【Diffusion实战】训练一个diffusion模型生成蝴蝶图像（Pytorch代码详解）

上一篇Diffusion实战是确确实实一步一步走的公式，这回采用一个更方便的库：diffusers，来实现Diffusion模型训练。

---

Diffusion实战篇：

【Diffusion实战】训练一个diffusion模型生成S曲线（Pytorch代码详解）

Diffusion综述篇：

【Diffusion综述】医学图像分析中的扩散模型（一）

【Diffusion综述】医学图像分析中的扩散模型（二）

---

0、所需安装

pip install diffusers  # diffusers库
pip install datasets  

---

1、数据集下载

下载地址：蝴蝶数据集

下载好后的文件夹中包括以下文件，放在当前目录下就可以了。

加载数据集，并对一批数据进行可视化：

import torch
import torchvision
from datasets import load_dataset
from torchvision import transforms
import numpy as np
import torch.nn.functional as F
from matplotlib import pyplot as plt
from PIL import Image

def show_images(x):
    """Given a batch of images x, make a grid and convert to PIL"""
    x = x * 0.5 + 0.5  # Map from (-1, 1) back to (0, 1)
    grid = torchvision.utils.make_grid(x)
    grid_im = grid.detach().cpu().permute(1, 2, 0).clip(0, 1) * 255
    grid_im = Image.fromarray(np.array(grid_im).astype(np.uint8))
    return grid_im

def transform(examples):
    images = [preprocess(image.convert("RGB")) for image in examples["image"]]
    return {"images": images}

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)

# 数据加载
dataset = load_dataset("./smithsonian_butterflies_subset", split='train')

image_size = 32
batch_size = 64

# 数据增强
preprocess = transforms.Compose(
    [
        transforms.Resize((image_size, image_size)),  # Resize
        transforms.RandomHorizontalFlip(),  # Randomly flip (data augmentation)
        transforms.ToTensor(),  # Convert to tensor (0, 1)
        transforms.Normalize([0.5], [0.5]),  # Map to (-1, 1)
    ]
)

dataset.set_transform(transform)

# 数据装载
train_dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True)

# 抽取一批数据可视化
xb = next(iter(train_dataloader))["images"].to(device)[:8]
print("X shape:", xb.shape)
show_images(xb).resize((8 * 64, 64), resample=Image.NEAREST)

输出可视化结果：

---

2、加噪调度器

即DDPM论文中需要预定义的 β t {\beta_t } βt ，可使用DDPMScheduler类来定义，其中num_train_timesteps参数为时间步 t {t} t 。

from diffusers import DDPMScheduler

# βt值
noise_scheduler = DDPMScheduler(num_train_timesteps=1000)

plt.figure(dpi=300)
plt.plot(noise_scheduler.alphas_cumprod.cpu() ** 0.5, label=r"${\sqrt{\bar{\alpha}_t}}$")
plt.plot((1 - noise_scheduler.alphas_cumprod.cpu()) ** 0.5, label=r"$\sqrt{(1 - \bar{\alpha}_t)}$")
plt.legend(fontsize="x-large");

根据定义的 β t {\beta_t } βt ，可视化 α ˉ t {\sqrt {{{\bar \alpha }_t}}} αˉt 和 1 − α ˉ t {\sqrt {1 - {{\bar \alpha }_t}}} 1−αˉt ：

通过设置beta_start、beta_end和beta_schedule三个参数来控制噪声调度器的超参数 β t {\beta_t } βt。

noise_scheduler = DDPMScheduler(num_train_timesteps=1000, beta_start=0.001, beta_end=0.004)

beta_schedule可以通过一个函数映射来为模型推理的每一步生成一个 β t {\beta_t } βt值。

noise_scheduler = DDPMScheduler(num_train_timesteps=1000, beta_schedule='squaredcos_cap_v2')

x t = α ˉ t x 0 + 1 − α ˉ t ε {{x_t} = \sqrt {{{\bar \alpha }_t}} {x_0} + \sqrt {1 - {{\bar \alpha }_t}} \varepsilon } xt=αˉt x0+1−αˉt ε 加噪前向过程可视化：

timesteps = torch.linspace(0, 999, 8).long().to(device)  # 随机采样时间步
noise = torch.randn_like(xb)
noisy_xb = noise_scheduler.add_noise(xb, noise, timesteps)  # 加噪
print("Noisy X shape", noisy_xb.shape)
show_images(noisy_xb).resize((8 * 64, 64), resample=Image.NEAREST)

输出为：

---

3、扩散模型定义

diffusers库中模型的定义也非常简洁：

# 创建模型
from diffusers import UNet2DModel

model = UNet2DModel(
    sample_size=image_size,  # the target image resolution
    in_channels=3,  # the number of input channels, 3 for RGB images
    out_channels=3,  # the number of output channels
    layers_per_block=2,  # how many ResNet layers to use per UNet block
    block_out_channels=(64, 128, 128, 256),  # More channels -> more parameters
    down_block_types=(
        "DownBlock2D",  # a regular ResNet downsampling block
        "DownBlock2D",
        "AttnDownBlock2D",  # a ResNet downsampling block with spatial self-attention
        "AttnDownBlock2D",
    ),
    up_block_types=(
        "AttnUpBlock2D",
        "AttnUpBlock2D",  # a ResNet upsampling block with spatial self-attention
        "UpBlock2D",
        "UpBlock2D",  # a regular ResNet upsampling block
    ),
)

model.to(device)
with torch.no_grad():
    model_prediction = model(noisy_xb, timesteps).sample
model_prediction.shape  # 验证输出与输出尺寸相同

---

4、扩散模型训练

定义优化器，和传统模型一样的训练写法：

# 定义噪声调度器
noise_scheduler = DDPMScheduler(
    num_train_timesteps=1000, beta_schedule="squaredcos_cap_v2"
)

# 优化器
optimizer = torch.optim.AdamW(model.parameters(), lr=4e-4)

losses = []

for epoch in range(30):
    for step, batch in enumerate(train_dataloader):
        clean_images = batch["images"].to(device)
        
        # 为图像添加随机噪声
        noise = torch.randn(clean_images.shape).to(clean_images.device)  # eps
        bs = clean_images.shape[0]

        # 为每一张图像随机选择一个时间步
        timesteps = torch.randint(
            0, noise_scheduler.num_train_timesteps, (bs,), device=clean_images.device
        ).long()  

        # 根据时间步,向清晰的图像中加噪声, 前向过程：根号下αt^ * x0 + 根号下(1-αt^) * eps
        noisy_images = noise_scheduler.add_noise(clean_images, noise, timesteps)

        # 获得模型预测结果
        noise_pred = model(noisy_images, timesteps, return_dict=False)[0]

        # 计算损失, 损失回传
        loss = F.mse_loss(noise_pred, noise)  
        loss.backward(loss)
        losses.append(loss.item())

        # 更新模型参数
        optimizer.step()
        optimizer.zero_grad()

    if (epoch + 1) % 5 == 0:
        loss_last_epoch = sum(losses[-len(train_dataloader) :]) / len(train_dataloader)
        print(f"Epoch:{epoch+1}, loss: {loss_last_epoch}")

30个epoch训练过程如下所示：

可用以下代码查看损失曲线：

# 损失曲线可视化
fig, axs = plt.subplots(1, 2, figsize=(12, 4))
axs[0].plot(losses)
axs[1].plot(np.log(losses))  # 对数坐标
plt.show()

损失曲线可视化：

---

5、图像生成

（1）通过建立pipeline生成图像：

# 图像生成
# 方法一：建立一个pipeline, 打包模型和噪声调度器
from diffusers import DDPMPipeline
image_pipe = DDPMPipeline(unet=model, scheduler=noise_scheduler)

pipeline_output = image_pipe()
plt.figure()
plt.imshow(pipeline_output.images[0])
plt.axis('off')
plt.show()

# 保存pipeline
image_pipe.save_pretrained("my_pipeline")  # 在当前目录下保存了一个 my_pipeline 的文件夹

生成的蝴蝶图像如下：

生成的my_pipeline文件夹如下：

（2）通过随机采样循环生成图像：

# 方法二：模型调用, 写采样循环 
# 随机初始化8张图像:
sample = torch.randn(8, 3, 32, 32).to(device)

for i, t in enumerate(noise_scheduler.timesteps):

    # 获得模型预测结果
    with torch.no_grad():
        residual = model(sample, t).sample

    # 根据预测结果更新图像
    sample = noise_scheduler.step(residual, t, sample).prev_sample

show_images(sample)

8张生成图像如下：

---

6、代码汇总

import torch
import torchvision
from datasets import load_dataset
from torchvision import transforms
import numpy as np
import torch.nn.functional as F
from matplotlib import pyplot as plt
from PIL import Image


def show_images(x):
    """Given a batch of images x, make a grid and convert to PIL"""
    x = x * 0.5 + 0.5  # Map from (-1, 1) back to (0, 1)
    grid = torchvision.utils.make_grid(x)
    grid_im = grid.detach().cpu().permute(1, 2, 0).clip(0, 1) * 255
    grid_im = Image.fromarray(np.array(grid_im).astype(np.uint8))
    return grid_im


def transform(examples):
    images = [preprocess(image.convert("RGB")) for image in examples["image"]]
    return {"images": images}

# --------------------------------------------------------------------------------
# 1、数据集加载与可视化
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)

# 数据加载
dataset = load_dataset("./smithsonian_butterflies_subset", split='train')

image_size = 32
batch_size = 64

# 数据增强
preprocess = transforms.Compose(
    [
        transforms.Resize((image_size, image_size)),  # Resize
        transforms.RandomHorizontalFlip(),  # Randomly flip (data augmentation)
        transforms.ToTensor(),  # Convert to tensor (0, 1)
        transforms.Normalize([0.5], [0.5]),  # Map to (-1, 1)
    ]
)

dataset.set_transform(transform)

# 数据装载
train_dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True)
# --------------------------------------------------------------------------------

# --------------------------------------------------------------------------------
# 抽取一批数据可视化
xb = next(iter(train_dataloader))["images"].to(device)[:8]
print("X shape:", xb.shape)
show_images(xb).resize((8 * 64, 64), resample=Image.NEAREST)
# --------------------------------------------------------------------------------

# --------------------------------------------------------------------------------
# 2、噪声调度器
from diffusers import DDPMScheduler

# 加噪声的系数βt
# noise_scheduler = DDPMScheduler(num_train_timesteps=1000)
# noise_scheduler = DDPMScheduler(num_train_timesteps=1000, beta_start=0.001, beta_end=0.004)
noise_scheduler = DDPMScheduler(num_train_timesteps=1000, beta_schedule='squaredcos_cap_v2')

plt.figure(dpi=300)
plt.plot(noise_scheduler.alphas_cumprod.cpu() ** 0.5, label=r"${\sqrt{\bar{\alpha}_t}}$")
plt.plot((1 - noise_scheduler.alphas_cumprod.cpu()) ** 0.5, label=r"$\sqrt{(1 - \bar{\alpha}_t)}$")
plt.legend(fontsize="x-large");
# --------------------------------------------------------------------------------

# --------------------------------------------------------------------------------
# 加噪声可视化
timesteps = torch.linspace(0, 999, 8).long().to(device)  # 随机采样时间步
noise = torch.randn_like(xb)
noisy_xb = noise_scheduler.add_noise(xb, noise, timesteps)  # 加噪
print("Noisy X shape", noisy_xb.shape)
show_images(noisy_xb).resize((8 * 64, 64), resample=Image.NEAREST)
# --------------------------------------------------------------------------------

# --------------------------------------------------------------------------------
# 3、创建模型
from diffusers import UNet2DModel

model = UNet2DModel(
    sample_size=image_size,  # the target image resolution
    in_channels=3,  # the number of input channels, 3 for RGB images
    out_channels=3,  # the number of output channels
    layers_per_block=2,  # how many ResNet layers to use per UNet block
    block_out_channels=(64, 128, 128, 256),  # More channels -> more parameters
    down_block_types=(
        "DownBlock2D",  # a regular ResNet downsampling block
        "DownBlock2D",
        "AttnDownBlock2D",  # a ResNet downsampling block with spatial self-attention
        "AttnDownBlock2D",
    ),
    up_block_types=(
        "AttnUpBlock2D",
        "AttnUpBlock2D",  # a ResNet upsampling block with spatial self-attention
        "UpBlock2D",
        "UpBlock2D",  # a regular ResNet upsampling block
    ),
)

model.to(device)
with torch.no_grad():
    model_prediction = model(noisy_xb, timesteps).sample
model_prediction.shape  # 验证输出与输出尺寸相同
# --------------------------------------------------------------------------------

# --------------------------------------------------------------------------------
# 4、扩散模型训练
# 定义噪声调度器
noise_scheduler = DDPMScheduler(
    num_train_timesteps=1000, beta_schedule="squaredcos_cap_v2"
)

# 优化器
optimizer = torch.optim.AdamW(model.parameters(), lr=4e-4)

losses = []

for epoch in range(30):
    for step, batch in enumerate(train_dataloader):
        clean_images = batch["images"].to(device)
        
        # 为图像添加随机噪声
        noise = torch.randn(clean_images.shape).to(clean_images.device)  # eps
        bs = clean_images.shape[0]

        # 为每一张图像随机选择一个时间步
        timesteps = torch.randint(
            0, noise_scheduler.num_train_timesteps, (bs,), device=clean_images.device
        ).long()  

        # 根据时间步,向清晰的图像中加噪声, 前向过程：根号下αt^ * x0 + 根号下(1-αt^) * eps
        noisy_images = noise_scheduler.add_noise(clean_images, noise, timesteps)

        # 获得模型预测结果
        noise_pred = model(noisy_images, timesteps, return_dict=False)[0]

        # 计算损失, 损失回传
        loss = F.mse_loss(noise_pred, noise)  
        loss.backward(loss)
        losses.append(loss.item())

        # 更新模型参数
        optimizer.step()
        optimizer.zero_grad()

    if (epoch + 1) % 5 == 0:
        loss_last_epoch = sum(losses[-len(train_dataloader) :]) / len(train_dataloader)
        print(f"Epoch:{epoch+1}, loss: {loss_last_epoch}")
# --------------------------------------------------------------------------------

# --------------------------------------------------------------------------------
# 损失曲线可视化
fig, axs = plt.subplots(1, 2, figsize=(12, 4))
axs[0].plot(losses)
axs[1].plot(np.log(losses))  # 对数坐标
plt.show()
# --------------------------------------------------------------------------------

# --------------------------------------------------------------------------------
# 5、图像生成
# 方法一：建立一个pipeline, 打包模型和噪声调度器
from diffusers import DDPMPipeline
image_pipe = DDPMPipeline(unet=model, scheduler=noise_scheduler)

pipeline_output = image_pipe()

plt.figure()
plt.imshow(pipeline_output.images[0])
plt.axis('off')
plt.show()

image_pipe.save_pretrained("my_pipeline")  # 在当前目录下保存了一个 my_pipeline 的文件夹

# 方法二：模型调用, 写采样循环 
# 随机初始化8张图像:
sample = torch.randn(8, 3, 32, 32).to(device)

for i, t in enumerate(noise_scheduler.timesteps):

    # 获得模型预测结果
    with torch.no_grad():
        residual = model(sample, t).sample

    # 根据预测结果更新图像
    sample = noise_scheduler.step(residual, t, sample).prev_sample

show_images(sample)

grid_im = show_images(sample).resize((8 * 64, 64), resample=Image.NEAREST)
plt.figure(dpi=300)
plt.imshow(grid_im)
plt.axis('off')
plt.show()
# --------------------------------------------------------------------------------

---

参考资料：扩散模型从原理到实践. 人民邮电出版社. 李忻玮, 苏步升等.

diffusers确实很方便使用，有点子PyCaret的感觉了~