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, itertools
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(description="Tiny LeNet-like auto-encoder.")
29 parser.add_argument("--nb_epochs", type=int, default=100)
31 parser.add_argument("--batch_size", type=int, default=100)
33 parser.add_argument("--data_dir", type=str, default="./data/")
35 parser.add_argument("--log_filename", type=str, default="train.log")
37 parser.add_argument("--latent_dim", type=int, default=32)
39 parser.add_argument("--nb_channels", type=int, default=128)
41 parser.add_argument("--no_dkl", action="store_true")
43 args = parser.parse_args()
45 log_file = open(args.log_filename, "w")
47 ######################################################################
51 t = time.strftime("%Y-%m-%d_%H:%M:%S - ", time.localtime())
53 if log_file is not None:
54 log_file.write(t + s + "\n")
61 ######################################################################
64 def sample_gaussian(mu, log_var):
65 std = log_var.mul(0.5).exp()
66 return torch.randn(mu.size(), device=mu.device) * std + mu
69 def log_p_gaussian(x, mu, log_var):
72 (-0.5 * ((x - mu).pow(2) / var) - 0.5 * log_var - 0.5 * math.log(2 * math.pi))
78 def dkl_gaussians(mean_a, log_var_a, mean_b, log_var_b):
79 mean_a, log_var_a = mean_a.flatten(1), log_var_a.flatten(1)
80 mean_b, log_var_b = mean_b.flatten(1), log_var_b.flatten(1)
81 var_a = log_var_a.exp()
82 var_b = log_var_b.exp()
84 log_var_b - log_var_a - 1 + (mean_a - mean_b).pow(2) / var_b + var_a / var_b
88 ######################################################################
91 class LatentGivenImageNet(nn.Module):
92 def __init__(self, nb_channels, latent_dim):
95 self.model = nn.Sequential(
96 nn.Conv2d(1, nb_channels, kernel_size=1), # to 28x28
97 nn.ReLU(inplace=True),
98 nn.Conv2d(nb_channels, nb_channels, kernel_size=5), # to 24x24
99 nn.ReLU(inplace=True),
100 nn.Conv2d(nb_channels, nb_channels, kernel_size=5), # to 20x20
101 nn.ReLU(inplace=True),
102 nn.Conv2d(nb_channels, nb_channels, kernel_size=4, stride=2), # to 9x9
103 nn.ReLU(inplace=True),
104 nn.Conv2d(nb_channels, nb_channels, kernel_size=3, stride=2), # to 4x4
105 nn.ReLU(inplace=True),
106 nn.Conv2d(nb_channels, 2 * latent_dim, kernel_size=4),
109 def forward(self, x):
110 output = self.model(x).view(x.size(0), 2, -1)
111 mu, log_var = output[:, 0], output[:, 1]
115 class ImageGivenLatentNet(nn.Module):
116 def __init__(self, nb_channels, latent_dim):
119 self.model = nn.Sequential(
120 nn.ConvTranspose2d(latent_dim, nb_channels, kernel_size=4),
121 nn.ReLU(inplace=True),
123 nb_channels, nb_channels, kernel_size=3, stride=2
125 nn.ReLU(inplace=True),
127 nb_channels, nb_channels, kernel_size=4, stride=2
129 nn.ReLU(inplace=True),
130 nn.ConvTranspose2d(nb_channels, nb_channels, kernel_size=5), # from 20x20
131 nn.ReLU(inplace=True),
132 nn.ConvTranspose2d(nb_channels, 2, kernel_size=5), # from 24x24
135 def forward(self, z):
136 output = self.model(z.view(z.size(0), -1, 1, 1))
137 mu, log_var = output[:, 0:1], output[:, 1:2]
141 ######################################################################
143 data_dir = os.path.join(args.data_dir, "mnist")
145 train_set = torchvision.datasets.MNIST(data_dir, train=True, download=True)
146 train_input = train_set.data.view(-1, 1, 28, 28).float()
148 test_set = torchvision.datasets.MNIST(data_dir, train=False, download=True)
149 test_input = test_set.data.view(-1, 1, 28, 28).float()
151 ######################################################################
153 model_q_Z_given_x = LatentGivenImageNet(
154 nb_channels=args.nb_channels, latent_dim=args.latent_dim
157 model_p_X_given_z = ImageGivenLatentNet(
158 nb_channels=args.nb_channels, latent_dim=args.latent_dim
161 optimizer = optim.Adam(
162 itertools.chain(model_p_X_given_z.parameters(), model_q_Z_given_x.parameters()),
166 model_p_X_given_z.to(device)
167 model_q_Z_given_x.to(device)
169 ######################################################################
171 train_input, test_input = train_input.to(device), test_input.to(device)
173 train_mu, train_std = train_input.mean(), train_input.std()
174 train_input.sub_(train_mu).div_(train_std)
175 test_input.sub_(train_mu).div_(train_std)
177 ######################################################################
179 mean_p_Z = train_input.new_zeros(1, args.latent_dim)
180 log_var_p_Z = mean_p_Z
182 for epoch in range(args.nb_epochs):
185 for x in train_input.split(args.batch_size):
186 mean_q_Z_given_x, log_var_q_Z_given_x = model_q_Z_given_x(x)
187 z = sample_gaussian(mean_q_Z_given_x, log_var_q_Z_given_x)
188 mean_p_X_given_z, log_var_p_X_given_z = model_p_X_given_z(z)
191 log_q_z_given_x = log_p_gaussian(z, mean_q_Z_given_x, log_var_q_Z_given_x)
192 log_p_x_z = log_p_gaussian(
193 x, mean_p_X_given_z, log_var_p_X_given_z
194 ) + log_p_gaussian(z, mean_p_Z, log_var_p_Z)
195 loss = -(log_p_x_z - log_q_z_given_x).mean()
197 log_p_x_given_z = log_p_gaussian(x, mean_p_X_given_z, log_var_p_X_given_z)
198 dkl_q_Z_given_x_from_p_Z = dkl_gaussians(
199 mean_q_Z_given_x, log_var_q_Z_given_x, mean_p_Z, log_var_p_Z
201 loss = (-log_p_x_given_z + dkl_q_Z_given_x_from_p_Z).mean()
203 optimizer.zero_grad()
207 acc_loss += loss.item() * x.size(0)
209 log_string(f"acc_loss {epoch} {acc_loss/train_input.size(0)}")
211 ######################################################################
214 def save_image(x, filename):
215 x = x * train_std + train_mu
216 x = x.clamp(min=0, max=255) / 255
217 torchvision.utils.save_image(1 - x, filename, nrow=16, pad_value=0.8)
220 # Save a bunch of test images
223 save_image(x, "input.png")
225 # Save the same images after encoding / decoding
227 mean_q_Z_given_x, log_var_q_Z_given_x = model_q_Z_given_x(x)
228 z = sample_gaussian(mean_q_Z_given_x, log_var_q_Z_given_x)
229 mean_p_X_given_z, log_var_p_X_given_z = model_p_X_given_z(z)
230 x = sample_gaussian(mean_p_X_given_z, log_var_p_X_given_z)
231 save_image(x, "output.png")
233 # Generate a bunch of images
235 z = sample_gaussian(mean_p_Z.expand(x.size(0), -1), log_var_p_Z.expand(x.size(0), -1))
236 mean_p_X_given_z, log_var_p_X_given_z = model_p_X_given_z(z)
237 x = sample_gaussian(mean_p_X_given_z, log_var_p_X_given_z)
238 save_image(x, "synth.png")
240 ######################################################################