3 # Any copyright is dedicated to the Public Domain.
4 # https://creativecommons.org/publicdomain/zero/1.0/
6 # Written by Francois Fleuret <francois@fleuret.org>
10 import matplotlib.pyplot as plt
12 import torch, torchvision
15 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
17 ######################################################################
19 def sample_gaussian_mixture(nb):
21 result = torch.empty(nb, 1).normal_(0, std)
22 result = result + torch.sign(torch.rand(result.size()) - p) / 2
25 def sample_two_discs(nb):
26 a = torch.rand(nb) * math.pi * 2
27 b = torch.rand(nb).sqrt()
28 q = (torch.rand(nb) <= 0.5).long()
29 b = b * (0.3 + 0.2 * q)
30 result = torch.empty(nb, 2)
31 result[:, 0] = a.cos() * b - 0.5 + q
32 result[:, 1] = a.sin() * b - 0.5 + q
35 def sample_disc_grid(nb):
36 a = torch.rand(nb) * math.pi * 2
37 b = torch.rand(nb).sqrt()
38 q = torch.randint(5, (nb,)) / 2.5 - 2 / 2.5
39 r = torch.randint(5, (nb,)) / 2.5 - 2 / 2.5
41 result = torch.empty(nb, 2)
42 result[:, 0] = a.cos() * b + q
43 result[:, 1] = a.sin() * b + r
46 def sample_spiral(nb):
48 rho = u * 0.65 + 0.25 + torch.rand(nb) * 0.15
49 theta = u * math.pi * 3
50 result = torch.empty(nb, 2)
51 result[:, 0] = theta.cos() * rho
52 result[:, 1] = theta.sin() * rho
56 train_set = torchvision.datasets.MNIST(root = './data/', train = True, download = True)
57 result = train_set.data[:nb].to(device).view(-1, 1, 28, 28).float()
61 'gaussian_mixture': sample_gaussian_mixture,
62 'two_discs': sample_two_discs,
63 'disc_grid': sample_disc_grid,
64 'spiral': sample_spiral,
65 'mnist': sample_mnist,
68 ######################################################################
70 parser = argparse.ArgumentParser(
71 description = '''A minimal implementation of Jonathan Ho, Ajay Jain, Pieter Abbeel
72 "Denoising Diffusion Probabilistic Models" (2020)
73 https://arxiv.org/abs/2006.11239''',
75 formatter_class = argparse.ArgumentDefaultsHelpFormatter
78 parser.add_argument('--seed',
79 type = int, default = 0,
80 help = 'Random seed, < 0 is no seeding')
82 parser.add_argument('--nb_epochs',
83 type = int, default = 100,
84 help = 'How many epochs')
86 parser.add_argument('--batch_size',
87 type = int, default = 25,
90 parser.add_argument('--nb_samples',
91 type = int, default = 25000,
92 help = 'Number of training examples')
94 parser.add_argument('--learning_rate',
95 type = float, default = 1e-3,
96 help = 'Learning rate')
98 parser.add_argument('--ema_decay',
99 type = float, default = 0.9999,
100 help = 'EMA decay, < 0 is no EMA')
102 data_list = ', '.join( [ str(k) for k in samplers ])
104 parser.add_argument('--data',
105 type = str, default = 'gaussian_mixture',
106 help = f'Toy data-set to use: {data_list}')
108 args = parser.parse_args()
111 # torch.backends.cudnn.deterministic = True
112 # torch.backends.cudnn.benchmark = False
113 # torch.use_deterministic_algorithms(True)
114 torch.manual_seed(args.seed)
115 if torch.cuda.is_available():
116 torch.cuda.manual_seed_all(args.seed)
118 ######################################################################
121 def __init__(self, model, decay):
124 if self.decay < 0: return
126 with torch.no_grad():
127 for p in model.parameters():
128 self.ema[p] = p.clone()
131 if self.decay < 0: return
132 with torch.no_grad():
133 for p in self.model.parameters():
134 self.ema[p].copy_(self.decay * self.ema[p] + (1 - self.decay) * p)
137 if self.decay < 0: return
138 with torch.no_grad():
139 for p in self.model.parameters():
142 ######################################################################
144 class ConvNet(nn.Module):
145 def __init__(self, in_channels, out_channels):
150 self.core = nn.Sequential(
151 nn.Conv2d(in_channels, nc, ks, padding = ks//2),
153 nn.Conv2d(nc, nc, ks, padding = ks//2),
155 nn.Conv2d(nc, nc, ks, padding = ks//2),
157 nn.Conv2d(nc, nc, ks, padding = ks//2),
159 nn.Conv2d(nc, nc, ks, padding = ks//2),
161 nn.Conv2d(nc, out_channels, ks, padding = ks//2),
164 def forward(self, x):
167 ######################################################################
171 train_input = samplers[args.data](args.nb_samples).to(device)
173 print(f'unknown data {args.data}')
176 train_mean, train_std = train_input.mean(), train_input.std()
178 ######################################################################
181 if train_input.dim() == 2:
184 model = nn.Sequential(
185 nn.Linear(train_input.size(1) + 1, nh),
191 nn.Linear(nh, train_input.size(1)),
194 elif train_input.dim() == 4:
196 model = ConvNet(train_input.size(1) + 1, train_input.size(1))
200 ######################################################################
204 beta = torch.linspace(1e-4, 0.02, T, device = device)
206 alpha_bar = alpha.log().cumsum(0).exp()
209 ema = EMA(model, decay = args.ema_decay)
211 for k in range(args.nb_epochs):
214 optimizer = torch.optim.Adam(model.parameters(), lr = args.learning_rate)
216 for x0 in train_input.split(args.batch_size):
217 x0 = (x0 - train_mean) / train_std
218 t = torch.randint(T, (x0.size(0),) + (1,) * (x0.dim() - 1), device = x0.device)
219 eps = torch.randn_like(x0)
220 input = torch.sqrt(alpha_bar[t]) * x0 + torch.sqrt(1 - alpha_bar[t]) * eps
221 input = torch.cat((input, t.expand_as(x0[:,:1]) / (T - 1) - 0.5), 1)
222 loss = (eps - model(input)).pow(2).mean()
223 acc_loss += loss.item() * x0.size(0)
225 optimizer.zero_grad()
231 if k%10 == 0: print(f'{k} {acc_loss / train_input.size(0)}')
235 ######################################################################
238 def generate(size, model):
239 with torch.no_grad():
240 x = torch.randn(size, device = device)
242 for t in range(T-1, -1, -1):
243 z = torch.zeros_like(x) if t == 0 else torch.randn_like(x)
244 input = torch.cat((x, torch.full_like(x[:,:1], t / (T - 1) - 0.5)), 1)
245 x = 1/torch.sqrt(alpha[t]) \
246 * (x - (1-alpha[t]) / torch.sqrt(1-alpha_bar[t]) * model(input)) \
249 x = x * train_std + train_mean
253 ######################################################################
258 if train_input.dim() == 2:
260 ax = fig.add_subplot(1, 1, 1)
262 if train_input.size(1) == 1:
264 x = generate((10000, 1), model)
266 ax.set_xlim(-1.25, 1.25)
268 d = train_input.flatten().detach().to('cpu').numpy()
269 ax.hist(d, 25, (-1, 1),
271 histtype = 'stepfilled', color = 'lightblue', label = 'Train')
273 d = x.flatten().detach().to('cpu').numpy()
274 ax.hist(d, 25, (-1, 1),
276 histtype = 'step', color = 'red', label = 'Synthesis')
278 ax.legend(frameon = False, loc = 2)
280 elif train_input.size(1) == 2:
282 x = generate((1000, 2), model)
284 ax.set_xlim(-1.25, 1.25)
285 ax.set_ylim(-1.25, 1.25)
288 d = train_input[:x.size(0)].detach().to('cpu').numpy()
289 ax.scatter(d[:, 0], d[:, 1],
290 color = 'lightblue', label = 'Train')
292 d = x.detach().to('cpu').numpy()
293 ax.scatter(d[:, 0], d[:, 1],
294 facecolors = 'none', color = 'red', label = 'Synthesis')
296 ax.legend(frameon = False, loc = 2)
298 filename = f'diffusion_{args.data}.pdf'
299 print(f'saving {filename}')
300 fig.savefig(filename, bbox_inches='tight')
302 if hasattr(plt.get_current_fig_manager(), 'window'):
303 plt.get_current_fig_manager().window.setGeometry(2, 2, 1024, 768)
306 elif train_input.dim() == 4:
307 x = generate((128,) + train_input.size()[1:], model)
308 x = 1 - x.clamp(min = 0, max = 255) / 255
309 torchvision.utils.save_image(x, f'diffusion_{args.data}.png', nrow = 16, pad_value = 0.8)
311 ######################################################################