|
| 1 | +""" Shifted Window Attn |
| 2 | +
|
| 3 | +This is a WIP experiment to apply windowed attention from the Swin Transformer |
| 4 | +to a stand-alone module for use as an attn block in conv nets. |
| 5 | +
|
| 6 | +Based on original swin window code at https://github.com/microsoft/Swin-Transformer |
| 7 | +Swin Transformer paper: https://arxiv.org/pdf/2103.14030.pdf |
| 8 | +""" |
| 9 | +from typing import Optional |
| 10 | + |
| 11 | +import torch |
| 12 | +import torch.nn as nn |
| 13 | + |
| 14 | +from .drop import DropPath |
| 15 | +from .helpers import to_2tuple |
| 16 | +from .weight_init import trunc_normal_ |
| 17 | + |
| 18 | + |
| 19 | +def window_partition(x, win_size: int): |
| 20 | + """ |
| 21 | + Args: |
| 22 | + x: (B, H, W, C) |
| 23 | + win_size (int): window size |
| 24 | +
|
| 25 | + Returns: |
| 26 | + windows: (num_windows*B, window_size, window_size, C) |
| 27 | + """ |
| 28 | + B, H, W, C = x.shape |
| 29 | + x = x.view(B, H // win_size, win_size, W // win_size, win_size, C) |
| 30 | + windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, win_size, win_size, C) |
| 31 | + return windows |
| 32 | + |
| 33 | + |
| 34 | +def window_reverse(windows, win_size: int, H: int, W: int): |
| 35 | + """ |
| 36 | + Args: |
| 37 | + windows: (num_windows*B, window_size, window_size, C) |
| 38 | + win_size (int): Window size |
| 39 | + H (int): Height of image |
| 40 | + W (int): Width of image |
| 41 | +
|
| 42 | + Returns: |
| 43 | + x: (B, H, W, C) |
| 44 | + """ |
| 45 | + B = int(windows.shape[0] / (H * W / win_size / win_size)) |
| 46 | + x = windows.view(B, H // win_size, W // win_size, win_size, win_size, -1) |
| 47 | + x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) |
| 48 | + return x |
| 49 | + |
| 50 | + |
| 51 | +class WindowAttention(nn.Module): |
| 52 | + r""" Window based multi-head self attention (W-MSA) module with relative position bias. |
| 53 | + It supports both of shifted and non-shifted window. |
| 54 | +
|
| 55 | + Args: |
| 56 | + dim (int): Number of input channels. |
| 57 | + win_size (int): The height and width of the window. |
| 58 | + num_heads (int): Number of attention heads. |
| 59 | + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True |
| 60 | + attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 |
| 61 | + """ |
| 62 | + |
| 63 | + def __init__( |
| 64 | + self, dim, dim_out=None, feat_size=None, stride=1, win_size=8, shift_size=None, num_heads=8, |
| 65 | + qkv_bias=True, attn_drop=0.): |
| 66 | + |
| 67 | + super().__init__() |
| 68 | + self.dim_out = dim_out or dim |
| 69 | + self.feat_size = to_2tuple(feat_size) |
| 70 | + self.win_size = win_size |
| 71 | + self.shift_size = shift_size or win_size // 2 |
| 72 | + if min(self.feat_size) <= win_size: |
| 73 | + # if window size is larger than input resolution, we don't partition windows |
| 74 | + self.shift_size = 0 |
| 75 | + self.win_size = min(self.feat_size) |
| 76 | + assert 0 <= self.shift_size < self.win_size, "shift_size must in 0-window_size" |
| 77 | + self.num_heads = num_heads |
| 78 | + head_dim = self.dim_out // num_heads |
| 79 | + self.scale = head_dim ** -0.5 |
| 80 | + |
| 81 | + if self.shift_size > 0: |
| 82 | + # calculate attention mask for SW-MSA |
| 83 | + H, W = self.feat_size |
| 84 | + img_mask = torch.zeros((1, H, W, 1)) # 1 H W 1 |
| 85 | + h_slices = ( |
| 86 | + slice(0, -self.win_size), |
| 87 | + slice(-self.win_size, -self.shift_size), |
| 88 | + slice(-self.shift_size, None)) |
| 89 | + w_slices = ( |
| 90 | + slice(0, -self.win_size), |
| 91 | + slice(-self.win_size, -self.shift_size), |
| 92 | + slice(-self.shift_size, None)) |
| 93 | + cnt = 0 |
| 94 | + for h in h_slices: |
| 95 | + for w in w_slices: |
| 96 | + img_mask[:, h, w, :] = cnt |
| 97 | + cnt += 1 |
| 98 | + mask_windows = window_partition(img_mask, self.win_size) # num_win, window_size, window_size, 1 |
| 99 | + mask_windows = mask_windows.view(-1, self.win_size * self.win_size) |
| 100 | + attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) |
| 101 | + attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) |
| 102 | + else: |
| 103 | + attn_mask = None |
| 104 | + self.register_buffer("attn_mask", attn_mask) |
| 105 | + |
| 106 | + # define a parameter table of relative position bias |
| 107 | + self.relative_position_bias_table = nn.Parameter( |
| 108 | + # 2 * Wh - 1 * 2 * Ww - 1, nH |
| 109 | + torch.zeros((2 * self.win_size - 1) * (2 * self.win_size - 1), num_heads)) |
| 110 | + |
| 111 | + # get pair-wise relative position index for each token inside the window |
| 112 | + coords_h = torch.arange(self.win_size) |
| 113 | + coords_w = torch.arange(self.win_size) |
| 114 | + coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww |
| 115 | + coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww |
| 116 | + relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww |
| 117 | + relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 |
| 118 | + relative_coords[:, :, 0] += self.win_size - 1 # shift to start from 0 |
| 119 | + relative_coords[:, :, 1] += self.win_size - 1 |
| 120 | + relative_coords[:, :, 0] *= 2 * self.win_size - 1 |
| 121 | + relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww |
| 122 | + self.register_buffer("relative_position_index", relative_position_index) |
| 123 | + trunc_normal_(self.relative_position_bias_table, std=.02) |
| 124 | + |
| 125 | + self.qkv = nn.Linear(dim, self.dim_out * 3, bias=qkv_bias) |
| 126 | + self.attn_drop = nn.Dropout(attn_drop) |
| 127 | + self.softmax = nn.Softmax(dim=-1) |
| 128 | + self.pool = nn.AvgPool2d(2, 2) if stride == 2 else nn.Identity() |
| 129 | + |
| 130 | + def forward(self, x): |
| 131 | + B, C, H, W = x.shape |
| 132 | + x = x.permute(0, 2, 3, 1) |
| 133 | + |
| 134 | + # cyclic shift |
| 135 | + if self.shift_size > 0: |
| 136 | + shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) |
| 137 | + else: |
| 138 | + shifted_x = x |
| 139 | + |
| 140 | + # partition windows |
| 141 | + win_size_sq = self.win_size * self.win_size |
| 142 | + x_windows = window_partition(shifted_x, self.win_size) # num_win * B, window_size, window_size, C |
| 143 | + x_windows = x_windows.view(-1, win_size_sq, C) # num_win * B, window_size*window_size, C |
| 144 | + BW, N, _ = x_windows.shape |
| 145 | + |
| 146 | + qkv = self.qkv(x_windows) |
| 147 | + qkv = qkv.reshape(BW, N, 3, self.num_heads, self.dim_out // self.num_heads).permute(2, 0, 3, 1, 4) |
| 148 | + q, k, v = qkv[0], qkv[1], qkv[2] |
| 149 | + q = q * self.scale |
| 150 | + attn = (q @ k.transpose(-2, -1)) |
| 151 | + |
| 152 | + relative_position_bias = self.relative_position_bias_table[ |
| 153 | + self.relative_position_index.view(-1)].view(win_size_sq, win_size_sq, -1) |
| 154 | + relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh * Ww, Wh * Ww |
| 155 | + attn = attn + relative_position_bias.unsqueeze(0) |
| 156 | + if self.attn_mask is not None: |
| 157 | + num_win = self.attn_mask.shape[0] |
| 158 | + attn = attn.view(B, num_win, self.num_heads, N, N) + self.attn_mask.unsqueeze(1).unsqueeze(0) |
| 159 | + attn = attn.view(-1, self.num_heads, N, N) |
| 160 | + attn = self.softmax(attn) |
| 161 | + attn = self.attn_drop(attn) |
| 162 | + |
| 163 | + x = (attn @ v).transpose(1, 2).reshape(BW, N, self.dim_out) |
| 164 | + |
| 165 | + # merge windows |
| 166 | + x = x.view(-1, self.win_size, self.win_size, self.dim_out) |
| 167 | + shifted_x = window_reverse(x, self.win_size, H, W) # B H' W' C |
| 168 | + |
| 169 | + # reverse cyclic shift |
| 170 | + if self.shift_size > 0: |
| 171 | + x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)) |
| 172 | + else: |
| 173 | + x = shifted_x |
| 174 | + x = x.view(B, H, W, self.dim_out).permute(0, 3, 1, 2) |
| 175 | + x = self.pool(x) |
| 176 | + return x |
| 177 | + |
| 178 | + |
0 commit comments