物流企业网站建设方案,招标网站有哪些,做电影网站一年赚多少钱,平泉网站建设我在上篇博客深入浅出 diffusion#xff08;1#xff09;#xff1a;白话 diffusion 原理#xff08;无公式#xff09;中介绍了 diffusion 的一些基本原理#xff0c;其中谈到了 diffusion 的加噪过程#xff0c;本文用pytorch 实现下到底是怎么加噪的。
import torch… 我在上篇博客深入浅出 diffusion1白话 diffusion 原理无公式中介绍了 diffusion 的一些基本原理其中谈到了 diffusion 的加噪过程本文用pytorch 实现下到底是怎么加噪的。
import torch
import math
import numpy as np
from PIL import Image
import requests
import matplotlib.pyplot as plot
import cv2def linear_beta_schedule(timesteps):linear schedule, proposed in original ddpm paperscale 1000 / timestepsbeta_start scale * 0.0001beta_end scale * 0.02return torch.linspace(beta_start, beta_end, timesteps, dtype torch.float64)def cosine_beta_schedule(timesteps, s 0.008):cosine scheduleas proposed in https://openreview.net/forum?id-NEXDKk8gZsteps timesteps 1t torch.linspace(0, timesteps, steps, dtype torch.float64) / timestepsalphas_cumprod torch.cos((t s) / (1 s) * math.pi * 0.5) ** 2alphas_cumprod alphas_cumprod / alphas_cumprod[0]betas 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])return torch.clip(betas, 0, 0.999)# 时间步(timestep)定义为1000
timesteps 1000# 定义Beta Schedule, 选择线性版本同DDPM原文一致当然也可以换成cosine_beta_schedule
betas linear_beta_schedule(timestepstimesteps)# 根据beta定义alpha
alphas 1. - betas
alphas_cumprod torch.cumprod(alphas, axis0)
sqrt_recip_alphas torch.sqrt(1.0 / alphas)# 计算前向过程 diffusion q(x_t | x_{t-1}) 中所需的
sqrt_alphas_cumprod torch.sqrt(alphas_cumprod)
sqrt_one_minus_alphas_cumprod torch.sqrt(1. - alphas_cumprod)def extract(a, t, x_shape):batch_size t.shape[0]out a.gather(-1, t.cpu())return out.reshape(batch_size, *((1,) * (len(x_shape) - 1))).to(t.device)# 前向加噪过程: forward diffusion process
def q_sample(x_start, t, noiseNone):if noise is None:noise torch.randn_like(x_start)cv2.imwrite(noise.png, noise.numpy()*255)sqrt_alphas_cumprod_t extract(sqrt_alphas_cumprod, t, x_start.shape)sqrt_one_minus_alphas_cumprod_t extract(sqrt_one_minus_alphas_cumprod, t, x_start.shape)print(sqrt_alphas_cumprod_t :, sqrt_alphas_cumprod_t)print(sqrt_one_minus_alphas_cumprod_t :, sqrt_one_minus_alphas_cumprod_t)return sqrt_alphas_cumprod_t * x_start sqrt_one_minus_alphas_cumprod_t * noise# 图像后处理
def get_noisy_image(x_start, t):# add noisex_noisy q_sample(x_start, tt)# turn back into PIL imagenoisy_image x_noisy.squeeze().numpy()return noisy_image...# 展示图像, t0, 50, 100, 500的效果
x_start cv2.imread(img.png) / 255.0
x_start torch.tensor(x_start, dtypetorch.float)
cv2.imwrite(img_0.png, get_noisy_image(x_start, torch.tensor([0])) * 255.0)
cv2.imwrite(img_50.png, get_noisy_image(x_start, torch.tensor([50])) * 255.0)
cv2.imwrite(img_100.png, get_noisy_image(x_start, torch.tensor([100])) * 255.0)
cv2.imwrite(img_500.png, get_noisy_image(x_start, torch.tensor([500])) * 255.0)
cv2.imwrite(img_999.png, get_noisy_image(x_start, torch.tensor([999])) * 255.0)sqrt_alphas_cumprod_t : tensor([[[0.9999]]], dtypetorch.float64)
sqrt_one_minus_alphas_cumprod_t : tensor([[[0.0100]]], dtypetorch.float64)
sqrt_alphas_cumprod_t : tensor([[[0.9849]]], dtypetorch.float64)
sqrt_one_minus_alphas_cumprod_t : tensor([[[0.1733]]], dtypetorch.float64)
sqrt_alphas_cumprod_t : tensor([[[0.9461]]], dtypetorch.float64)
sqrt_one_minus_alphas_cumprod_t : tensor([[[0.3238]]], dtypetorch.float64)
sqrt_alphas_cumprod_t : tensor([[[0.2789]]], dtypetorch.float64)
sqrt_one_minus_alphas_cumprod_t : tensor([[[0.9603]]], dtypetorch.float64)
sqrt_alphas_cumprod_t : tensor([[[0.0064]]], dtypetorch.float64)
sqrt_one_minus_alphas_cumprod_t : tensor([[[1.0000]]], dtypetorch.float64) 以下分别为原图t 0, 50, 100, 500, 999 的结果。 可见随着 t 的加大原图对应的比例系数减小噪声的强度系数加大t 500的时候隐约可见人脸轮廓t 999 的时候人脸彻底淹没在噪声里面了。
