File size: 14,930 Bytes
0161e74 | 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 | import torch
import torch.nn as nn
import numpy as np
import math
from typing import Dict, Mapping, Optional, Tuple, Any, Union
import pdb
from timm.models.vision_transformer import PatchEmbed, Attention, Mlp
from src.flow_matching.utils.model_wrapper import ModelWrapper
def modulate(x, shift, scale):
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
#################################################################################
# Embedding Layers for Timesteps and Class Labels #
#################################################################################
class TimestepEmbedder(nn.Module):
"""
Embeds scalar timesteps into vector representations.
"""
def __init__(self, hidden_size, frequency_embedding_size=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, dim, max_period=10000):
"""
Create sinusoidal timestep embeddings.
:param t: a 1-D Tensor of N indices, one per batch element.
These may be fractional.
:param dim: the dimension of the output.
:param max_period: controls the minimum frequency of the embeddings.
:return: an (N, D) Tensor of positional embeddings.
"""
# https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
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):
t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
t_emb = self.mlp(t_freq)
return t_emb
class PerturbationEmbedder(nn.Module):
"""
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
"""
def __init__(self, num_classes, hidden_size, dropout_prob):
super().__init__()
use_cfg_embedding = dropout_prob > 0
self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)
self.num_classes = num_classes
self.dropout_prob = dropout_prob
def token_drop(self, labels, force_drop_ids=None):
"""
Drops labels to enable classifier-free guidance.
"""
if force_drop_ids is None:
drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
else:
drop_ids = force_drop_ids == 1
labels = torch.where(drop_ids, self.num_classes, labels)
return labels
def forward(self, labels, train=True, force_drop_ids=None):
use_dropout = self.dropout_prob > 0
if (train and use_dropout) or (force_drop_ids is not None):
labels = self.token_drop(labels, force_drop_ids)
embeddings = self.embedding_table(labels)
return embeddings
class CrossAttentionModule(nn.Module):
def __init__(self, embed_dim, num_heads, dropout=0.0):
super().__init__()
self.attn = nn.MultiheadAttention(embed_dim=embed_dim,
num_heads=num_heads,
dropout=dropout,
batch_first=True) # using (batch, seq, dim)
self.norm = nn.LayerNorm(embed_dim)
self.ff = nn.Sequential(
nn.Linear(embed_dim, embed_dim * 4),
nn.GELU(),
nn.Linear(embed_dim * 4, embed_dim),
nn.Dropout(dropout),
)
def forward(self, x1, x2, attn_mask=None, key_padding_mask=None):
# x1: (b, l1, dim), x2: (b, l2, dim)
residual = x1
attn_out, attn_weights = self.attn(query=x1, key=x2, value=x2,
attn_mask=attn_mask,
key_padding_mask=key_padding_mask)
x = self.norm(attn_out + residual)
# add FFN
residual2 = x
x = self.ff(x)
x = self.norm(x + residual2)
return x, attn_weights
#################################################################################
# Core DiT Model #
#################################################################################
class DiTBlock(nn.Module):
"""
A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning.
"""
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, **block_kwargs):
super().__init__()
self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, **block_kwargs)
self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
mlp_hidden_dim = int(hidden_size * mlp_ratio)
approx_gelu = lambda: nn.GELU(approximate="tanh")
self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)
self.adaLN_modulation = nn.Sequential(
nn.SiLU(),
nn.Linear(hidden_size, 6 * hidden_size, bias=True)
)
def forward(self, x, c):
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=1)
x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa))
x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
return x
class HierarchicalBlock(nn.Module):
def __init__(self,d_model,nhead,dropout):
super().__init__()
self.cell1_block = DiTBlock(d_model,nhead,dropout)
self.cell2_block = DiTBlock(d_model,nhead,dropout)
self.cross_attention = CrossAttentionModule(d_model,nhead,dropout)
def forward(self,cell1,cell2,perturbation,t):
cell1 = self.cell1_block(cell1,perturbation+t)
cell2 = self.cell2_block(cell2,t)
cell1, _ = self.cross_attention(cell2,cell1)
return cell1,cell2,
class SymmetricBlock(nn.Module):
def __init__(self,d_model,nhead,dropout):
super().__init__()
# self.hierarchical_block1 = HierarchicalBlock(d_model,nhead,dropout)
# self.hierarchical_block2 = HierarchicalBlock(d_model,nhead,dropout)
self.hierarchical_block = HierarchicalBlock(d_model,nhead,dropout)
self.norm = nn.LayerNorm(d_model)
def forward(self,cell1,cell2,perturbation,t):
# cell1_1,cell2_1 = self.hierarchical_block1(cell1,cell2,perturbation,t)
# cell2_2,cell1_2 = self.hierarchical_block2(cell2,cell1,-perturbation,t)
cell1_1,cell2_1 = self.hierarchical_block(cell1,cell2,perturbation,t)
cell2_2,cell1_2 = self.hierarchical_block(cell2,cell1,-perturbation,t)
cell1 = self.norm(cell1_1 + cell1_2)
cell2 = self.norm(cell2_1 + cell2_2)
return cell1,cell2
class Model(nn.Module):
def __init__(
self,
model_type: str = 'hierarchical',
ntoken: int = 512,
d_model: int = 512,
nhead: int = 8,
nlayers: int = 6 ,
n_combination: int = 2,
nlayers_cls: int = 3,
d_perturbation: int = 512,
vocab: Any = None,
dropout: float = 0.5,
):
super().__init__()
self.n_combination = n_combination
self.d_model = d_model
self.model_type = model_type
self.ntoken = ntoken
self.d_perturbation = d_perturbation
# self.p_embedder = PerturbationEmbedder(n_perturbation, d_model, dropout_prob=0)
self.p_embedder = nn.Sequential(
nn.Linear(d_perturbation,d_model),
nn.SiLU(),
nn.Linear(d_model,d_model),
)
self.t_embedder = TimestepEmbedder(d_model)
self.blocks = nn.ModuleList([
DiTBlock(d_model, nhead, dropout) for _ in range(nlayers)
])
if model_type == 'hierarchical':
self.blocks = nn.ModuleList([
SymmetricBlock(d_model, nhead, dropout) for _ in range(nlayers)
])
else:
self.blocks = nn.ModuleList([
DiTBlock(d_model, nhead, dropout) for _ in range(nlayers)
])
self.perturbation_decoder = nn.Sequential(
nn.Linear(d_model,d_model*4),
nn.SiLU(),
nn.Linear(d_model*4,d_model),
)
def forward(self,cell1,cell2,t,perturbation=None):
"""_summary_
Args:
cell1 (tensor): (B,ntoken,d_model) for control cell
cell2 (tensor): (B,ntoken,d_model) for perturbed cell
t (long): (B,) when inference is a number
perturbation (long): (B,n_combination)
"""
t = self.t_embedder(t) # (N, D)
# pdb.set_trace()
if perturbation is not None:
perturbation_ = perturbation.reshape(-1,self.d_perturbation)
p = self.p_embedder(perturbation_) # (N, D)
# p = p.reshape(-1,self.n_combination,self.d_model)
# p = p.sum(dim=1)
c = t + p
else:
c = t
cells = torch.cat([cell1,cell2],dim=1)
for block in self.blocks:
if self.model_type == 'hierarchical':
cell1,cell2 = block(cell1,cell2,p,t)
else:
cells = block(cells,c)
if self.model_type != 'hierarchical':
cells = cells.reshape(-1,2,self.ntoken,self.d_model)
cell1 = cells[:,0,:,:]
cell2 = cells[:,1,:,:]
p = self.perturbation_decoder((cell1 + cell2).mean(dim=1))
# p = p.reshape(-1,self.n_combination)
return cell1, cell2, p
class TimedTransformer(nn.Module):
def __init__(
self,
model_type: str = 'hierarchical',
ntoken: int = 512,
d_model: int = 512,
nhead: int = 8,
nlayers: int = 8 ,
n_combination: int = 2,
nlayers_cls: int = 3,
d_perturbation: int = 512,
vocab: Any = None,
dropout: float = 0.5,
):
super().__init__()
self.n_combination = n_combination
self.d_model = d_model
self.model_type = model_type
self.ntoken = ntoken
# self.p_embedder = PerturbationEmbedder(n_perturbation, d_model, dropout_prob=0)
self.t_embedder = TimestepEmbedder(d_model)
self.blocks = nn.ModuleList([
DiTBlock(d_model, nhead, dropout) for _ in range(nlayers)
])
def forward(self,cell1,t):
"""_summary_
Args:
cell1 (tensor): (B,ntoken,d_model) for control cell
cell2 (tensor): (B,ntoken,d_model) for perturbed cell
t (long): (B,) when inference is a number
perturbation (long): (B,n_combination)
"""
t = self.t_embedder(t) # (N, D)
c = t
cells = cell1
for block in self.blocks:
cells = block(cells,c)
return cells
# class WrappedModel(ModelWrapper):
# def __init__(self, model, mode="noisy_y"):
# super().__init__(model)
# assert mode in ["noisy_y", "noisy_p"]
# self.mode = mode
# def forward(self, x: torch.Tensor, t: torch.Tensor, x_t: torch.Tensor, condition_vec: torch.Tensor, mode=None):
# """
# x: x_0
# x_t: path sample at time t
# t: time step
# condition_vec: condition vector
# """
# if mode is not None:
# self.mode = mode
# if self.mode == "noisy_y":
# _, predicted_x_t, _ = self.model(x, x_t, t, condition_vec)
# return (x, predicted_x_t, condition_vec)
# elif self.mode == "noisy_p":
# _, _, predicted_p_t = self.model(x, x_t, t, condition_vec)
# return (x, x_t, predicted_p_t)
# else:
# raise ValueError(f"Unsupported mode: {self.mode}")
class WrappedModel(ModelWrapper):
def __init__(self, model, mode="noisy_y"):
super().__init__(model)
assert mode in ["noisy_y", "noisy_p"]
self.mode = mode
self.model = model
def change_mode(self, mode):
self.mode = mode
def forward(self, x: torch.Tensor, t: torch.Tensor, **extras):
if t.ndim == 0:
t = t.expand(x.shape[0]).to(x.device)
if self.mode == "noisy_y":
condition_vec = extras["condition_vec"]
x_0 = extras["x_0"]
_, predicted_x_t, _ = self.model(x_0, x, t, condition_vec)
return predicted_x_t
elif self.mode == "noisy_p":
target_y = extras["target_y"]
x_0 = extras["x_0"]
_, _, predicted_p_t = self.model(x_0, target_y, t, x)
return predicted_p_t
else:
raise ValueError(f"Unsupported mode: {self.mode}")
if __name__ == '__main__':
batch_size = 128
model = Model(model_type='hs',ntoken=512).cuda()
cell1 = torch.randn(batch_size, 1024,512).cuda()
cell2 = torch.randn(batch_size, 1024,512).cuda()
t = torch.randint(0, 1000, (batch_size,)).cuda()
# p = torch.randint(0,2,(5,2)).cuda()
p = torch.randn(batch_size,512).cuda()
model = nn.DataParallel(model,device_ids=[0,1,2,3,4,5,6,7])
from tqdm import tqdm
for i in tqdm(range(5000)):
out = model(cell1,cell2, t,p)
out[0].sum().backward()
pdb.set_trace()
print(out.shape)
# from accelerate import Accelerator
# import time
# accelerator = Accelerator()
# batch_size = 128
# model = Model(model_type='hs', ntoken=512)
# cell1 = torch.randn(batch_size, 1024, 512)
# cell2 = torch.randn(batch_size, 1024, 512)
# t = torch.randint(0, 1000, (batch_size,))
# p = torch.randn(batch_size, 512)
# # 用 accelerator.prepare 包装
# model, cell1, cell2, t, p = accelerator.prepare(model, cell1, cell2, t, p)
# from tqdm import tqdm
# start = time.time()
# for i in tqdm(range(5000)):
# out = model(cell1, cell2, t, p)
# loss = out[0].sum() + out[1].sum() + out[2].sum()
# accelerator.backward(loss)
# end = time.time()
# print(f"Total time: {end - start:.2f} seconds")
# import pdb; pdb.set_trace()
# print(out[0].shape) |