Spaces:
Runtime error
Runtime error
File size: 11,979 Bytes
da3eeba |
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 |
# 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.
import math
import warnings
from functools import partial
from typing import Tuple, Type
import torch
import torch.nn.functional as F
from torch import nn, Tensor
from sam2.modeling.position_encoding import apply_rotary_enc, compute_axial_cis
from sam2.modeling.sam2_utils import MLP
from sam2.utils.misc import get_sdpa_settings
warnings.simplefilter(action="ignore", category=FutureWarning)
OLD_GPU, USE_FLASH_ATTN, MATH_KERNEL_ON = get_sdpa_settings()
class TwoWayTransformer(nn.Module):
def __init__(
self,
depth: int,
embedding_dim: int,
num_heads: int,
mlp_dim: int,
activation: Type[nn.Module] = nn.ReLU,
attention_downsample_rate: int = 2,
) -> None:
"""
A transformer decoder that attends to an input image using
queries whose positional embedding is supplied.
Args:
depth (int): number of layers in the transformer
embedding_dim (int): the channel dimension for the input embeddings
num_heads (int): the number of heads for multihead attention. Must
divide embedding_dim
mlp_dim (int): the channel dimension internal to the MLP block
activation (nn.Module): the activation to use in the MLP block
"""
super().__init__()
self.depth = depth
self.embedding_dim = embedding_dim
self.num_heads = num_heads
self.mlp_dim = mlp_dim
self.layers = nn.ModuleList()
for i in range(depth):
self.layers.append(
TwoWayAttentionBlock(
embedding_dim=embedding_dim,
num_heads=num_heads,
mlp_dim=mlp_dim,
activation=activation,
attention_downsample_rate=attention_downsample_rate,
skip_first_layer_pe=(i == 0),
)
)
self.final_attn_token_to_image = Attention(
embedding_dim, num_heads, downsample_rate=attention_downsample_rate
)
self.norm_final_attn = nn.LayerNorm(embedding_dim)
def forward(
self,
image_embedding: Tensor,
image_pe: Tensor,
point_embedding: Tensor,
) -> Tuple[Tensor, Tensor]:
"""
Args:
image_embedding (torch.Tensor): image to attend to. Should be shape
B x embedding_dim x h x w for any h and w.
image_pe (torch.Tensor): the positional encoding to add to the image. Must
have the same shape as image_embedding.
point_embedding (torch.Tensor): the embedding to add to the query points.
Must have shape B x N_points x embedding_dim for any N_points.
Returns:
torch.Tensor: the processed point_embedding
torch.Tensor: the processed image_embedding
"""
# BxCxHxW -> BxHWxC == B x N_image_tokens x C
bs, c, h, w = image_embedding.shape
image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
image_pe = image_pe.flatten(2).permute(0, 2, 1)
# Prepare queries
queries = point_embedding
keys = image_embedding
# Apply transformer blocks and final layernorm
for layer in self.layers:
queries, keys = layer(
queries=queries,
keys=keys,
query_pe=point_embedding,
key_pe=image_pe,
)
# Apply the final attention layer from the points to the image
q = queries + point_embedding
k = keys + image_pe
attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
queries = queries + attn_out
queries = self.norm_final_attn(queries)
return queries, keys
class TwoWayAttentionBlock(nn.Module):
def __init__(
self,
embedding_dim: int,
num_heads: int,
mlp_dim: int = 2048,
activation: Type[nn.Module] = nn.ReLU,
attention_downsample_rate: int = 2,
skip_first_layer_pe: bool = False,
) -> None:
"""
A transformer block with four layers: (1) self-attention of sparse
inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
block on sparse inputs, and (4) cross attention of dense inputs to sparse
inputs.
Arguments:
embedding_dim (int): the channel dimension of the embeddings
num_heads (int): the number of heads in the attention layers
mlp_dim (int): the hidden dimension of the mlp block
activation (nn.Module): the activation of the mlp block
skip_first_layer_pe (bool): skip the PE on the first layer
"""
super().__init__()
self.self_attn = Attention(embedding_dim, num_heads)
self.norm1 = nn.LayerNorm(embedding_dim)
self.cross_attn_token_to_image = Attention(
embedding_dim, num_heads, downsample_rate=attention_downsample_rate
)
self.norm2 = nn.LayerNorm(embedding_dim)
self.mlp = MLP(
embedding_dim, mlp_dim, embedding_dim, num_layers=2, activation=activation
)
self.norm3 = nn.LayerNorm(embedding_dim)
self.norm4 = nn.LayerNorm(embedding_dim)
self.cross_attn_image_to_token = Attention(
embedding_dim, num_heads, downsample_rate=attention_downsample_rate
)
self.skip_first_layer_pe = skip_first_layer_pe
def forward(
self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
) -> Tuple[Tensor, Tensor]:
# Self attention block
if self.skip_first_layer_pe:
queries = self.self_attn(q=queries, k=queries, v=queries)
else:
q = queries + query_pe
attn_out = self.self_attn(q=q, k=q, v=queries)
queries = queries + attn_out
queries = self.norm1(queries)
# Cross attention block, tokens attending to image embedding
q = queries + query_pe
k = keys + key_pe
attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
queries = queries + attn_out
queries = self.norm2(queries)
# MLP block
mlp_out = self.mlp(queries)
queries = queries + mlp_out
queries = self.norm3(queries)
# Cross attention block, image embedding attending to tokens
q = queries + query_pe
k = keys + key_pe
attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
keys = keys + attn_out
keys = self.norm4(keys)
return queries, keys
class Attention(nn.Module):
"""
An attention layer that allows for downscaling the size of the embedding
after projection to queries, keys, and values.
"""
def __init__(
self,
embedding_dim: int,
num_heads: int,
downsample_rate: int = 1,
dropout: float = 0.0,
kv_in_dim: int = None,
) -> None:
super().__init__()
self.embedding_dim = embedding_dim
self.kv_in_dim = kv_in_dim if kv_in_dim is not None else embedding_dim
self.internal_dim = embedding_dim // downsample_rate
self.num_heads = num_heads
assert (
self.internal_dim % num_heads == 0
), "num_heads must divide embedding_dim."
self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
self.k_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
self.v_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
self.dropout_p = dropout
def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
b, n, c = x.shape
x = x.reshape(b, n, num_heads, c // num_heads)
return x.transpose(1, 2) # B x N_heads x N_tokens x C_per_head
def _recombine_heads(self, x: Tensor) -> Tensor:
b, n_heads, n_tokens, c_per_head = x.shape
x = x.transpose(1, 2)
return x.reshape(b, n_tokens, n_heads * c_per_head) # B x N_tokens x C
def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
# Input projections
q = self.q_proj(q)
k = self.k_proj(k)
v = self.v_proj(v)
# Separate into heads
q = self._separate_heads(q, self.num_heads)
k = self._separate_heads(k, self.num_heads)
v = self._separate_heads(v, self.num_heads)
dropout_p = self.dropout_p if self.training else 0.0
# Attention
# with torch.backends.cuda.sdp_kernel(
# enable_flash=USE_FLASH_ATTN,
# # if Flash attention kernel is off, then math kernel needs to be enabled
# enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
# enable_mem_efficient=OLD_GPU,
# ):
# out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
out = self._recombine_heads(out)
out = self.out_proj(out)
return out
class RoPEAttention(Attention):
"""Attention with rotary position encoding."""
def __init__(
self,
*args,
rope_theta=10000.0,
# whether to repeat q rope to match k length
# this is needed for cross-attention to memories
rope_k_repeat=False,
feat_sizes=(32, 32), # [w, h] for stride 16 feats at 512 resolution
**kwargs,
):
super().__init__(*args, **kwargs)
self.compute_cis = partial(
compute_axial_cis, dim=self.internal_dim // self.num_heads, theta=rope_theta
)
freqs_cis = self.compute_cis(end_x=feat_sizes[0], end_y=feat_sizes[1])
self.freqs_cis = freqs_cis
self.rope_k_repeat = rope_k_repeat
def forward(
self, q: Tensor, k: Tensor, v: Tensor, num_k_exclude_rope: int = 0
) -> Tensor:
# Input projections
q = self.q_proj(q)
k = self.k_proj(k)
v = self.v_proj(v)
# Separate into heads
q = self._separate_heads(q, self.num_heads)
k = self._separate_heads(k, self.num_heads)
v = self._separate_heads(v, self.num_heads)
# Apply rotary position encoding
w = h = math.sqrt(q.shape[-2])
self.freqs_cis = self.freqs_cis.to(q.device)
if self.freqs_cis.shape[0] != q.shape[-2]:
self.freqs_cis = self.compute_cis(end_x=w, end_y=h).to(q.device)
if q.shape[-2] != k.shape[-2]:
assert self.rope_k_repeat
num_k_rope = k.size(-2) - num_k_exclude_rope
q, k[:, :, :num_k_rope] = apply_rotary_enc(
q,
k[:, :, :num_k_rope],
freqs_cis=self.freqs_cis,
repeat_freqs_k=self.rope_k_repeat,
)
dropout_p = self.dropout_p if self.training else 0.0
# Attention
# with torch.backends.cuda.sdp_kernel(
# enable_flash=USE_FLASH_ATTN,
# # if Flash attention kernel is off, then math kernel needs to be enabled
# enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
# enable_mem_efficient=OLD_GPU,
# ):
# out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
out = self._recombine_heads(out)
out = self.out_proj(out)
return out
|