本文主要针对目标检测部分的代码。
开始之前，先上一下swin transformer 结构图
首先从模型训练开始，训练模型py文件位于项目根目录/tools/train.py，该文件中整体结构简单，仅有一个main函数。为了方便程序运行，我直接在配置项中将config配置成
\configs\swin\mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_1x_coco.py
main函数中主要关注161行后的代码，以下是代码片段
model = build_detector(
cfg.model,
train_cfg=cfg.get('train_cfg'),
test_cfg=cfg.get('test_cfg'))
datasets = [build_dataset(cfg.data.train)]
if len(cfg.workflow) == 2:
val_dataset = copy.deepcopy(cfg.data.val)
val_dataset.pipeline = cfg.data.train.pipeline
datasets.append(build_dataset(val_dataset))
if cfg.checkpoint_config is not None:
# save mmdet version, config file content and class names in
# checkpoints as meta data
cfg.checkpoint_config.meta = dict(
mmdet_version=__version__ + get_git_hash()[:7],
CLASSES=datasets[0].CLASSES)
# add an attribute for visualization convenience
model.CLASSES = datasets[0].CLASSES
train_detector(
model,
datasets,
cfg,
distributed=distributed,
validate=(not args.no_validate),
timestamp=timestamp,
meta=meta)这部分主要是构建模型，构建数据集，模型训练函数
这里用的Mask RCNN结构，构建模型的时候，会分别构建如下文件夹中的相应组件：
mmdet/models/detectors/mask_rcnn.py 中的Mask RCNN类
mmdet/models/detectors/two_stage.py 中的TwoStageDetector类
mmdet/models/backbones/swin_transformer.py 中的SwinTransformer类（算法关键）
mmdet/models/necks/fpn.py 中的FPN类
mmdet/models/dense_heads/rpn_head.py 中的RPNHead类，其中还会构建各种损失函数和一些功能组件
mmdet/models/roi_heads/base_roi_head.py 中的BaseRoIHead 类
mmdet/models/roi_heads/bbox_heads/convfc_bbox_head.py 中的ConvFCBBoxHead类
mmdet/models/losses/cross_entropy_loss.py 中的CrossEntropyLoss类
mmdet/models/losses/smooth_l1_loss.py 中的L1Loss类
mmdet/core/bbox/assigners/max_iou_assigner.py 中的MaxIoUAssigner类
mmdet/core/bbox/samplers/random_sampler.py中的RandomSampler类
构建完的模型：
MaskRCNN(
(backbone): SwinTransformer(
(patch_embed): PatchEmbed(
(proj): Conv2d(3, 96, kernel_size=(4, 4), stride=(4, 4))
(norm): LayerNorm((96,), eps=1e-05, elementwise_affine=True)
)
(pos_drop): Dropout(p=0.0, inplace=False)
(layers): ModuleList(
(0): BasicLayer(
(blocks): ModuleList(
(0): SwinTransformerBlock(
(norm1): LayerNorm((96,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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(
(norm1): LayerNorm((96,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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(
(reduction): Linear(in_features=384, out_features=192, bias=False)
(norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
)
)
(1): BasicLayer(
(blocks): ModuleList(
(0): SwinTransformerBlock(
(norm1): LayerNorm((192,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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(
(norm1): LayerNorm((192,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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(
(reduction): Linear(in_features=768, out_features=384, bias=False)
(norm): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
)
)
(2): BasicLayer(
(blocks): ModuleList(
(0): SwinTransformerBlock(
(norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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): SwinTransformerBlock(
(norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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)
)
)
(2): SwinTransformerBlock(
(norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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)
)
)
(3): SwinTransformerBlock(
(norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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)
)
)
(4): SwinTransformerBlock(
(norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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)
)
)
(5): SwinTransformerBlock(
(norm1): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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)
)
)
)
(downsample): PatchMerging(
(reduction): Linear(in_features=1536, out_features=768, bias=False)
(norm): LayerNorm((1536,), eps=1e-05, elementwise_affine=True)
)
)
(3): BasicLayer(
(blocks): ModuleList(
(0): SwinTransformerBlock(
(norm1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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(
(norm1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
(attn): WindowAttention(
(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()
(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)
)
)
)
)
)
(norm0): LayerNorm((96,), eps=1e-05, elementwise_affine=True)
(norm1): LayerNorm((192,), eps=1e-05, elementwise_affine=True)
(norm2): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
(norm3): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
)
(neck): FPN(
(lateral_convs): ModuleList(
(0): ConvModule(
(conv): Conv2d(96, 256, kernel_size=(1, 1), stride=(1, 1))
)
(1): ConvModule(
(conv): Conv2d(192, 256, kernel_size=(1, 1), stride=(1, 1))
)
(2): ConvModule(
(conv): Conv2d(384, 256, kernel_size=(1, 1), stride=(1, 1))
)
(3): ConvModule(
(conv): Conv2d(768, 256, kernel_size=(1, 1), stride=(1, 1))
)
)
(fpn_convs): ModuleList(
(0): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
)
(1): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
)
(2): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
)
(3): ConvModule(
(conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
)
)
)
(rpn_head): RPNHead(
(loss_cls): CrossEntropyLoss()
(loss_bbox): L1Loss()
(rpn_conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
(rpn_cls): Conv2d(256, 3, kernel_size=(1, 1), stride=(1, 1))
(rpn_reg): Conv2d(256, 12, kernel_size=(1, 1), stride=(1, 1))
)
(roi_head): StandardRoIHead(
(bbox_roi_extractor): SingleRoIExtractor(
(roi_layers): ModuleList(
(0): RoIAlign(output_size=(7, 7), spatial_scale=0.25, sampling_ratio=0, pool_mode=avg, aligned=True, use_torchvision=False)
(1): RoIAlign(output_size=(7, 7), spatial_scale=0.125, sampling_ratio=0, pool_mode=avg, aligned=True, use_torchvision=False)
(2): RoIAlign(output_size=(7, 7), spatial_scale=0.0625, sampling_ratio=0, pool_mode=avg, aligned=True, use_torchvision=False)
(3): RoIAlign(output_size=(7, 7), spatial_scale=0.03125, sampling_ratio=0, pool_mode=avg, aligned=True, use_torchvision=False)
)
)
(bbox_head): Shared2FCBBoxHead(
(loss_cls): CrossEntropyLoss()
(loss_bbox): L1Loss()
(fc_cls): Linear(in_features=1024, out_features=21, bias=True)
(fc_reg): Linear(in_features=1024, out_features=80, bias=True)
(shared_convs): ModuleList()
(shared_fcs): ModuleList(
(0): Linear(in_features=12544, out_features=1024, bias=True)
(1): Linear(in_features=1024, out_features=1024, bias=True)
)
(cls_convs): ModuleList()
(cls_fcs): ModuleList()
(reg_convs): ModuleList()
(reg_fcs): ModuleList()
(relu): ReLU(inplace=True)
)
)
)backbone部分的就是swin transformer的精髓，下图是mask rcnn的结构图
Swin Transformer就是用transformer块替换了图中CNN的结构，作为特征采集器。扯了这么多，是时候进入正题了，代码的关键算法都位于mmdet/models/backbones/swin_transformer.py文件中。主体位于SwinTransformer类中。
class SwinTransformer(nn.Module):
""" Swin Transformer backbone.
A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` -
https://arxiv.org/pdf/2103.14030
Args:
pretrain_img_size (int): Input image size for training the pretrained model,
used in absolute postion embedding. Default 224.
patch_size (int | tuple(int)): Patch size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
depths (tuple[int]): Depths of each Swin Transformer stage.
num_heads (tuple[int]): Number of attention head of each stage.
window_size (int): Window size. Default: 7.
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
drop_rate (float): Dropout rate.
attn_drop_rate (float): Attention dropout rate. Default: 0.
drop_path_rate (float): Stochastic depth rate. Default: 0.2.
norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
patch_norm (bool): If True, add normalization after patch embedding. Default: True.
out_indices (Sequence[int]): Output from which stages.
frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
-1 means not freezing any parameters.
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
"""
def __init__(self,
pretrain_img_size=224,
patch_size=4,
in_chans=3,
embed_dim=96,
depths=[2, 2, 6, 2],
num_heads=[3, 6, 12, 24],
window_size=7,
mlp_ratio=4.,
qkv_bias=True,
qk_scale=None,
drop_rate=0.,
attn_drop_rate=0.,
drop_path_rate=0.2,
norm_layer=nn.LayerNorm,
ape=False,
patch_norm=True,
out_indices=(0, 1, 2, 3),
frozen_stages=-1,
use_checkpoint=False):
super().__init__()
self.pretrain_img_size = pretrain_img_size
self.num_layers = len(depths)
self.embed_dim = embed_dim
self.ape = ape
self.patch_norm = patch_norm
self.out_indices = out_indices
self.frozen_stages = frozen_stages
# split image into non-overlapping patches
self.patch_embed = PatchEmbed(
patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
# absolute position embedding
if self.ape:
pretrain_img_size = to_2tuple(pretrain_img_size)
patch_size = to_2tuple(patch_size)
patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]
# 对应网络结构中的 linear embedding 网络结构
self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
# 绝对位置编码参数初始化
trunc_normal_(self.absolute_pos_embed, std=.02)
self.pos_drop = nn.Dropout(p=drop_rate)
# stochastic depth
# 给网络层数每层设置随机dropout rate
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule
# build layers
self.layers = nn.ModuleList()
# 构建四层网络结构
# mlp_ratio Ratio of mlp hidden dim to embedding dim.
# downsample 下采样 前三个block 会进行下采样 第四个block 不会在进行下采样
for i_layer in range(self.num_layers):
layer = BasicLayer(
dim=int(embed_dim * 2 ** i_layer),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
norm_layer=norm_layer,
downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
use_checkpoint=use_checkpoint)
self.layers.append(layer)
num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
self.num_features = num_features
# add a norm layer for each output
for i_layer in out_indices:
layer = norm_layer(num_features[i_layer])
layer_name = f'norm{i_layer}'
self.add_module(layer_name, layer)
self._freeze_stages()
def _freeze_stages(self):
if self.frozen_stages >= 0:
self.patch_embed.eval()
for param in self.patch_embed.parameters():
param.requires_grad = False
if self.frozen_stages >= 1 and self.ape:
self.absolute_pos_embed.requires_grad = False
if self.frozen_stages >= 2:
self.pos_drop.eval()
for i in range(0, self.frozen_stages - 1):
m = self.layers[i]
m.eval()
for param in m.parameters():
param.requires_grad = False
def init_weights(self, pretrained=None):
"""Initialize the weights in backbone.
Args:
pretrained (str, optional): Path to pre-trained weights.
Defaults to None.
"""
def _init_weights(m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
if isinstance(pretrained, str):
self.apply(_init_weights)
logger = get_root_logger()
load_checkpoint(self, pretrained, strict=False, logger=logger)
elif pretrained is None:
self.apply(_init_weights)
else:
raise TypeError('pretrained must be a str or None')
def forward(self, x):
"""Forward function."""
x = self.patch_embed(x)
Wh, Ww = x.size(2), x.size(3)
if self.ape:
# interpolate the position embedding to the corresponding size
absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
x = (x + absolute_pos_embed).flatten(2).transpose(1, 2) # B Wh*Ww C
else:
x = x.flatten(2).transpose(1, 2)
x = self.pos_drop(x)
outs = []
for i in range(self.num_layers):
layer = self.layers[i]
x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
if i in self.out_indices:
norm_layer = getattr(self, f'norm{i}')
x_out = norm_layer(x_out)
out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
outs.append(out)
return tuple(outs)
def train(self, mode=True):
"""Convert the model into training mode while keep layers freezed."""
super(SwinTransformer, self).train(mode)
self._freeze_stages()
为了方便将img_scale设置为[(224,224)]，此时的数据集中的图片会进行resize，并且将短边padding成32的倍数，按最长边与224的比例缩放最短边，即 x=短边 / (500/224)，x即为缩放后的实际长度，比如输入时500*287（w*h）的图片，缩放后的短边长为287 * 224 / 500 = 129，由于输入的图像尺寸需要是32的倍数，此时需要向上取整，所以padding后的短边长为160。
吃了贫穷的亏，电脑只有单卡3080ti，batch size设置为2，使用未修改的img_scale训练coco数据集时还经常贴着最大显存跑，生怕训练的时候爆显存。。。所以单卡训练batch size就设置成2吧，每个batch中的数据会根据边长最大的图片再进行一次padding，比如其中一张第一次padding后大小为(160, 224, 3)，另一张为(192, 224, 3)，那么最终输入网络的数据尺寸为[2, 3, 192, 224] (B,C,H,W)
之后将会以输入[2, 3, 192, 224]为例进行讲解，按网络结构的顺序来。
PatchEmbed 通过一个size为4，stride为4的2d卷积来达到图像缩小4倍，并将维度升到embed_dim，PatchEmbed可以简单理解为包含了网络结构中的patch partition和linear embedding，所以embed_dim为输入transformer block的维度而不是论文中的48。
class PatchEmbed(nn.Module):
""" Image to Patch Embedding
Args:
patch_size (int): Patch token size. Default: 4.
in_chans (int): Number of input image channels. Default: 3.
embed_dim (int): Number of linear projection output channels. Default: 96.
norm_layer (nn.Module, optional): Normalization layer. Default: None
"""
def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
super().__init__()
patch_size = to_2tuple(patch_size)
self.patch_size = patch_size
self.in_chans = in_chans
self.embed_dim = embed_dim
# 用2d卷积实现图像缩小四倍 kernel_size = stride = patch_size = 4
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
if norm_layer is not None:
self.norm = norm_layer(embed_dim)
else:
self.norm = None
def forward(self, x):
"""Forward function."""
# padding
_, _, H, W = x.size()
if W % self.patch_size[1] != 0:
x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
if H % self.patch_size[0] != 0:
x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
x = self.proj(x) # B C Wh Ww
if self.norm is not None:
# x的shape变化 B, C, H, W --> B, C, h, w --> B, C, h * w --> B, h * w, c
# x shape 为 [2, 96, 48, 56]
Wh, Ww = x.size(2), x.size(3)
# x shape 为 [2, 2688, 96]
x = x.flatten(2).transpose(1, 2)
x = self.norm(x)
# x shape 为 [2, 96, 48, 56]
x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
return x经过PatchEmbed后的输出维度为[2, 96, 48, 56]，之后会再经过一个x.flatten(2).transpose(1, 2)将输入维度转换为transformer block能够接收的输入，即[2, 2688, 96]。
接下来是BasicLayer
class BasicLayer(nn.Module):
""" A basic Swin Transformer layer for one stage.
Args:
dim (int): Number of feature channels
depth (int): Depths of this stage.
num_heads (int): Number of attention head.
window_size (int): Local window size. Default: 7.
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
"""
def __init__(self,
dim,
depth,
num_heads,
window_size=7,
mlp_ratio=4.,
qkv_bias=True,
qk_scale=None,
drop=0.,
attn_drop=0.,
drop_path=0.,
norm_layer=nn.LayerNorm,
downsample=None,
use_checkpoint=False):
super().__init__()
self.window_size = window_size
self.shift_size = window_size // 2
self.depth = depth
self.use_checkpoint = use_checkpoint
# build blocks
# 序号为偶数的block进行W-MSA，奇数进行SW-MSA
# 这样让输出特征包含local window attention和跨窗口的 window attention
self.blocks = nn.ModuleList([
SwinTransformerBlock(
dim=dim,
num_heads=num_heads,
window_size=window_size,
shift_size=0 if (i % 2 == 0) else window_size // 2,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop,
attn_drop=attn_drop,
drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
norm_layer=norm_layer)
for i in range(depth)])
# patch merging layer
# 只有前三个block执行patch merging，最后一个block不会执行 patch merging
if downsample is not None:
self.downsample = downsample(dim=dim, norm_layer=norm_layer)
else:
self.downsample = None
def forward(self, x, H, W):
""" Forward function.
Args:
x: Input feature, tensor size (B, H*W, C).
H, W: Spatial resolution of the input feature.
"""
# calculate attention mask for SW-MSA
Hp = int(np.ceil(H / self.window_size)) * self.window_size
Wp = int(np.ceil(W / self.window_size)) * self.window_size
img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device) # 1 Hp Wp 1
h_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 = 0
for h in h_slices:
for w in w_slices:
img_mask[:, h, w, :] = cnt
cnt += 1
mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1
mask_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))
for blk in self.blocks:
blk.H, blk.W = H, W
if self.use_checkpoint:
x = checkpoint.checkpoint(blk, x, attn_mask)
else:
x = blk(x, attn_mask)
if self.downsample is not None:
x_down = self.downsample(x, H, W)
Wh, Ww = (H + 1) // 2, (W + 1) // 2
return x, H, W, x_down, Wh, Ww
else:
return x, H, W, x, H, WBasicLayer构建了一个stage的swin transformer基本结构，包含了带窗（SW-MSA）和不带窗（W-MSA）的transformer block以及一个PatchMerging，可以理解为网络结构图中的swin transformer block + patch merging。
在训练的过程中序号为偶数的block进行W-MSA，奇数进行SW-MSA，比如说第一个transformer block 有一个W-MSA和一个SW-MSA，先计算W-MSA，再计算SW-MSA，这样让输出特征包含local window attention和跨窗口的 window attention，而patch merging layer 仅在前三个block执行。接下来把这部分代码拆分出来解读。
def window_partition(x, window_size):
"""
Args:
x: (B, H, W, C)
window_size (int): window size
Returns:
windows: (num_windows*B, window_size, window_size, C)
"""
# B, H, W, C = x.shape
# x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
# windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
H, W = x.shape
x = x.view(H // window_size, window_size, W // window_size, window_size)
windows = x.permute(0, 2, 1, 3).contiguous().view(-1,window_size, window_size)
return windows
Hp = 14
Wp = 14
window_size = 7
shift_size = window_size //2
# img_mask = torch.zeros((1, Hp, Wp, 1)) # 1 Hp Wp 1
img_mask = torch.zeros((Hp, Wp))
h_slices = (slice(0, -window_size),
slice(-window_size, -shift_size),
slice(-shift_size, None))
w_slices = (slice(0, -window_size),
slice(-window_size, -shift_size),
slice(-shift_size, None))
cnt = 0
for h in h_slices:
for w in w_slices:
# img_mask[:, h, w, :] = cnt
img_mask[h, w] = cnt
cnt += 1
print(img_mask)
# tensor([[0., 0., 0., 0., 0., 0., 0.,| 1., 1., 1., 1., 2., 2., 2.],
# [0., 0., 0., 0., 0., 0., 0.,| 1., 1., 1., 1., 2., 2., 2.],
# [0., 0., 0., 0., 0., 0., 0.,| 1., 1., 1., 1., 2., 2., 2.],
# [0., 0., 0., 0., 0., 0., 0.,| 1., 1., 1., 1., 2., 2., 2.],
# [0., 0., 0., 0., 0., 0., 0.,| 1., 1., 1., 1., 2., 2., 2.],
# [0., 0., 0., 0., 0., 0., 0.,| 1., 1., 1., 1., 2., 2., 2.],
# [0., 0., 0., 0., 0., 0., 0.,| 1., 1., 1., 1., 2., 2., 2.],
# - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
# [3., 3., 3., 3., 3., 3., 3.,| 4., 4., 4., 4., 5., 5., 5.],
# [3., 3., 3., 3., 3., 3., 3.,| 4., 4., 4., 4., 5., 5., 5.],
# [3., 3., 3., 3., 3., 3., 3.,| 4., 4., 4., 4., 5., 5., 5.],
# [3., 3., 3., 3., 3., 3., 3.,| 4., 4., 4., 4., 5., 5., 5.],
# [6., 6., 6., 6., 6., 6., 6.,| 7., 7., 7., 7., 8., 8., 8.],
# [6., 6., 6., 6., 6., 6., 6.,| 7., 7., 7., 7., 8., 8., 8.],
# [6., 6., 6., 6., 6., 6., 6.,| 7., 7., 7., 7., 8., 8., 8.]])
mask_windows = window_partition(img_mask, window_size) # nW, window_size, window_size, 1
mask_windows = mask_windows.view(-1, window_size * 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)) 这部分代码是用于生成mask的，源码中的维度比较高，不是很直观，这里将其降为后输出成低维张量，就很直观了，张量的高宽设置为window_size的两倍，这样正好能够完整的看出其中张量做了什么操作，shift_size保持不变，打印出的img_mask可以看出，mask大体上分为四个部分。
window_partition函数则是将img_mask按照每个部分展开，即将张量分成N个[window_size，window_size]的小窗张量，此时的张量shape为[4,7,7]。之后view成[4，49]，再在扩充对应的维度，再相减，张量中不为0的填充为-100，最后得到的attn_mask的shape为[4,49,49]，这里的尺寸就和后面自注意力中的大小对应上了。
具体形式如下图
其中：
# 白色色块 # [0., 0., 0., 0., 0., 0., 0., # 0., 0., 0., 0., 0., 0., 0., # 0., 0., 0., 0., 0., 0., 0., # 0., 0., 0., 0., 0., 0., 0., # 0., 0., 0., 0., 0., 0., 0., # 0., 0., 0., 0., 0., 0., 0., # 0., 0., 0., 0., 0., 0., 0.],
# 蓝色色块 # [ 0., 0., 0., 0., -100., -100., -100., # 0., 0., 0., 0., -100., -100., -100., # 0., 0., 0., 0., -100., -100., -100., # 0., 0., 0., 0., -100., -100., -100., # -100., -100., -100., -100., 0., 0., 0., # -100., -100., -100., -100., 0., 0., 0., # -100., -100., -100., -100., 0., 0., 0.,],
# 黄色色块 # [-100., -100., -100., -100., -100., -100., -100., # -100., -100., -100., -100., -100., -100., -100., # -100., -100., -100., -100., -100., -100., -100., # -100., -100., -100., -100., -100., -100., -100., # -100., -100., -100., -100., -100., -100., -100., # -100., -100., -100., -100., -100., -100., -100., # -100., -100., -100., -100., -100., -100., -100.,]
实际上输出的attn_mask的shape为[56,49,49]，其中的56=(Hp / window_size) * (Wp / window_size)
之后就是transformer block
class SwinTransformerBlock(nn.Module):
""" Swin Transformer Block.
Args:
dim (int): Number of input channels.
num_heads (int): Number of attention heads.
window_size (int): Window size.
shift_size (int): Shift size for SW-MSA.
mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
drop (float, optional): Dropout rate. Default: 0.0
attn_drop (float, optional): Attention dropout rate. Default: 0.0
drop_path (float, optional): Stochastic depth rate. Default: 0.0
act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self, dim, num_heads, window_size=7, shift_size=0,
mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
act_layer=nn.GELU, norm_layer=nn.LayerNorm):
super().__init__()
self.dim = dim
self.num_heads = num_heads
# window_size默认大小 7
self.window_size = window_size
# 进行 SW-MSA shift-size 7//2=3
# 进行 W-MSA shift-size 0
self.shift_size = shift_size
# multi self attention 最后神经网络的隐藏层的维度的倍率
self.mlp_ratio = mlp_ratio
assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
self.norm1 = norm_layer(dim)
# local window multi head self attention
self.attn = WindowAttention(
dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
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(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
self.H = None
self.W = None
def forward(self, x, mask_matrix):
""" Forward function.
Args:
x: Input feature, tensor size (B, H*W, C).
H, W: Spatial resolution of the input feature.
mask_matrix: Attention mask for cyclic shift.
"""
# x的shape为[2,2688,96]
B, L, C = x.shape
H, W = self.H, self.W
assert L == H * W, "input feature has wrong size"
shortcut = x
# 进行LN，再将x展开为[2,48,56,96]
x = self.norm1(x)
x = x.view(B, H, W, C)
# pad feature maps to multiples of window size
# 此处需要根据窗的大小对特征图进行pad操作，pad之后的shape为[2,49,56,96]
pad_l = pad_t = 0
pad_r = (self.window_size - W % self.window_size) % self.window_size
pad_b = (self.window_size - H % self.window_size) % self.window_size
x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
_, Hp, Wp, _ = x.shape
# cyclic shift
if self.shift_size > 0:
# 如果是进行 sw-msa 将数据进行变换
shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
attn_mask = mask_matrix
else:
shifted_x = x
attn_mask = None
# partition windows
# x_windows 的shape为[112,7,7,96]
x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C
x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C
# W-MSA/SW-MSA
attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C
# merge windows
attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
# shifted_x的shape为[2,49,56,96]
shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C
# reverse cyclic shift
if self.shift_size > 0:
x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
else:
x = shifted_x
if pad_r > 0 or pad_b > 0:
# 映射回输入时的大小
x = x[:, :H, :W, :].contiguous()
x = x.view(B, H * W, C)
# FFN
x = shortcut + self.drop_path(x)
x = x + self.drop_path(self.mlp(self.norm2(x)))
return xSwinTransformerBlock由LN，W-MSA/SW-MSA以及MLP组成。在W-MSA中输入为[2,2688,96]，通过window_partition函数，将输入划分为window_size大小的窗，此时的shape为[112,7,7,96]，输入自注意力层时打平成[112,49,96]，其中自注意力层就是标准的自注意力结构，这里就不多说了，要是不知道的话可以参考我的另一篇博文：ref
计算自注意力的代码如下：
class WindowAttention(nn.Module):
""" Window based multi-head self attention (W-MSA) module with relative position bias.
It supports both of shifted and non-shifted window.
Args:
dim (int): Number of input channels.
window_size (tuple[int]): The height and width of the window.
num_heads (int): Number of attention heads.
qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
proj_drop (float, optional): Dropout ratio of output. Default: 0.0
"""
def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
super().__init__()
self.dim = dim
self.window_size = window_size # Wh, Ww
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim ** -0.5
# define a parameter table of relative position bias
self.relative_position_bias_table = nn.Parameter(
torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH
# get pair-wise relative position index for each token inside the window
coords_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, Ww
coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww
relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0
relative_coords[:, :, 1] += self.window_size[1] - 1
relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww
self.register_buffer("relative_position_index", relative_position_index)
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
trunc_normal_(self.relative_position_bias_table, std=.02)
self.softmax = nn.Softmax(dim=-1)
def forward(self, x, mask=None):
""" Forward function.
Args:
x: input features with shape of (num_windows*B, N, C)
mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
"""
B_, N, C = x.shape
qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)
q = q * self.scale
attn = (q @ k.transpose(-2, -1))
relative_position_bias = self.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,nH
relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww
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)
x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x这里有个细节就是在计算自注意力后，加了个截断正太分布的relative_position_bias
回归transformer block，算法为了使每个窗之间有相互联系，设计了shift window，这一设计能够大幅降低计算量。在计算SW-MSA时，先要使用torch.roll将输入[2,49,56,96]在1，2维上（也就是高宽上）做滚动。
下面有个小例子，还是将维度降低，方便看数据形式，容易理解
import torch
a = torch.arange(0, 49).reshape((7, 7))
print(a)
b = torch.roll(a, shifts=(-3, -3), dims=(0, 1))
print(b)
# tensor([[ 0, 1, 2, 3, 4, 5, 6],
# [ 7, 8, 9, 10, 11, 12, 13],
# [14, 15, 16, 17, 18, 19, 20],
# [21, 22, 23, 24, 25, 26, 27],
# [28, 29, 30, 31, 32, 33, 34],
# [35, 36, 37, 38, 39, 40, 41],
# [42, 43, 44, 45, 46, 47, 48]])
# tensor([[24, 25, 26, 27, |21, 22, 23],
# [31, 32, 33, 34, |28, 29, 30],
# [38, 39, 40, 41, |35, 36, 37],
# [45, 46, 47, 48, |42, 43, 44],
# — — - - - - - - - - - - -- - -
# [ 3, 4, 5, 6, | 0, 1, 2],
# [10, 11, 12, 13, | 7, 8, 9],
# [17, 18, 19, 20, |14, 15, 16]])也是就论文中的这张图中间部分：
进入自注意力层后，基本操作和W-MSA一致，就是多加了个之前生成的mask
最后就是PatchMerging
class PatchMerging(nn.Module):
""" Patch Merging Layer
Args:
dim (int): Number of input channels.
norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
"""
def __init__(self, dim, norm_layer=nn.LayerNorm):
super().__init__()
self.dim = dim
self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
self.norm = norm_layer(4 * dim)
def forward(self, x, H, W):
""" Forward function.
Args:
x: Input feature, tensor size (B, H*W, C).
H, W: Spatial resolution of the input feature.
"""
B, L, C = x.shape
assert L == H * W, "input feature has wrong size"
x = x.view(B, H, W, C)
# padding
pad_input = (H % 2 == 1) or (W % 2 == 1)
if pad_input:
x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
# 这里实现path merging 图片缩小一半
# 0::2 从 0 开始 隔一个点取一个值
# 1::2 从 1 开始 隔一个点取一个值
x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C
x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C
x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C
x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C
x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C
x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C
x = self.norm(x)
# 降维到2 * dim 图片缩小一半 通道维度增加一倍
x = self.reduction(x)
return xPatchMerging实现的功能有点像yolov5中的focus操作，将图像缩小一半，再将通道数增加一倍。
到这里基本将swin transformer的主体结构讲完了。
文章出处登录后可见！
已经登录？立即刷新
