- 前置知识： Transformer
- 作者：Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo
- 文章链接
- 代码链接
- 视频推荐
在计算机视觉中，CNN长期以来一直是主导模型。然而，在NLP领域中，Transformer架构已经成为主流。这篇文章核心想法就是：1.解决在计算机视觉中，物体的尺度不同这个问题。2.将Transformer应用在视觉领域。
Swin Transformer的架构如图1所示。可以看到，其实它的核心module是三个：Patch Partition，Patch Merging以及Swin Transformer Block。
图1：Architecture
输入的 RGB 图像首先被分割成固定大小的非重叠块，（例如4x4或16x16像素）。每个图像块被展平成一个一维向量，并且通过一个线性层映射到一个高维空间。这样输入就从$H\times W \times 3$变成了$\frac{H}{4} \times \frac{W}{4} \times 48$。
Patch Merging会将相邻的图像块合并成更大的块，例如将2x2的块看作一个大的像素块。通过合并操作，特征图的分辨率减半，同时特征的通道数翻倍。Merging后，特征会进行一个Linear层，将特征数减半，从而减少计算量。这也就是为什么每一个stage输出后，分辨率会减半，而通道数会加倍的原因。注意，swin transformer block不改变特征的shape。
swin transformer block中包括两个attention，分别是Window Multi-head Self Attention和Shifted Window Multi-head Self Attention
W-MSA是在固定窗口内计算自注意力。图像会被划分为多个不重叠的局部窗口，每个窗口内进行自注意力的计算。
SW-MSA是W-MSA的扩展，通过引入窗口的移位从而实现跨窗口的信息交流。这部分论文解释并不清楚，我们通过代码进行理解。
先来看PatchEmbedding，其核心思路就是通过卷积进行嵌入。
class PatchEmbed(BaseModule):
"""Image to Patch Embedding.
We use a conv layer to implement PatchEmbed.
Args:
in_channels (int): The num of input channels. Default: 3
embed_dims (int): The dimensions of embedding. Default: 768
conv_type (str): The config dict for embedding
conv layer type selection. Default: "Conv2d.
kernel_size (int): The kernel_size of embedding conv. Default: 16.
stride (int): The slide stride of embedding conv.
Default: None (Would be set as `kernel_size`).
padding (int | tuple | string ): The padding length of
embedding conv. When it is a string, it means the mode
of adaptive padding, support "same" and "corner" now.
Default: "corner".
dilation (int): The dilation rate of embedding conv. Default: 1.
bias (bool): Bias of embed conv. Default: True.
norm_cfg (dict, optional): Config dict for normalization layer.
Default: None.
input_size (int | tuple | None): The size of input, which will be
used to calculate the out size. Only work when `dynamic_size`
is False. Default: None.
init_cfg (`mmengine.ConfigDict`, optional): The Config for
initialization. Default: None.
"""
def __init__(self,
in_channels: int = 3,
embed_dims: int = 768,
conv_type: str = 'Conv2d',
kernel_size: int = 16,
stride: int = 16,
padding: Union[int, tuple, str] = 'corner',
dilation: int = 1,
bias: bool = True,
norm_cfg: OptConfigType = None,
input_size: Union[int, tuple] = None,
init_cfg: OptConfigType = None) -> None:
super(PatchEmbed, self).__init__(init_cfg=init_cfg)
self.embed_dims = embed_dims
if stride is None:
stride = kernel_size
kernel_size = to_2tuple(kernel_size)
stride = to_2tuple(stride)
dilation = to_2tuple(dilation)
if isinstance(padding, str):
self.adap_padding = AdaptivePadding(
kernel_size=kernel_size,
stride=stride,
dilation=dilation,
padding=padding)
# disable the padding of conv
padding = 0
else:
self.adap_padding = None
padding = to_2tuple(padding)
self.projection = build_conv_layer(
dict(type=conv_type),
in_channels=in_channels,
out_channels=embed_dims,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
bias=bias)
if norm_cfg is not None:
self.norm = build_norm_layer(norm_cfg, embed_dims)[1]
else:
self.norm = None
if input_size:
input_size = to_2tuple(input_size)
# `init_out_size` would be used outside to
# calculate the num_patches
# when `use_abs_pos_embed` outside
self.init_input_size = input_size
if self.adap_padding:
pad_h, pad_w = self.adap_padding.get_pad_shape(input_size)
input_h, input_w = input_size
input_h = input_h + pad_h
input_w = input_w + pad_w
input_size = (input_h, input_w)
# https://pytorch.org/docs/stable/generated/torch.nn.Conv2d.html
h_out = (input_size[0] + 2 * padding[0] - dilation[0] *
(kernel_size[0] - 1) - 1) // stride[0] + 1
w_out = (input_size[1] + 2 * padding[1] - dilation[1] *
(kernel_size[1] - 1) - 1) // stride[1] + 1
self.init_out_size = (h_out, w_out)
else:
self.init_input_size = None
self.init_out_size = None
def forward(self, x: Tensor) -> Tuple[Tensor, Tuple[int]]:
"""
Args:
x (Tensor): Has shape (B, C, H, W). In most case, C is 3.
Returns:
tuple: Contains merged results and its spatial shape.
- x (Tensor): Has shape (B, out_h * out_w, embed_dims)
- out_size (tuple[int]): Spatial shape of x, arrange as
(out_h, out_w).
"""
if self.adap_padding:
x = self.adap_padding(x)
x = self.projection(x)
out_size = (x.shape[2], x.shape[3])
x = x.flatten(2).transpose(1, 2)
if self.norm is not None:
x = self.norm(x)
return x, out_sizePatch Merging也是类似的，通过卷积嵌入。
class PatchMerging(BaseModule):
"""Merge patch feature map.
This layer groups feature map by kernel_size, and applies norm and linear
layers to the grouped feature map. Our implementation uses `nn.Unfold` to
merge patch, which is about 25% faster than original implementation.
Instead, we need to modify pretrained models for compatibility.
Args:
in_channels (int): The num of input channels.
to gets fully covered by filter and stride you specified..
Default: True.
out_channels (int): The num of output channels.
kernel_size (int | tuple, optional): the kernel size in the unfold
layer. Defaults to 2.
stride (int | tuple, optional): the stride of the sliding blocks in the
unfold layer. Default: None. (Would be set as `kernel_size`)
padding (int | tuple | string ): The padding length of
embedding conv. When it is a string, it means the mode
of adaptive padding, support "same" and "corner" now.
Default: "corner".
dilation (int | tuple, optional): dilation parameter in the unfold
layer. Default: 1.
bias (bool, optional): Whether to add bias in linear layer or not.
Defaults: False.
norm_cfg (dict, optional): Config dict for normalization layer.
Default: dict(type='LN').
init_cfg (dict, optional): The extra config for initialization.
Default: None.
"""
def __init__(self,
in_channels: int,
out_channels: int,
kernel_size: Optional[Union[int, tuple]] = 2,
stride: Optional[Union[int, tuple]] = None,
padding: Union[int, tuple, str] = 'corner',
dilation: Optional[Union[int, tuple]] = 1,
bias: Optional[bool] = False,
norm_cfg: OptConfigType = dict(type='LN'),
init_cfg: OptConfigType = None) -> None:
super().__init__(init_cfg=init_cfg)
self.in_channels = in_channels
self.out_channels = out_channels
if stride:
stride = stride
else:
stride = kernel_size
kernel_size = to_2tuple(kernel_size)
stride = to_2tuple(stride)
dilation = to_2tuple(dilation)
if isinstance(padding, str):
self.adap_padding = AdaptivePadding(
kernel_size=kernel_size,
stride=stride,
dilation=dilation,
padding=padding)
# disable the padding of unfold
padding = 0
else:
self.adap_padding = None
padding = to_2tuple(padding)
self.sampler = nn.Unfold(
kernel_size=kernel_size,
dilation=dilation,
padding=padding,
stride=stride)
sample_dim = kernel_size[0] * kernel_size[1] * in_channels
if norm_cfg is not None:
self.norm = build_norm_layer(norm_cfg, sample_dim)[1]
else:
self.norm = None
self.reduction = nn.Linear(sample_dim, out_channels, bias=bias)
def forward(self, x: Tensor,
input_size: Tuple[int]) -> Tuple[Tensor, Tuple[int]]:
"""
Args:
x (Tensor): Has shape (B, H*W, C_in).
input_size (tuple[int]): The spatial shape of x, arrange as (H, W).
Default: None.
Returns:
tuple: Contains merged results and its spatial shape.
- x (Tensor): Has shape (B, Merged_H * Merged_W, C_out)
- out_size (tuple[int]): Spatial shape of x, arrange as
(Merged_H, Merged_W).
"""
B, L, C = x.shape
assert isinstance(input_size, Sequence), f'Expect ' \
f'input_size is ' \
f'`Sequence` ' \
f'but get {input_size}'
H, W = input_size
assert L == H * W, 'input feature has wrong size'
x = x.view(B, H, W, C).permute([0, 3, 1, 2]) # B, C, H, W
# Use nn.Unfold to merge patch. About 25% faster than original method,
# but need to modify pretrained model for compatibility
if self.adap_padding:
x = self.adap_padding(x)
H, W = x.shape[-2:]
x = self.sampler(x)
# if kernel_size=2 and stride=2, x should has shape (B, 4*C, H/2*W/2)
out_h = (H + 2 * self.sampler.padding[0] - self.sampler.dilation[0] *
(self.sampler.kernel_size[0] - 1) -
1) // self.sampler.stride[0] + 1
out_w = (W + 2 * self.sampler.padding[1] - self.sampler.dilation[1] *
(self.sampler.kernel_size[1] - 1) -
1) // self.sampler.stride[1] + 1
output_size = (out_h, out_w)
x = x.transpose(1, 2) # B, H/2*W/2, 4*C
x = self.norm(x) if self.norm else x
x = self.reduction(x)
return x, output_sizeW-MSA和普通的self attention一样，但是要注意把每个window展开。
class WindowMSA(BaseModule):
"""Window based multi-head self-attention (W-MSA) module with relative
position bias.
Args:
embed_dims (int): Number of input channels.
num_heads (int): Number of attention heads.
window_size (tuple[int]): The height and width of the window.
qkv_bias (bool, optional): If True, add a learnable bias to q, k, v.
Default: True.
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Default: None.
attn_drop_rate (float, optional): Dropout ratio of attention weight.
Default: 0.0
proj_drop_rate (float, optional): Dropout ratio of output. Default: 0.
init_cfg (dict | None, optional): The Config for initialization.
Default: None.
"""
def __init__(self,
embed_dims,
num_heads,
window_size,
qkv_bias=True,
qk_scale=None,
attn_drop_rate=0.,
proj_drop_rate=0.,
init_cfg=None):
super().__init__()
self.embed_dims = embed_dims
self.window_size = window_size # Wh, Ww
self.num_heads = num_heads
head_embed_dims = embed_dims // num_heads
self.scale = qk_scale or head_embed_dims**-0.5
self.init_cfg = init_cfg
# 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
# About 2x faster than original impl
Wh, Ww = self.window_size
rel_index_coords = self.double_step_seq(2 * Ww - 1, Wh, 1, Ww)
rel_position_index = rel_index_coords + rel_index_coords.T
rel_position_index = rel_position_index.flip(1).contiguous()
self.register_buffer('relative_position_index', rel_position_index)
self.qkv = nn.Linear(embed_dims, embed_dims * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop_rate)
self.proj = nn.Linear(embed_dims, embed_dims)
self.proj_drop = nn.Dropout(proj_drop_rate)
self.softmax = nn.Softmax(dim=-1)
def init_weights(self):
trunc_normal_(self.relative_position_bias_table, std=0.02)
def forward(self, x, mask=None):
"""
Args:
x (tensor): input features with shape of (num_windows*B, N, C)
mask (tensor | None, Optional): mask with shape of (num_windows,
Wh*Ww, Wh*Ww), value should be between (-inf, 0].
"""
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)
# make torchscript happy (cannot use tensor as tuple)
q, k, v = qkv[0], qkv[1], qkv[2]
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)
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 xSW-MSA稍微复杂一些。
class ShiftWindowMSA(BaseModule):
"""Shifted Window Multihead Self-Attention Module.
Args:
embed_dims (int): Number of input channels.
num_heads (int): Number of attention heads.
window_size (int): The height and width of the window.
shift_size (int, optional): The shift step of each window towards
right-bottom. If zero, act as regular window-msa. Defaults to 0.
qkv_bias (bool, optional): If True, add a learnable bias to q, k, v.
Default: True
qk_scale (float | None, optional): Override default qk scale of
head_dim ** -0.5 if set. Defaults: None.
attn_drop_rate (float, optional): Dropout ratio of attention weight.
Defaults: 0.
proj_drop_rate (float, optional): Dropout ratio of output.
Defaults: 0.
dropout_layer (dict, optional): The dropout_layer used before output.
Defaults: dict(type='DropPath', drop_prob=0.).
init_cfg (dict, optional): The extra config for initialization.
Default: None.
"""
def __init__(self,
embed_dims,
num_heads,
window_size,
shift_size=0,
qkv_bias=True,
qk_scale=None,
attn_drop_rate=0,
proj_drop_rate=0,
dropout_layer=dict(type='DropPath', drop_prob=0.),
init_cfg=None):
super().__init__(init_cfg)
self.window_size = window_size
self.shift_size = shift_size
assert 0 <= self.shift_size < self.window_size
self.w_msa = WindowMSA(
embed_dims=embed_dims,
num_heads=num_heads,
window_size=to_2tuple(window_size),
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop_rate=attn_drop_rate,
proj_drop_rate=proj_drop_rate,
init_cfg=None)
self.drop = build_dropout(dropout_layer)
def forward(self, query, hw_shape):
B, L, C = query.shape
H, W = hw_shape
assert L == H * W, 'input feature has wrong size'
query = query.view(B, H, W, C)
# pad feature maps to multiples of window size
pad_r = (self.window_size - W % self.window_size) % self.window_size
pad_b = (self.window_size - H % self.window_size) % self.window_size
query = F.pad(query, (0, 0, 0, pad_r, 0, pad_b))
H_pad, W_pad = query.shape[1], query.shape[2]
# cyclic shift
if self.shift_size > 0:
shifted_query = torch.roll(
query,
shifts=(-self.shift_size, -self.shift_size),
dims=(1, 2))
# calculate attention mask for SW-MSA
img_mask = torch.zeros((1, H_pad, W_pad, 1), device=query.device)
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
# nW, window_size, window_size, 1
mask_windows = self.window_partition(img_mask)
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))
else:
shifted_query = query
attn_mask = None
# nW*B, window_size, window_size, C
query_windows = self.window_partition(shifted_query)
# nW*B, window_size*window_size, C
query_windows = query_windows.view(-1, self.window_size**2, C)
# W-MSA/SW-MSA (nW*B, window_size*window_size, C)
attn_windows = self.w_msa(query_windows, mask=attn_mask)
# merge windows
attn_windows = attn_windows.view(-1, self.window_size,
self.window_size, C)
# B H' W' C
shifted_x = self.window_reverse(attn_windows, H_pad, W_pad)
# 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:
x = x[:, :H, :W, :].contiguous()
x = x.view(B, H * W, C)
x = self.drop(x)
return x
def window_reverse(self, windows, H, W):
"""
Args:
windows: (num_windows*B, window_size, window_size, C)
H (int): Height of image
W (int): Width of image
Returns:
x: (B, H, W, C)
"""
window_size = self.window_size
B = int(windows.shape[0] / (H * W / window_size / window_size))
x = windows.view(B, H // window_size, W // window_size, window_size,
window_size, -1)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x
def window_partition(self, x):
"""
Args:
x: (B, H, W, C)
Returns:
windows: (num_windows*B, window_size, window_size, C)
"""
B, H, W, C = x.shape
window_size = self.window_size
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()
windows = windows.view(-1, window_size, window_size, C)
return windows通过多层级架构以及窗口注意力提取图像特征。