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=1e-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=64)
  44
  45 parser.add_argument("--no_dkl", action="store_true")
  46
  47 parser.add_argument("--beta", type=float, default=1.0)
  48
  49 args = parser.parse_args()
  50
  51 log_file = open(args.log_filename, "w")
  52
  53 ######################################################################
  54
  55
  56 def log_string(s):
  57     t = time.strftime("%Y-%m-%d_%H:%M:%S - ", time.localtime())
  58
  59     if log_file is not None:
  60         log_file.write(t + s + "\n")
  61         log_file.flush()
  62
  63     print(t + s)
  64     sys.stdout.flush()
  65
  66
  67 ######################################################################
  68
  69
  70 def sample_gaussian(param):
  71     mu, log_var = param
  72     std = log_var.mul(0.5).exp()
  73     return torch.randn(mu.size(), device=mu.device) * std + mu
  74
  75
  76 def log_p_gaussian(x, param):
  77     mu, log_var = param
  78     var = log_var.exp()
  79     return (
  80         (-0.5 * ((x - mu).pow(2) / var) - 0.5 * log_var - 0.5 * math.log(2 * math.pi))
  81         .flatten(1)
  82         .sum(1)
  83     )
  84
  85
  86 def dkl_gaussians(param_a, param_b):
  87     mean_a, log_var_a = param_a[0].flatten(1), param_a[1].flatten(1)
  88     mean_b, log_var_b = param_b[0].flatten(1), param_b[1].flatten(1)
  89     var_a = log_var_a.exp()
  90     var_b = log_var_b.exp()
  91     return 0.5 * (
  92         log_var_b - log_var_a - 1 + (mean_a - mean_b).pow(2) / var_b + var_a / var_b
  93     ).sum(1)
  94
  95
  96 ######################################################################
  97
  98
  99 class LatentGivenImageNet(nn.Module):
 100     def __init__(self, nb_channels, latent_dim):
 101         super().__init__()
 102
 103         self.model = nn.Sequential(
 104             nn.Conv2d(1, nb_channels, kernel_size=1),  # to 28x28
 105             nn.ReLU(inplace=True),
 106             nn.Conv2d(nb_channels, nb_channels, kernel_size=5),  # to 24x24
 107             nn.ReLU(inplace=True),
 108             nn.Conv2d(nb_channels, nb_channels, kernel_size=5),  # to 20x20
 109             nn.ReLU(inplace=True),
 110             nn.Conv2d(nb_channels, nb_channels, kernel_size=4, stride=2),  # to 9x9
 111             nn.ReLU(inplace=True),
 112             nn.Conv2d(nb_channels, nb_channels, kernel_size=3, stride=2),  # to 4x4
 113             nn.ReLU(inplace=True),
 114             nn.Conv2d(nb_channels, 2 * latent_dim, kernel_size=4),
 115         )
 116
 117     def forward(self, x):
 118         output = self.model(x).view(x.size(0), 2, -1)
 119         mu, log_var = output[:, 0], output[:, 1]
 120         return mu, log_var
 121
 122
 123 class ImageGivenLatentNet(nn.Module):
 124     def __init__(self, nb_channels, latent_dim):
 125         super().__init__()
 126
 127         self.model = nn.Sequential(
 128             nn.ConvTranspose2d(latent_dim, nb_channels, kernel_size=4),
 129             nn.ReLU(inplace=True),
 130             nn.ConvTranspose2d(
 131                 nb_channels, nb_channels, kernel_size=3, stride=2
 132             ),  # from 4x4
 133             nn.ReLU(inplace=True),
 134             nn.ConvTranspose2d(
 135                 nb_channels, nb_channels, kernel_size=4, stride=2
 136             ),  # from 9x9
 137             nn.ReLU(inplace=True),
 138             nn.ConvTranspose2d(nb_channels, nb_channels, kernel_size=5),  # from 20x20
 139             nn.ReLU(inplace=True),
 140             nn.ConvTranspose2d(nb_channels, 2, kernel_size=5),  # from 24x24
 141         )
 142
 143     def forward(self, z):
 144         output = self.model(z.view(z.size(0), -1, 1, 1))
 145         mu, log_var = output[:, 0:1], output[:, 1:2]
 146         # log_var.flatten(1)[...] = log_var.flatten(1)[:, :1]
 147         return mu, log_var
 148
 149
 150 ######################################################################
 151
 152 data_dir = os.path.join(args.data_dir, "mnist")
 153
 154 train_set = torchvision.datasets.MNIST(data_dir, train=True, download=True)
 155 train_input = train_set.data.view(-1, 1, 28, 28).float()
 156
 157 test_set = torchvision.datasets.MNIST(data_dir, train=False, download=True)
 158 test_input = test_set.data.view(-1, 1, 28, 28).float()
 159
 160 ######################################################################
 161
 162
 163 def save_images(model_q_Z_given_x, model_p_X_given_z, prefix=""):
 164     def save_image(x, filename):
 165         x = x * train_std + train_mu
 166         x = x.clamp(min=0, max=255) / 255
 167         torchvision.utils.save_image(1 - x, filename, nrow=16, pad_value=0.8)
 168         log_string(f"wrote {filename}")
 169
 170     # Save a bunch of train images
 171
 172     x = train_input[:256]
 173     save_image(x, f"{prefix}train_input.png")
 174
 175     # Save the same images after encoding / decoding
 176
 177     param_q_Z_given_x = model_q_Z_given_x(x)
 178     z = sample_gaussian(param_q_Z_given_x)
 179     param_p_X_given_z = model_p_X_given_z(z)
 180     x = sample_gaussian(param_p_X_given_z)
 181     save_image(x, f"{prefix}train_output.png")
 182     save_image(param_p_X_given_z[0], f"{prefix}train_output_mean.png")
 183
 184     # Save a bunch of test images
 185
 186     x = test_input[:256]
 187     save_image(x, f"{prefix}input.png")
 188
 189     # Save the same images after encoding / decoding
 190
 191     param_q_Z_given_x = model_q_Z_given_x(x)
 192     z = sample_gaussian(param_q_Z_given_x)
 193     param_p_X_given_z = model_p_X_given_z(z)
 194     x = sample_gaussian(param_p_X_given_z)
 195     save_image(x, f"{prefix}output.png")
 196     save_image(param_p_X_given_z[0], f"{prefix}output_mean.png")
 197
 198     # Generate a bunch of images
 199
 200     z = sample_gaussian(
 201         (param_p_Z[0].expand(x.size(0), -1), param_p_Z[1].expand(x.size(0), -1))
 202     )
 203     param_p_X_given_z = model_p_X_given_z(z)
 204     x = sample_gaussian(param_p_X_given_z)
 205     save_image(x, f"{prefix}synth.png")
 206     save_image(param_p_X_given_z[0], f"{prefix}synth_mean.png")
 207
 208
 209 ######################################################################
 210
 211 model_q_Z_given_x = LatentGivenImageNet(
 212     nb_channels=args.nb_channels, latent_dim=args.latent_dim
 213 )
 214
 215 model_p_X_given_z = ImageGivenLatentNet(
 216     nb_channels=args.nb_channels, latent_dim=args.latent_dim
 217 )
 218
 219 optimizer = optim.Adam(
 220     itertools.chain(model_p_X_given_z.parameters(), model_q_Z_given_x.parameters()),
 221     lr=args.learning_rate,
 222 )
 223
 224 model_p_X_given_z.to(device)
 225 model_q_Z_given_x.to(device)
 226
 227 ######################################################################
 228
 229 train_input, test_input = train_input.to(device), test_input.to(device)
 230
 231 train_mu, train_std = train_input.mean(), train_input.std()
 232 train_input.sub_(train_mu).div_(train_std)
 233 test_input.sub_(train_mu).div_(train_std)
 234
 235 ######################################################################
 236
 237 zeros = train_input.new_zeros(1, args.latent_dim)
 238
 239 param_p_Z = zeros, zeros
 240
 241 for n_epoch in range(args.nb_epochs):
 242     acc_loss = 0
 243
 244     for x in train_input.split(args.batch_size):
 245         param_q_Z_given_x = model_q_Z_given_x(x)
 246         z = sample_gaussian(param_q_Z_given_x)
 247         param_p_X_given_z = model_p_X_given_z(z)
 248         log_p_x_given_z = log_p_gaussian(x, param_p_X_given_z)
 249
 250         if args.no_dkl:
 251             log_q_z_given_x = log_p_gaussian(z, param_q_Z_given_x)
 252             log_p_z = log_p_gaussian(z, param_p_Z)
 253             log_p_x_z = log_p_x_given_z + log_p_x_z
 254             loss = -(log_p_x_z - log_q_z_given_x).mean()
 255         else:
 256             dkl_q_Z_given_x_from_p_Z = dkl_gaussians(param_q_Z_given_x, param_p_Z)
 257             loss = -(log_p_x_given_z - args.beta * dkl_q_Z_given_x_from_p_Z).mean()
 258
 259         optimizer.zero_grad()
 260         loss.backward()
 261         optimizer.step()
 262
 263         acc_loss += loss.item() * x.size(0)
 264
 265     log_string(f"acc_loss {n_epoch} {acc_loss/train_input.size(0)}")
 266
 267     if (n_epoch + 1) % 25 == 0:
 268         save_images(model_q_Z_given_x, model_p_X_given_z, f"epoch_{n_epoch+1:04d}_")
 269
 270 ######################################################################