3 # @XREMOTE_HOST: elk.fleuret.org
4 # @XREMOTE_EXEC: python
5 # @XREMOTE_PRE: source ${HOME}/misc/venv/pytorch/bin/activate
6 # @XREMOTE_PRE: ln -sf ${HOME}/data/pytorch ./data
9 # Any copyright is dedicated to the Public Domain.
10 # https://creativecommons.org/publicdomain/zero/1.0/
12 # Written by Francois Fleuret <francois@fleuret.org>
14 import sys, os, argparse, time, math
16 import torch, torchvision
18 from torch import optim, nn
19 from torch.nn import functional as F
21 ######################################################################
23 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
25 ######################################################################
27 parser = argparse.ArgumentParser(
28 description="Very simple implementation of a VAE for teaching."
31 parser.add_argument("--nb_epochs", type=int, default=25)
33 parser.add_argument("--learning_rate", type=float, default=1e-3)
35 parser.add_argument("--batch_size", type=int, default=100)
37 parser.add_argument("--data_dir", type=str, default="./data/")
39 parser.add_argument("--log_filename", type=str, default="train.log")
41 parser.add_argument("--latent_dim", type=int, default=32)
43 parser.add_argument("--nb_channels", type=int, default=32)
45 parser.add_argument("--no_dkl", action="store_true")
47 parser.add_argument("--beta", type=float, default=1.0)
49 args = parser.parse_args()
51 log_file = open(args.log_filename, "w")
53 ######################################################################
57 t = time.strftime("%Y-%m-%d_%H:%M:%S ", time.localtime())
59 if log_file is not None:
60 log_file.write(t + s + "\n")
67 ######################################################################
70 def sample_categorical(param):
71 dist = torch.distributions.Categorical(logits=param)
72 return (dist.sample().unsqueeze(1).float() - train_mu) / train_std
75 def log_p_categorical(x, param):
76 x = (x.squeeze(1) * train_std + train_mu).long().clamp(min=0, max=255)
77 param = param.permute(0, 3, 1, 2)
78 return -F.cross_entropy(param, x, reduction="none").flatten(1).sum(dim=1)
81 def sample_gaussian(param):
83 std = log_var.mul(0.5).exp()
84 return torch.randn(mean.size(), device=mean.device) * std + mean
87 def log_p_gaussian(x, param):
88 mean, log_var, x = param[0].flatten(1), param[1].flatten(1), x.flatten(1)
90 return -0.5 * (((x - mean).pow(2) / var) + log_var + math.log(2 * math.pi)).sum(1)
93 def dkl_gaussians(param_a, param_b):
94 mean_a, log_var_a = param_a[0].flatten(1), param_a[1].flatten(1)
95 mean_b, log_var_b = param_b[0].flatten(1), param_b[1].flatten(1)
96 var_a = log_var_a.exp()
97 var_b = log_var_b.exp()
99 log_var_b - log_var_a - 1 + (mean_a - mean_b).pow(2) / var_b + var_a / var_b
103 def dup_param(param, nb):
104 mean, log_var = param
105 s = (nb,) + (-1,) * (mean.dim() - 1)
106 return (mean.expand(s), log_var.expand(s))
109 ######################################################################
112 class VariationalAutoEncoder(nn.Module):
113 def __init__(self, nb_channels, latent_dim):
116 self.encoder = nn.Sequential(
117 nn.Conv2d(1, nb_channels, kernel_size=1), # to 28x28
118 nn.ReLU(inplace=True),
119 nn.Conv2d(nb_channels, nb_channels, kernel_size=5), # to 24x24
120 nn.ReLU(inplace=True),
121 nn.Conv2d(nb_channels, nb_channels, kernel_size=5), # to 20x20
122 nn.ReLU(inplace=True),
123 nn.Conv2d(nb_channels, nb_channels, kernel_size=4, stride=2), # to 9x9
124 nn.ReLU(inplace=True),
125 nn.Conv2d(nb_channels, nb_channels, kernel_size=3, stride=2), # to 4x4
126 nn.ReLU(inplace=True),
127 nn.Conv2d(nb_channels, 2 * latent_dim, kernel_size=4),
130 self.decoder = nn.Sequential(
131 nn.ConvTranspose2d(latent_dim, nb_channels, kernel_size=4),
132 nn.ReLU(inplace=True),
134 nb_channels, nb_channels, kernel_size=3, stride=2
136 nn.ReLU(inplace=True),
138 nb_channels, nb_channels, kernel_size=4, stride=2
140 nn.ReLU(inplace=True),
141 nn.ConvTranspose2d(nb_channels, nb_channels, kernel_size=5), # from 20x20
142 nn.ReLU(inplace=True),
143 nn.ConvTranspose2d(nb_channels, 2, kernel_size=5), # from 24x24
147 output = self.encoder(x).view(x.size(0), 2, -1)
148 mu, log_var = output[:, 0], output[:, 1]
152 # return self.decoder(z.view(z.size(0), -1, 1, 1)).permute(0, 2, 3, 1)
153 output = self.decoder(z.view(z.size(0), -1, 1, 1))
154 mu, log_var = output[:, 0:1], output[:, 1:2]
155 log_var.flatten(1)[...] = 1 # math.log(1e-1)
156 # log_var.flatten(1)[...] = log_var.flatten(1)[:, :1]
157 # log_var = log_var.clamp(min=2*math.log(1/256))
161 ######################################################################
163 data_dir = os.path.join(args.data_dir, "mnist")
165 train_set = torchvision.datasets.MNIST(data_dir, train=True, download=True)
166 train_input = train_set.data.view(-1, 1, 28, 28).float()
168 test_set = torchvision.datasets.MNIST(data_dir, train=False, download=True)
169 test_input = test_set.data.view(-1, 1, 28, 28).float()
171 ######################################################################
174 def save_images(model, prefix=""):
175 def save_image(x, filename):
176 x = x * train_std + train_mu
177 x = x.clamp(min=0, max=255) / 255
178 torchvision.utils.save_image(1 - x, filename, nrow=16, pad_value=0.8)
179 log_string(f"wrote {filename}")
181 # Save a bunch of train images
183 x = train_input[:256]
184 save_image(x, f"{prefix}train_input.png")
186 # Save the same images after encoding / decoding
188 param_q_Z_given_x = model.encode(x)
189 z = sample_gaussian(param_q_Z_given_x)
190 param_p_X_given_z = model.decode(z)
191 x = sample_gaussian(param_p_X_given_z)
192 save_image(x, f"{prefix}train_output.png")
193 save_image(param_p_X_given_z[0], f"{prefix}train_output_mean.png")
195 # Save a bunch of test images
198 save_image(x, f"{prefix}input.png")
200 # Save the same images after encoding / decoding
202 param_q_Z_given_x = model.encode(x)
203 z = sample_gaussian(param_q_Z_given_x)
204 param_p_X_given_z = model.decode(z)
205 x = sample_gaussian(param_p_X_given_z)
206 save_image(x, f"{prefix}output.png")
207 save_image(param_p_X_given_z[0], f"{prefix}output_mean.png")
209 # Generate a bunch of images
211 z = sample_gaussian(dup_param(param_p_Z, x.size(0)))
212 param_p_X_given_z = model.decode(z)
213 x = sample_gaussian(param_p_X_given_z)
214 save_image(x, f"{prefix}synth.png")
215 save_image(param_p_X_given_z[0], f"{prefix}synth_mean.png")
218 ######################################################################
220 model = VariationalAutoEncoder(nb_channels=args.nb_channels, latent_dim=args.latent_dim)
224 ######################################################################
226 train_input, test_input = train_input.to(device), test_input.to(device)
228 train_mu, train_std = train_input.mean(), train_input.std()
229 train_input.sub_(train_mu).div_(train_std)
230 test_input.sub_(train_mu).div_(train_std)
232 ######################################################################
234 zeros = train_input.new_zeros(1, args.latent_dim)
236 param_p_Z = zeros, zeros
238 for n_epoch in range(args.nb_epochs):
239 optimizer = optim.Adam(
241 lr=args.learning_rate,
246 for x in train_input.split(args.batch_size):
247 param_q_Z_given_x = model.encode(x)
248 z = sample_gaussian(param_q_Z_given_x)
249 param_p_X_given_z = model.decode(z)
250 log_p_x_given_z = log_p_gaussian(x, param_p_X_given_z)
253 log_q_z_given_x = log_p_gaussian(z, param_q_Z_given_x)
254 log_p_z = log_p_gaussian(z, param_p_Z)
255 log_p_x_z = log_p_x_given_z + log_p_z
256 loss = -(log_p_x_z - log_q_z_given_x).mean()
258 dkl_q_Z_given_x_from_p_Z = dkl_gaussians(param_q_Z_given_x, param_p_Z)
259 loss = -(log_p_x_given_z - args.beta * dkl_q_Z_given_x_from_p_Z).mean()
261 optimizer.zero_grad()
265 acc_loss += loss.item() * x.size(0)
267 log_string(f"acc_loss {n_epoch} {acc_loss/train_input.size(0)}")
269 if (n_epoch + 1) % 25 == 0:
270 save_images(model, f"epoch_{n_epoch+1:04d}_")
272 ######################################################################