Spaces:
Sleeping
Sleeping
File size: 15,979 Bytes
31726e5 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 |
"""Collection of generative models."""
import torch as th
import ttools
import rendering
import modules
LOG = ttools.get_logger(__name__)
class BaseModel(th.nn.Module):
def sample_z(self, bs, device="cpu"):
return th.randn(bs, self.zdim).to(device)
class BaseVectorModel(BaseModel):
def get_vector(self, z):
_, scenes = self._forward(z)
return scenes
def _forward(self, x):
raise NotImplementedError()
def forward(self, z):
# Only return the raster
return self._forward(z)[0]
class BezierVectorGenerator(BaseVectorModel):
NUM_SEGMENTS = 2
def __init__(self, num_strokes=4,
zdim=128, width=32, imsize=32,
color_output=False,
stroke_width=None):
super(BezierVectorGenerator, self).__init__()
if stroke_width is None:
self.stroke_width = (0.5, 3.0)
LOG.warning("Setting default stroke with %s", self.stroke_width)
else:
self.stroke_width = stroke_width
self.imsize = imsize
self.num_strokes = num_strokes
self.zdim = zdim
self.trunk = th.nn.Sequential(
th.nn.Linear(zdim, width),
th.nn.SELU(inplace=True),
th.nn.Linear(width, 2*width),
th.nn.SELU(inplace=True),
th.nn.Linear(2*width, 4*width),
th.nn.SELU(inplace=True),
th.nn.Linear(4*width, 8*width),
th.nn.SELU(inplace=True),
)
# 4 points bezier with n_segments -> 3*n_segments + 1 points
self.point_predictor = th.nn.Sequential(
th.nn.Linear(8*width,
2*self.num_strokes*(
BezierVectorGenerator.NUM_SEGMENTS*3 + 1)),
th.nn.Tanh() # bound spatial extent
)
self.width_predictor = th.nn.Sequential(
th.nn.Linear(8*width, self.num_strokes),
th.nn.Sigmoid()
)
self.alpha_predictor = th.nn.Sequential(
th.nn.Linear(8*width, self.num_strokes),
th.nn.Sigmoid()
)
self.color_predictor = None
if color_output:
self.color_predictor = th.nn.Sequential(
th.nn.Linear(8*width, 3*self.num_strokes),
th.nn.Sigmoid()
)
def _forward(self, z):
bs = z.shape[0]
feats = self.trunk(z)
all_points = self.point_predictor(feats)
all_alphas = self.alpha_predictor(feats)
if self.color_predictor:
all_colors = self.color_predictor(feats)
all_colors = all_colors.view(bs, self.num_strokes, 3)
else:
all_colors = None
all_widths = self.width_predictor(feats)
min_width = self.stroke_width[0]
max_width = self.stroke_width[1]
all_widths = (max_width - min_width) * all_widths + min_width
all_points = all_points.view(
bs, self.num_strokes, BezierVectorGenerator.NUM_SEGMENTS*3+1, 2)
output, scenes = rendering.bezier_render(all_points, all_widths, all_alphas,
colors=all_colors,
canvas_size=self.imsize)
# map to [-1, 1]
output = output*2.0 - 1.0
return output, scenes
class VectorGenerator(BaseVectorModel):
def __init__(self, num_strokes=4,
zdim=128, width=32, imsize=32,
color_output=False,
stroke_width=None):
super(VectorGenerator, self).__init__()
if stroke_width is None:
self.stroke_width = (0.5, 3.0)
LOG.warning("Setting default stroke with %s", self.stroke_width)
else:
self.stroke_width = stroke_width
self.imsize = imsize
self.num_strokes = num_strokes
self.zdim = zdim
self.trunk = th.nn.Sequential(
th.nn.Linear(zdim, width),
th.nn.SELU(inplace=True),
th.nn.Linear(width, 2*width),
th.nn.SELU(inplace=True),
th.nn.Linear(2*width, 4*width),
th.nn.SELU(inplace=True),
th.nn.Linear(4*width, 8*width),
th.nn.SELU(inplace=True),
)
# straight lines so n_segments -> n_segments - 1 points
self.point_predictor = th.nn.Sequential(
th.nn.Linear(8*width, 2*(self.num_strokes*2)),
th.nn.Tanh() # bound spatial extent
)
self.width_predictor = th.nn.Sequential(
th.nn.Linear(8*width, self.num_strokes),
th.nn.Sigmoid()
)
self.alpha_predictor = th.nn.Sequential(
th.nn.Linear(8*width, self.num_strokes),
th.nn.Sigmoid()
)
self.color_predictor = None
if color_output:
self.color_predictor = th.nn.Sequential(
th.nn.Linear(8*width, 3*self.num_strokes),
th.nn.Sigmoid()
)
def _forward(self, z):
bs = z.shape[0]
feats = self.trunk(z)
all_points = self.point_predictor(feats)
all_alphas = self.alpha_predictor(feats)
if self.color_predictor:
all_colors = self.color_predictor(feats)
all_colors = all_colors.view(bs, self.num_strokes, 3)
else:
all_colors = None
all_widths = self.width_predictor(feats)
min_width = self.stroke_width[0]
max_width = self.stroke_width[1]
all_widths = (max_width - min_width) * all_widths + min_width
all_points = all_points.view(bs, self.num_strokes, 2, 2)
output, scenes = rendering.line_render(all_points, all_widths, all_alphas,
colors=all_colors,
canvas_size=self.imsize)
# map to [-1, 1]
output = output*2.0 - 1.0
return output, scenes
class RNNVectorGenerator(BaseVectorModel):
def __init__(self, num_strokes=64,
zdim=128, width=32, imsize=32,
hidden_size=512, dropout=0.9,
color_output=False,
num_layers=3, stroke_width=None):
super(RNNVectorGenerator, self).__init__()
if stroke_width is None:
self.stroke_width = (0.5, 3.0)
LOG.warning("Setting default stroke with %s", self.stroke_width)
else:
self.stroke_width = stroke_width
self.num_layers = num_layers
self.imsize = imsize
self.num_strokes = num_strokes
self.hidden_size = hidden_size
self.zdim = zdim
self.hidden_cell_predictor = th.nn.Linear(
zdim, 2*hidden_size*num_layers)
self.lstm = th.nn.LSTM(
zdim, hidden_size,
num_layers=self.num_layers, dropout=dropout,
batch_first=True)
# straight lines so n_segments -> n_segments - 1 points
self.point_predictor = th.nn.Sequential(
th.nn.Linear(hidden_size, 2*2), # 2 points, (x,y)
th.nn.Tanh() # bound spatial extent
)
self.width_predictor = th.nn.Sequential(
th.nn.Linear(hidden_size, 1),
th.nn.Sigmoid()
)
self.alpha_predictor = th.nn.Sequential(
th.nn.Linear(hidden_size, 1),
th.nn.Sigmoid()
)
def _forward(self, z, hidden_and_cell=None):
steps = self.num_strokes
# z is passed at each step, duplicate it
bs = z.shape[0]
expanded_z = z.unsqueeze(1).repeat(1, steps, 1)
# First step in the RNN
if hidden_and_cell is None:
# Initialize from latent vector
hidden_and_cell = self.hidden_cell_predictor(th.tanh(z))
hidden = hidden_and_cell[:, :self.hidden_size*self.num_layers]
hidden = hidden.view(-1, self.num_layers, self.hidden_size)
hidden = hidden.permute(1, 0, 2).contiguous()
cell = hidden_and_cell[:, self.hidden_size*self.num_layers:]
cell = cell.view(-1, self.num_layers, self.hidden_size)
cell = cell.permute(1, 0, 2).contiguous()
hidden_and_cell = (hidden, cell)
feats, hidden_and_cell = self.lstm(expanded_z, hidden_and_cell)
hidden, cell = hidden_and_cell
feats = feats.reshape(bs*steps, self.hidden_size)
all_points = self.point_predictor(feats).view(bs, steps, 2, 2)
all_alphas = self.alpha_predictor(feats).view(bs, steps)
all_widths = self.width_predictor(feats).view(bs, steps)
min_width = self.stroke_width[0]
max_width = self.stroke_width[1]
all_widths = (max_width - min_width) * all_widths + min_width
output, scenes = rendering.line_render(all_points, all_widths, all_alphas,
canvas_size=self.imsize)
# map to [-1, 1]
output = output*2.0 - 1.0
return output, scenes
class ChainRNNVectorGenerator(BaseVectorModel):
"""Strokes form a single long chain."""
def __init__(self, num_strokes=64,
zdim=128, width=32, imsize=32,
hidden_size=512, dropout=0.9,
color_output=False,
num_layers=3, stroke_width=None):
super(ChainRNNVectorGenerator, self).__init__()
if stroke_width is None:
self.stroke_width = (0.5, 3.0)
LOG.warning("Setting default stroke with %s", self.stroke_width)
else:
self.stroke_width = stroke_width
self.num_layers = num_layers
self.imsize = imsize
self.num_strokes = num_strokes
self.hidden_size = hidden_size
self.zdim = zdim
self.hidden_cell_predictor = th.nn.Linear(
zdim, 2*hidden_size*num_layers)
self.lstm = th.nn.LSTM(
zdim, hidden_size,
num_layers=self.num_layers, dropout=dropout,
batch_first=True)
# straight lines so n_segments -> n_segments - 1 points
self.point_predictor = th.nn.Sequential(
th.nn.Linear(hidden_size, 2), # 1 point, (x,y)
th.nn.Tanh() # bound spatial extent
)
self.width_predictor = th.nn.Sequential(
th.nn.Linear(hidden_size, 1),
th.nn.Sigmoid()
)
self.alpha_predictor = th.nn.Sequential(
th.nn.Linear(hidden_size, 1),
th.nn.Sigmoid()
)
def _forward(self, z, hidden_and_cell=None):
steps = self.num_strokes
# z is passed at each step, duplicate it
bs = z.shape[0]
expanded_z = z.unsqueeze(1).repeat(1, steps, 1)
# First step in the RNN
if hidden_and_cell is None:
# Initialize from latent vector
hidden_and_cell = self.hidden_cell_predictor(th.tanh(z))
hidden = hidden_and_cell[:, :self.hidden_size*self.num_layers]
hidden = hidden.view(-1, self.num_layers, self.hidden_size)
hidden = hidden.permute(1, 0, 2).contiguous()
cell = hidden_and_cell[:, self.hidden_size*self.num_layers:]
cell = cell.view(-1, self.num_layers, self.hidden_size)
cell = cell.permute(1, 0, 2).contiguous()
hidden_and_cell = (hidden, cell)
feats, hidden_and_cell = self.lstm(expanded_z, hidden_and_cell)
hidden, cell = hidden_and_cell
feats = feats.reshape(bs*steps, self.hidden_size)
# Construct the chain
end_points = self.point_predictor(feats).view(bs, steps, 1, 2)
start_points = th.cat([
# first point is canvas center
th.zeros(bs, 1, 1, 2, device=feats.device),
end_points[:, 1:, :, :]], 1)
all_points = th.cat([start_points, end_points], 2)
all_alphas = self.alpha_predictor(feats).view(bs, steps)
all_widths = self.width_predictor(feats).view(bs, steps)
min_width = self.stroke_width[0]
max_width = self.stroke_width[1]
all_widths = (max_width - min_width) * all_widths + min_width
output, scenes = rendering.line_render(all_points, all_widths, all_alphas,
canvas_size=self.imsize)
# map to [-1, 1]
output = output*2.0 - 1.0
return output, scenes
class Generator(BaseModel):
def __init__(self, width=64, imsize=32, zdim=128,
stroke_width=None,
color_output=False,
num_strokes=4):
super(Generator, self).__init__()
assert imsize == 32
self.imsize = imsize
self.zdim = zdim
num_in_chans = self.zdim // (2*2)
num_out_chans = 3 if color_output else 1
self.net = th.nn.Sequential(
th.nn.ConvTranspose2d(num_in_chans, width*8, 4, padding=1,
stride=2),
th.nn.LeakyReLU(0.2, inplace=True),
th.nn.Conv2d(width*8, width*8, 3, padding=1),
th.nn.BatchNorm2d(width*8),
th.nn.LeakyReLU(0.2, inplace=True),
# 4x4
th.nn.ConvTranspose2d(8*width, 4*width, 4, padding=1, stride=2),
th.nn.LeakyReLU(0.2, inplace=True),
th.nn.Conv2d(4*width, 4*width, 3, padding=1),
th.nn.BatchNorm2d(width*4),
th.nn.LeakyReLU(0.2, inplace=True),
# 8x8
th.nn.ConvTranspose2d(4*width, 2*width, 4, padding=1, stride=2),
th.nn.LeakyReLU(0.2, inplace=True),
th.nn.Conv2d(2*width, 2*width, 3, padding=1),
th.nn.BatchNorm2d(width*2),
th.nn.LeakyReLU(0.2, inplace=True),
# 16x16
th.nn.ConvTranspose2d(2*width, width, 4, padding=1, stride=2),
th.nn.LeakyReLU(0.2, inplace=True),
th.nn.Conv2d(width, width, 3, padding=1),
th.nn.BatchNorm2d(width),
th.nn.LeakyReLU(0.2, inplace=True),
# 32x32
th.nn.Conv2d(width, width, 3, padding=1),
th.nn.BatchNorm2d(width),
th.nn.LeakyReLU(0.2, inplace=True),
th.nn.Conv2d(width, width, 3, padding=1),
th.nn.LeakyReLU(0.2, inplace=True),
th.nn.Conv2d(width, num_out_chans, 1),
th.nn.Tanh(),
)
def forward(self, z):
bs = z.shape[0]
num_in_chans = self.zdim // (2*2)
raster = self.net(z.view(bs, num_in_chans, 2, 2))
return raster
class Discriminator(th.nn.Module):
def __init__(self, conditional=False, width=64, color_output=False):
super(Discriminator, self).__init__()
self.conditional = conditional
sn = th.nn.utils.spectral_norm
num_chan_in = 3 if color_output else 1
self.net = th.nn.Sequential(
th.nn.Conv2d(num_chan_in, width, 3, padding=1),
th.nn.LeakyReLU(0.2, inplace=True),
th.nn.Conv2d(width, 2*width, 4, padding=1, stride=2),
th.nn.LeakyReLU(0.2, inplace=True),
# 16x16
sn(th.nn.Conv2d(2*width, 2*width, 3, padding=1)),
th.nn.LeakyReLU(0.2, inplace=True),
sn(th.nn.Conv2d(2*width, 4*width, 4, padding=1, stride=2)),
th.nn.LeakyReLU(0.2, inplace=True),
# 8x8
sn(th.nn.Conv2d(4*width, 4*width, 3, padding=1)),
th.nn.LeakyReLU(0.2, inplace=True),
sn(th.nn.Conv2d(4*width, width*4, 4, padding=1, stride=2)),
th.nn.LeakyReLU(0.2, inplace=True),
# 4x4
sn(th.nn.Conv2d(4*width, 4*width, 3, padding=1)),
th.nn.LeakyReLU(0.2, inplace=True),
sn(th.nn.Conv2d(4*width, width*4, 4, padding=1, stride=2)),
th.nn.LeakyReLU(0.2, inplace=True),
# 2x2
modules.Flatten(),
th.nn.Linear(width*4*2*2, 1),
)
def forward(self, x):
out = self.net(x)
return out
|