|
| 1 | +#!/usr/bin/env python |
| 2 | +# -*- coding:utf-8 -*- |
| 3 | +""" |
| 4 | + @Date : 2025/9/2 16:18 |
| 5 | + @Author : Elizabeth |
| 6 | + @Site : https://github.com/RachelElizaUK |
| 7 | +""" |
| 8 | +from typing import Optional, List, Union |
| 9 | + |
| 10 | +import torch |
| 11 | +import logging |
| 12 | +import coloredlogs |
| 13 | +from torch import nn |
| 14 | + |
| 15 | +from tqdm import tqdm |
| 16 | + |
| 17 | +from iddm.model.samples.base import BaseDiffusion |
| 18 | + |
| 19 | +logger = logging.getLogger(__name__) |
| 20 | +coloredlogs.install(level="INFO") |
| 21 | + |
| 22 | + |
| 23 | +class DPM2Diffusion(BaseDiffusion): |
| 24 | + """ |
| 25 | + DPM2 sampler class |
| 26 | + """ |
| 27 | + |
| 28 | + def __init__( |
| 29 | + self, |
| 30 | + noise_steps: int = 1000, |
| 31 | + sample_steps: int = 50, |
| 32 | + beta_start: float = 1e-4, |
| 33 | + beta_end: float = 2e-2, |
| 34 | + img_size: Optional[List[int]] = None, |
| 35 | + device: Union[str, torch.device] = "cpu", |
| 36 | + schedule_name: str = "linear", |
| 37 | + latent: bool = False, |
| 38 | + latent_channel: int = 8, |
| 39 | + autoencoder: Optional[nn.Module] = None |
| 40 | + ): |
| 41 | + """ |
| 42 | + The implement of DPM2 |
| 43 | + :param noise_steps: Total noise steps |
| 44 | + :param sample_steps: Sampling steps (DPM2 typically uses fewer steps than full noise steps) |
| 45 | + :param beta_start: β start value |
| 46 | + :param beta_end: β end value |
| 47 | + :param img_size: Image size |
| 48 | + :param device: Device type |
| 49 | + :param schedule_name: Noise schedule name |
| 50 | + :param latent: Whether to use latent diffusion |
| 51 | + :param latent_channel: Latent channel size |
| 52 | + :param autoencoder: Autoencoder model for latent diffusion |
| 53 | + """ |
| 54 | + super().__init__( |
| 55 | + noise_steps=noise_steps, |
| 56 | + beta_start=beta_start, |
| 57 | + beta_end=beta_end, |
| 58 | + img_size=img_size, |
| 59 | + device=device, |
| 60 | + schedule_name=schedule_name, |
| 61 | + latent=latent, |
| 62 | + latent_channel=latent_channel, |
| 63 | + autoencoder=autoencoder |
| 64 | + ) |
| 65 | + self.sample_steps = sample_steps |
| 66 | + self.eta = 0.0 # DPM2 uses deterministic path by default |
| 67 | + self._init_time_steps() |
| 68 | + |
| 69 | + def _init_time_steps(self): |
| 70 | + """ |
| 71 | + Initialize time steps for DPM2 sampling |
| 72 | + Creates a sequence of time steps from T down to 0 with equal intervals |
| 73 | + """ |
| 74 | + step_ratio = self.noise_steps // self.sample_steps |
| 75 | + self.time_steps = torch.arange(0, self.noise_steps, step_ratio).long() + 1 |
| 76 | + self.time_steps = reversed(torch.cat((torch.tensor([0], dtype=torch.long), self.time_steps))) |
| 77 | + self.time_steps = list(zip(self.time_steps[:-1], self.time_steps[1:])) |
| 78 | + |
| 79 | + def sample( |
| 80 | + self, |
| 81 | + model: nn.Module, |
| 82 | + x: Optional[torch.Tensor] = None, |
| 83 | + n: int = 1, |
| 84 | + labels: Optional[torch.Tensor] = None, |
| 85 | + cfg_scale: Optional[float] = None |
| 86 | + ) -> torch.Tensor: |
| 87 | + """ |
| 88 | + DPM2 sampling method |
| 89 | + :param model: Diffusion model |
| 90 | + :param x: Initial input tensor (optional) |
| 91 | + :param n: Number of samples to generate |
| 92 | + :param labels: Conditional labels (optional) |
| 93 | + :param cfg_scale: Classifier-free guidance scale (optional) |
| 94 | + :return: Generated images tensor |
| 95 | + """ |
| 96 | + # Get initial input image |
| 97 | + x, n = self._get_input_image(n=n, x=x) |
| 98 | + logger.info(msg=f"DPM2 Sampling {n} new images....") |
| 99 | + model.eval() |
| 100 | + |
| 101 | + with torch.no_grad(): |
| 102 | + for i, p_i in tqdm(self.time_steps, position=0, total=len(self.time_steps)): |
| 103 | + # Current and previous time steps |
| 104 | + t = (torch.ones(n) * i).long().to(self.device) |
| 105 | + p_t = (torch.ones(n) * p_i).long().to(self.device) |
| 106 | + |
| 107 | + # Get alpha values for current and previous steps |
| 108 | + alpha_curr = self.alpha_hat[t][:, None, None, None] |
| 109 | + alpha_prev = self.alpha_hat[p_t][:, None, None, None] |
| 110 | + |
| 111 | + # Step 1: Predict noise at current time step |
| 112 | + predicted_noise = self._get_predicted_noise(model, x, t, labels, cfg_scale) |
| 113 | + |
| 114 | + # Step 2: Compute x0 from current x and predicted noise |
| 115 | + x0 = (x - torch.sqrt(1 - alpha_curr) * predicted_noise) / torch.sqrt(alpha_curr) |
| 116 | + x0 = torch.clamp(x0, -1.0, 1.0) # Stabilize x0 prediction |
| 117 | + |
| 118 | + # Step 3: Midpoint prediction (DPM2 uses 2nd-order method) |
| 119 | + t_mid = (t + p_t) // 2 |
| 120 | + alpha_mid = self.alpha_hat[t_mid][:, None, None, None] |
| 121 | + |
| 122 | + # Compute midpoint x |
| 123 | + x_mid = torch.sqrt(alpha_mid) * x0 + torch.sqrt(1 - alpha_mid) * predicted_noise |
| 124 | + |
| 125 | + # Predict noise at midpoint |
| 126 | + predicted_noise_mid = self._get_predicted_noise(model, x_mid, t_mid, labels, cfg_scale) |
| 127 | + |
| 128 | + # Step 4: Update x using midpoint correction |
| 129 | + x = torch.sqrt(alpha_prev) * x0 + torch.sqrt(1 - alpha_prev) * predicted_noise_mid |
| 130 | + |
| 131 | + # Add noise if using stochastic sampling (eta > 0) |
| 132 | + if self.eta > 0 and i > 1: |
| 133 | + sigma = self.eta * torch.sqrt( |
| 134 | + (1 - alpha_prev) / (1 - alpha_curr) * (1 - alpha_curr / alpha_prev) |
| 135 | + ) |
| 136 | + x += sigma * torch.randn_like(x) |
| 137 | + |
| 138 | + x = self.post_process(x=x) |
| 139 | + |
| 140 | + model.train() |
| 141 | + return x |
0 commit comments