Spaces:
Running
on
Zero
Running
on
Zero
File size: 16,684 Bytes
1ae4e5b 279a838 1ae4e5b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 |
import torch
import torch.nn as nn
import torch.nn.init as init
import logging
from ddiffusers.models.lora import LoRALinearLayer
from ddiffusers.models.attention import Attention
from ddiffusers.utils import USE_PEFT_BACKEND
from typing import Optional
from einops import rearrange
logger = logging.getLogger(__name__)
class AttnProcessor:
r"""
Default processor for performing attention-related computations.
"""
def __call__(
self,
attn: Attention,
hidden_states: torch.FloatTensor,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
temb: Optional[torch.FloatTensor] = None,
scale: float = 1.0,
camera_feature=None
) -> torch.Tensor:
residual = hidden_states
args = () if USE_PEFT_BACKEND else (scale,)
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states, *args)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states, *args)
value = attn.to_v(encoder_hidden_states, *args)
query = attn.head_to_batch_dim(query)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
attention_probs = attn.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states, *args)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class LoRAAttnProcessor(nn.Module):
r"""
Default processor for performing attention-related computations.
"""
def __init__(
self,
hidden_size=None,
cross_attention_dim=None,
rank=4,
network_alpha=None,
lora_scale=1.0,
):
super().__init__()
self.rank = rank
self.lora_scale = lora_scale
self.to_q_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
self.to_v_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
self.to_out_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
camera_feature=None,
scale=None
):
lora_scale = self.lora_scale if scale is None else scale
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states) + lora_scale * self.to_q_lora(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states) + lora_scale * self.to_k_lora(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states) + lora_scale * self.to_v_lora(encoder_hidden_states)
query = attn.head_to_batch_dim(query)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
attention_probs = attn.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states) + lora_scale * self.to_out_lora(hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class CameraAdaptorAttnProcessor(nn.Module):
def __init__(self,
hidden_size, # dimension of hidden state
camera_feature_dim=None, # dimension of the camera feature
cross_attention_dim=None, # dimension of the text embedding
query_condition=False,
key_value_condition=False,
scale=1.0):
super().__init__()
self.hidden_size = hidden_size
self.camera_feature_dim = camera_feature_dim
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.query_condition = query_condition
self.key_value_condition = key_value_condition
assert hidden_size == camera_feature_dim
if self.query_condition and self.key_value_condition:
self.qkv_merge = nn.Linear(hidden_size, hidden_size)
init.zeros_(self.qkv_merge.weight)
init.zeros_(self.qkv_merge.bias)
elif self.query_condition:
self.q_merge = nn.Linear(hidden_size, hidden_size)
init.zeros_(self.q_merge.weight)
init.zeros_(self.q_merge.bias)
else:
self.kv_merge = nn.Linear(hidden_size, hidden_size)
init.zeros_(self.kv_merge.weight)
init.zeros_(self.kv_merge.bias)
def forward(self,
attn,
hidden_states,
camera_feature,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
scale=None,):
assert camera_feature is not None
camera_embedding_scale = (scale or self.scale)
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
if hidden_states.dim == 5:
hidden_states = rearrange(hidden_states, 'b c f h w -> (b f) (h w) c')
elif hidden_states.ndim == 4:
hidden_states = rearrange(hidden_states, 'b c h w -> b (h w) c')
else:
assert hidden_states.ndim == 3
if self.query_condition and self.key_value_condition:
assert encoder_hidden_states is None
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
if encoder_hidden_states.ndim == 5:
encoder_hidden_states = rearrange(encoder_hidden_states, 'b c f h w -> (b f) (h w) c')
elif encoder_hidden_states.ndim == 4:
encoder_hidden_states = rearrange(encoder_hidden_states, 'b c h w -> b (h w) c')
else:
assert encoder_hidden_states.ndim == 3
if camera_feature.ndim == 5:
camera_feature = rearrange(camera_feature, "b c f h w -> (b f) (h w) c")
elif camera_feature.ndim == 4:
camera_feature = rearrange(camera_feature, "b c h w -> b (h w) c")
else:
assert camera_feature.ndim == 3
batch_size, ehs_sequence_length, _ = encoder_hidden_states.shape
attention_mask = attn.prepare_attention_mask(attention_mask, ehs_sequence_length, batch_size)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
if self.query_condition and self.key_value_condition: # only self attention
query_hidden_state = self.qkv_merge(hidden_states + camera_feature) * camera_embedding_scale + hidden_states
key_value_hidden_state = query_hidden_state
elif self.query_condition:
query_hidden_state = self.q_merge(hidden_states + camera_feature) * camera_embedding_scale + hidden_states
key_value_hidden_state = encoder_hidden_states
else:
key_value_hidden_state = self.kv_merge(encoder_hidden_states + camera_feature) * camera_embedding_scale + encoder_hidden_states
query_hidden_state = hidden_states
# original attention
query = attn.to_q(query_hidden_state)
key = attn.to_k(key_value_hidden_state)
value = attn.to_v(key_value_hidden_state)
query = attn.head_to_batch_dim(query)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
attention_probs = attn.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class LORACameraAdaptorAttnProcessor(nn.Module):
def __init__(self,
hidden_size, # dimension of hidden state
camera_feature_dim=None, # dimension of the camera feature
cross_attention_dim=None, # dimension of the text embedding
query_condition=False,
key_value_condition=False,
scale=1.0,
# lora keywords
rank=4,
network_alpha=None,
lora_scale=1.0):
super().__init__()
self.hidden_size = hidden_size
self.camera_feature_dim = camera_feature_dim
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.query_condition = query_condition
self.key_value_condition = key_value_condition
assert hidden_size == camera_feature_dim
if self.query_condition and self.key_value_condition:
self.qkv_merge = nn.Linear(hidden_size, hidden_size)
init.zeros_(self.qkv_merge.weight)
init.zeros_(self.qkv_merge.bias)
elif self.query_condition:
self.q_merge = nn.Linear(hidden_size, hidden_size)
init.zeros_(self.q_merge.weight)
init.zeros_(self.q_merge.bias)
else:
self.kv_merge = nn.Linear(hidden_size, hidden_size)
init.zeros_(self.kv_merge.weight)
init.zeros_(self.kv_merge.bias)
# lora
self.rank = rank
self.lora_scale = lora_scale
self.to_q_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
self.to_v_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
self.to_out_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
def __call__(self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
scale=1.0,
camera_feature=None,
):
assert camera_feature is not None
lora_scale = self.lora_scale if scale is None else scale
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
if hidden_states.dim == 5:
hidden_states = rearrange(hidden_states, 'b c f h w -> (b f) (h w) c')
elif hidden_states.ndim == 4:
hidden_states = rearrange(hidden_states, 'b c h w -> b (h w) c')
else:
assert hidden_states.ndim == 3
if self.query_condition and self.key_value_condition:
assert encoder_hidden_states is None
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
if encoder_hidden_states.ndim == 5:
encoder_hidden_states = rearrange(encoder_hidden_states, 'b c f h w -> (b f) (h w) c')
elif encoder_hidden_states.ndim == 4:
encoder_hidden_states = rearrange(encoder_hidden_states, 'b c h w -> b (h w) c')
else:
assert encoder_hidden_states.ndim == 3
if camera_feature.ndim == 5:
camera_feature = rearrange(camera_feature, "b c f h w -> (b f) (h w) c")
elif camera_feature.ndim == 4:
camera_feature = rearrange(camera_feature, "b c h w -> b (h w) c")
else:
assert camera_feature.ndim == 3
batch_size, ehs_sequence_length, _ = encoder_hidden_states.shape
attention_mask = attn.prepare_attention_mask(attention_mask, ehs_sequence_length, batch_size)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
if self.query_condition and self.key_value_condition: # only self attention
query_hidden_state = self.qkv_merge(hidden_states + camera_feature) * self.scale + hidden_states
key_value_hidden_state = query_hidden_state
elif self.query_condition:
query_hidden_state = self.q_merge(hidden_states + camera_feature) * self.scale + hidden_states
key_value_hidden_state = encoder_hidden_states
else:
key_value_hidden_state = self.kv_merge(encoder_hidden_states + camera_feature) * self.scale + encoder_hidden_states
query_hidden_state = hidden_states
# original attention
query = attn.to_q(query_hidden_state) + lora_scale * self.to_q_lora(query_hidden_state)
key = attn.to_k(key_value_hidden_state) + lora_scale * self.to_k_lora(key_value_hidden_state)
value = attn.to_v(key_value_hidden_state) + lora_scale * self.to_v_lora(key_value_hidden_state)
query = attn.head_to_batch_dim(query)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
attention_probs = attn.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states) + lora_scale * self.to_out_lora(hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
|