diffusers/src/diffusers/models/normalization.py at main · koji/diffusers · GitHub

History

361 lines (296 loc) · 13.2 KB

Raw

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

# coding=utf-8

# Licensed under the Apache License, Version 2.0 (the "License");

# you may not use this file except in compliance with the License.

# You may obtain a copy of the License at

# http://www.apache.org/licenses/LICENSE-2.0

# Unless required by applicable law or agreed to in writing, software

# distributed under the License is distributed on an "AS IS" BASIS,

# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

# See the License for the specific language governing permissions and

# limitations under the License.

import numbers

from typing import Dict, Optional, Tuple

import torch

import torch.nn as nn

import torch.nn.functional as F

from ..utils import is_torch_version

from .activations import get_activation

from .embeddings import (

CombinedTimestepLabelEmbeddings,

PixArtAlphaCombinedTimestepSizeEmbeddings,

)

class AdaLayerNorm(nn.Module):

r"""

Norm layer modified to incorporate timestep embeddings.

Parameters:

embedding_dim (`int`): The size of each embedding vector.

num_embeddings (`int`): The size of the embeddings dictionary.

"""

def __init__(self, embedding_dim: int, num_embeddings: int):

super().__init__()

self.emb = nn.Embedding(num_embeddings, embedding_dim)

self.silu = nn.SiLU()

self.linear = nn.Linear(embedding_dim, embedding_dim * 2)

self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False)

def forward(self, x: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor:

emb = self.linear(self.silu(self.emb(timestep)))

scale, shift = torch.chunk(emb, 2)

x = self.norm(x) * (1 + scale) + shift

return x

class FP32LayerNorm(nn.LayerNorm):

def forward(self, inputs: torch.Tensor) -> torch.Tensor:

origin_dtype = inputs.dtype

return F.layer_norm(

inputs.float(),

self.normalized_shape,

self.weight.float() if self.weight is not None else None,

self.bias.float() if self.bias is not None else None,

self.eps,

).to(origin_dtype)

class AdaLayerNormZero(nn.Module):

r"""

Norm layer adaptive layer norm zero (adaLN-Zero).

Parameters:

embedding_dim (`int`): The size of each embedding vector.

num_embeddings (`int`): The size of the embeddings dictionary.

"""

def __init__(self, embedding_dim: int, num_embeddings: Optional[int] = None, norm_type="layer_norm", bias=True):

super().__init__()

if num_embeddings is not None:

self.emb = CombinedTimestepLabelEmbeddings(num_embeddings, embedding_dim)

else:

self.emb = None

self.silu = nn.SiLU()

self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=bias)

if norm_type == "layer_norm":

self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)

elif norm_type == "fp32_layer_norm":

self.norm = FP32LayerNorm(embedding_dim, elementwise_affine=False, bias=False)

else:

raise ValueError(

f"Unsupported `norm_type` ({norm_type}) provided. Supported ones are: 'layer_norm', 'fp32_layer_norm'."

)

def forward(

self,

x: torch.Tensor,

timestep: Optional[torch.Tensor] = None,

class_labels: Optional[torch.LongTensor] = None,

hidden_dtype: Optional[torch.dtype] = None,

emb: Optional[torch.Tensor] = None,

) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:

if self.emb is not None:

emb = self.emb(timestep, class_labels, hidden_dtype=hidden_dtype)

emb = self.linear(self.silu(emb))

shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1)

x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None]

return x, gate_msa, shift_mlp, scale_mlp, gate_mlp

class LuminaRMSNormZero(nn.Module):

"""

Norm layer adaptive RMS normalization zero.

Parameters:

embedding_dim (`int`): The size of each embedding vector.

"""

def __init__(self, embedding_dim: int, norm_eps: float, norm_elementwise_affine: bool):

super().__init__()

self.silu = nn.SiLU()

self.linear = nn.Linear(

min(embedding_dim, 1024),

4 * embedding_dim,

bias=True,

)

self.norm = RMSNorm(embedding_dim, eps=norm_eps, elementwise_affine=norm_elementwise_affine)

def forward(

self,

x: torch.Tensor,

emb: Optional[torch.Tensor] = None,

) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:

# emb = self.emb(timestep, encoder_hidden_states, encoder_mask)

emb = self.linear(self.silu(emb))

scale_msa, gate_msa, scale_mlp, gate_mlp = emb.chunk(4, dim=1)

x = self.norm(x) * (1 + scale_msa[:, None])

return x, gate_msa, scale_mlp, gate_mlp

class AdaLayerNormSingle(nn.Module):

r"""

Norm layer adaptive layer norm single (adaLN-single).

As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3).

Parameters:

embedding_dim (`int`): The size of each embedding vector.

use_additional_conditions (`bool`): To use additional conditions for normalization or not.

"""

def __init__(self, embedding_dim: int, use_additional_conditions: bool = False):

super().__init__()

self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings(

embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions

)

self.silu = nn.SiLU()

self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)

def forward(

self,

timestep: torch.Tensor,

added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,

batch_size: Optional[int] = None,

hidden_dtype: Optional[torch.dtype] = None,

) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:

# No modulation happening here.

embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype)

return self.linear(self.silu(embedded_timestep)), embedded_timestep

class AdaGroupNorm(nn.Module):

r"""

GroupNorm layer modified to incorporate timestep embeddings.

Parameters:

embedding_dim (`int`): The size of each embedding vector.

num_embeddings (`int`): The size of the embeddings dictionary.

num_groups (`int`): The number of groups to separate the channels into.

act_fn (`str`, *optional*, defaults to `None`): The activation function to use.

eps (`float`, *optional*, defaults to `1e-5`): The epsilon value to use for numerical stability.

"""

def __init__(

self, embedding_dim: int, out_dim: int, num_groups: int, act_fn: Optional[str] = None, eps: float = 1e-5

super().__init__()

self.num_groups = num_groups

self.eps = eps

if act_fn is None:

self.act = None

else:

self.act = get_activation(act_fn)

self.linear = nn.Linear(embedding_dim, out_dim * 2)

def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor:

if self.act:

emb = self.act(emb)

emb = self.linear(emb)

emb = emb[:, :, None, None]

scale, shift = emb.chunk(2, dim=1)

x = F.group_norm(x, self.num_groups, eps=self.eps)

x = x * (1 + scale) + shift

return x

class AdaLayerNormContinuous(nn.Module):

def __init__(

self,

embedding_dim: int,

conditioning_embedding_dim: int,

# NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters

# because the output is immediately scaled and shifted by the projected conditioning embeddings.

# Note that AdaLayerNorm does not let the norm layer have scale and shift parameters.

# However, this is how it was implemented in the original code, and it's rather likely you should

# set `elementwise_affine` to False.

elementwise_affine=True,

eps=1e-5,

bias=True,

norm_type="layer_norm",

super().__init__()

self.silu = nn.SiLU()

self.linear = nn.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias)

if norm_type == "layer_norm":

self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias)

elif norm_type == "rms_norm":

self.norm = RMSNorm(embedding_dim, eps, elementwise_affine)

else:

raise ValueError(f"unknown norm_type {norm_type}")

def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:

# convert back to the original dtype in case `conditioning_embedding`` is upcasted to float32 (needed for hunyuanDiT)

emb = self.linear(self.silu(conditioning_embedding).to(x.dtype))

scale, shift = torch.chunk(emb, 2, dim=1)

x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]

return x

class LuminaLayerNormContinuous(nn.Module):

def __init__(

self,

embedding_dim: int,

conditioning_embedding_dim: int,

# NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters

# because the output is immediately scaled and shifted by the projected conditioning embeddings.

# Note that AdaLayerNorm does not let the norm layer have scale and shift parameters.

# However, this is how it was implemented in the original code, and it's rather likely you should

# set `elementwise_affine` to False.

elementwise_affine=True,

eps=1e-5,

bias=True,

norm_type="layer_norm",

out_dim: Optional[int] = None,

super().__init__()

# AdaLN

self.silu = nn.SiLU()

self.linear_1 = nn.Linear(conditioning_embedding_dim, embedding_dim, bias=bias)

if norm_type == "layer_norm":

self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias)

else:

raise ValueError(f"unknown norm_type {norm_type}")

# linear_2

if out_dim is not None:

self.linear_2 = nn.Linear(

embedding_dim,

out_dim,

bias=bias,

)

def forward(

self,

x: torch.Tensor,

conditioning_embedding: torch.Tensor,

) -> torch.Tensor:

# convert back to the original dtype in case `conditioning_embedding`` is upcasted to float32 (needed for hunyuanDiT)

emb = self.linear_1(self.silu(conditioning_embedding).to(x.dtype))

scale = emb

x = self.norm(x) * (1 + scale)[:, None, :]

if self.linear_2 is not None:

x = self.linear_2(x)

return x

if is_torch_version(">=", "2.1.0"):

LayerNorm = nn.LayerNorm

else:

# Has optional bias parameter compared to torch layer norm

# TODO: replace with torch layernorm once min required torch version >= 2.1

class LayerNorm(nn.Module):

def __init__(self, dim, eps: float = 1e-5, elementwise_affine: bool = True, bias: bool = True):

super().__init__()

self.eps = eps

if isinstance(dim, numbers.Integral):

dim = (dim,)

self.dim = torch.Size(dim)

if elementwise_affine:

self.weight = nn.Parameter(torch.ones(dim))

self.bias = nn.Parameter(torch.zeros(dim)) if bias else None

else:

self.weight = None

self.bias = None

def forward(self, input):

return F.layer_norm(input, self.dim, self.weight, self.bias, self.eps)

class RMSNorm(nn.Module):

def __init__(self, dim, eps: float, elementwise_affine: bool = True):

super().__init__()

self.eps = eps

if isinstance(dim, numbers.Integral):

dim = (dim,)

self.dim = torch.Size(dim)

if elementwise_affine:

self.weight = nn.Parameter(torch.ones(dim))

else:

self.weight = None

def forward(self, hidden_states):

input_dtype = hidden_states.dtype

variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)

hidden_states = hidden_states * torch.rsqrt(variance + self.eps)

if self.weight is not None:

# convert into half-precision if necessary

if self.weight.dtype in [torch.float16, torch.bfloat16]:

hidden_states = hidden_states.to(self.weight.dtype)

hidden_states = hidden_states * self.weight

else:

hidden_states = hidden_states.to(input_dtype)

return hidden_states

class GlobalResponseNorm(nn.Module):

# Taken from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105

def __init__(self, dim):

super().__init__()

self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim))

self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim))

def forward(self, x):

gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True)

nx = gx / (gx.mean(dim=-1, keepdim=True) + 1e-6)

return self.gamma * (x * nx) + self.beta + x

diffusers/src/diffusers/models/normalization.py at main · koji/diffusers · GitHub

Expand file tree

Breadcrumbs

normalization.py

Latest commit

History

Breadcrumbs

normalization.py

File metadata and controls