|
|
import torch
|
|
|
import torch.nn as nn
|
|
|
import torch.nn.functional as F
|
|
|
from timm.models.layers import to_2tuple, trunc_normal_
|
|
|
|
|
|
def window_partition(x, window_size):
|
|
|
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)
|
|
|
return windows
|
|
|
|
|
|
|
|
|
def window_reverse(windows, window_size, H, W):
|
|
|
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
|
|
|
|
|
|
class StandardUnifiedAttention(nn.Module):
|
|
|
|
|
|
def __init__(self, dim, num_heads, qkv_bias=True, attn_drop=0., proj_drop=0.):
|
|
|
super().__init__()
|
|
|
self.num_heads = num_heads
|
|
|
head_dim = dim // num_heads
|
|
|
self.scale = head_dim ** -0.5
|
|
|
|
|
|
self.q_proj = nn.Linear(dim, dim, bias=qkv_bias)
|
|
|
self.k_proj = nn.Linear(dim, dim, bias=qkv_bias)
|
|
|
self.v_proj = nn.Linear(dim, dim, bias=qkv_bias)
|
|
|
|
|
|
self.attn_drop = nn.Dropout(attn_drop)
|
|
|
self.proj = nn.Linear(dim, dim)
|
|
|
self.proj_drop = nn.Dropout(proj_drop)
|
|
|
|
|
|
def forward(self, query, key, value, mask=None):
|
|
|
B, N, C = query.shape
|
|
|
|
|
|
q = self.q_proj(query).view(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
|
|
k = self.k_proj(key).view(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
|
|
v = self.v_proj(value).view(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
|
|
|
|
|
attn = (q @ k.transpose(-2, -1)) * self.scale
|
|
|
|
|
|
if mask is not None:
|
|
|
attn = attn.masked_fill(mask == 0, float('-inf'))
|
|
|
|
|
|
attn_map = attn.softmax(dim=-1)
|
|
|
attn_map_dropped = self.attn_drop(attn_map)
|
|
|
|
|
|
x = (attn_map_dropped @ v).transpose(1, 2).reshape(B, N, C)
|
|
|
|
|
|
|
|
|
x = self.proj(x)
|
|
|
x = self.proj_drop(x)
|
|
|
|
|
|
return x, attn_map
|
|
|
|
|
|
class GuidedResampler(nn.Module):
|
|
|
|
|
|
def __init__(self, dim, downsample_ratio=4, k_top_samples=1):
|
|
|
super().__init__()
|
|
|
self.dim = dim
|
|
|
self.ratio = downsample_ratio
|
|
|
self.k_samples = k_top_samples
|
|
|
|
|
|
def forward(self, v_high_feat, coarse_attn_map):
|
|
|
|
|
|
B, C, H, W = v_high_feat.shape
|
|
|
H_low, W_low = H // self.ratio, W // self.ratio
|
|
|
N_high = H * W
|
|
|
N_low = H_low * W_low
|
|
|
|
|
|
assert coarse_attn_map.shape == (B, N_low, N_low), \
|
|
|
f"Coarse map shape mismatch. Expected {(B, N_low, N_low)}, but got {coarse_attn_map.shape}"
|
|
|
|
|
|
v_high_seq = v_high_feat.flatten(2).transpose(1, 2)
|
|
|
|
|
|
topk_values, topk_indices_low = torch.topk(coarse_attn_map, k=self.k_samples, dim=-1)
|
|
|
|
|
|
topk_indices_low_row = topk_indices_low // W_low
|
|
|
topk_indices_low_col = topk_indices_low % W_low
|
|
|
|
|
|
topk_indices_high_topleft_row = topk_indices_low_row * self.ratio
|
|
|
topk_indices_high_topleft_col = topk_indices_low_col * self.ratio
|
|
|
|
|
|
delta = torch.stack(torch.meshgrid(
|
|
|
torch.arange(self.ratio, device=v_high_feat.device),
|
|
|
torch.arange(self.ratio, device=v_high_feat.device),
|
|
|
indexing='ij'
|
|
|
), dim=-1).view(-1, 2)
|
|
|
|
|
|
topleft = torch.stack([topk_indices_high_topleft_row, topk_indices_high_topleft_col], dim=-1)
|
|
|
sparse_indices_2d = topleft.unsqueeze(-2) + delta.view(1, 1, 1, -1, 2)
|
|
|
|
|
|
sparse_indices_1d = sparse_indices_2d[..., 0] * W + sparse_indices_2d[..., 1]
|
|
|
sparse_indices_1d = sparse_indices_1d.view(B, N_low, -1)
|
|
|
|
|
|
high_res_q_coords = torch.stack(torch.meshgrid(
|
|
|
torch.arange(H, device=v_high_feat.device),
|
|
|
torch.arange(W, device=v_high_feat.device),
|
|
|
indexing='ij'
|
|
|
), dim=-1).view(-1, 2)
|
|
|
|
|
|
low_res_grid_indices = (high_res_q_coords[:, 0] // self.ratio) * W_low + (high_res_q_coords[:, 1] // self.ratio)
|
|
|
|
|
|
K_sparse_len = sparse_indices_1d.shape[-1]
|
|
|
low_res_grid_indices_expanded = low_res_grid_indices.view(1, N_high, 1).expand(B, -1, K_sparse_len)
|
|
|
|
|
|
final_sparse_indices = torch.gather(sparse_indices_1d, 1, low_res_grid_indices_expanded)
|
|
|
|
|
|
batch_indices = torch.arange(B, device=v_high_feat.device).view(B, 1, 1)
|
|
|
v_sparse_seq = v_high_seq[batch_indices, final_sparse_indices]
|
|
|
|
|
|
normalized_weights_low = F.softmax(topk_values, dim=-1)
|
|
|
low_res_grid_indices_weights_expanded = low_res_grid_indices.view(1, N_high, 1).expand(B, -1, self.k_samples)
|
|
|
weights_high = torch.gather(normalized_weights_low, 1, low_res_grid_indices_weights_expanded)
|
|
|
|
|
|
v_reshaped = v_sparse_seq.view(B, N_high, self.k_samples, self.ratio**2, C)
|
|
|
|
|
|
weights_for_broadcast = (weights_high / (self.ratio**2)).view(B, N_high, self.k_samples, 1, 1)
|
|
|
|
|
|
warped_seq = (v_reshaped * weights_for_broadcast).sum(dim=(2, 3))
|
|
|
|
|
|
warped_feat = warped_seq.transpose(1, 2).view(B, C, H, W)
|
|
|
|
|
|
return warped_feat
|
|
|
|
|
|
|
|
|
class SwinUnifiedAttention(nn.Module):
|
|
|
def __init__(self, dim, num_heads, window_size, qkv_bias=True, attn_drop=0., proj_drop=0.):
|
|
|
super().__init__()
|
|
|
self.num_heads = num_heads
|
|
|
head_dim = dim // num_heads
|
|
|
self.scale = head_dim ** -0.5
|
|
|
|
|
|
self.q = nn.Linear(dim, dim, bias=qkv_bias)
|
|
|
self.k = nn.Linear(dim, dim, bias=qkv_bias)
|
|
|
self.v = nn.Linear(dim, dim, bias=qkv_bias)
|
|
|
self.relative_position_bias_table = nn.Parameter(
|
|
|
torch.zeros((2 * window_size - 1) * (2 * window_size - 1), num_heads))
|
|
|
|
|
|
coords_h = torch.arange(window_size)
|
|
|
coords_w = torch.arange(window_size)
|
|
|
coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing="ij"))
|
|
|
coords_flatten = torch.flatten(coords, 1)
|
|
|
relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
|
|
|
relative_coords = relative_coords.permute(1, 2, 0).contiguous()
|
|
|
relative_coords[:, :, 0] += window_size - 1
|
|
|
relative_coords[:, :, 1] += window_size - 1
|
|
|
relative_coords[:, :, 0] *= 2 * window_size - 1
|
|
|
relative_position_index = relative_coords.sum(-1)
|
|
|
self.register_buffer("relative_position_index", relative_position_index)
|
|
|
|
|
|
trunc_normal_(self.relative_position_bias_table, std=.02)
|
|
|
self.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, query, key, value, mask=None):
|
|
|
B_, N, C = query.shape
|
|
|
q = self.q(query).reshape(B_, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
|
|
k = self.k(key).reshape(B_, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
|
|
v = self.v(value).reshape(B_, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
|
|
|
|
|
|
attn = (q * self.scale) @ k.transpose(-2, -1)
|
|
|
|
|
|
relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
|
|
|
N, N, -1).permute(2, 0, 1).contiguous()
|
|
|
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 x
|
|
|
|
|
|
class UnifiedTransformerBlock(nn.Module):
|
|
|
"""统一的标准 Transformer Block。"""
|
|
|
def __init__(self, dim, input_resolution, num_heads, mlp_ratio=2.0, qkv_bias=True,
|
|
|
drop=0., attn_drop=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
|
|
|
super().__init__()
|
|
|
self.norm_q = norm_layer(dim)
|
|
|
self.norm_kv = norm_layer(dim)
|
|
|
self.attn = StandardUnifiedAttention(dim, num_heads, qkv_bias, attn_drop, drop)
|
|
|
|
|
|
H, W = to_2tuple(input_resolution)
|
|
|
dim_spatial = H * W
|
|
|
|
|
|
self.q_pos_embedding = nn.Parameter(torch.randn(1, dim_spatial, dim))
|
|
|
self.k_pos_embedding = nn.Parameter(torch.randn(1, dim_spatial, dim))
|
|
|
|
|
|
self.norm_ffn = norm_layer(dim)
|
|
|
mlp_hidden_dim = int(dim * mlp_ratio)
|
|
|
self.mlp = nn.Sequential(
|
|
|
nn.Linear(dim, mlp_hidden_dim), act_layer(), nn.Dropout(drop),
|
|
|
nn.Linear(mlp_hidden_dim, dim), nn.Dropout(drop)
|
|
|
)
|
|
|
|
|
|
def forward(self, query, key=None, value=None):
|
|
|
B, C, H, W = query.shape
|
|
|
is_cross_attention = key is not None
|
|
|
if not is_cross_attention:
|
|
|
key, value = query, query
|
|
|
|
|
|
q_in = query.flatten(2).transpose(1, 2)
|
|
|
k_in = key.flatten(2).transpose(1, 2)
|
|
|
v_in = value.flatten(2).transpose(1, 2)
|
|
|
shortcut = v_in
|
|
|
|
|
|
q_norm = self.norm_q(q_in + self.q_pos_embedding)
|
|
|
k_norm = self.norm_kv(k_in + self.k_pos_embedding)
|
|
|
v_norm = self.norm_kv(v_in)
|
|
|
|
|
|
attn_output, _ = self.attn(query=q_norm, key=k_norm, value=v_norm)
|
|
|
|
|
|
x = shortcut + attn_output
|
|
|
x = x + self.mlp(self.norm_ffn(x))
|
|
|
return x.transpose(1, 2).view(B, C, H, W)
|
|
|
|
|
|
class UnifiedSwinBlock(nn.Module):
|
|
|
def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
|
|
|
mlp_ratio=2., qkv_bias=True, drop=0., attn_drop=0.,
|
|
|
act_layer=nn.GELU, norm_layer=nn.LayerNorm):
|
|
|
super().__init__()
|
|
|
self.input_resolution = to_2tuple(input_resolution)
|
|
|
self.window_size = window_size
|
|
|
self.shift_size = shift_size
|
|
|
if min(self.input_resolution) <= self.window_size:
|
|
|
self.shift_size = 0
|
|
|
self.window_size = min(self.input_resolution)
|
|
|
|
|
|
self.norm_q = norm_layer(dim)
|
|
|
self.norm_kv = norm_layer(dim)
|
|
|
self.attn = SwinUnifiedAttention(
|
|
|
dim, num_heads, self.window_size, qkv_bias, attn_drop, drop)
|
|
|
|
|
|
self.norm_ffn = norm_layer(dim)
|
|
|
mlp_hidden_dim = int(dim * mlp_ratio)
|
|
|
self.mlp = nn.Sequential(
|
|
|
nn.Linear(dim, mlp_hidden_dim), act_layer(),
|
|
|
nn.Linear(mlp_hidden_dim, dim), nn.Dropout(drop))
|
|
|
|
|
|
if self.shift_size > 0:
|
|
|
H, W = self.input_resolution
|
|
|
img_mask = torch.zeros((1, H, W, 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.view(1,H,W,1), self.window_size).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:
|
|
|
attn_mask = None
|
|
|
self.register_buffer("attn_mask", attn_mask)
|
|
|
|
|
|
def forward(self, query, key=None, value=None):
|
|
|
B, C, H, W = query.shape
|
|
|
is_cross_attention = key is not None
|
|
|
if not is_cross_attention:
|
|
|
key, value = query, query
|
|
|
|
|
|
q = query.flatten(2).transpose(1, 2)
|
|
|
k = key.flatten(2).transpose(1, 2)
|
|
|
v = value.flatten(2).transpose(1, 2)
|
|
|
shortcut = v
|
|
|
|
|
|
q = self.norm_q(q).view(B, H, W, C)
|
|
|
k = self.norm_kv(k).view(B, H, W, C)
|
|
|
v = self.norm_kv(v).view(B, H, W, C)
|
|
|
|
|
|
if self.shift_size > 0:
|
|
|
shifted_q, shifted_k, shifted_v = [torch.roll(t, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
|
|
|
for t in (q, k, v)]
|
|
|
else:
|
|
|
shifted_q, shifted_k, shifted_v = q, k, v
|
|
|
|
|
|
q_win = window_partition(shifted_q, self.window_size)
|
|
|
k_win = window_partition(shifted_k, self.window_size)
|
|
|
v_win = window_partition(shifted_v, self.window_size)
|
|
|
|
|
|
attn_windows = self.attn(q_win, k_win, v_win, mask=self.attn_mask)
|
|
|
shifted_x = window_reverse(attn_windows, self.window_size, H, W)
|
|
|
|
|
|
if self.shift_size > 0:
|
|
|
x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
|
|
|
else:
|
|
|
x = shifted_x
|
|
|
|
|
|
x = x.view(B, H * W, C)
|
|
|
x = shortcut + x
|
|
|
x = x + self.mlp(self.norm_ffn(x))
|
|
|
return x.transpose(1, 2).view(B, C, H, W)
|
|
|
|
|
|
class CrossAttention(nn.Module):
|
|
|
|
|
|
def __init__(self, args, dim, resolution):
|
|
|
super().__init__()
|
|
|
self.is_standard_attention = resolution[0] < args.swin_res_threshold
|
|
|
|
|
|
if self.is_standard_attention:
|
|
|
self.block_efc = StandardUnifiedAttention(dim=dim, num_heads=args.num_heads)
|
|
|
else:
|
|
|
anchor_resolution = args.swin_res_threshold
|
|
|
ratio = 2 * (resolution[0] / anchor_resolution)
|
|
|
assert ratio >= 1 and ratio.is_integer(), "Fine resolution must be a multiple of anchor resolution"
|
|
|
self.block = GuidedResampler(dim=dim, downsample_ratio=int(ratio))
|
|
|
|
|
|
def coarse_stage(self, A, B, C, attn=None):
|
|
|
B_, C_, H, W = A.shape
|
|
|
A_seq = A.flatten(2).transpose(1, 2)
|
|
|
B_seq = B.flatten(2).transpose(1, 2)
|
|
|
C_seq = C.flatten(2).transpose(1, 2)
|
|
|
out_seq, attn_map = self.block_efc(A_seq, B_seq, C_seq)
|
|
|
out = out_seq.transpose(1, 2).view(B_, C_, H, W)
|
|
|
return out, attn_map
|
|
|
|
|
|
def fine_stage(self, C, attn=None):
|
|
|
out = self.block(C, attn.mean(dim=1))
|
|
|
return out
|
|
|
|
|
|
def forward(self, A, B, C, D, attn=None):
|
|
|
|
|
|
if not self.is_standard_attention:
|
|
|
out = self.block(C, attn.mean(dim=1))
|
|
|
return out
|
|
|
|
|
|
else:
|
|
|
B_, C_, H, W = A.shape
|
|
|
A_seq = A.flatten(2).transpose(1, 2)
|
|
|
B_seq = B.flatten(2).transpose(1, 2)
|
|
|
C_seq = C.flatten(2).transpose(1, 2)
|
|
|
out_seq, attn_map = self.block_efc(A_seq, B_seq, C_seq)
|
|
|
out = out_seq.transpose(1, 2).view(B_, C_, H, W)
|
|
|
return out, attn_map
|
|
|
|
|
|
class SelfAttention(nn.Module):
|
|
|
def __init__(self, args, dim, resolution):
|
|
|
super().__init__()
|
|
|
self.blocks = nn.ModuleList()
|
|
|
|
|
|
common_kwargs = {
|
|
|
'dim': dim,
|
|
|
'input_resolution': resolution,
|
|
|
'num_heads': args.num_heads,
|
|
|
}
|
|
|
|
|
|
if resolution[0] >= args.swin_res_threshold:
|
|
|
self.blocks.append(UnifiedSwinBlock(window_size=args.window_size, shift_size=0, **common_kwargs))
|
|
|
self.blocks.append(UnifiedSwinBlock(window_size=args.window_size, shift_size=args.window_size // 2, **common_kwargs))
|
|
|
else:
|
|
|
self.blocks.append(UnifiedTransformerBlock(mlp_ratio=2.0, **common_kwargs))
|
|
|
|
|
|
def forward(self, query, key=None, value=None):
|
|
|
is_cross_attention = key is not None
|
|
|
if is_cross_attention:
|
|
|
v_out = value
|
|
|
for block in self.blocks:
|
|
|
v_out = block(query, key, v_out)
|
|
|
return v_out
|
|
|
else:
|
|
|
x_out = query
|
|
|
for block in self.blocks:
|
|
|
x_out = block(x_out)
|
|
|
return x_out |
