• YOLOv5、YOLOv8改进：C3STR（Swin Transformer）

    目录

    1.介绍

    2. YOLOv5、YOLOv8改进

    2.1 common.py配置

    2.2 yolo.py配置

    2.3 yaml配置文件

    1.介绍

    视觉领域正在见证从 CNN 到 Transformers 的建模转变，纯 Transformer 架构在主要视频识别基准测试中达到了最高准确度。这些视频模型都建立在 Transformer 层之上，Transformer 层在空间和时间维度上全局连接块。在本文中，我们提倡视频 Transformer 中的局部归纳偏差，与以前的方法相比，即使使用时空分解，也可以在全局范围内计算自注意力，从而实现更好的速度-精度权衡。所提出的视频架构的局部性是通过调整为图像域设计的 Swin Transformer 实现的，同时继续利用预训练图像模型的力量。我们的方法在广泛的视频识别基准测试中实现了最先进的准确度，包括动作识别（Kinetics-400 上的 84.9 top-1 准确度和 Kinetics-600 上的 85.9 top-1 准确度，减少了约 20× 预训练数据和小模型尺寸的 3 倍）和时间建模（SomethingSomething v2 上的 69.6 top-1 准确率）。

    论文地址Swin-Transformer论文下载论文地址

     

    该论文介绍了一种名为 Swin Transformer 的新视觉 Transformer，它能够作为计算机视觉的通用主干。将 Transformer 从语言适应到视觉的挑战来自两个领域之间的差异，例如视觉实体的规模变化很大，以及与文本中的单词相比，图像中像素的高分辨率。为了解决这些差异，我们提出了一种分层 Transformer，其表示是用移位窗口计算的。移位窗口方案通过将 self-attention 计算限制在不重叠的本地窗口上，同时还允许跨窗口连接，从而带来更高的效率。这种分层架构具有在各种尺度上建模的灵活性，并且具有相对于图像大小的线性计算复杂度。Swin Transformer 的这些特性使其与广泛的视觉任务兼容，包括图像分类（ImageNet-1K 上 86.4 top-1 准确度）和密集预测任务，例如对象检测（COCO 测试上 58.7 box AP 和 51.1 mask AP dev）和语义分割（ADE20K val 为 53.5 mIoU）。它的性能大大超过了之前的 state-of-the-art，在 COCO 上 +2.7 box AP 和 +2.6 mask AP，在 ADE20K 上 +3.2 mIoU，展示了基于 Transformer 的模型作为视觉骨干的潜力。代码和模型将在 它的性能大大超过了之前的 state-of-the-art，在 COCO 上 +2.7 box AP 和 +2.6 mask AP，在 ADE20K 上 +3.2 mIoU，展示了基于 Transformer 的模型作为视觉骨干的潜力。代码和模型将在 它的性能大大超过了之前的 state-of-the-art，在 COCO 上 +2.7 box AP 和 +2.6 mask AP，在 ADE20K 上 +3.2 mIoU，展示了基于 Transformer 的模型作为视觉骨干的潜力。

    面临问题：
    作者提出了将Swin Transformer缩放到30亿个参数的技术 ，并使其能够使用高达1536×1536分辨率的图像进行训练。在很多方面达到了SOTA。

    目前，视觉模型尚未像NLP语言模型那样被广泛探索，部分原因是训练和应用中的以下差异:

    （1）视觉模型通常在规模上面临不稳定性问题；

    （2）许多下游视觉任务需要高分辨率图像，如何有效地将低分辨率预训练的模型转换为高分辨率模型尚未被有效探索,也就是跨窗口分辨率迁移模型时性能下降。

    （3）当图像分辨率较高时，GPU显存消耗也是一个问题。

    解决思路：
    为了解决这些问题，作者提出了几种技术，并在本文中以Swin Transformer进行了说明：

    （1）提高大视觉模型稳定性的后归一化（post normalization） 技术和缩放余弦注意力（scaled cosine attention）方法，以提高大型视觉模型的稳定性；

    （2）一种对数间隔连续位置偏差技术（log-spaced continuous position bias technique） ，用于有效地将在低分辨率图像中预训练的模型转换为其高分辨率对应模型。

    （3）分享节约GPU内存消耗方法，使得训练大分辨率模型可行；


    2. YOLOv5、YOLOv8改进

    2.1 common.py配置

    在./models/common.py文件中增加以下模块，直接复制即可

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

    2.2 yolo.py配置

    不需要

    2.3 yaml配置文件

    增加以下yolov5_swin_transfomrer.yaml文件

    1. 代码
    2. # YOLOv5 🚀 by Ultralytics, GPL-3.0 license
    3.  
    4. # Parameters
    5. nc: 80  # number of classes
    6. depth_multiple: 0.33  # model depth multiple
    7. width_multiple: 0.50  # layer channel multiple
    8. anchors:
    9.   - [10,13, 16,30, 33,23]  # P3/8
    10.   - [30,61, 62,45, 59,119]  # P4/16
    11.   - [116,90, 156,198, 373,326]  # P5/32
    12.  
    13. # YOLOv5 v6.0 backbone by yoloair
    14. backbone:
    15.   # [from, number, module, args]
    16.   [[-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2
    17.    [-1, 1, Conv, [128, 3, 2]],  # 1-P2/4
    18.    [-1, 3, C3, [128]],
    19.    [-1, 1, Conv, [256, 3, 2]],  # 3-P3/8
    20.    [-1, 6, C3, [256]],
    21.    [-1, 1, Conv, [512, 3, 2]],  # 5-P4/16
    22.    [-1, 9, C3STR, [256]],
    23.    [-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32
    24.    [-1, 3, C3STR, [512]],  # 9 <--- ST2CSPB() Transformer module
    25.    [-1, 1, SPPF, [1024, 5]],  # 9
    26.   ]
    27.  
    28. # YOLOv5 v6.0 head
    29. head:
    30.   [[-1, 1, Conv, [512, 1, 1]],
    31.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
    32.    [[-1, 6], 1, Concat, [1]],  # cat backbone P4
    33.    [-1, 3, C3, [512, False]],  # 13
    34.  
    35.    [-1, 1, Conv, [256, 1, 1]],
    36.    [-1, 1, nn.Upsample, [None, 2, 'nearest']],
    37.    [[-1, 4], 1, Concat, [1]],  # cat backbone P3
    38.    [-1, 3, C3, [256, False]],  # 17 (P3/8-small)
    39.  
    40.    [-1, 1, Conv, [256, 3, 2]],
    41.    [[-1, 14], 1, Concat, [1]],  # cat head P4
    42.    [-1, 3, C3, [512, False]],  # 20 (P4/16-medium)
    43.  
    44.    [-1, 1, Conv, [512, 3, 2]],
    45.    [[-1, 10], 1, Concat, [1]],  # cat head P5
    46.    [-1, 3, C3, [1024, False]],  # 23 (P5/32-large)
    47.  
    48.    [[17, 20, 23], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)
    49.   ]

    修改完成

  • 相关阅读:
    如何实现一个状态机？
    Kotlin调用Java代码时引起空指针异常，要怎么避免？
    百趣土壤非靶标代谢组学文献分享，来自Microbiome的灵感
    [手写spring](5)实现AOP机制（完结）
    python 综合练习
    智慧校园管理在疫情防控中的作用有哪些？
    【XTDrone Ubuntu20.04】XTDrone+ Ubuntu20.04 + PX4安装
    Linux /etc/passwd和/etc/shadow
    代季峰对话张祥雨 | 自动驾驶感知新时代！新一代环视感知算法BEVFormer有哪些优势...
    数据中心的防雷接地
  • 原文地址：https://blog.csdn.net/weixin_45303602/article/details/133215639