File size: 11,963 Bytes
6de6ae4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 |
import pytorch_lightning as pl
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Dict, List, Optional, OrderedDict, Tuple
class Discriminator(nn.Module):
def __init__(
self,
hidden_size: Optional[int] = 64,
channels: Optional[int] = 3,
kernel_size: Optional[int] = 4,
stride: Optional[int] = 2,
padding: Optional[int] = 1,
negative_slope: Optional[float] = 0.2,
bias: Optional[bool] = False,
):
"""
Initializes the discriminator.
Parameters
----------
hidden_size : int, optional
The input size. (the default is 64)
channels : int, optional
The number of channels. (default: 3)
kernel_size : int, optional
The kernal size. (default: 4)
stride : int, optional
The stride. (default: 2)
padding : int, optional
The padding. (default: 1)
negative_slope : float, optional
The negative slope. (default: 0.2)
bias : bool, optional
Whether to use bias. (default: False)
"""
super().__init__()
self.hidden_size = hidden_size
self.channels = channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.negative_slope = negative_slope
self.bias = bias
self.model = nn.Sequential(
nn.utils.spectral_norm(
nn.Conv2d(
self.channels, self.hidden_size,
kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, bias=self.bias
),
),
nn.LeakyReLU(self.negative_slope, inplace=True),
nn.utils.spectral_norm(
nn.Conv2d(
hidden_size, hidden_size * 2,
kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, bias=self.bias
),
),
nn.BatchNorm2d(hidden_size * 2),
nn.LeakyReLU(self.negative_slope, inplace=True),
nn.utils.spectral_norm(
nn.Conv2d(
hidden_size * 2, hidden_size * 4,
kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, bias=self.bias
),
),
nn.BatchNorm2d(hidden_size * 4),
nn.LeakyReLU(self.negative_slope, inplace=True),
nn.utils.spectral_norm(
nn.Conv2d(
hidden_size * 4, hidden_size * 8,
kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, bias=self.bias
),
),
nn.BatchNorm2d(hidden_size * 8),
nn.LeakyReLU(self.negative_slope, inplace=True),
nn.utils.spectral_norm(
nn.Conv2d(hidden_size * 8, 1, kernel_size=4, stride=1, padding=0, bias=self.bias), # output size: (1, 1, 1)
),
nn.Flatten(),
nn.Sigmoid(),
)
def forward(self, input_img: torch.Tensor) -> torch.Tensor:
"""
Forward propagation.
Parameters
----------
input_img : torch.Tensor
The input image.
Returns
-------
torch.Tensor
The output.
"""
logits = self.model(input_img)
return logits
class Generator(nn.Module):
def __init__(
self,
hidden_size: Optional[int] = 64,
latent_size: Optional[int] = 128,
channels: Optional[int] = 3,
kernel_size: Optional[int] = 4,
stride: Optional[int] = 2,
padding: Optional[int] = 1,
bias: Optional[bool] = False,
):
"""
Initializes the generator.
Parameters
----------
hidden_size : int, optional
The hidden size. (default: 64)
latent_size : int, optional
The latent size. (default: 128)
channels : int, optional
The number of channels. (default: 3)
kernel_size : int, optional
The kernel size. (default: 4)
stride : int, optional
The stride. (default: 2)
padding : int, optional
The padding. (default: 1)
bias : bool, optional
Whether to use bias. (default: False)
"""
super().__init__()
self.hidden_size = hidden_size
self.latent_size = latent_size
self.channels = channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.bias = bias
self.model = nn.Sequential(
nn.ConvTranspose2d(
self.latent_size, self.hidden_size * 8,
kernel_size=self.kernel_size, stride=1, padding=0, bias=self.bias
),
nn.BatchNorm2d(self.hidden_size * 8),
nn.ReLU(inplace=True),
nn.ConvTranspose2d(
self.hidden_size * 8, self.hidden_size * 4,
kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, bias=self.bias
),
nn.BatchNorm2d(self.hidden_size * 4),
nn.ReLU(inplace=True),
nn.ConvTranspose2d(
self.hidden_size * 4, self.hidden_size * 2,
kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, bias=self.bias
),
nn.BatchNorm2d(self.hidden_size * 2),
nn.ReLU(inplace=True),
nn.ConvTranspose2d(
self.hidden_size * 2, self.hidden_size,
kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, bias=self.bias
),
nn.BatchNorm2d(self.hidden_size),
nn.ReLU(inplace=True),
nn.ConvTranspose2d(
self.hidden_size, self.channels,
kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, bias=self.bias
),
nn.Tanh() # output size: (channels, 64, 64)
)
def forward(self, input_noise: torch.Tensor) -> torch.Tensor:
"""
Forward propagation.
Parameters
----------
input_noise : torch.Tensor
The input image.
Returns
-------
torch.Tensor
The output.
"""
fake_img = self.model(input_noise)
return fake_img
class DocuGAN(pl.LightningModule):
def __init__(
self,
hidden_size: Optional[int] = 64,
latent_size: Optional[int] = 128,
num_channel: Optional[int] = 3,
learning_rate: Optional[float] = 0.0002,
batch_size: Optional[int] = 128,
bias1: Optional[float] = 0.5,
bias2: Optional[float] = 0.999,
):
"""
Initializes the LightningGan.
Parameters
----------
hidden_size : int, optional
The hidden size. (default: 64)
latent_size : int, optional
The latent size. (default: 128)
num_channel : int, optional
The number of channels. (default: 3)
learning_rate : float, optional
The learning rate. (default: 0.0002)
batch_size : int, optional
The batch size. (default: 128)
bias1 : float, optional
The bias1. (default: 0.5)
bias2 : float, optional
The bias2. (default: 0.999)
"""
super().__init__()
self.hidden_size = hidden_size
self.latent_size = latent_size
self.num_channel = num_channel
self.learning_rate = learning_rate
self.batch_size = batch_size
self.bias1 = bias1
self.bias2 = bias2
self.criterion = nn.BCELoss()
self.validation = torch.randn(self.batch_size, self.latent_size, 1, 1)
self.save_hyperparameters()
self.generator = Generator(
latent_size=self.latent_size, channels=self.num_channel, hidden_size=self.hidden_size
)
self.generator.apply(self.weights_init)
self.discriminator = Discriminator(channels=self.num_channel, hidden_size=self.hidden_size)
self.discriminator.apply(self.weights_init)
# self.model = InceptionV3() # For FID metric
def weights_init(self, m: nn.Module) -> None:
"""
Initializes the weights.
Parameters
----------
m : nn.Module
The module.
"""
classname = m.__class__.__name__
if classname.find("Conv") != -1:
nn.init.normal_(m.weight.data, 0.0, 0.02)
elif classname.find("BatchNorm") != -1:
nn.init.normal_(m.weight.data, 1.0, 0.02)
nn.init.constant_(m.bias.data, 0)
def configure_optimizers(self) -> Tuple[List[torch.optim.Optimizer], List]:
"""
Configures the optimizers.
Returns
-------
Tuple[List[torch.optim.Optimizer], List]
The optimizers and the LR schedulers.
"""
opt_generator = torch.optim.Adam(
self.generator.parameters(), lr=self.learning_rate, betas=(self.bias1, self.bias2)
)
opt_discriminator = torch.optim.Adam(
self.discriminator.parameters(), lr=self.learning_rate, betas=(self.bias1, self.bias2)
)
return [opt_generator, opt_discriminator], []
def forward(self, z: torch.Tensor) -> torch.Tensor:
"""
Forward propagation.
Parameters
----------
z : torch.Tensorh
The latent vector.
Returns
-------
torch.Tensor
The output.
"""
return self.generator(z)
def training_step(
self, batch: Tuple[torch.Tensor, torch.Tensor], batch_idx: int, optimizer_idx: int
) -> Dict:
"""
Training step.
Parameters
----------
batch : Tuple[torch.Tensor, torch.Tensor]
The batch.
batch_idx : int
The batch index.
optimizer_idx : int
The optimizer index.
Returns
-------
Dict
The training loss.
"""
real_images = batch["tr_image"]
if optimizer_idx == 0: # Only train the generator
fake_random_noise = torch.randn(self.batch_size, self.latent_size, 1, 1)
fake_random_noise = fake_random_noise.type_as(real_images)
fake_images = self(fake_random_noise)
# Try to fool the discriminator
preds = self.discriminator(fake_images)
loss = self.criterion(preds, torch.ones_like(preds))
self.log("g_loss", loss, on_step=False, on_epoch=True)
tqdm_dict = {"g_loss": loss}
output = OrderedDict({"loss": loss, "progress_bar": tqdm_dict, "log": tqdm_dict})
return output
elif optimizer_idx == 1: # Only train the discriminator
real_preds = self.discriminator(real_images)
real_loss = self.criterion(real_preds, torch.ones_like(real_preds))
# Generate fake images
real_random_noise = torch.randn(self.batch_size, self.latent_size, 1, 1)
real_random_noise = real_random_noise.type_as(real_images)
fake_images = self(real_random_noise)
# Pass fake images though discriminator
fake_preds = self.discriminator(fake_images)
fake_loss = self.criterion(fake_preds, torch.zeros_like(fake_preds))
# Update discriminator weights
loss = real_loss + fake_loss
self.log("d_loss", loss, on_step=False, on_epoch=True)
tqdm_dict = {"d_loss": loss}
output = OrderedDict({"loss": loss, "progress_bar": tqdm_dict, "log": tqdm_dict})
return output
|