1.介绍

论文地址：https://arxiv.org/abs/2111.09883

综述

该论文作者提出了缩放 Swin Transformer 的技术 多达 30 亿个参数，使其能够使用多达 1,536 个图像进行训练1,536 分辨率。通过扩大容量和分辨率，Swin Transformer 在四个具有代表性的视觉基准上创造了新记录：ImageNet-V2 图像分类的84.0% top-1 准确率，COCO 对象检测的63.1 / 54.4 box / mask mAP，ADE20K 语义分割的59.9 mIoU，和86.8%Kinetics-400 视频动作分类的前 1 准确率。我们的技术通常适用于扩大视觉模型，但尚未像 NLP 语言模型那样被广泛探索，部分原因是在训练和应用方面存在以下困难：1）视觉模型经常面临大规模的不稳定性问题和 2）许多下游视觉任务需要高分辨率图像或窗口，目前尚不清楚如何有效地将低分辨率预训练的模型转移到更高分辨率的模型。当图像分辨率很高时，GPU 内存消耗也是一个问题。为了解决这些问题，我们提出了几种技术，并通过使用 Swin Transformer 作为案例研究来说明：1）后归一化技术和缩放余弦注意方法，以提高大型视觉模型的稳定性；2) 一种对数间隔的连续位置偏差技术，可有效地将在低分辨率图像和窗口上预训练的模型转移到其更高分辨率的对应物上。此外，我们分享了我们的关键实现细节，这些细节可以显着节省 GPU 内存消耗，从而使使用常规 GPU 训练大型视觉模型变得可行。使用这些技术和自我监督的预训练，我们成功训练了一个强大的 30 亿个 Swin Transformer 模型，并有效地将其转移到涉及高分辨率图像或窗口的各种视觉任务中，在各种的基准。代码将在 我们分享了我们的关键实现细节，这些细节可以显着节省 GPU 内存消耗，从而使使用常规 GPU 训练大型视觉模型变得可行。使用这些技术和自我监督的预训练，我们成功训练了一个强大的 30 亿个 Swin Transformer 模型，并有效地将其转移到涉及高分辨率图像或窗口的各种视觉任务中，在各种的基准。代码将在 我们分享了我们的关键实现细节，这些细节可以显着节省 GPU 内存消耗，从而使使用常规 GPU 训练大型视觉模型变得可行。使用这些技术和自我监督的预训练，我们成功训练了一个强大的 30 亿个 Swin Transformer 模型，并有效地将其转移到涉及高分辨率图像或窗口的各种视觉任务中，在各种的基准。代码将在 我们成功训练了一个强大的 30 亿个 Swin Transformer 模型，并将其有效地转移到涉及高分辨率图像或窗口的各种视觉任务中，在各种基准测试中达到了最先进的精度。代码将在 我们成功训练了一个强大的 30 亿个 Swin Transformer 模型，并将其有效地转移到涉及高分辨率图像或窗口的各种视觉任务中，在各种基准测试中达到了最先进的精度。

要解决的问题
视觉模型通常面临尺度不稳定问题；

下游任务需要高分辨率图像，尚不明确如何将低分辨率预训练模型迁移为高分辨率版本 ；当图像分辨率非常大时，GPU显存占用也是个问题。

改进方案
提出后规范化(Post Normalization)技术与可缩放(Scaled)cosine注意力提升大视觉模型的稳定性;
提出log空间连续位置偏置技术进行低分辨率预训练模型向高分辨率模型迁移；
我们还共享了至关重要的实现细节 ，它可以大幅节省GPU显存占用以使得大视觉模型训练变得可行。

2.  YOLOv5改进方法

2.1 YOLOv5的yaml配置文件

首先增加以下yolov5_swin_transfomrer.yaml文件

# YOLOv5 🚀 by Ultralytics, GPL-3.0 license# Parameters
nc: 80  # number of classes
depth_multiple: 0.33  # model depth multiple
width_multiple: 0.50  # layer channel multiple
anchors:- [10,13, 16,30, 33,23]  # P3/8- [30,61, 62,45, 59,119]  # P4/16- [116,90, 156,198, 373,326]  # P5/32# YOLOv5 v6.0 backbone by yoloair
backbone:# [from, number, module, args][[-1, 1, Conv, [64, 6, 2, 2]],  # 0-P1/2[-1, 1, Conv, [128, 3, 2]],  # 1-P2/4[-1, 3, C3, [128]],[-1, 1, Conv, [256, 3, 2]],  # 3-P3/8[-1, 6, C3, [256]],[-1, 1, Conv, [512, 3, 2]],  # 5-P4/16[-1, 9, SwinV2_CSPB, [256, 256]],[-1, 1, Conv, [1024, 3, 2]],  # 7-P5/32[-1, 3, SwinV2_CSPB, [512, 512]],  # 9 <--- ST2CSPB() Transformer module[-1, 1, SPPF, [1024, 5]],  # 9]# YOLOv5 v6.0 head
head:[[-1, 1, Conv, [512, 1, 1]],[-1, 1, nn.Upsample, [None, 2, 'nearest']],[[-1, 6], 1, Concat, [1]],  # cat backbone P4[-1, 3, C3, [512, False]],  # 13[-1, 1, Conv, [256, 1, 1]],[-1, 1, nn.Upsample, [None, 2, 'nearest']],[[-1, 4], 1, Concat, [1]],  # cat backbone P3[-1, 3, C3, [256, False]],  # 17 (P3/8-small)[-1, 1, Conv, [256, 3, 2]],[[-1, 14], 1, Concat, [1]],  # cat head P4[-1, 3, C3, [512, False]],  # 20 (P4/16-medium)[-1, 1, Conv, [512, 3, 2]],[[-1, 10], 1, Concat, [1]],  # cat head P5[-1, 3, C3, [1024, False]],  # 23 (P5/32-large)[[17, 20, 23], 1, Detect, [nc, anchors]],  # Detect(P3, P4, P5)]

2.2 common.py配置

在./models/common.py文件中增加以下模块，直接复制即可

class WindowAttention_v2(nn.Module):def __init__(self, dim, window_size, num_heads, qkv_bias=True, attn_drop=0., proj_drop=0.,pretrained_window_size=[0, 0]):super().__init__()self.dim = dimself.window_size = window_size  # Wh, Wwself.pretrained_window_size = pretrained_window_sizeself.num_heads = num_headsself.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True)# mlp to generate continuous relative position biasself.cpb_mlp = nn.Sequential(nn.Linear(2, 512, bias=True),nn.ReLU(inplace=True),nn.Linear(512, num_heads, bias=False))# get relative_coords_tablerelative_coords_h = torch.arange(-(self.window_size[0] - 1), self.window_size[0], dtype=torch.float32)relative_coords_w = torch.arange(-(self.window_size[1] - 1), self.window_size[1], dtype=torch.float32)relative_coords_table = torch.stack(torch.meshgrid([relative_coords_h,relative_coords_w])).permute(1, 2, 0).contiguous().unsqueeze(0)  # 1, 2*Wh-1, 2*Ww-1, 2if pretrained_window_size[0] > 0:relative_coords_table[:, :, :, 0] /= (pretrained_window_size[0] - 1)relative_coords_table[:, :, :, 1] /= (pretrained_window_size[1] - 1)else:relative_coords_table[:, :, :, 0] /= (self.window_size[0] - 1)relative_coords_table[:, :, :, 1] /= (self.window_size[1] - 1)relative_coords_table *= 8  # normalize to -8, 8relative_coords_table = torch.sign(relative_coords_table) * torch.log2(torch.abs(relative_coords_table) + 1.0) / np.log2(8)self.register_buffer("relative_coords_table", relative_coords_table)# get pair-wise relative position index for each token inside the windowcoords_h = torch.arange(self.window_size[0])coords_w = torch.arange(self.window_size[1])coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Wwcoords_flatten = torch.flatten(coords, 1)  # 2, Wh*Wwrelative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Wwrelative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0relative_coords[:, :, 1] += self.window_size[1] - 1relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Wwself.register_buffer("relative_position_index", relative_position_index)self.qkv = nn.Linear(dim, dim * 3, bias=False)if qkv_bias:self.q_bias = nn.Parameter(torch.zeros(dim))self.v_bias = nn.Parameter(torch.zeros(dim))else:self.q_bias = Noneself.v_bias = Noneself.attn_drop = nn.Dropout(attn_drop)self.proj = nn.Linear(dim, dim)self.proj_drop = nn.Dropout(proj_drop)self.softmax = nn.Softmax(dim=-1)def forward(self, x, mask=None):B_, N, C = x.shapeqkv_bias = Noneif self.q_bias is not None:qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)# cosine attentionattn = (F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1))logit_scale = torch.clamp(self.logit_scale, max=torch.log(torch.tensor(1. / 0.01))).exp()attn = attn * logit_scalerelative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view(-1, self.num_heads)relative_position_bias = relative_position_bias_table[self.relative_position_index.view(-1)].view(self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nHrelative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Wwrelative_position_bias = 16 * torch.sigmoid(relative_position_bias)attn = attn + relative_position_bias.unsqueeze(0)if mask is not None:nW = mask.shape[0]attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)attn = attn.view(-1, self.num_heads, N, N)attn = self.softmax(attn)else:attn = self.softmax(attn)attn = self.attn_drop(attn)try:x = (attn @ v).transpose(1, 2).reshape(B_, N, C)except:x = (attn.half() @ v).transpose(1, 2).reshape(B_, N, C)x = self.proj(x)x = self.proj_drop(x)return xdef extra_repr(self) -> str:return f'dim={self.dim}, window_size={self.window_size}, ' \f'pretrained_window_size={self.pretrained_window_size}, num_heads={self.num_heads}'def flops(self, N):# calculate flops for 1 window with token length of Nflops = 0# qkv = self.qkv(x)flops += N * self.dim * 3 * self.dim# attn = (q @ k.transpose(-2, -1))flops += self.num_heads * N * (self.dim // self.num_heads) * N#  x = (attn @ v)flops += self.num_heads * N * N * (self.dim // self.num_heads)# x = self.proj(x)flops += N * self.dim * self.dimreturn flopsclass Mlp_v2(nn.Module):def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0.):super().__init__()out_features = out_features or in_featureshidden_features = hidden_features or in_featuresself.fc1 = nn.Linear(in_features, hidden_features)self.act = act_layer()self.fc2 = nn.Linear(hidden_features, out_features)self.drop = nn.Dropout(drop)def forward(self, x):x = self.fc1(x)x = self.act(x)x = self.drop(x)x = self.fc2(x)x = self.drop(x)return x
# add 2 functions
class SwinTransformerLayer_v2(nn.Module):def __init__(self, dim, num_heads, window_size=7, shift_size=0,mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,act_layer=nn.SiLU, norm_layer=nn.LayerNorm, pretrained_window_size=0):super().__init__()self.dim = dim#self.input_resolution = input_resolutionself.num_heads = num_headsself.window_size = window_sizeself.shift_size = shift_sizeself.mlp_ratio = mlp_ratio#if min(self.input_resolution) <= self.window_size:#    # if window size is larger than input resolution, we don't partition windows#    self.shift_size = 0#    self.window_size = min(self.input_resolution)assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"self.norm1 = norm_layer(dim)self.attn = WindowAttention_v2(dim, window_size=(self.window_size, self.window_size), num_heads=num_heads,qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,pretrained_window_size=(pretrained_window_size, pretrained_window_size))self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()self.norm2 = norm_layer(dim)mlp_hidden_dim = int(dim * mlp_ratio)self.mlp = Mlp_v2(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)def create_mask(self, H, W):# calculate attention mask for SW-MSAimg_mask = torch.zeros((1, H, W, 1))  # 1 H W 1h_slices = (slice(0, -self.window_size),slice(-self.window_size, -self.shift_size),slice(-self.shift_size, None))w_slices = (slice(0, -self.window_size),slice(-self.window_size, -self.shift_size),slice(-self.shift_size, None))cnt = 0for h in h_slices:for w in w_slices:img_mask[:, h, w, :] = cntcnt += 1mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1mask_windows = mask_windows.view(-1, self.window_size * self.window_size)attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))return attn_maskdef forward(self, x):# reshape x[b c h w] to x[b l c]_, _, H_, W_ = x.shapePadding = Falseif min(H_, W_) < self.window_size or H_ % self.window_size!=0 or W_ % self.window_size!=0:Padding = True# print(f'img_size {min(H_, W_)} is less than (or not divided by) window_size {self.window_size}, Padding.')pad_r = (self.window_size - W_ % self.window_size) % self.window_sizepad_b = (self.window_size - H_ % self.window_size) % self.window_sizex = F.pad(x, (0, pad_r, 0, pad_b))# print('2', x.shape)B, C, H, W = x.shapeL = H * Wx = x.permute(0, 2, 3, 1).contiguous().view(B, L, C)  # b, L, c# create mask from init to forwardif self.shift_size > 0:attn_mask = self.create_mask(H, W).to(x.device)else:attn_mask = Noneshortcut = xx = x.view(B, H, W, C)# cyclic shiftif self.shift_size > 0:shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))else:shifted_x = x# partition windowsx_windows = window_partition_v2(shifted_x, self.window_size)  # nW*B, window_size, window_size, Cx_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C# W-MSA/SW-MSAattn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C# merge windowsattn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)shifted_x = window_reverse_v2(attn_windows, self.window_size, H, W)  # B H' W' C# reverse cyclic shiftif self.shift_size > 0:x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))else:x = shifted_xx = x.view(B, H * W, C)x = shortcut + self.drop_path(self.norm1(x))# FFNx = x + self.drop_path(self.norm2(self.mlp(x)))x = x.permute(0, 2, 1).contiguous().view(-1, C, H, W)  # b c h wif Padding:x = x[:, :, :H_, :W_]  # reverse paddingreturn xdef extra_repr(self) -> str:return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"def flops(self):flops = 0H, W = self.input_resolution# norm1flops += self.dim * H * W# W-MSA/SW-MSAnW = H * W / self.window_size / self.window_sizeflops += nW * self.attn.flops(self.window_size * self.window_size)# mlpflops += 2 * H * W * self.dim * self.dim * self.mlp_ratio# norm2flops += self.dim * H * Wreturn flopsclass SwinTransformer2Block(nn.Module):def __init__(self, c1, c2, num_heads, num_layers, window_size=7):super().__init__()self.conv = Noneif c1 != c2:self.conv = Conv(c1, c2)# remove input_resolutionself.blocks = nn.Sequential(*[SwinTransformerLayer_v2(dim=c2, num_heads=num_heads, window_size=window_size,shift_size=0 if (i % 2 == 0) else window_size // 2) for i in range(num_layers)])def forward(self, x):if self.conv is not None:x = self.conv(x)x = self.blocks(x)return xclass SwinV2_CSPB(nn.Module):# CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworksdef __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansionsuper(SwinV2_CSPB, self).__init__()c_ = int(c2)  # hidden channelsself.cv1 = Conv(c1, c_, 1, 1)self.cv2 = Conv(c_, c_, 1, 1)self.cv3 = Conv(2 * c_, c2, 1, 1)num_heads = c_ // 32self.m = SwinTransformer2Block(c_, c_, num_heads, n)#self.m = nn.Sequential(*[Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)])def forward(self, x):x1 = self.cv1(x)y1 = self.m(x1)y2 = self.cv2(x1)return self.cv3(torch.cat((y1, y2), dim=1))

2.3 yolo.py配置

不需要

修改完成