首页 > 其他分享 >(5-4-05)基于Stable Diffusion的文生图系统:(5)概率分布+编码和嵌入

(5-4-05)基于Stable Diffusion的文生图系统:(5)概率分布+编码和嵌入

时间:2024-09-04 16:54:58浏览次数:13  
标签:Diffusion __ return 05 文生 self torch device def

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:

  1. 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。
  2. 数据预处理:对输入数据(文本或图像)进行必要的预处理。
  3. 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。
  4. 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。

标签:Diffusion,__,return,05,文生,self,torch,device,def
From: https://blog.csdn.net/asd343442/article/details/141898132

相关文章

  • 20240907_051745 python 正则表达式 常见元字符
    •.:匹配任意单个字符•\d:匹配数字(等价于[0-9])•\w:匹配字母、数字、下划线(等价于[a-zA-Z0-9_])•\s:匹配空格、制表符、换行符等空白字符•^:匹配开头•$:匹配结尾•*:匹配前面的字符零次或多次•+:匹配前面的字符一次或多次•?:匹配前面的字符零次或一次•[]:匹配方括......
  • AI绘画Stable Diffusion:从新手到高手,漫画小说创收不是梦,几条视频变现几k(Ai工具+教程)
    许多人都被大量的小说推文项目所淹没,看着别人收益高、账号做得好,很多人讲述这个项目时,要么不透露具体AI工具名称,要么不提供推文授权渠道,让人无从下手,干着急。今天,向阳将带给大家一期全新的纯AI制作小说推文项目,这个新玩法将让你轻松掌握保姆级的详细教程。在本文的结尾......
  • Stable Diffusion【XL Lora】推荐!AI助力服装设计,让服装拆分设计就是这么高效!
    今天给大家介绍一个服装饰品分类背景的基于SDXL的Lora模型:分类背景XUER。该模型是由作者(B站绪儿已成精)炼制,非常适合饰品服装分类背景。绪儿大佬其实推出了很多非常棒的模型,比如之前非常受大家喜欢的敦煌飞天、超梦幻场景等模型。下面我们来实际体验一下,看使用这个模型出来的图片......
  • How to perform Lot Translate Transactions using Transaction Open Interface (MTI)
        HowtoperformLotTranslateTransactionsusingTransactionOpenInterface(MTI)(DocID838305.1)ToBottomInthisDocument Goal Solution......
  • 『SD』Stable Diffusion WebUI 安装插件(以汉化为例)
    本文简介点赞+关注+收藏=学会了StableDiffusionWebUI是允许用户自行安装插件的,插件的种类有很多,有将页面翻译成中文的插件,也有提示词补全插件,也有精细控制出图的插件。以汉化为例,StableDiffusionWebUI默认是英文的,我们只需装个汉化插件然后重启一下项目就能......
  • Stable Diffusion进阶篇,ComfyUI文字生成视频的一条龙服务(附完整工作流)
    有些小伙伴发现了一个问题,那就是根据图片生成出来的视频看着怪怪的:前半段看着好像还可以,但是后面部分的面部就开始崩坏了。而今天这篇笔记则是要简单了解一下一些进阶参数以及一些文生图生视频的内容,不然的话我担心内容太少字数都凑不够。我这里准备了ComfyUI文字生成......
  • Stable Diffusion教程,直接一键安装使用!(附SD整合包)
    ▼今天给大家分享的是**开源免费StableDiffusion教程无需魔法一键安装使用!**下载包里面有安装教程,大家如果有需要自行下载就可以了。StableDiffusion是基于LatentDiffusionModel(LDM)的,LDM是一款顶尖的文转图合成技术。在了解LDM的工作原理之前,让我们先看看什么是......
  • [1055]
    Togettheindexofthe“True”valuesinaPandasSeries,youcanusetheindexattributealongwithbooleanindexing.Here’sasimplewaytodoit:ImportPandas:importpandasaspdCreateyourSeries:series=pd.Series([True,False,True,False,......
  • AI绘画实操 Stable Diffusion 到底怎么玩儿,新手必看的AI绘画入门安装使用教程
    大家好,我是灵魂画师向阳2024年,是AI绘画技术飞速发展的一年,各种AI绘画工具层出不穷,为了让大家在了解和学习AI绘画的过程中少走弯路,今天我将详细介绍目前世界上使用用户最多,社区最大,生态最丰富的免费图像生成模型——StableDiffusion,并为你提供详细的安装教程,让你轻松踏入AI......
  • AGRI10051 2024 Prac
    AGRI100512024Prac1PRACTICAL INFORMATIONANDAPPENDICESWelcome to the Biology Laboratory!This first page contains important information that will help you duringyour practicalclasses,soyou maywishto readthrough it carefully so......