File size: 5,843 Bytes
9f5a022 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 |
import torch
import torch.nn as nn
from enum import Enum
from tqdm import trange
Schedule = Enum('Schedule', ['LINEAR', 'COSINE'])
class DiffusionManager(nn.Module):
def __init__(self, model: nn.Module, noise_steps=1000, start=0.0001, end=0.02, device="cpu", **kwargs ) -> None:
super().__init__(**kwargs)
self.model = model
self.noise_steps = noise_steps
self.start = start
self.end = end
self.device = device
self.schedule = None
self.set_schedule()
#model.set_parent(self)
def _get_schedule(self, schedule_type: Schedule = Schedule.LINEAR):
if schedule_type == Schedule.LINEAR:
return torch.linspace(self.start, self.end, self.noise_steps)
elif schedule_type == Schedule.COSINE:
# https://arxiv.org/pdf/2102.09672 page 4
#https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
#line 18
def get_alphahat_at(t):
def f(t):
s=self.start
return torch.cos((t/self.noise_steps + s)/(1+s) * torch.pi/2) ** 2
return f(t)/f(torch.zeros_like(t))
t = torch.Tensor(range(self.noise_steps))
t = 1-(get_alphahat_at(t + 1)/get_alphahat_at(t))
t = torch.minimum(t, torch.ones_like(t) * 0.999) #"In practice, we clip β_t to be no larger than 0.999 to prevent singularities at the end of the diffusion process n"
return t
def set_schedule(self, schedule: Schedule = Schedule.LINEAR):
self.schedule = self._get_schedule(schedule).to(self.device)
def get_schedule_at(self, step):
beta = self.schedule
alpha = 1 - beta
alpha_hat = torch.cumprod(alpha, dim=0)
return self._unsqueezify(beta.data[step]), self._unsqueezify(alpha.data[step]), self._unsqueezify(alpha_hat.data[step])
@staticmethod
def _unsqueezify(value):
return value.view(-1, 1, 1, 1)#.unsqueeze(-1).unsqueeze(-1).unsqueeze(-1)
def noise_image(self, image, step):
image = image.to(self.device)
beta, alpha, alpha_hat = self.get_schedule_at(step)
epsilon = torch.randn_like(image)
# print(alpha_hat)
# print(alpha_hat.size())
# print(image.size())
noised_img = torch.sqrt(alpha_hat) * image + torch.sqrt(1 - alpha_hat) * epsilon
return noised_img, epsilon
def random_timesteps(self, amt=1):
return torch.randint(low=1, high=self.noise_steps, size=(amt,))
def sample(self, img_size, condition, amt=5, use_tqdm=True):
if tuple(condition.shape)[0] < amt:
condition = condition.repeat(amt, 1)
self.model.eval()
condition = condition.to(self.device)
my_trange = lambda x, y, z: trange(x,y, z, leave=False,dynamic_ncols=True)
fn = my_trange if use_tqdm else range
with torch.no_grad():
cur_img = torch.randn((amt, 3, img_size, img_size)).to(self.device)
for i in fn(self.noise_steps-1, 0, -1):
timestep = torch.ones(amt) * (i)
timestep = timestep.to(self.device)
predicted_noise = self.model(cur_img, timestep, condition)
beta, alpha, alpha_hat = self.get_schedule_at(i)
cur_img = (1/torch.sqrt(alpha))*(cur_img - (beta/torch.sqrt(1-alpha_hat))*predicted_noise)
if i > 1:
cur_img = cur_img + torch.sqrt(beta)*torch.randn_like(cur_img)
self.model.train()
return cur_img
def sample_multicond(self, img_size, condition, use_tqdm=True):
num_conditions = condition.shape[0]
amt = num_conditions
self.model.eval()
condition = condition.to(self.device)
my_trange = lambda x, y, z: trange(x, y, z, leave=False, dynamic_ncols=True)
fn = my_trange if use_tqdm else range
with torch.no_grad():
cur_img = torch.randn((amt, 3, img_size, img_size)).to(self.device)
for i in fn(self.noise_steps-1, 0, -1):
timestep = torch.ones(amt) * i
timestep = timestep.to(self.device)
predicted_noise = self.model(cur_img, timestep, condition)
beta, alpha, alpha_hat = self.get_schedule_at(i)
cur_img = (1 / torch.sqrt(alpha)) * (cur_img - (beta / torch.sqrt(1 - alpha_hat)) * predicted_noise)
if i > 1:
cur_img = cur_img + torch.sqrt(beta) * torch.randn_like(cur_img)
self.model.train()
# Return images sampled for each condition
return cur_img
def training_loop_iteration(self, optimizer, batch, label, criterion):
def print_(string):
for i in range(10):
print(string)
batch = batch.to(self.device)
#label = label.long() # uncomment for nn.Embedding
label = label.to(self.device)
timesteps = self.random_timesteps(batch.shape[0]).to(self.device)
noisy_batch, real_noise = self.noise_image(batch, timesteps)
if torch.isnan(noisy_batch).any() or torch.isnan(real_noise).any():
print_("NaNs detected in the noisy batch or real noise")
pred_noise = self.model(noisy_batch, timesteps, label)
if torch.isnan(pred_noise).any():
print_("NaNs detected in the predicted noise")
loss = criterion(real_noise, pred_noise)
if torch.isnan(loss).any():
print_("NaNs detected in the loss")
loss.backward()
optimizer.step()
return loss.item()
|