential(*[SwinTransformerLayer(dim=c2, num_heads=num_heads, window_size=window_size,
    11.                                  shift_size=0 if (i % 2 == 0) else window_size // 2) for i in range(num_layers)])
    12.  
    13.     def forward(self, x):
    14.         if self.conv is not None:
    15.             x = self.conv(x)
    16.         x = self.blocks(x)
    17.         return x
    18. class WindowAttention(nn.Module):
    19.  
    20.     def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
    21.  
    22.         super().__init__()
    23.         self.dim = dim
    24.         self.window_size = window_size  # Wh, Ww
    25.         self.num_heads = num_heads
    26.         head_dim = dim // num_heads
    27.         self.scale = qk_scale or head_dim ** -0.5
    28.  
    29.         # define a parameter table of relative position bias
    30.         self.relative_position_bias_table = nn.Parameter(
    31.             torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
    32.  
    33.         # get pair-wise relative position index for each token inside the window
    34.         coords_h = torch.arange(self.window_size[0])
    35.         coords_w = torch.arange(self.window_size[1])
    36.         coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
    37.         coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
    38.         relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
    39.         relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
    40.         relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
    41.         relative_coords[:, :, 1] += self.window_size[1] - 1
    42.         relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
    43.         relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
    44.         self.register_buffer("relative_position_index", relative_position_index)
    45.  
    46.         self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
    47.         self.attn_drop = nn.Dropout(attn_drop)
    48.         self.proj = nn.Linear(dim, dim)
    49.         self.proj_drop = nn.Dropout(proj_drop)
    50.  
    51.         nn.init.normal_(self.relative_position_bias_table, std=.02)
    52.         self.softmax = nn.Softmax(dim=-1)
    53.  
    54.     def forward(self, x, mask=None):
    55.  
    56.         B_, N, C = x.shape
    57.         qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
    58.         q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
    59.  
    60.         q = q * self.scale
    61.         attn = (q @ k.transpose(-2, -1))
    62.  
    63.         relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
    64.             self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
    65.         relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
    66.         attn = attn + relative_position_bias.unsqueeze(0)
    67.  
    68.         if mask is not None:
    69.             nW = mask.shape[0]
    70.             attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
    71.             attn = attn.view(-1, self.num_heads, N, N)
    72.             attn = self.softmax(attn)
    73.         else:
    74.             attn = self.softmax(attn)
    75.  
    76.         attn = self.attn_drop(attn)
    77.  
    78.         # print(attn.dtype, v.dtype)
    79.         try:
    80.             x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
    81.         except:
    82.             #print(attn.dtype, v.dtype)
    83.             x = (attn.half() @ v).transpose(1, 2).reshape(B_, N, C)
    84.         x = self.proj(x)
    85.         x = self.proj_drop(x)
    86.         return x
    87.  
    88. class Mlp(nn.Module):
    89.  
    90.     def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0.):
    91.         super().__init__()
    92.         out_features = out_features or in_features
    93.         hidden_features = hidden_features or in_features
    94.         self.fc1 = nn.Linear(in_features, hidden_features)
    95.         self.act = act_layer()
    96.         self.fc2 = nn.Linear(hidden_features, out_features)
    97.         self.drop = nn.Dropout(drop)
    98.  
    99.     def forward(self, x):
    100.         x = self.fc1(x)
    101.         x = self.act(x)
    102.         x = self.drop(x)
    103.         x = self.fc2(x)
    104.         x = self.drop(x)
    105.         return x
    106.  
    107. class SwinTransformerLayer(nn.Module):
    108.  
    109.     def __init__(self, dim, num_heads, window_size=8, shift_size=0,
    110.                  mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
    111.                  act_layer=nn.SiLU, norm_layer=nn.LayerNorm):
    112.         super().__init__()
    113.         self.dim = dim
    114.         self.num_heads = num_heads
    115.         self.window_size = window_size
    116.         self.shift_size = shift_size
    117.         self.mlp_ratio = mlp_ratio
    118.         # if min(self.input_resolution) <= self.window_size:
    119.         #     # if window size is larger than input resolution, we don't partition windows
    120.         #     self.shift_size = 0
    121.         #     self.window_size = min(self.input_resolution)
    122.         assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
    123.  
    124.         self.norm1 = norm_layer(dim)
    125.         self.attn = WindowAttention(
    126.             dim, window_size=(self.window_size, self.window_size), num_heads=num_heads,
    127.             qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
    128.  
    129.         self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
    130.         self.norm2 = norm_layer(dim)
    131.         mlp_hidden_dim = int(dim * mlp_ratio)
    132.         self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
    133.  
    134.     def create_mask(self, H, W):
    135.         # calculate attention mask for SW-MSA
    136.         img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
    137.         h_slices = (slice(0, -self.window_size),
    138.                     slice(-self.window_size, -self.shift_size),
    139.                     slice(-self.shift_size, None))
    140.         w_slices = (slice(0, -self.window_size),
    141.                     slice(-self.window_size, -self.shift_size),
    142.                     slice(-self.shift_size, None))
    143.         cnt = 0
    144.         for h in h_slices:
    145.             for w in w_slices:
    146.                 img_mask[:, h, w, :] = cnt
    147.                 cnt += 1
    148.  
    149.         def window_partition(x, window_size):
    150.             """
    151.             Args:
    152.                 x: (B, H, W, C)
    153.                 window_size (int): window size
    155.             Returns:
    156.                 windows: (num_windows*B, window_size, window_size, C)
    157.             """
    158.             B, H, W, C = x.shape
    159.             x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
    160.             windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    161.             return windows
    162.  
    163.         def window_reverse(windows, window_size, H, W):
    164.             """
    165.             Args:
    166.                 windows: (num_windows*B, window_size, window_size, C)
    167.                 window_size (int): Window size
    168.                 H (int): Height of image
    169.                 W (int): Width of image
    171.             Returns:
    172.                 x: (B, H, W, C)
    173.             """
    174.             B = int(windows.shape[0] / (H * W / window_size / window_size))
    175.             x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
    176.             x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
    177.             return x
    178.  
    179.         mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
    180.         mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
    181.         attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
    182.         attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
    183.  
    184.         return attn_mask
    185.  
    186.     def forward(self, x):
    187.         # reshape x[b c h w] to x[b l c]
    188.         _, _, H_, W_ = x.shape
    189.  
    190.         Padding = False
    191.         if min(H_, W_) < self.window_size or H_ % self.window_size!=0 or W_ % self.window_size!=0:
    192.             Padding = True
    193.             # print(f'img_size {min(H_, W_)} is less than (or not divided by) window_size {self.window_size}, Padding.')
    194.             pad_r = (self.window_size - W_ % self.window_size) % self.window_size
    195.             pad_b = (self.window_size - H_ % self.window_size) % self.window_size
    196.             x = F.pad(x, (0, pad_r, 0, pad_b))
    197.  
    198.         # print('2', x.shape)
    199.         B, C, H, W = x.shape
    200.         L = H * W
    201.         x = x.permute(0, 2, 3, 1).contiguous().view(B, L, C)  # b, L, c
    202.  
    203.         # create mask from init to forward
    204.         if self.shift_size > 0:
    205.             attn_mask = self.create_mask(H, W).to(x.device)
    206.         else:
    207.             attn_mask = None
    208.  
    209.         shortcut = x
    210.         x = self.norm1(x)
    211.         x = x.view(B, H, W, C)
    212.  
    213.         # cyclic shift
    214.         if self.shift_size > 0:
    215.             shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
    216.         else:
    217.             shifted_x = x
    218.  
    219.         def window_partition(x, window_size):
    220.             """
    221.             Args:
    222.                 x: (B, H, W, C)
    223.                 window_size (int): window size
    225.             Returns:
    226.                 windows: (num_windows*B, window_size, window_size, C)
    227.             """
    228.             B, H, W, C = x.shape
    229.             x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
    230.             windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    231.             return windows
    232.  
    233.         def window_reverse(windows, window_size, H, W):
    234.             """
    235.             Args:
    236.                 windows: (num_windows*B, window_size, window_size, C)
    237.                 window_size (int): Window size
    238.                 H (int): Height of image
    239.                 W (int): Width of image
    241.             Returns:
    242.                 x: (B, H, W, C)
    243.             """
    244.             B = int(windows.shape[0] / (H * W / window_size / window_size))
    245.             x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
    246.             x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
    247.             return x
    248.  
    249.         # partition windows
    250.         x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
    251.         x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
    252.  
    253.         # W-MSA/SW-MSA
    254.         attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C
    255.  
    256.         # merge windows
    257.         attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
    258.         shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
    259.  
    260.         # reverse cyclic shift
    261.         if self.shift_size > 0:
    262.             x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
    263.         else:
    264.             x = shifted_x
    265.         x = x.view(B, H * W, C)
    266.  
    267.         # FFN
    268.         x = shortcut + self.drop_path(x)
    269.         x = x + self.drop_path(self.mlp(self.norm2(x)))
    270.  
    271.         x = x.permute(0, 2, 1).contiguous().view(-1, C, H, W)  # b c h w
    272.  
    273.         if Padding:
    274.             x = x[:, :, :H_, :W_]  # reverse padding
    275.  
    276.         return x
    277.  
    278. class C3STR(C3):
    279.     # C3 module with SwinTransformerBlock()
    280.     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
    281.         super().__init__(c1, c2, n, shortcut, g, e)
    282.         c_ = int(c2 * e)
    283.         num_heads = c_ // 32
    284.         self.m = SwinTransformerBlock(c_, c_, num_heads, n)

    然后是在目录的init.py里面把这个模块注册进去，参考我的另一篇博客的后半部分

    YOLOv8+swin_transfomerv2_不会写代码！！的博客-CSDN博客

    task.py需要修改两处

    1.         n = n_ = max(round(n * depth), 1) if n > 1 else n  # depth gain
    2.         if m in (Classify, Conv, ConvTranspose, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, Focus,
    3.                  BottleneckCSP, C1, C2, C2f, C3, C3TR, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x, RepC3
    4.                  ,CBAM , GAM_Attention ,ResBlock_CBAM,GCT,C3STR,SwinV2_CSPB):
    1.             args = [c1, c2, *args[1:]]
    2.             if m in (BottleneckCSP, C1, C2, C2f, C3, C3TR, C3Ghost, C3x, RepC3,C3STR):
    3.                 args.insert(2, n)  # number of repeats
    4.                 n = 1

    有问题私信

    结构图如下，可以对模块排列组合涨点

  • 相关阅读:
    Java 内部类 面试“变态题”
    解释一下React中的钩子（hooks），例如useState和useEffect。
    【趣味测试】
    python第三方库-requests的使用
    Pytorch API
    使用DIV+CSS技术设计的非遗文化网页与实现制作（web前端网页制作课作业）
    【DDPM论文解读】Denoising Diffusion Probabilistic Models
    Java文件流练习
    Lingolingo
    Blazor VS Vue
  • 原文地址：https://blog.csdn.net/xty123abc/article/details/133429072