File size: 10,479 Bytes
8b4c6c7 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
from functools import partial
from typing import List, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from sam2.modeling.backbones.utils import (
PatchEmbed,
window_partition,
window_unpartition,
)
from sam2.modeling.sam2_utils import DropPath, MLP
def do_pool(x: torch.Tensor, pool: nn.Module, norm: nn.Module = None) -> torch.Tensor:
if pool is None:
return x
# (B, H, W, C) -> (B, C, H, W)
x = x.permute(0, 3, 1, 2)
x = pool(x)
# (B, C, H', W') -> (B, H', W', C)
x = x.permute(0, 2, 3, 1)
if norm:
x = norm(x)
return x
class MultiScaleAttention(nn.Module):
def __init__(
self,
dim: int,
dim_out: int,
num_heads: int,
q_pool: nn.Module = None,
):
super().__init__()
self.dim = dim
self.dim_out = dim_out
self.num_heads = num_heads
head_dim = dim_out // num_heads
self.scale = head_dim**-0.5
self.q_pool = q_pool
self.qkv = nn.Linear(dim, dim_out * 3)
self.proj = nn.Linear(dim_out, dim_out)
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, H, W, _ = x.shape
# qkv with shape (B, H * W, 3, nHead, C)
qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1)
# q, k, v with shape (B, H * W, nheads, C)
q, k, v = torch.unbind(qkv, 2)
# Q pooling (for downsample at stage changes)
if self.q_pool:
q = do_pool(q.reshape(B, H, W, -1), self.q_pool)
H, W = q.shape[1:3] # downsampled shape
q = q.reshape(B, H * W, self.num_heads, -1)
# Torch's SDPA expects [B, nheads, H*W, C] so we transpose
x = F.scaled_dot_product_attention(
q.transpose(1, 2),
k.transpose(1, 2),
v.transpose(1, 2),
)
# Transpose back
x = x.transpose(1, 2)
x = x.reshape(B, H, W, -1)
x = self.proj(x)
return x
class MultiScaleBlock(nn.Module):
def __init__(
self,
dim: int,
dim_out: int,
num_heads: int,
mlp_ratio: float = 4.0,
drop_path: float = 0.0,
norm_layer: Union[nn.Module, str] = "LayerNorm",
q_stride: Tuple[int, int] = None,
act_layer: nn.Module = nn.GELU,
window_size: int = 0,
):
super().__init__()
if isinstance(norm_layer, str):
norm_layer = partial(getattr(nn, norm_layer), eps=1e-6)
self.dim = dim
self.dim_out = dim_out
self.norm1 = norm_layer(dim)
self.window_size = window_size
self.pool, self.q_stride = None, q_stride
if self.q_stride:
self.pool = nn.MaxPool2d(
kernel_size=q_stride, stride=q_stride, ceil_mode=False
)
self.attn = MultiScaleAttention(
dim,
dim_out,
num_heads=num_heads,
q_pool=self.pool,
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.norm2 = norm_layer(dim_out)
self.mlp = MLP(
dim_out,
int(dim_out * mlp_ratio),
dim_out,
num_layers=2,
activation=act_layer,
)
if dim != dim_out:
self.proj = nn.Linear(dim, dim_out)
def forward(self, x: torch.Tensor) -> torch.Tensor:
shortcut = x # B, H, W, C
x = self.norm1(x)
# Skip connection
if self.dim != self.dim_out:
shortcut = do_pool(self.proj(x), self.pool)
# Window partition
window_size = self.window_size
if window_size > 0:
H, W = x.shape[1], x.shape[2]
x, pad_hw = window_partition(x, window_size)
# Window Attention + Q Pooling (if stage change)
x = self.attn(x)
if self.q_stride:
# Shapes have changed due to Q pooling
window_size = self.window_size // self.q_stride[0]
H, W = shortcut.shape[1:3]
pad_h = (window_size - H % window_size) % window_size
pad_w = (window_size - W % window_size) % window_size
pad_hw = (H + pad_h, W + pad_w)
# Reverse window partition
if self.window_size > 0:
x = window_unpartition(x, window_size, pad_hw, (H, W))
x = shortcut + self.drop_path(x)
# MLP
x = x + self.drop_path(self.mlp(self.norm2(x)))
return x
class Hiera(nn.Module):
"""
Reference: https://arxiv.org/abs/2306.00989
"""
def __init__(
self,
embed_dim: int = 96, # initial embed dim
num_heads: int = 1, # initial number of heads
drop_path_rate: float = 0.0, # stochastic depth
q_pool: int = 3, # number of q_pool stages
q_stride: Tuple[int, int] = (2, 2), # downsample stride bet. stages
stages: Tuple[int, ...] = (2, 3, 16, 3), # blocks per stage
dim_mul: float = 2.0, # dim_mul factor at stage shift
head_mul: float = 2.0, # head_mul factor at stage shift
window_pos_embed_bkg_spatial_size: Tuple[int, int] = (14, 14),
# window size per stage, when not using global att.
window_spec: Tuple[int, ...] = (
8,
4,
14,
7,
),
# global attn in these blocks
global_att_blocks: Tuple[int, ...] = (
12,
16,
20,
),
return_interm_layers=True, # return feats from every stage
):
super().__init__()
assert len(stages) == len(window_spec)
self.window_spec = window_spec
depth = sum(stages)
self.q_stride = q_stride
self.stage_ends = [sum(stages[:i]) - 1 for i in range(1, len(stages) + 1)]
assert 0 <= q_pool <= len(self.stage_ends[:-1])
self.q_pool_blocks = [x + 1 for x in self.stage_ends[:-1]][:q_pool]
self.return_interm_layers = return_interm_layers
self.patch_embed = PatchEmbed(
embed_dim=embed_dim,
)
# Which blocks have global att?
self.global_att_blocks = global_att_blocks
# Windowed positional embedding (https://arxiv.org/abs/2311.05613)
self.window_pos_embed_bkg_spatial_size = window_pos_embed_bkg_spatial_size
self.pos_embed = nn.Parameter(
torch.zeros(1, embed_dim, *self.window_pos_embed_bkg_spatial_size)
)
self.pos_embed_window = nn.Parameter(
torch.zeros(1, embed_dim, self.window_spec[0], self.window_spec[0])
)
dpr = [
x.item() for x in torch.linspace(0, drop_path_rate, depth)
] # stochastic depth decay rule
cur_stage = 1
self.blocks = nn.ModuleList()
for i in range(depth):
dim_out = embed_dim
# lags by a block, so first block of
# next stage uses an initial window size
# of previous stage and final window size of current stage
window_size = self.window_spec[cur_stage - 1]
if self.global_att_blocks is not None:
window_size = 0 if i in self.global_att_blocks else window_size
if i - 1 in self.stage_ends:
dim_out = int(embed_dim * dim_mul)
num_heads = int(num_heads * head_mul)
cur_stage += 1
block = MultiScaleBlock(
dim=embed_dim,
dim_out=dim_out,
num_heads=num_heads,
drop_path=dpr[i],
q_stride=self.q_stride if i in self.q_pool_blocks else None,
window_size=window_size,
)
embed_dim = dim_out
self.blocks.append(block)
self.channel_list = (
[self.blocks[i].dim_out for i in self.stage_ends[::-1]]
if return_interm_layers
else [self.blocks[-1].dim_out]
)
def _get_pos_embed(self, hw: Tuple[int, int]) -> torch.Tensor:
h, w = hw
window_embed = self.pos_embed_window
pos_embed = F.interpolate(self.pos_embed, size=(h, w), mode="bicubic")
pos_embed = pos_embed + window_embed.tile(
[x // y for x, y in zip(pos_embed.shape, window_embed.shape)]
)
pos_embed = pos_embed.permute(0, 2, 3, 1)
return pos_embed
def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
x = self.patch_embed(x)
# x: (B, H, W, C)
# Add pos embed
x = x + self._get_pos_embed(x.shape[1:3])
outputs = []
for i, blk in enumerate(self.blocks):
x = blk(x)
if (i == self.stage_ends[-1]) or (
i in self.stage_ends and self.return_interm_layers
):
feats = x.permute(0, 3, 1, 2)
outputs.append(feats)
return outputs
class HieraBBoxMask(Hiera):
def __init__(
self,
**kwargs,
) -> None:
super().__init__(**kwargs)
self.bbox_mask_patch_embed = PatchEmbed(
in_chans=4,
embed_dim=self.patch_embed.proj.out_channels,
)
def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
# x = self.patch_embed(x)
img, condition = x[0], x[1]
if condition is not None:
# concat mask and img as condition
bbox_mask = torch.zeros_like(img)[:, 0:1]
for i in range(condition.shape[0]):
l, u, r, d = condition[i, 0, :]
bbox_mask[i, :, int(u): int(d), int(l): int(r)] = 1.0
condition_input = torch.concat((img, bbox_mask), dim=1)
x = self.patch_embed(img) + self.bbox_mask_patch_embed(condition_input)
else:
x = self.patch_embed(img)
# x: (B, H, W, C)
# Add pos embed
x = x + self._get_pos_embed(x.shape[1:3])
outputs = []
for i, blk in enumerate(self.blocks):
x = blk(x)
if (i == self.stage_ends[-1]) or (
i in self.stage_ends and self.return_interm_layers
):
feats = x.permute(0, 3, 1, 2)
outputs.append(feats)
return outputs |