File size: 10,615 Bytes
b2cb4a0 | 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 | #
# For licensing see accompanying LICENSE file.
# Copyright (c) 2025 Apple Inc. Licensed under MIT License.
#
import math
from einops import rearrange
import torch
from torch import nn
import torch.nn.functional as F
def modulate(x, shift, scale):
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
#################################################################################
# Attention Layers #
#################################################################################
class SelfAttentionLayer(nn.Module):
def __init__(
self,
hidden_size,
num_heads=8,
qkv_bias=False,
qk_scale=None,
attn_drop=0.0,
proj_drop=0.0,
use_bias=True,
qk_norm=True,
pos_embedder=None,
linear_target: nn.Module = nn.Linear,
):
super().__init__()
self.num_heads = num_heads
head_dim = hidden_size // num_heads
self.scale = qk_scale or head_dim**-0.5
self.qkv = linear_target(hidden_size, hidden_size * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = linear_target(hidden_size, hidden_size, bias=use_bias)
self.proj_drop = nn.Dropout(proj_drop)
self.q_norm = RMSNorm(head_dim) if qk_norm else nn.Identity()
self.k_norm = RMSNorm(head_dim) if qk_norm else nn.Identity()
self.pos_embedder = pos_embedder
def forward(self, x, **kwargs):
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
pos = kwargs.get("pos")
qkv = rearrange(qkv, "b n t h c -> t b h n c")
q, k, v = (
qkv[0],
qkv[1],
qkv[2],
) # make torchscript happy (cannot use tensor as tuple)
q, k = self.q_norm(q), self.k_norm(k)
if self.pos_embedder and pos is not None:
q, k = self.pos_embedder(q, k, pos)
attn = (q @ k.transpose(-2, -1)) * self.scale
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class EfficientSelfAttentionLayer(SelfAttentionLayer):
"""Started from https://github.com/facebookresearch/dinov2/blob/main/dinov2/layers/attention.py"""
def __init__(
self,
*args,
**kwargs,
):
super().__init__(*args, **kwargs)
def forward(self, x, **kwargs):
B, N, C = x.shape
attn_mask = kwargs.get("attention_mask")
pos = kwargs.get("pos")
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
qkv = rearrange(qkv, "b n t h c -> t b h n c")
q, k, v = qkv.unbind(0)
if attn_mask is not None:
attn_mask = attn_mask.to(dtype=q.dtype)
if self.pos_embedder and pos is not None:
q, k = self.pos_embedder(q, k, pos)
q, k = self.q_norm(q), self.k_norm(k)
v1 = v.to(dtype = q.dtype)
x = nn.functional.scaled_dot_product_attention(q, k, v1, attn_mask=attn_mask)
x = x.transpose(1, 2).reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
#################################################################################
# FeedForward Layer #
#################################################################################
class SwiGLUFeedForward(nn.Module):
def __init__(self, dim, hidden_dim, multiple_of=256):
super().__init__()
hidden_dim = int(2 * hidden_dim / 3)
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
self.w1 = nn.Linear(dim, hidden_dim, bias=False)
self.w2 = nn.Linear(hidden_dim, dim, bias=True)
self.w3 = nn.Linear(dim, hidden_dim, bias=False)
self.reset_parameters()
def reset_parameters(self):
torch.nn.init.xavier_uniform_(self.w1.weight)
torch.nn.init.xavier_uniform_(self.w2.weight)
torch.nn.init.xavier_uniform_(self.w3.weight)
if self.w1.bias is not None:
torch.nn.init.constant_(self.w1.bias, 0)
if self.w2.bias is not None:
torch.nn.init.constant_(self.w2.bias, 0)
if self.w3.bias is not None:
torch.nn.init.constant_(self.w3.bias, 0)
def forward(self, x):
return self.w2(F.silu(self.w1(x)) * self.w3(x))
#################################################################################
# Utility Layers #
#################################################################################
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
self.initialize_weights()
def initialize_weights(self):
nn.init.normal_(self.mlp[0].weight, std=0.02)
nn.init.normal_(self.mlp[2].weight, std=0.02)
@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 ConditionEmbedder(nn.Module):
"""
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
"""
def __init__(self, input_dim, hidden_size, dropout_prob):
super().__init__()
self.proj = nn.Sequential(
nn.Linear(input_dim, hidden_size),
nn.LayerNorm(hidden_size),
nn.SiLU(),
)
self.dropout_prob = dropout_prob
self.null_token = nn.Parameter(torch.randn(input_dim), requires_grad=True)
def token_drop(self, cond, force_drop_ids=None):
"""
cond: (B, N, D)
Drops conditions to enable classifier-free guidance.
"""
if force_drop_ids is None:
drop_ids = torch.rand(cond.shape[0], device=cond.device) < self.dropout_prob
else:
drop_ids = force_drop_ids
cond[drop_ids] = self.null_token[None, None, :]
return cond
def forward(self, cond, train, force_drop_ids=None):
use_dropout = self.dropout_prob > 0
if (train and use_dropout) or (force_drop_ids is not None):
cond = self.token_drop(cond, force_drop_ids)
embeddings = self.proj(cond)
return embeddings
class FinalLayer(nn.Module):
"""
The final layer of DiT.
"""
def __init__(self, hidden_size, out_channels, c_dim=None):
super().__init__()
self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.linear = nn.Linear(hidden_size, out_channels, bias=True)
self.adaLN_modulation = nn.Sequential(
nn.SiLU(), nn.Linear(c_dim, 2 * hidden_size, bias=True)
)
self.initialize_weights()
def initialize_weights(self):
# Initialize transformer layers:
def _basic_init(module):
if isinstance(module, nn.Linear):
torch.nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.constant_(module.bias, 0)
self.apply(_basic_init)
# Zero-out output layers:
nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
nn.init.constant_(self.linear.weight, 0)
nn.init.constant_(self.linear.bias, 0)
def forward(self, x, c):
shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
x = modulate(self.norm_final(x), shift, scale)
x = self.linear(x)
return x
class RMSNorm(nn.Module):
def __init__(self, d, p=-1.0, eps=1e-8, bias=False):
"""
Root Mean Square Layer Normalization
:param d: model size
:param p: partial RMSNorm, valid value [0, 1], default -1.0 (disabled)
:param eps: epsilon value, default 1e-8
:param bias: whether use bias term for RMSNorm, disabled by
default because RMSNorm doesn't enforce re-centering invariance.
"""
super(RMSNorm, self).__init__()
self.eps = eps
self.d = d
self.p = p
self.bias = bias
self.scale = nn.Parameter(torch.ones(d))
self.register_parameter("scale", self.scale)
if self.bias:
self.offset = nn.Parameter(torch.zeros(d))
self.register_parameter("offset", self.offset)
def forward(self, x):
if self.p < 0.0 or self.p > 1.0:
norm_x = x.norm(2, dim=-1, keepdim=True, dtype=x.dtype)
d_x = self.d
else:
partial_size = int(self.d * self.p)
partial_x, _ = torch.split(x, [partial_size, self.d - partial_size], dim=-1)
norm_x = partial_x.norm(2, dim=-1, keepdim=True, dtype=x.dtype)
d_x = partial_size
rms_x = norm_x * d_x ** (-1.0 / 2)
x_normed = x / (rms_x + self.eps)
if self.bias:
return self.scale * x_normed + self.offset
return self.scale * x_normed
|