Spaces:
Running
on
Zero
Running
on
Zero
File size: 13,033 Bytes
b5ce381 |
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 |
import torch
from einops import repeat
from ..modules.attention import *
from ..modules.diffusionmodules.util import AlphaBlender, linear, timestep_embedding
class TimeMixSequential(nn.Sequential):
def forward(self, x, context=None, timesteps=None):
for layer in self:
x = layer(x, context, timesteps)
return x
class VideoTransformerBlock(nn.Module):
ATTENTION_MODES = {
"softmax": CrossAttention,
"softmax-xformers": MemoryEfficientCrossAttention,
}
def __init__(
self,
dim,
n_heads,
d_head,
dropout=0.0,
context_dim=None,
gated_ff=True,
checkpoint=True,
timesteps=None,
ff_in=False,
inner_dim=None,
attn_mode="softmax",
disable_self_attn=False,
disable_temporal_crossattention=False,
switch_temporal_ca_to_sa=False,
):
super().__init__()
attn_cls = self.ATTENTION_MODES[attn_mode]
self.ff_in = ff_in or inner_dim is not None
if inner_dim is None:
inner_dim = dim
assert int(n_heads * d_head) == inner_dim
self.is_res = inner_dim == dim
if self.ff_in:
self.norm_in = nn.LayerNorm(dim)
self.ff_in = FeedForward(dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff)
self.timesteps = timesteps
self.disable_self_attn = disable_self_attn
if self.disable_self_attn:
self.attn1 = attn_cls(
query_dim=inner_dim,
heads=n_heads,
dim_head=d_head,
context_dim=context_dim,
dropout=dropout,
) # is a cross-attention
else:
self.attn1 = attn_cls(
query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout
) # is a self-attention
self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff)
if disable_temporal_crossattention:
if switch_temporal_ca_to_sa:
raise ValueError
else:
self.attn2 = None
else:
self.norm2 = nn.LayerNorm(inner_dim)
if switch_temporal_ca_to_sa:
self.attn2 = attn_cls(
query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout
) # is a self-attention
else:
self.attn2 = attn_cls(
query_dim=inner_dim,
context_dim=context_dim,
heads=n_heads,
dim_head=d_head,
dropout=dropout,
) # is self-attn if context is none
self.norm1 = nn.LayerNorm(inner_dim)
self.norm3 = nn.LayerNorm(inner_dim)
self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa
self.checkpoint = checkpoint
if self.checkpoint:
print(f"{self.__class__.__name__} is using checkpointing")
def forward(
self, x: torch.Tensor, context: torch.Tensor = None, timesteps: int = None, skip_attention=False
) -> torch.Tensor:
if self.checkpoint:
return checkpoint(self._forward, x, context, timesteps, skip_attention)
else:
return self._forward(x, context, timesteps=timesteps, skip_attention=skip_attention)
def _forward(self, x, context=None, timesteps=None, skip_attention=None):
assert self.timesteps or timesteps
assert not (self.timesteps and timesteps) or self.timesteps == timesteps
timesteps = self.timesteps or timesteps
B, S, C = x.shape
x = rearrange(x, "(b t) s c -> (b s) t c", t=timesteps)
if skip_attention is not None:
skip_attention = repeat(skip_attention[: B // timesteps], "b ... -> (b s) ...", s=S)
if self.ff_in:
x_skip = x
x = self.ff_in(self.norm_in(x))
if self.is_res:
x += x_skip
if self.disable_self_attn:
x = self.attn1(self.norm1(x), context=context) + x
else:
x = self.attn1(self.norm1(x), skip_attention=skip_attention) + x
if self.attn2 is not None:
if self.switch_temporal_ca_to_sa:
x = self.attn2(self.norm2(x), skip_attention=skip_attention) + x
else:
x = self.attn2(self.norm2(x), context=context) + x
x_skip = x
x = self.ff(self.norm3(x))
if self.is_res:
x += x_skip
x = rearrange(x, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps)
return x
def get_last_layer(self):
return self.ff.net[-1].weight
class SpatialVideoTransformer(SpatialTransformer):
def __init__(
self,
in_channels,
n_heads,
d_head,
depth=1,
dropout=0.0,
use_linear=False,
context_dim=None,
use_spatial_context=False,
timesteps=None,
merge_strategy: str = "fixed",
merge_factor: float = 0.5,
merge_audio_factor: float = 5.0, # Almost 0 audio at first
time_context_dim=None,
audio_context_dim=None,
ff_in=False,
checkpoint=False,
time_depth=1,
attn_mode="softmax",
disable_self_attn=False,
disable_temporal_crossattention=False,
max_time_embed_period: int = 10000,
skip_time=False,
reference_to=None,
):
super().__init__(
in_channels,
n_heads,
d_head,
depth=depth,
dropout=dropout,
attn_type=attn_mode,
use_checkpoint=checkpoint,
context_dim=context_dim,
use_linear=use_linear,
disable_self_attn=disable_self_attn,
reference_to=reference_to,
)
self.time_depth = time_depth
self.depth = depth
self.max_time_embed_period = max_time_embed_period
self.skip_time = skip_time
time_mix_d_head = d_head
n_time_mix_heads = n_heads
time_mix_inner_dim = int(time_mix_d_head * n_time_mix_heads)
inner_dim = n_heads * d_head
if use_spatial_context:
time_context_dim = context_dim
if not self.skip_time:
self.time_stack = nn.ModuleList(
[
VideoTransformerBlock(
inner_dim,
n_time_mix_heads,
time_mix_d_head,
dropout=dropout,
context_dim=time_context_dim,
timesteps=timesteps,
checkpoint=checkpoint,
ff_in=ff_in,
inner_dim=time_mix_inner_dim,
attn_mode=attn_mode,
disable_self_attn=disable_self_attn,
disable_temporal_crossattention=disable_temporal_crossattention,
)
for _ in range(self.depth)
]
)
else:
self.time_stack = None
self.audio_stack = None
if audio_context_dim is not None:
self.audio_stack = nn.ModuleList(
[
VideoTransformerBlock(
inner_dim,
n_time_mix_heads,
time_mix_d_head,
dropout=dropout,
context_dim=audio_context_dim,
timesteps=timesteps,
checkpoint=checkpoint,
ff_in=ff_in,
inner_dim=time_mix_inner_dim,
attn_mode=attn_mode,
disable_self_attn=disable_self_attn,
disable_temporal_crossattention=disable_temporal_crossattention or self.skip_time,
)
for _ in range(self.depth)
]
)
self.audio_mixer = AlphaBlender(alpha=merge_audio_factor, merge_strategy=merge_strategy)
if self.time_stack is None:
self.time_stack = [None] * len(self.transformer_blocks)
assert len(self.time_stack) == len(self.transformer_blocks)
self.use_spatial_context = use_spatial_context
self.in_channels = in_channels
time_embed_dim = self.in_channels * 4
self.time_pos_embed = nn.Sequential(
linear(self.in_channels, time_embed_dim),
nn.SiLU(),
linear(time_embed_dim, self.in_channels),
)
self.time_mixer = AlphaBlender(alpha=merge_factor, merge_strategy=merge_strategy)
def forward(
self,
x: torch.Tensor,
context: Optional[torch.Tensor] = None,
reference_context: Optional[torch.Tensor] = None,
time_context: Optional[torch.Tensor] = None,
audio_context: Optional[torch.Tensor] = None,
timesteps: Optional[int] = None,
image_only_indicator: Optional[torch.Tensor] = None,
skip_spatial_attention: bool = False,
skip_temporal_attention: bool = False,
) -> torch.Tensor:
_, _, h, w = x.shape
x_in = x
spatial_context = None
if exists(context):
spatial_context = context
if not isinstance(spatial_context, list):
spatial_context = [spatial_context]
if reference_context is not None and not isinstance(reference_context, list):
reference_context = [reference_context]
# else:
# # spatial_context.reverse()
# print([c.shape for c in spatial_context])
if self.use_spatial_context and not self.skip_time:
assert (
isinstance(context, list) or context.ndim == 3
), f"n dims of spatial context should be 3 but are {context.ndim}"
time_context = context
if not isinstance(context, list):
time_context_first_timestep = time_context[::timesteps]
time_context = repeat(time_context_first_timestep, "b ... -> (b n) ...", n=h * w)
elif time_context is not None and not self.use_spatial_context:
time_context = repeat(time_context, "b ... -> (b n) ...", n=h * w)
if time_context.ndim == 2:
time_context = rearrange(time_context, "b c -> b 1 c")
if audio_context is not None:
audio_context = repeat(audio_context, "b ... -> (b n) ...", n=h * w)
if audio_context.ndim == 2:
audio_context = rearrange(audio_context, "b c -> b 1 c")
x = self.norm(x)
if not self.use_linear:
x = self.proj_in(x)
x = rearrange(x, "b c h w -> b (h w) c")
if self.use_linear:
x = self.proj_in(x)
if not self.skip_time:
num_frames = torch.arange(timesteps, device=x.device, dtype=x.dtype)
num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
num_frames = rearrange(num_frames, "b t -> (b t)")
t_emb = timestep_embedding(
num_frames,
self.in_channels,
repeat_only=False,
max_period=self.max_time_embed_period,
)
emb = self.time_pos_embed(t_emb)
emb = emb[:, None, :]
for it_, (block, mix_block) in enumerate(zip(self.transformer_blocks, self.time_stack)):
if it_ > 0 and len(spatial_context) == 1:
it_ = 0 # use same context for each block
x = block(
x,
context=spatial_context[it_],
reference_context=reference_context[it_] if reference_context is not None else None,
skip_attention=skip_spatial_attention,
)
if not self.skip_time:
x_mix = x
x_mix = x_mix + emb
x_mix = mix_block(
x_mix,
context=time_context,
timesteps=timesteps,
skip_attention=skip_temporal_attention,
)
x = self.time_mixer(
x_spatial=x,
x_temporal=x_mix,
image_only_indicator=image_only_indicator,
)
if self.audio_stack is not None:
audio_mix_block = self.audio_stack[it_]
x_audio = x
# x_audio = x_audio + emb
x_audio = audio_mix_block(x_audio, context=audio_context, timesteps=timesteps)
x = self.audio_mixer(x, x_audio, image_only_indicator=image_only_indicator)
if self.use_linear:
x = self.proj_out(x)
x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
if not self.use_linear:
x = self.proj_out(x)
out = x + x_in
return out
|