tiny_vae.py

   1 #!/usr/bin/env python
   2
   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
   7 # @XREMOTE_GET: *.png
   8
   9 # Any copyright is dedicated to the Public Domain.
  10 # https://creativecommons.org/publicdomain/zero/1.0/
  11
  12 # Written by Francois Fleuret <francois@fleuret.org>
  13
  14 import sys, os, argparse, time, math, itertools
  15
  16 import torch, torchvision
  17
  18 from torch import optim, nn
  19 from torch.nn import functional as F
  20
  21 ######################################################################
  22
  23 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  24
  25 ######################################################################
  26
  27 parser = argparse.ArgumentParser(
  28     description="Very simple implementation of a VAE for teaching."
  29 )
  30
  31 parser.add_argument("--nb_epochs", type=int, default=100)
  32
  33 parser.add_argument("--learning_rate", type=float, default=2e-4)
  34
  35 parser.add_argument("--batch_size", type=int, default=100)
  36
  37 parser.add_argument("--data_dir", type=str, default="./data/")
  38
  39 parser.add_argument("--log_filename", type=str, default="train.log")
  40
  41 parser.add_argument("--latent_dim", type=int, default=32)
  42
  43 parser.add_argument("--nb_channels", type=int, default=128)
  44
  45 parser.add_argument("--no_dkl", action="store_true")
  46
  47 args = parser.parse_args()
  48
  49 log_file = open(args.log_filename, "w")
  50
  51 ######################################################################
  52
  53
  54 def log_string(s):
  55     t = time.strftime("%Y-%m-%d_%H:%M:%S - ", time.localtime())
  56
  57     if log_file is not None:
  58         log_file.write(t + s + "\n")
  59         log_file.flush()
  60
  61     print(t + s)
  62     sys.stdout.flush()
  63
  64
  65 ######################################################################
  66
  67
  68 def sample_gaussian(param):
  69     mu, log_var = param
  70     std = log_var.mul(0.5).exp()
  71     return torch.randn(mu.size(), device=mu.device) * std + mu
  72
  73
  74 def log_p_gaussian(x, param):
  75     mu, log_var = param
  76     var = log_var.exp()
  77     return (
  78         (-0.5 * ((x - mu).pow(2) / var) - 0.5 * log_var - 0.5 * math.log(2 * math.pi))
  79         .flatten(1)
  80         .sum(1)
  81     )
  82
  83
  84 def dkl_gaussians(param_a, param_b):
  85     mean_a, log_var_a = param_a[0].flatten(1), param_a[1].flatten(1)
  86     mean_b, log_var_b = param_b[0].flatten(1), param_b[1].flatten(1)
  87     var_a = log_var_a.exp()
  88     var_b = log_var_b.exp()
  89     return 0.5 * (
  90         log_var_b - log_var_a - 1 + (mean_a - mean_b).pow(2) / var_b + var_a / var_b
  91     ).sum(1)
  92
  93
  94 ######################################################################
  95
  96
  97 class LatentGivenImageNet(nn.Module):
  98     def __init__(self, nb_channels, latent_dim):
  99         super().__init__()
 100
 101         self.model = nn.Sequential(
 102             nn.Conv2d(1, nb_channels, kernel_size=1),  # to 28x28
 103             nn.ReLU(inplace=True),
 104             nn.Conv2d(nb_channels, nb_channels, kernel_size=5),  # to 24x24
 105             nn.ReLU(inplace=True),
 106             nn.Conv2d(nb_channels, nb_channels, kernel_size=5),  # to 20x20
 107             nn.ReLU(inplace=True),
 108             nn.Conv2d(nb_channels, nb_channels, kernel_size=4, stride=2),  # to 9x9
 109             nn.ReLU(inplace=True),
 110             nn.Conv2d(nb_channels, nb_channels, kernel_size=3, stride=2),  # to 4x4
 111             nn.ReLU(inplace=True),
 112             nn.Conv2d(nb_channels, 2 * latent_dim, kernel_size=4),
 113         )
 114
 115     def forward(self, x):
 116         output = self.model(x).view(x.size(0), 2, -1)
 117         mu, log_var = output[:, 0], output[:, 1]
 118         return mu, log_var
 119
 120
 121 class ImageGivenLatentNet(nn.Module):
 122     def __init__(self, nb_channels, latent_dim):
 123         super().__init__()
 124
 125         self.model = nn.Sequential(
 126             nn.ConvTranspose2d(latent_dim, nb_channels, kernel_size=4),
 127             nn.ReLU(inplace=True),
 128             nn.ConvTranspose2d(
 129                 nb_channels, nb_channels, kernel_size=3, stride=2
 130             ),  # from 4x4
 131             nn.ReLU(inplace=True),
 132             nn.ConvTranspose2d(
 133                 nb_channels, nb_channels, kernel_size=4, stride=2
 134             ),  # from 9x9
 135             nn.ReLU(inplace=True),
 136             nn.ConvTranspose2d(nb_channels, nb_channels, kernel_size=5),  # from 20x20
 137             nn.ReLU(inplace=True),
 138             nn.ConvTranspose2d(nb_channels, 2, kernel_size=5),  # from 24x24
 139         )
 140
 141     def forward(self, z):
 142         output = self.model(z.view(z.size(0), -1, 1, 1))
 143         mu, log_var = output[:, 0:1], output[:, 1:2]
 144         # log_var.flatten(1)[...]=log_var.flatten(1)[:,:1]
 145         return mu, log_var
 146
 147
 148 ######################################################################
 149
 150 data_dir = os.path.join(args.data_dir, "mnist")
 151
 152 train_set = torchvision.datasets.MNIST(data_dir, train=True, download=True)
 153 train_input = train_set.data.view(-1, 1, 28, 28).float()
 154
 155 test_set = torchvision.datasets.MNIST(data_dir, train=False, download=True)
 156 test_input = test_set.data.view(-1, 1, 28, 28).float()
 157
 158 ######################################################################
 159
 160 model_q_Z_given_x = LatentGivenImageNet(
 161     nb_channels=args.nb_channels, latent_dim=args.latent_dim
 162 )
 163
 164 model_p_X_given_z = ImageGivenLatentNet(
 165     nb_channels=args.nb_channels, latent_dim=args.latent_dim
 166 )
 167
 168 optimizer = optim.Adam(
 169     itertools.chain(model_p_X_given_z.parameters(), model_q_Z_given_x.parameters()),
 170     lr=args.learning_rate,
 171 )
 172
 173 model_p_X_given_z.to(device)
 174 model_q_Z_given_x.to(device)
 175
 176 ######################################################################
 177
 178 train_input, test_input = train_input.to(device), test_input.to(device)
 179
 180 train_mu, train_std = train_input.mean(), train_input.std()
 181 train_input.sub_(train_mu).div_(train_std)
 182 test_input.sub_(train_mu).div_(train_std)
 183
 184 ######################################################################
 185
 186 zeros = train_input.new_zeros(1, args.latent_dim)
 187
 188 param_p_Z = zeros, zeros
 189
 190 for epoch in range(args.nb_epochs):
 191     acc_loss = 0
 192
 193     for x in train_input.split(args.batch_size):
 194         param_q_Z_given_x = model_q_Z_given_x(x)
 195         z = sample_gaussian(param_q_Z_given_x)
 196         param_p_X_given_z = model_p_X_given_z(z)
 197         log_p_x_given_z = log_p_gaussian(x, param_p_X_given_z)
 198
 199         if args.no_dkl:
 200             log_q_z_given_x = log_p_gaussian(z, param_q_Z_given_x)
 201             log_p_z = log_p_gaussian(z, param_p_Z)
 202             log_p_x_z = log_p_x_given_z + log_p_x_z
 203             loss = -(log_p_x_z - log_q_z_given_x).mean()
 204         else:
 205             dkl_q_Z_given_x_from_p_Z = dkl_gaussians(param_q_Z_given_x, param_p_Z)
 206             loss = (-log_p_x_given_z + dkl_q_Z_given_x_from_p_Z).mean()
 207
 208         optimizer.zero_grad()
 209         loss.backward()
 210         optimizer.step()
 211
 212         acc_loss += loss.item() * x.size(0)
 213
 214     log_string(f"acc_loss {epoch} {acc_loss/train_input.size(0)}")
 215
 216 ######################################################################
 217
 218
 219 def save_image(x, filename):
 220     x = x * train_std + train_mu
 221     x = x.clamp(min=0, max=255) / 255
 222     torchvision.utils.save_image(1 - x, filename, nrow=16, pad_value=0.8)
 223
 224
 225 # Save a bunch of test images
 226
 227 x = test_input[:256]
 228 save_image(x, "input.png")
 229
 230 # Save the same images after encoding / decoding
 231
 232 param_q_Z_given_x = model_q_Z_given_x(x)
 233 z = sample_gaussian(param_q_Z_given_x)
 234 param_p_X_given_z = model_p_X_given_z(z)
 235 x = sample_gaussian(param_p_X_given_z)
 236 save_image(x, "output.png")
 237
 238 # Generate a bunch of images
 239
 240 z = sample_gaussian(
 241     param_p_Z[0].expand(x.size(0), -1), param_p_Z[1].expand(x.size(0), -1)
 242 )
 243 param_p_X_given_z = model_p_X_given_z(z)
 244 x = sample_gaussian(param_p_X_given_z)
 245 save_image(x, "synth.png")
 246
 247 ######################################################################