Spaces:
Running
on
Zero
Running
on
Zero
File size: 14,587 Bytes
a930e1f |
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 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops.einops import rearrange
class Mlp(nn.Module):
"""Multi-Layer Perceptron (MLP)"""
def __init__(self,
in_dim,
hidden_dim=None,
out_dim=None,
act_layer=nn.GELU):
"""
Args:
in_dim: input features dimension
hidden_dim: hidden features dimension
out_dim: output features dimension
act_layer: activation function
"""
super().__init__()
out_dim = out_dim or in_dim
hidden_dim = hidden_dim or in_dim
self.fc1 = nn.Linear(in_dim, hidden_dim)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_dim, out_dim)
self.out_dim = out_dim
def forward(self, x):
x_size = x.size()
x = x.view(-1, x_size[-1])
x = self.fc1(x)
x = self.act(x)
x = self.fc2(x)
x = x.view(*x_size[:-1], self.out_dim)
return x
class VanillaAttention(nn.Module):
def __init__(self,
dim,
num_heads=8,
proj_bias=False):
super().__init__()
"""
Args:
dim: feature dimension
num_heads: number of attention head
proj_bias: bool use query, key, value bias
"""
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.softmax_temp = self.head_dim ** -0.5
self.kv_proj = nn.Linear(dim, dim * 2, bias=proj_bias)
self.q_proj = nn.Linear(dim, dim, bias=proj_bias)
self.merge = nn.Linear(dim, dim)
def forward(self, x_q, x_kv=None):
"""
Args:
x_q (torch.Tensor): [N, L, C]
x_kv (torch.Tensor): [N, S, C]
"""
if x_kv is None:
x_kv = x_q
bs, _, dim = x_q.shape
bs, _, dim = x_kv.shape
# [N, S, 2, H, D] => [2, N, H, S, D]
kv = self.kv_proj(x_kv).reshape(bs, -1, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
# [N, L, H, D] => [N, H, L, D]
q = self.q_proj(x_q).reshape(bs, -1, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
k, v = kv[0].transpose(-2, -1).contiguous(), kv[1].contiguous() # [N, H, D, S], [N, H, S, D]
attn = (q @ k) * self.softmax_temp # [N, H, L, S]
attn = attn.softmax(dim=-1)
x_q = (attn @ v).transpose(1, 2).reshape(bs, -1, dim)
x_q = self.merge(x_q)
return x_q
class CrossBidirectionalAttention(nn.Module):
def __init__(self, dim, num_heads, proj_bias = False):
super().__init__()
"""
Args:
dim: feature dimension
num_heads: number of attention head
proj_bias: bool use query, key, value bias
"""
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.softmax_temp = self.head_dim ** -0.5
self.qk_proj = nn.Linear(dim, dim, bias=proj_bias)
self.v_proj = nn.Linear(dim, dim, bias=proj_bias)
self.merge = nn.Linear(dim, dim, bias=proj_bias)
self.temperature = nn.Parameter(torch.tensor([0.0]), requires_grad=True)
# print(self.temperature)
def map_(self, func, x0, x1):
return func(x0), func(x1)
def forward(self, x0, x1):
"""
Args:
x0 (torch.Tensor): [N, L, C]
x1 (torch.Tensor): [N, S, C]
"""
bs = x0.size(0)
qk0, qk1 = self.map_(self.qk_proj, x0, x1)
v0, v1 = self.map_(self.v_proj, x0, x1)
qk0, qk1, v0, v1 = map(
lambda t: t.reshape(bs, -1, self.num_heads, self.head_dim).permute(0, 2, 1, 3).contiguous(),
(qk0, qk1, v0, v1))
qk0, qk1 = qk0 * self.softmax_temp**0.5, qk1 * self.softmax_temp**0.5
sim = qk0 @ qk1.transpose(-2,-1).contiguous()
attn01 = F.softmax(sim, dim=-1)
attn10 = F.softmax(sim.transpose(-2, -1).contiguous(), dim=-1)
x0 = attn01 @ v1
x1 = attn10 @ v0
x0, x1 = self.map_(lambda t: t.transpose(1, 2).flatten(start_dim=-2),
x0, x1)
x0, x1 = self.map_(self.merge, x0, x1)
return x0, x1
class SwinPosEmbMLP(nn.Module):
def __init__(self,
dim):
super().__init__()
self.pos_embed = None
self.pos_mlp = nn.Sequential(nn.Linear(2, 512, bias=True),
nn.ReLU(),
nn.Linear(512, dim, bias=False))
def forward(self, x):
seq_length = x.shape[1]
if self.pos_embed is None or self.training:
seq_length = int(seq_length**0.5)
coords = torch.arange(0, seq_length, device=x.device, dtype = x.dtype)
grid = torch.stack(torch.meshgrid([coords, coords])).contiguous().unsqueeze(0)
grid -= seq_length // 2
grid /= (seq_length // 2)
self.pos_embed = self.pos_mlp(grid.flatten(2).transpose(1,2))
x = x + self.pos_embed
return x
class WindowSelfAttention(nn.Module):
def __init__(self, dim, num_heads, mlp_hidden_coef, use_pre_pos_embed=False):
super().__init__()
self.mlp = Mlp(in_dim=dim*2, hidden_dim=dim*mlp_hidden_coef, out_dim=dim, act_layer=nn.GELU)
self.gamma = nn.Parameter(torch.ones(dim))
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.attn = VanillaAttention(dim, num_heads=num_heads)
self.pos_embed = SwinPosEmbMLP(dim)
self.pos_embed_pre = SwinPosEmbMLP(dim) if use_pre_pos_embed else nn.Identity()
def forward(self, x, x_pre):
ww = x.shape[1]
ww_pre = x_pre.shape[1]
x = self.pos_embed(x)
x_pre = self.pos_embed_pre(x_pre)
x = torch.cat((x, x_pre), dim=1)
x = x + self.gamma*self.norm1(self.mlp(torch.cat([x, self.attn(self.norm2(x))], dim=-1)))
x, x_pre = x.split([ww, ww_pre], dim=1)
return x, x_pre
class WindowCrossAttention(nn.Module):
def __init__(self, dim, num_heads, mlp_hidden_coef):
super().__init__()
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.mlp = Mlp(in_dim=dim*2, hidden_dim=dim*mlp_hidden_coef, out_dim=dim, act_layer=nn.GELU)
self.cross_attn = CrossBidirectionalAttention(dim, num_heads=num_heads, proj_bias=False)
self.gamma = nn.Parameter(torch.ones(dim))
def forward(self, x0, x1):
m_x0, m_x1 = self.cross_attn(self.norm1(x0), self.norm1(x1))
x0 = x0 + self.gamma*self.norm2(self.mlp(torch.cat([x0, m_x0], dim=-1)))
x1 = x1 + self.gamma*self.norm2(self.mlp(torch.cat([x1, m_x1], dim=-1)))
return x0, x1
class FineProcess(nn.Module):
def __init__(self, config):
super().__init__()
# Config
block_dims = config['resnet']['block_dims']
self.block_dims = block_dims
self.W_f = config['fine_window_size']
self.W_m = config['medium_window_size']
nhead_f = config["fine"]['nhead_fine_level']
nhead_m = config["fine"]['nhead_medium_level']
mlp_hidden_coef = config["fine"]['mlp_hidden_dim_coef']
# Networks
self.conv_merge = nn.Sequential(nn.Conv2d(block_dims[2]*2, block_dims[1], kernel_size=1, stride=1, padding=0, bias=False),
nn.Conv2d(block_dims[1], block_dims[1], kernel_size=3, stride=1, padding=1, groups=block_dims[1], bias=False),
nn.BatchNorm2d(block_dims[1])
)
self.out_conv_m = nn.Conv2d(block_dims[1], block_dims[1], kernel_size=1, stride=1, padding=0, bias=False)
self.out_conv_f = nn.Conv2d(block_dims[0], block_dims[0], kernel_size=1, stride=1, padding=0, bias=False)
self.self_attn_m = WindowSelfAttention(block_dims[1], num_heads=nhead_m,
mlp_hidden_coef=mlp_hidden_coef, use_pre_pos_embed=False)
self.cross_attn_m = WindowCrossAttention(block_dims[1], num_heads=nhead_m,
mlp_hidden_coef=mlp_hidden_coef)
self.self_attn_f = WindowSelfAttention(block_dims[0], num_heads=nhead_f,
mlp_hidden_coef=mlp_hidden_coef, use_pre_pos_embed=True)
self.cross_attn_f = WindowCrossAttention(block_dims[0], num_heads=nhead_f,
mlp_hidden_coef=mlp_hidden_coef)
self.down_proj_m_f = nn.Linear(block_dims[1], block_dims[0], bias=False)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def pre_process(self, feat_f0, feat_f1, feat_m0, feat_m1, feat_c0, feat_c1, feat_c0_pre, feat_c1_pre, data):
W_f = self.W_f
W_m = self.W_m
data.update({'W_f': W_f,
'W_m': W_m})
# merge coarse features before and after loftr layer, and down proj channel dimesions
feat_c0 = rearrange(feat_c0, 'n (h w) c -> n c h w', h =data["hw0_c"][0], w =data["hw0_c"][1])
feat_c1 = rearrange(feat_c1, 'n (h w) c -> n c h w', h =data["hw1_c"][0], w =data["hw1_c"][1])
feat_c0 = self.conv_merge(torch.cat([feat_c0, feat_c0_pre], dim=1))
feat_c1 = self.conv_merge(torch.cat([feat_c1, feat_c1_pre], dim=1))
feat_c0 = rearrange(feat_c0, 'n c h w -> n (h w) 1 c')
feat_c1 = rearrange(feat_c1, 'n c h w -> n (h w) 1 c')
stride_f = data['hw0_f'][0] // data['hw0_c'][0]
stride_m = data['hw0_m'][0] // data['hw0_c'][0]
if feat_m0.shape[2] == feat_m1.shape[2] and feat_m0.shape[3] == feat_m1.shape[3]:
feat_m = self.out_conv_m(torch.cat([feat_m0, feat_m1], dim=0))
feat_m0, feat_m1 = torch.chunk(feat_m, 2, dim=0)
feat_f = self.out_conv_f(torch.cat([feat_f0, feat_f1], dim=0))
feat_f0, feat_f1 = torch.chunk(feat_f, 2, dim=0)
else:
feat_m0 = self.out_conv_m(feat_m0)
feat_m1 = self.out_conv_m(feat_m1)
feat_f0 = self.out_conv_f(feat_f0)
feat_f1 = self.out_conv_f(feat_f1)
# 1. unfold (crop windows) all local windows
feat_m0_unfold = F.unfold(feat_m0, kernel_size=(W_m, W_m), stride=stride_m, padding=W_m//2)
feat_m0_unfold = rearrange(feat_m0_unfold, 'n (c ww) l -> n l ww c', ww=W_m**2)
feat_m1_unfold = F.unfold(feat_m1, kernel_size=(W_m, W_m), stride=stride_m, padding=W_m//2)
feat_m1_unfold = rearrange(feat_m1_unfold, 'n (c ww) l -> n l ww c', ww=W_m**2)
feat_f0_unfold = F.unfold(feat_f0, kernel_size=(W_f, W_f), stride=stride_f, padding=W_f//2)
feat_f0_unfold = rearrange(feat_f0_unfold, 'n (c ww) l -> n l ww c', ww=W_f**2)
feat_f1_unfold = F.unfold(feat_f1, kernel_size=(W_f, W_f), stride=stride_f, padding=W_f//2)
feat_f1_unfold = rearrange(feat_f1_unfold, 'n (c ww) l -> n l ww c', ww=W_f**2)
# 2. select only the predicted matches
feat_c0 = feat_c0[data['b_ids'], data['i_ids']] # [n, ww, cm]
feat_c1 = feat_c1[data['b_ids'], data['j_ids']]
feat_m0_unfold = feat_m0_unfold[data['b_ids'], data['i_ids']] # [n, ww, cm]
feat_m1_unfold = feat_m1_unfold[data['b_ids'], data['j_ids']]
feat_f0_unfold = feat_f0_unfold[data['b_ids'], data['i_ids']] # [n, ww, cf]
feat_f1_unfold = feat_f1_unfold[data['b_ids'], data['j_ids']]
return feat_c0, feat_c1, feat_m0_unfold, feat_m1_unfold, feat_f0_unfold, feat_f1_unfold
def forward(self, feat_f0, feat_f1, feat_m0, feat_m1, feat_c0, feat_c1, feat_c0_pre, feat_c1_pre, data):
"""
Args:
feat_f0 (torch.Tensor): [N, C, H, W]
feat_f1 (torch.Tensor): [N, C, H, W]
feat_m0 (torch.Tensor): [N, C, H, W]
feat_m1 (torch.Tensor): [N, C, H, W]
feat_c0 (torch.Tensor): [N, L, C]
feat_c1 (torch.Tensor): [N, S, C]
feat_c0_pre (torch.Tensor): [N, C, H, W]
feat_c1_pre (torch.Tensor): [N, C, H, W]
data (dict): with keys ['hw0_c', 'hw1_c', 'hw0_m', 'hw1_m', 'hw0_f', 'hw1_f', 'b_ids', 'j_ids']
"""
# TODO: Check for this case
if data['b_ids'].shape[0] == 0:
feat0 = torch.empty(0, self.W_f**2, self.block_dims[0], device=feat_f0.device)
feat1 = torch.empty(0, self.W_f**2, self.block_dims[0], device=feat_f0.device)
return feat0, feat1
feat_c0, feat_c1, feat_m0_unfold, feat_m1_unfold, \
feat_f0_unfold, feat_f1_unfold = self.pre_process(feat_f0, feat_f1, feat_m0, feat_m1,
feat_c0, feat_c1, feat_c0_pre, feat_c1_pre, data)
# self attention (c + m)
feat_m_unfold, _ = self.self_attn_m(torch.cat([feat_m0_unfold, feat_m1_unfold], dim=0),
torch.cat([feat_c0, feat_c1], dim=0))
feat_m0_unfold, feat_m1_unfold = torch.chunk(feat_m_unfold, 2, dim=0)
# cross attention (m0 <-> m1)
feat_m0_unfold, feat_m1_unfold = self.cross_attn_m(feat_m0_unfold, feat_m1_unfold)
# down proj m
feat_m_unfold = self.down_proj_m_f(torch.cat([feat_m0_unfold, feat_m1_unfold], dim=0))
feat_m0_unfold, feat_m1_unfold = torch.chunk(feat_m_unfold, 2, dim=0)
# self attention (m + f)
feat_f_unfold, _ = self.self_attn_f(torch.cat([feat_f0_unfold, feat_f1_unfold], dim=0),
torch.cat([feat_m0_unfold, feat_m1_unfold], dim=0))
feat_f0_unfold, feat_f1_unfold = torch.chunk(feat_f_unfold, 2, dim=0)
# cross attention (f0 <-> f1)
feat_f0_unfold, feat_f1_unfold = self.cross_attn_f(feat_f0_unfold, feat_f1_unfold)
return feat_f0_unfold, feat_f1_unfold
|