File size: 12,869 Bytes
e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f 3b2d4dc e505e1f | 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 | """Self-contained AFM generator for Hugging Face diffusers checkpoints."""
from __future__ import annotations
import math
from typing import Any, Dict, Optional
import torch
import torch.nn.functional as F
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.transformers.dit_transformer_2d import DiTTransformer2DModel
from torch import nn
def build_dit_config(
*,
depth: int = 28,
hidden_size: int = 1152,
patch_size: int = 2,
num_heads: int = 16,
learn_sigma: bool = False,
class_dropout_prob: float = 0.0,
input_size: int = 32,
num_classes: int = 1000,
in_channels: int = 4,
) -> Dict[str, Any]:
attention_head_dim = hidden_size // num_heads
num_embeds_ada_norm = num_classes + (1 if class_dropout_prob > 0 else 0)
out_channels = in_channels * 2 if learn_sigma else in_channels
return {
"sample_size": input_size,
"num_layers": depth,
"num_attention_heads": num_heads,
"attention_head_dim": attention_head_dim,
"in_channels": in_channels,
"out_channels": out_channels,
"patch_size": patch_size,
"attention_bias": True,
"activation_fn": "gelu-approximate",
"num_embeds_ada_norm": num_embeds_ada_norm,
"norm_type": "ada_norm_zero",
"norm_elementwise_affine": False,
"dropout": 0.0,
"norm_num_groups": 32,
"norm_eps": 1e-5,
"upcast_attention": False,
}
_PREFIX_KEYS = ("transformer.", "t_tgt_embedder.")
def _wrap_dit_state_dict(state_dict: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
return {k if k.startswith(_PREFIX_KEYS) else f"transformer.{k}": v for k, v in state_dict.items()}
def forward_dit_output(
transformer: DiTTransformer2DModel,
hidden_states: torch.Tensor,
timestep: Optional[torch.Tensor],
class_labels: torch.Tensor,
extra_cond: Optional[torch.Tensor] = None,
) -> torch.Tensor:
hidden_states = transformer.pos_embed(hidden_states)
for block in transformer.transformer_blocks:
if torch.is_grad_enabled() and transformer.gradient_checkpointing:
hidden_states = transformer._gradient_checkpointing_func(
block,
hidden_states,
None,
None,
None,
timestep,
None,
class_labels,
)
else:
hidden_states = block(
hidden_states,
attention_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
timestep=timestep,
cross_attention_kwargs=None,
class_labels=class_labels,
)
conditioning = transformer.transformer_blocks[0].norm1.emb(
timestep, class_labels, hidden_dtype=hidden_states.dtype
)
if extra_cond is not None:
conditioning = conditioning + extra_cond
shift, scale = transformer.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
hidden_states = transformer.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
hidden_states = transformer.proj_out_2(hidden_states)
height = width = int(hidden_states.shape[1] ** 0.5)
hidden_states = hidden_states.reshape(
shape=(-1, height, width, transformer.patch_size, transformer.patch_size, transformer.out_channels)
)
hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
return hidden_states.reshape(
shape=(-1, transformer.out_channels, height * transformer.patch_size, width * transformer.patch_size)
)
def split_output_channels(output: torch.Tensor, in_channels: int) -> torch.Tensor:
if output.shape[1] == in_channels:
return output
output, _ = output.chunk(2, dim=1)
return output
class TimestepEmbedder(nn.Module):
def __init__(self, hidden_size: int, frequency_embedding_size: int = 256):
super().__init__()
self.mlp = nn.Sequential(
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
nn.SiLU(),
nn.Linear(hidden_size, hidden_size, bias=True),
)
self.frequency_embedding_size = frequency_embedding_size
@staticmethod
def timestep_embedding(t: torch.Tensor, dim: int, max_period: int = 10000) -> torch.Tensor:
half = dim // 2
freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
device=t.device
)
args = t[:, None].float() * freqs[None]
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
if dim % 2:
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
return embedding
def forward(self, t: torch.Tensor) -> torch.Tensor:
emb = self.timestep_embedding(t.float(), self.frequency_embedding_size)
weight_dtype = self.mlp[0].weight.dtype
return self.mlp(emb.to(dtype=weight_dtype))
class _AFMCheckpointMixin(ModelMixin):
def _fix_state_dict_keys_on_load(self, state_dict):
wrapped = _wrap_dit_state_dict(state_dict)
state_dict.clear()
state_dict.update(wrapped)
super()._fix_state_dict_keys_on_load(state_dict)
class Generator(nn.Module):
def __init__(
self,
*,
use_t_src: bool = False,
use_t_tgt: bool = False,
depth: int = 28,
hidden_size: int = 1152,
patch_size: int = 2,
num_heads: int = 16,
learn_sigma: bool = False,
class_dropout_prob: float = 0.0,
input_size: int = 32,
num_classes: int = 1000,
in_channels: int = 4,
**_,
):
super().__init__()
self.use_t_src = use_t_src
self.use_t_tgt = use_t_tgt
self.transformer = DiTTransformer2DModel(
**build_dit_config(
depth=depth,
hidden_size=hidden_size,
patch_size=patch_size,
num_heads=num_heads,
learn_sigma=learn_sigma,
class_dropout_prob=class_dropout_prob,
input_size=input_size,
num_classes=num_classes,
in_channels=in_channels,
)
)
self.in_channels = in_channels
if use_t_tgt:
self.t_tgt_embedder = TimestepEmbedder(hidden_size)
nn.init.normal_(self.t_tgt_embedder.mlp[0].weight, std=0.02)
nn.init.normal_(self.t_tgt_embedder.mlp[2].weight, std=0.02)
def load_state_dict(self, state_dict, strict: bool = True):
if self.use_t_tgt and not any(key.startswith("t_tgt_embedder.") for key in state_dict):
strict = False
return super().load_state_dict(_wrap_dit_state_dict(state_dict), strict=strict)
def forward(self, x, y, t_src=None, t_tgt=None):
if self.use_t_src and t_src is not None:
timestep = (t_src * 1000).long()
else:
timestep = torch.zeros(x.shape[0], device=x.device, dtype=torch.long)
extra_cond = None
if self.use_t_tgt and t_tgt is not None:
extra_cond = self.t_tgt_embedder(t_tgt * 1000)
output = forward_dit_output(self.transformer, x, timestep, y, extra_cond=extra_cond)
return split_output_channels(output, self.in_channels)
class GeneratorDeep(nn.Module):
def __init__(
self,
*,
repeat: int,
depth: int = 28,
hidden_size: int = 1152,
patch_size: int = 2,
num_heads: int = 16,
learn_sigma: bool = False,
class_dropout_prob: float = 0.0,
input_size: int = 32,
num_classes: int = 1000,
in_channels: int = 4,
**_,
):
super().__init__()
self.repeat = repeat
self.transformer = DiTTransformer2DModel(
**build_dit_config(
depth=depth,
hidden_size=hidden_size,
patch_size=patch_size,
num_heads=num_heads,
learn_sigma=learn_sigma,
class_dropout_prob=class_dropout_prob,
input_size=input_size,
num_classes=num_classes,
in_channels=in_channels,
)
)
self.in_channels = in_channels
def load_state_dict(self, state_dict, strict: bool = True):
return super().load_state_dict(_wrap_dit_state_dict(state_dict), strict=strict)
def forward(self, x, y, *args, **kwargs):
hidden_states = self.transformer.pos_embed(x)
for t in torch.arange(1, 0, -1 / self.repeat).tolist():
timestep = torch.full([x.shape[0]], int(t * 1000), device=x.device, dtype=torch.long)
for block in self.transformer.transformer_blocks:
hidden_states = block(
hidden_states,
attention_mask=None,
encoder_hidden_states=None,
encoder_attention_mask=None,
timestep=timestep,
cross_attention_kwargs=None,
class_labels=y,
)
conditioning = self.transformer.transformer_blocks[0].norm1.emb(
torch.zeros(x.shape[0], device=x.device, dtype=torch.long),
y,
hidden_dtype=hidden_states.dtype,
)
shift, scale = self.transformer.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
hidden_states = self.transformer.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
hidden_states = self.transformer.proj_out_2(hidden_states)
height = width = int(hidden_states.shape[1] ** 0.5)
hidden_states = hidden_states.reshape(
shape=(
-1,
height,
width,
self.transformer.patch_size,
self.transformer.patch_size,
self.transformer.out_channels,
)
)
hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
output = hidden_states.reshape(
shape=(-1, self.transformer.out_channels, height * self.transformer.patch_size, width * self.transformer.patch_size)
)
return split_output_channels(output, self.in_channels)
class AFMGenerator2DModel(_AFMCheckpointMixin, Generator, ConfigMixin):
config_name = "config.json"
@register_to_config
def __init__(
self,
model_type: str = "AFM-XL/2",
architecture: str = "standard",
repeat: int = 1,
use_t_src: bool = False,
use_t_tgt: bool = False,
depth: int = 28,
hidden_size: int = 1152,
patch_size: int = 2,
num_heads: int = 16,
learn_sigma: bool = False,
class_dropout_prob: float = 0.0,
input_size: int = 32,
num_classes: int = 1000,
in_channels: int = 4,
pred_type: str = "x",
num_inference_steps: int = 1,
) -> None:
if architecture == "deep":
raise ValueError("Use AFMGeneratorDeep2DModel for deep AFM checkpoints.")
Generator.__init__(
self,
use_t_src=use_t_src,
use_t_tgt=use_t_tgt,
depth=depth,
hidden_size=hidden_size,
patch_size=patch_size,
num_heads=num_heads,
learn_sigma=learn_sigma,
class_dropout_prob=class_dropout_prob,
input_size=input_size,
num_classes=num_classes,
in_channels=in_channels,
)
class AFMGeneratorDeep2DModel(_AFMCheckpointMixin, GeneratorDeep, ConfigMixin):
config_name = "config.json"
@register_to_config
def __init__(
self,
model_type: str = "AFM-XL/2",
architecture: str = "deep",
repeat: int = 2,
use_t_src: bool = False,
use_t_tgt: bool = False,
depth: int = 28,
hidden_size: int = 1152,
patch_size: int = 2,
num_heads: int = 16,
learn_sigma: bool = False,
class_dropout_prob: float = 0.0,
input_size: int = 32,
num_classes: int = 1000,
in_channels: int = 4,
pred_type: str = "x",
num_inference_steps: int = 1,
) -> None:
GeneratorDeep.__init__(
self,
repeat=repeat,
depth=depth,
hidden_size=hidden_size,
patch_size=patch_size,
num_heads=num_heads,
learn_sigma=learn_sigma,
class_dropout_prob=class_dropout_prob,
input_size=input_size,
num_classes=num_classes,
in_channels=in_channels,
)
__all__ = ["AFMGenerator2DModel", "AFMGeneratorDeep2DModel"]
|