5. 概率分布
文件distributions.py定义了与概率分布相关的抽象类和具体实现,包括抽象分布类 AbstractDistribution、狄拉克分布 DiracDistribution 和对角高斯分布 DiagonalGaussianDistribution。这些类提供了样本生成、模式计算和 KL 散度等功能,支持概率模型中的采样和分布计算。此外,还定义了一个计算两个高斯分布之间 KL 散度的函数 normal_kl,用于评估不同分布之间的相似性。整体上,该文件为概率建模和生成任务提供了基础设施。
import torch
import numpy as np
class AbstractDistribution:
def sample(self):
raise NotImplementedError()
def mode(self):
raise NotImplementedError()
class DiracDistribution(AbstractDistribution):
def __init__(self, value):
self.value = value
def sample(self):
return self.value
def mode(self):
return self.value
class DiagonalGaussianDistribution(object):
def __init__(self, parameters, deterministic=False):
self.parameters = parameters
self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
self.deterministic = deterministic
self.std = torch.exp(0.5 * self.logvar)
self.var = torch.exp(self.logvar)
if self.deterministic:
self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
def sample(self):
x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
return x
def kl(self, other=None):
if self.deterministic:
return torch.Tensor([0.])
else:
if other is None:
return 0.5 * torch.sum(torch.pow(self.mean, 2)
+ self.var - 1.0 - self.logvar,
dim=[1, 2, 3])
else:
return 0.5 * torch.sum(
torch.pow(self.mean - other.mean, 2) / other.var
+ self.var / other.var - 1.0 - self.logvar + other.logvar,
dim=[1, 2, 3])
def nll(self, sample, dims=[1,2,3]):
if self.deterministic:
return torch.Tensor([0.])
logtwopi = np.log(2.0 * np.pi)
return 0.5 * torch.sum(
logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
dim=dims)
def mode(self):
return self.mean
def normal_kl(mean1, logvar1, mean2, logvar2):
"""
source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
Compute the KL divergence between two gaussians.
Shapes are automatically broadcasted, so batches can be compared to
scalars, among other use cases.
"""
tensor = None
for obj in (mean1, logvar1, mean2, logvar2):
if isinstance(obj, torch.Tensor):
tensor = obj
break
assert tensor is not None, "at least one argument must be a Tensor"
# Force variances to be Tensors. Broadcasting helps convert scalars to
# Tensors, but it does not work for torch.exp().
logvar1, logvar2 = [
x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
for x in (logvar1, logvar2)
]
return 0.5 * (
-1.0
+ logvar2
- logvar1
+ torch.exp(logvar1 - logvar2)
+ ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
)
6. 编码和嵌入
文件modules.py定义了一系列用于实现编码和嵌入操作的神经网络模块,主要包括文本和图像的嵌入器。这些模块继承自 AbstractEncoder 类,提供了不同的编码方法,支持 T5 和 CLIP 模型的冻结版本,以便进行特征提取。关键类包括 IdentityEncoder、ClassEmbedder、FrozenT5Embedder、FrozenCLIPEmbedder、ClipImageEmbedder 和 FrozenOpenCLIPEmbedder,每个类都有其特定的前向传播和编码逻辑。此外,还实现了 CLIP 嵌入的噪声增强功能 CLIPEmbeddingNoiseAugmentation,用于处理输入数据的标准化和去噪。这些模块为图像和文本数据的处理和融合提供了灵活的基础架构,适用于多模态学习和生成任务。
import torch
import torch.nn as nn
import kornia
from torch.utils.checkpoint import checkpoint
from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel
import open_clip
from ldm.util import default, count_params, autocast
class AbstractEncoder(nn.Module):
def __init__(self):
super().__init__()
def encode(self, *args, **kwargs):
raise NotImplementedError
class IdentityEncoder(AbstractEncoder):
def encode(self, x):
return x
class ClassEmbedder(nn.Module):
def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1):
super().__init__()
self.key = key
self.embedding = nn.Embedding(n_classes, embed_dim)
self.n_classes = n_classes
self.ucg_rate = ucg_rate
def forward(self, batch, key=None, disable_dropout=False):
if key is None:
key = self.key
# this is for use in crossattn
c = batch[key][:, None]
if self.ucg_rate > 0. and not disable_dropout:
mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate)
c = mask * c + (1 - mask) * torch.ones_like(c) * (self.n_classes - 1)
c = c.long()
c = self.embedding(c)
return c
def get_unconditional_conditioning(self, bs, device="cuda"):
uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000)
uc = torch.ones((bs,), device=device) * uc_class
uc = {self.key: uc}
return uc
def disabled_train(self, mode=True):
"""Overwrite model.train with this function to make sure train/eval mode
does not change anymore."""
return self
class FrozenT5Embedder(AbstractEncoder):
"""Uses the T5 transformer encoder for text"""
def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77,
freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
super().__init__()
self.tokenizer = T5Tokenizer.from_pretrained(version)
self.transformer = T5EncoderModel.from_pretrained(version)
self.device = device
self.max_length = max_length # TODO: typical value?
if freeze:
self.freeze()
def freeze(self):
self.transformer = self.transformer.eval()
# self.train = disabled_train
for param in self.parameters():
param.requires_grad = False
def forward(self, text):
batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
tokens = batch_encoding["input_ids"].to(self.device)
outputs = self.transformer(input_ids=tokens)
z = outputs.last_hidden_state
return z
def encode(self, text):
return self(text)
class FrozenCLIPEmbedder(AbstractEncoder):
"""Uses the CLIP transformer encoder for text (from huggingface)"""
LAYERS = [
"last",
"pooled",
"hidden"
]
def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77,
freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32
super().__init__()
assert layer in self.LAYERS
self.tokenizer = CLIPTokenizer.from_pretrained(version)
self.transformer = CLIPTextModel.from_pretrained(version)
self.device = device
self.max_length = max_length
if freeze:
self.freeze()
self.layer = layer
self.layer_idx = layer_idx
if layer == "hidden":
assert layer_idx is not None
assert 0 <= abs(layer_idx) <= 12
def freeze(self):
self.transformer = self.transformer.eval()
# self.train = disabled_train
for param in self.parameters():
param.requires_grad = False
def forward(self, text):
batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
tokens = batch_encoding["input_ids"].to(self.device)
outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer == "hidden")
if self.layer == "last":
z = outputs.last_hidden_state
elif self.layer == "pooled":
z = outputs.pooler_output[:, None, :]
else:
z = outputs.hidden_states[self.layer_idx]
return z
def encode(self, text):
return self(text)
class ClipImageEmbedder(nn.Module):
def __init__(
self,
model,
jit=False,
device='cuda' if torch.cuda.is_available() else 'cpu',
antialias=True,
ucg_rate=0.
):
super().__init__()
from clip import load as load_clip
self.model, _ = load_clip(name=model, device=device, jit=jit)
self.antialias = antialias
self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
self.ucg_rate = ucg_rate
def preprocess(self, x):
# normalize to [0,1]
x = kornia.geometry.resize(x, (224, 224),
interpolation='bicubic', align_corners=True,
antialias=self.antialias)
x = (x + 1.) / 2.
# re-normalize according to clip
x = kornia.enhance.normalize(x, self.mean, self.std)
return x
def forward(self, x, no_dropout=False):
# x is assumed to be in range [-1,1]
out = self.model.encode_image(self.preprocess(x))
out = out.to(x.dtype)
if self.ucg_rate > 0. and not no_dropout:
out = torch.bernoulli((1. - self.ucg_rate) * torch.ones(out.shape[0], device=out.device))[:, None] * out
return out
class FrozenOpenCLIPEmbedder(AbstractEncoder):
"""
Uses the OpenCLIP transformer encoder for text
"""
LAYERS = [
# "pooled",
"last",
"penultimate"
]
def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
freeze=True, layer="last"):
super().__init__()
assert layer in self.LAYERS
model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'), pretrained=version)
del model.visual
self.model = model
self.device = device
self.max_length = max_length
if freeze:
self.freeze()
self.layer = layer
if self.layer == "last":
self.layer_idx = 0
elif self.layer == "penultimate":
self.layer_idx = 1
else:
raise NotImplementedError()
def freeze(self):
self.model = self.model.eval()
for param in self.parameters():
param.requires_grad = False
def forward(self, text):
tokens = open_clip.tokenize(text)
z = self.encode_with_transformer(tokens.to(self.device))
return z
def encode_with_transformer(self, text):
x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model]
x = x + self.model.positional_embedding
x = x.permute(1, 0, 2) # NLD -> LND
x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
x = x.permute(1, 0, 2) # LND -> NLD
x = self.model.ln_final(x)
return x
def text_transformer_forward(self, x: torch.Tensor, attn_mask=None):
for i, r in enumerate(self.model.transformer.resblocks):
if i == len(self.model.transformer.resblocks) - self.layer_idx:
break
if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting():
x = checkpoint(r, x, attn_mask)
else:
x = r(x, attn_mask=attn_mask)
return x
def encode(self, text):
return self(text)
class FrozenOpenCLIPImageEmbedder(AbstractEncoder):
"""
Uses the OpenCLIP vision transformer encoder for images
"""
def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
freeze=True, layer="pooled", antialias=True, ucg_rate=0.):
super().__init__()
model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'),
pretrained=version, )
del model.transformer
self.model = model
self.device = device
self.max_length = max_length
if freeze:
self.freeze()
self.layer = layer
if self.layer == "penultimate":
raise NotImplementedError()
self.layer_idx = 1
self.antialias = antialias
self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
self.ucg_rate = ucg_rate
def preprocess(self, x):
# normalize to [0,1]
x = kornia.geometry.resize(x, (224, 224),
interpolation='bicubic', align_corners=True,
antialias=self.antialias)
x = (x + 1.) / 2.
# renormalize according to clip
x = kornia.enhance.normalize(x, self.mean, self.std)
return x
def freeze(self):
self.model = self.model.eval()
for param in self.parameters():
param.requires_grad = False
@autocast
def forward(self, image, no_dropout=False):
z = self.encode_with_vision_transformer(image)
if self.ucg_rate > 0. and not no_dropout:
z = torch.bernoulli((1. - self.ucg_rate) * torch.ones(z.shape[0], device=z.device))[:, None] * z
return z
def encode_with_vision_transformer(self, img):
img = self.preprocess(img)
x = self.model.visual(img)
return x
def encode(self, text):
return self(text)
class FrozenCLIPT5Encoder(AbstractEncoder):
def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda",
clip_max_length=77, t5_max_length=77):
super().__init__()
self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length)
self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length)
print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, "
f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params.")
def encode(self, text):
return self(text)
def forward(self, text):
clip_z = self.clip_encoder.encode(text)
t5_z = self.t5_encoder.encode(text)
return [clip_z, t5_z]
from ldm.modules.diffusionmodules.upscaling import ImageConcatWithNoiseAugmentation
from ldm.modules.diffusionmodules.openaimodel import Timestep
class CLIPEmbeddingNoiseAugmentation(ImageConcatWithNoiseAugmentation):
def __init__(self, *args, clip_stats_path=None, timestep_dim=256, **kwargs):
super().__init__(*args, **kwargs)
if clip_stats_path is None:
clip_mean, clip_std = torch.zeros(timestep_dim), torch.ones(timestep_dim)
else:
clip_mean, clip_std = torch.load(clip_stats_path, map_location="cpu")
self.register_buffer("data_mean", clip_mean[None, :], persistent=False)
self.register_buffer("data_std", clip_std[None, :], persistent=False)
self.time_embed = Timestep(timestep_dim)
def scale(self, x):
# re-normalize to centered mean and unit variance
x = (x - self.data_mean) * 1. / self.data_std
return x
def unscale(self, x):
# back to original data stats
x = (x * self.data_std) + self.data_mean
return x
def forward(self, x, noise_level=None):
if noise_level is None:
noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long()
else:
assert isinstance(noise_level, torch.Tensor)
x = self.scale(x)
z = self.q_sample(x, noise_level)
z = self.unscale(z)
noise_level = self.time_embed(noise_level)
return z, noise_level
对上述代码的具体说明如下所示:
(1)实现类 AbstractEncoder:这是一个抽象基类,定义了 encode 方法的接口,用于后续的具体实现。
(2)IdentityEncoder:继承自 AbstractEncoder,实现了一个恒等编码器,即直接返回输入数据。
(3)ClassEmbedder:这个类用于将类别标签嵌入到一个固定维度的空间中。它使用一个嵌入层来实现这一点,并支持在嵌入时使用不确定性(通过dropout模拟)。
(4)FrozenT5Embedder:这个类使用 T5 模型(一种预训练的文本转换器模型)来编码文本数据。它在初始化时加载预训练的 T5 模型,并可以选择冻结模型的参数,使其在训练过程中不更新。
(5) FrozenCLIPEmbedder:类似于 FrozenT5Embedder,这个类使用 CLIP 模型来编码文本数据。CLIP 是另一种预训练的文本和图像转换器模型。它同样可以选择冻结模型参数。
(6)ClipImageEmbedder:这个类用于使用 CLIP 模型的图像编码器部分来处理图像数据。它首先对图像进行预处理(包括缩放和归一化),然后使用 CLIP 模型进行编码。
(7)FrozenOpenCLIPEmbedder:这个类使用 OpenCLIP 的文本编码器来处理文本数据。OpenCLIP 是一个开源的 CLIP 实现,它允许更灵活的模型配置和扩展。
(8)FrozenCLIPT5Encoder:这个类结合了 CLIP 和 T5 编码器,用于同时从两个不同的角度编码文本数据。
(9)CLIPEmbeddingNoiseAugmentation:这个类是一个图像处理模块,用于在图像数据中添加噪声,并结合时间步嵌入进行数据增强。它使用了 CLIP 模型的统计数据来标准化和反标准化图像数据。
(10)辅助函数disabled_train:这是一个用于禁用模型训练模式改变的函数。
(11)辅助函数count_params:用于计算模型参数的数量。
文件modules.py的实现流程如下所示:
- 初始化:根据不同的编码器需求(文本或图像),选择合适的编码器类并初始化。
- 数据预处理:对输入数据(文本或图像)进行必要的预处理。
- 编码:使用编码器对预处理后的数据进行编码,得到嵌入向量。
- 后处理:对编码结果进行进一步的处理,如应用噪声增强等。
文件modules.py中的类和函数提供了一套完整的工具,用于在深度学习模型中处理和编码文本和图像数据。