Spaces:
Sleeping
Sleeping
File size: 26,051 Bytes
0917e8d |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 |
import os
import torch
import time
import json
import torch.backends.cudnn as cudnn
import torch.optim as optim
import torch.nn as nn
import numpy as np
import matplotlib.pyplot as plt
from typing import Optional
from networks import Generator, Discriminator
from torch import autograd
from matplotlib.pyplot import cm
from matplotlib.patches import Rectangle
class Config:
"""Config class"""
def __init__(self, tag, root=""):
self.tag = tag
self.cli = False
# self.wandb = True
self.path = os.path.join(root, f"runs/{self.tag}")
self.cm = "gray"
self.data_path = ""
self.mask_coords = []
self.net_type = "conv-resize"
self.image_type = "n-phase"
self.l = 80
self.n_phases = 2
# Training hyperparams
self.batch_size = 4
self.beta1 = 0.9
self.beta2 = 0.999
self.max_iters = 400e3
self.timeout = 1e12
self.lrg = 0.0005
self.lr = 0.0005
self.Lambda = 10
self.critic_iters = 10
self.pw_coeff = 1
self.ngpu = torch.cuda.device_count()
if self.ngpu > 0:
self.device_name = "cuda:0"
else:
self.device_name = "cpu"
self.conv_resize = True
self.nz = 100
# Architecture
self.lays = 4
self.laysd = 5
# kernel sizes
self.dk, self.gk = [4] * self.laysd, [4] * self.lays
self.ds, self.gs = [2] * self.laysd, [2] * self.lays
self.df, self.gf = [self.n_phases, 64, 128, 256, 512, 1], [
self.nz,
512,
256,
128,
self.n_phases,
]
self.dp, self.gp = [1] * self.laysd, [2] * self.lays
# Last two layers conv resize (3,1,0)
self.gk[-2:], self.gs[-2:], self.gp[-2:] = [3, 3], [1, 1], [0, 0]
def update_params(self):
self.df[0] = self.n_phases
self.gf[-1] = self.n_phases
def save(self):
j = {}
for k, v in self.__dict__.items():
j[k] = v
with open(f"{self.path}/config.json", "w") as f:
json.dump(j, f)
def load(self):
with open(f"{self.path}/config.json", "r") as f:
j = json.load(f)
for k, v in j.items():
setattr(self, k, v)
def get_net_params(self):
return self.dk, self.ds, self.df, self.dp, self.gk, self.gs, self.gf, self.gp
def get_train_params(self):
return (
self.l,
self.batch_size,
self.beta1,
self.beta2,
self.lrg,
self.lr,
self.Lambda,
self.critic_iters,
self.nz,
)
# check for existing models and folders
def check_existence(tag, root):
"""Checks if model exists, then asks for user input. Returns True for overwrite, False for load.
:param tag: [description]
:type tag: [type]
:raises SystemExit: [description]
:raises AssertionError: [description]
:return: True for overwrite, False for load
:rtype: [type]
"""
check_D = os.path.exists(f"{root}/runs/{tag}/Disc.pt")
check_G = os.path.exists(f"{root}/runs/{tag}/Gen.pt")
if check_G or check_D:
print(f"Models already exist for tag {tag}.")
x = input(
"To overwrite existing model enter 'o', to load existing model enter 'l' or to cancel enter 'c'.\n"
)
if x == "o":
print("Overwriting")
return True
if x == "l":
print("Loading previous model")
return False
elif x == "c":
raise SystemExit
else:
raise AssertionError("Incorrect argument entered.")
return True
# set-up util
def initialise_folders(tag, overwrite, root):
"""[summary]
:param tag: [description]
:type tag: [type]
"""
if overwrite:
try:
os.mkdir(f"{root}/runs")
except:
pass
try:
os.mkdir(f"{root}/runs/{tag}")
except:
pass
# training util
def preprocess(data_path, imtype, load=True):
"""[summary]
:param imgs: [description]
:type imgs: [type]
:return: [description]
:rtype: [type]
"""
# img = tifffile.imread(data_path)
img = plt.imread(data_path)
if imtype == "colour":
img = img[:, :, :3]
img = torch.tensor(img)
if torch.max(img) > 1:
img = img / torch.max(img)
return img.permute(2, 0, 1), 3
else:
if len(img.shape) > 2:
img = img[..., 0]
if imtype == "n-phase":
phases = np.unique(img)
if len(phases) > 10:
raise AssertionError("Image not one hot encoded.")
x, y = img.shape
img_oh = torch.zeros(len(phases), x, y)
for i, ph in enumerate(phases):
img_oh[i][img == ph] = 1
return img_oh, len(phases)
elif imtype == "grayscale":
img = np.expand_dims(img, 0)
img = torch.tensor(img)
if torch.max(img) > 1:
img = img / torch.max(img)
return img, 1
def calculate_size_from_seed(seed, c):
imsize = seed
count = 0
no_layers = len(c.gk)
for k, s, p in zip(c.gk, c.gs, c.gp):
if count < no_layers - 2:
imsize = (imsize - 1) * s - 2 * p + k
elif count == no_layers - 2:
imsize = ((imsize - k + 2 * p) / s + 1).to(int)
imsize = imsize * 2 + 2
else:
imsize = ((imsize - k + 2 * p) / s + 1).to(int)
count += 1
return imsize
def calculate_seed_from_size(imsize, c):
count = 0
no_layers = len(c.gk)
for k, s, p in zip(c.gk, c.gs, c.gp):
if count < no_layers - 2:
imsize = ((imsize - k + 2 * p) / s + 1).to(int)
elif count == no_layers - 2:
imsize = (imsize - 1) * s - 2 * p + k
imsize = ((imsize - 2) / 2).to(int)
else:
imsize = (imsize - 1) * s - 2 * p + k
count += 1
return imsize
def make_mask(training_imgs, c):
y1, y2, x1, x2 = c.mask_coords
ydiff, xdiff = y2 - y1, x2 - x1
# seed for size of inpainting region
seed = calculate_seed_from_size(torch.tensor([xdiff, ydiff]).to(int), c)
# add 2 seed to each side to make the MSE region, the total G region
img_seed = seed + 4
G_out_size = calculate_size_from_seed(img_seed, c)
mask_size = calculate_size_from_seed(seed, c)
# THIS IS WHERE WE TELL D WHAT SIZE TO BE
D_seed = img_seed
x2, y2 = x1 + mask_size[0].item(), y1 + mask_size[1].item()
xmid, ymid = (x2 + x1) // 2, (y2 + y1) // 2
x1_bound, x2_bound, y1_bound, y2_bound = (
xmid - G_out_size[0].item() // 2,
xmid + G_out_size[0].item() // 2,
ymid - G_out_size[1].item() // 2,
ymid + G_out_size[1].item() // 2,
)
unmasked = training_imgs[:, x1_bound:x2_bound, y1_bound:y2_bound].clone()
training_imgs[:, x1:x2, y1:y2] = 0
mask = training_imgs[:, x1_bound:x2_bound, y1_bound:y2_bound]
mask_layer = torch.zeros_like(training_imgs[0]).unsqueeze(0)
unmasked = torch.cat([unmasked, torch.zeros_like(unmasked[0]).unsqueeze(0)])
mask_layer[:, x1:x2, y1:y2] = 1
mask = torch.cat((mask, mask_layer[:, x1_bound:x2_bound, y1_bound:y2_bound]))
# save coords to c
c.img_seed_x, c.img_seed_y = (img_seed[0].item(), img_seed[1].item())
c.mask_coords = (x1, x2, y1, y2)
c.G_out_size = (G_out_size[0].item(), G_out_size[1].item())
c.mask_size = (mask_size[0].item(), mask_size[1].item())
c.D_seed_x = D_seed[0].item()
c.D_seed_y = D_seed[1].item()
return mask, unmasked, G_out_size, img_seed, c
def update_pixmap_rect(raw, img, c, save_path=None, border=False):
updated_pixmap = raw.clone().unsqueeze(0)
x1, x2, y1, y2 = c.mask_coords
lx, ly = c.mask_size
x_1, x_2, y_1, y_2 = (
(img.shape[2] - lx) // 2,
(img.shape[2] + lx) // 2,
(img.shape[3] - ly) // 2,
(img.shape[3] + ly) // 2,
)
updated_pixmap[:, :, x1:x2, y1:y2] = img[:, :, x_1:x_2, y_1:y_2]
updated_pixmap = post_process(updated_pixmap, c).permute(0, 2, 3, 1)
if c.image_type == "grayscale":
pm = updated_pixmap[0, ...]
else:
pm = updated_pixmap[0].numpy()
if save_path:
fig, ax = plt.subplots()
if c.image_type == "grayscale":
ax.imshow(pm, cmap="gray")
rect_col = "#CC2825"
else:
ax.imshow(pm)
rect_col = "#CC2825"
# rect_col = 'white'
if border:
rect = Rectangle(
(y1, x1),
ly,
lx,
linewidth=1,
ls="--",
edgecolor=rect_col,
facecolor="none",
)
ax.add_patch(rect)
ax.set_axis_off()
plt.tight_layout()
plt.savefig("data/temp/temp_fig.png", transparent=True, pad_inches=0)
plt.close()
if c.image_type == "grayscale":
plt.imsave(c.temp_path, np.concatenate([pm for i in range(3)], -1))
else:
plt.imsave(c.temp_path, pm)
return fig
else:
if c.image_type == "grayscale":
pm = np.concatenate([pm for i in range(3)], -1)
plt.imsave(c.temp_path, pm)
return pm
def calc_gradient_penalty(
netD: Discriminator,
real_data: torch.Tensor,
fake_data: torch.Tensor,
batch_size: int,
lx: int,
ly: int,
device,
gp_lambda: float,
nc: int,
) -> torch.Tensor:
"""
Calculate gradient penalty used in WGAN-GP.
"""
# randomly weight real and fake data
alpha = torch.rand(batch_size, 1)
alpha = alpha.expand(
batch_size, int(real_data.nelement() / batch_size)
).contiguous()
alpha = alpha.view(batch_size, nc, lx, ly)
alpha = alpha.to(device)
# compute interpolate sample: (real + fake)
interpolates = alpha * real_data.detach() + ((1 - alpha) * fake_data.detach())
interpolates = interpolates.to(device)
interpolates.requires_grad_(True)
disc_interpolates = netD(interpolates)
# compute gradient of discriminator w.r.t. interpolated samples
gradients = autograd.grad(
outputs=disc_interpolates,
inputs=interpolates,
grad_outputs=torch.ones(disc_interpolates.size()).to(device),
create_graph=True,
only_inputs=True,
)[0]
# calculate gradient penalty
gradients = gradients.view(gradients.size(0), -1)
gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean() * gp_lambda
return gradient_penalty
def batch_real_poly(img, l, bs, real_seeds):
n_ph, _, _ = img.shape
max_idx = len(real_seeds[0])
idxs = torch.randint(max_idx, (bs,))
data = torch.zeros((bs, n_ph, l, l))
for i, idx in enumerate(idxs):
x, y = real_seeds[0][idx], real_seeds[1][idx]
data[i] = img[:, x : x + l, y : y + l]
return data
def batch_real(img, lx, ly, bs, mask_coords):
"""[summary]
:param training_imgs: [description]
:type training_imgs: [type]
:return: [description]
:rtype: [type]
"""
x1, x2, y1, y2 = mask_coords
n_ph, x_max, y_max = img.shape
data = torch.zeros((bs, n_ph, lx, ly))
for i in range(bs):
x, y = torch.randint(x_max - lx, (1,)), torch.randint(y_max - ly, (1,))
while (x1 < x + lx and x1 > x - lx) and (y1 < y + ly and y1 > y - ly):
x, y = torch.randint(x_max - lx, (1,)), torch.randint(y_max - ly, (1,))
data[i] = img[:, x : x + lx, y : y + ly]
return data
def pixel_wise_loss(
fake_img: torch.Tensor, real_img: torch.Tensor, unmasked, mode="mse", device=None
):
"""
Parameters
---
:param unmasked: unused?
"""
# create a mask to partially obstruct `real_img`
mask = real_img.clone().permute(1, 2, 0)
# mask out all pixels in LAST COLOR CHANNEL
# [H, W, C] -> [1, H, W, C]
mask = (mask[..., -1] == 0).unsqueeze(0)
# num of pixels not in last color channel
number_valid_pixels = mask.sum()
# pad mask
mask = mask.repeat(fake_img.shape[0], fake_img.shape[1], 1, 1)
# ???
fake_img = torch.where(mask == True, fake_img, torch.tensor(0).float().to(device))
real_img = real_img.unsqueeze(0).repeat(fake_img.shape[0], 1, 1, 1)[:, 0:-1]
real_img = torch.where(mask == True, real_img, torch.tensor(0).float().to(device))
if mode == "mse":
loss = torch.nn.MSELoss(reduction="sum")(fake_img, real_img) / (
number_valid_pixels * fake_img.shape[0] * fake_img.shape[1]
)
elif mode == "ce":
loss = -(
real_img * torch.log(fake_img) + (1 - real_img) * torch.log(1 - fake_img)
).nanmean()
return loss
# Evaluation util
def post_process(img: torch.Tensor, c: Config):
"""Turns a n phase image (bs, n, imsize, imsize) into a plottable euler image (bs, 3, imsize, imsize, imsize)
:param img: a tensor of the n phase img
:type img: torch.Tensor
:return:
:rtype:
"""
img = img.detach().cpu()
if c.image_type == "n-phase":
phases = np.arange(c.n_phases)
color = iter(cm.get_cmap(c.cm)(np.linspace(0, 1, c.n_phases)))
# color = iter([[0,0,0],[0.5,0.5,0.5], [1,1,1]])
img = torch.argmax(img, dim=1)
if len(phases) > 10:
raise AssertionError("Image not one hot encoded.")
bs, x, y = img.shape
out = torch.zeros((bs, 3, x, y))
for b in range(bs):
for i, ph in enumerate(phases):
col = next(color)
col = torch.tile(
torch.Tensor(col[0:3]).unsqueeze(1).unsqueeze(1), (x, y)
)
out[b] = torch.where((img[b] == ph), col, out[b])
out = out
else:
out = img
return out
def crop(fake_data, l, miniD=False, l_mini=16, offset=8):
w = fake_data.shape[2]
h = fake_data.shape[3]
x1, x2 = (w - l) // 2, (w + l) // 2
y1, y2 = (h - l) // 2, (h + l) // 2
out = fake_data[:, :, x1:x2, y1:y2]
return out
def init_noise(batch_size: int, nz: int, c: Config, device) -> torch.Tensor:
"""
Create and return noise tensor.
TODO: what is the shape?
Parameters
---
:param nz: number of channels
"""
noise = torch.randn(1, nz, c.seed_x, c.seed_y, device=device)
noise = torch.tile(noise, (batch_size, 1, 1, 1))
noise.requires_grad = True
return noise
def make_noise(noise, device, mask_noise=False, delta=[1, 1]):
# zeros in mask are fixed, ones are random
mask = torch.zeros_like(noise).to(device)
_, _, x, y = mask.shape
if mask_noise:
dx = torch.div(delta[0], 2, rounding_mode="floor")
dy = torch.div(delta[1], 2, rounding_mode="floor")
if dx > 0 and dy > 0:
mask[:, :, x // 2 - dx : x // 2 + dx, y // 2 - dy : y // 2 + dy] = 1
elif dx == 0:
mask[:, :, x // 2, y // 2 - dy : y // 2 + dy] = 1
elif dy == 0:
mask[:, :, x // 2 - dx : x // 2 + dx, y // 2] = 1
rand = torch.randn_like(noise).to(device) * mask
noise = noise * (mask == 0) + rand
else:
noise = torch.randn_like(noise).to(device)
return noise
def rgb2gray(rgb):
return np.dot(rgb[..., :3], [0.2989, 0.5870, 0.1140])
class RectWorker:
"""
Code: https://github.com/tldr-group/microstructure-inpainter
Paper: https://arxiv.org/pdf/2210.06997
"""
def __init__(
self,
c: Config,
netG: Generator,
netD: Discriminator,
training_imgs: torch.Tensor,
nc: int,
mask: Optional[torch.Tensor] = None,
unmasked=None,
):
super().__init__()
self.c: Config = c
self.netG: Generator = netG
self.netD: Discriminator = netD
self.training_imgs: torch.Tensor = training_imgs
self.nc: int = nc
self.mask: torch.Tensor = mask
self.unmasked: torch.Tensor = unmasked
self.quit_flag = False
self.opt_whilst_train = True
# self.opt_whilst_train = not c.cli
def stop(self):
self.quit_flag = True
def train(self, wandb=None):
"""
...
"""
# NOTE: really bad code...
overwrite = True
c: Config = self.c
Gen: Generator = self.netG
Disc: Discriminator = self.netD
training_imgs: torch.Tensor = self.training_imgs
nc: int = self.nc
mask: torch.Tensor = self.mask
unmasked = self.unmasked
ngpu = c.ngpu
tag = c.tag
path = c.path
device = torch.device(
c.device_name if (torch.cuda.is_available() and ngpu > 0) else "cpu"
)
# print(f"Using {ngpu} GPUs")
# print(device, " will be used.\n")
print(
f"Data shape: {training_imgs.shape}. Inpainting shape: {c.mask_size} Seed size: {c.img_seed_x, c.img_seed_y}"
)
cudnn.benchmark = True
# train parameters
(
l,
batch_size,
beta1,
beta2,
lrg,
lr,
Lambda,
critic_iters,
nz,
) = c.get_train_params()
mask = mask.to(device)
unmasked = unmasked.to(device)
# init noise
noise: torch.Tensor = init_noise(1, nz, c, device)
# TODO: we pass in fns; should just be model objects
netG = Gen.to(device)
netD = Disc.to(device)
# NOTE: we remove this wonky support for multiple GPUs
# -------------------------------------------------------
# if ("cuda" in str(device)) and (ngpu > 1):
# Dnet = (nn.DataParallel(netD, list(range(ngpu)))).to(device)
# netG = nn.DataParallel(netG, list(range(ngpu))).to(device)
# optimizer for discriminator/generator
optD = optim.Adam(netD.parameters(), lr=lr, betas=(beta1, beta2))
optG = optim.Adam(netG.parameters(), lr=lrg, betas=(beta1, beta2))
# NOTE: here we load model + noise from memory; I think we can just disable this
# -------------------------------------------------------
# if not overwrite:
# netG.load_state_dict(torch.load(f"{path}/Gen.pt"))
# netD.load_state_dict(torch.load(f"{path}/Disc.pt"))
# noise = torch.load(f"{c.path}/noise.pt")
# NOTE: disable wandb logging
# if c.wandb:
# wandb.wandb_init(tag, netG, netD, offline=False)
# NOTE: remove timing logging
# # start timing training
# if ("cuda" in str(device)) and (ngpu > 1):
# start_overall = torch.cuda.Event(enable_timing=True)
# end_overall = torch.cuda.Event(enable_timing=True)
# start_overall.record()
# else:
# start_overall = time.time()
i = 0
t = 0
# main training loop
while i < c.max_iters:
# discriminator training
netD.zero_grad()
netG.train()
d_noise = torch.randn_like(noise).to(device)
# generate fake sample from `d_noise` input
fake_data: torch.Tensor = netG(d_noise).detach()
# fake_data = crop(fake_data,dl)
# generate a batch of real data
real_data = batch_real(
training_imgs,
fake_data.shape[-2],
fake_data.shape[-1],
batch_size,
c.mask_coords,
).to(device)
# discriminator: predict on real data
out_real = netD(real_data).mean()
# discriminator: predict on fake data
out_fake = netD(fake_data).mean()
# calculate WGAN-GP penalty
gradient_penalty = calc_gradient_penalty(
netD,
real_data,
fake_data,
batch_size,
fake_data.shape[-2],
fake_data.shape[-1],
device,
Lambda,
nc,
)
# Compute the discriminator loss and backprop
wass = out_fake - out_real
disc_cost = wass + gradient_penalty
disc_cost.backward()
# take optimization step on discriminator
optD.step()
# if c.wandb:
# wandb.log(
# {"D_real": out_real.item(), "D_fake": out_fake.item()}, step=i
# )
# generator training
if (i % int(critic_iters)) == 0:
netG.zero_grad()
noise_G = torch.randn_like(noise).to(device)
# create a sample with generator
fake_data = netG(noise_G)
# discriminator guesses (is this data real)?
# -output ~ likelyhood this data is FAKE
output = -netD(fake_data).mean()
# hmm... how is make_noise method different from torch.rand_like?
noise_G = make_noise(noise, device, mask_noise=True, delta=[-1, -1])
# create another piece of fake data?
fake_data = netG(noise_G)
# ...
pw = pixel_wise_loss(
fake_data, mask, unmasked, mode="mse", device=device
)
output += pw * c.pw_coeff
# Calculate loss for G and backprop
output.backward(retain_graph=True)
optG.step()
# Every 100 iters log images and useful metrics
if i % 100 == 0:
netG.eval()
with torch.no_grad():
torch.save(netG.state_dict(), f"{path}/Gen.pt")
torch.save(netD.state_dict(), f"{path}/Disc.pt")
torch.save(noise, f"{path}/noise.pt")
if ("cuda" in str(device)) and (ngpu > 1):
end_overall.record()
torch.cuda.synchronize()
t = start_overall.elapsed_time(end_overall)
else:
end_overall = time.time()
t = end_overall - start_overall
if self.opt_whilst_train:
plot_noise = make_noise(
noise.detach().clone(),
device,
mask_noise=True,
delta=[-1, -1],
)
img = netG(plot_noise).detach()
pixmap = update_pixmap_rect(training_imgs, img, c)
if c.cli:
print(
f"Iter: {i}, Time: {t:.1f}, MSE: {pw.sum().item():.2g}, Wass: {abs(wass.item()):.2g}"
)
if c.wandb:
wandb.log(
{
"mse": pw.nanmean().item(),
"wass": wass.item(),
"gp": gradient_penalty.item(),
"raw out": wandb.Image(img[0].cpu()),
"inpaint out": wandb.Image(pixmap),
},
step=i,
)
else:
self.progress.emit(i, t, pw.item(), abs(wass.item()))
else:
print(f"Iter: {i}, Time {t:.1f}")
i += 1
if i == c.max_iters:
print(f"Max iterations reached: {i}")
if self.quit_flag:
self.finished.emit()
print("Quitting training")
if t > c.timeout:
print(f"Timeout: {t:.2g}")
self.finished.emit()
print("TRAINING FINISHED")
def generate(self, save_path=None, border=False, delta=None):
if self.verbose:
print("Generating new inpainted image")
device = torch.device(
self.c.device_name
if (torch.cuda.is_available() and self.c.ngpu > 0)
else "cpu"
)
netG = self.netG().to(device)
netD = self.netD().to(device)
if ("cuda" in str(device)) and (self.c.ngpu > 1):
netD = (nn.DataParallel(netD, list(range(self.c.ngpu)))).to(device)
netG = nn.DataParallel(netG, list(range(self.c.ngpu))).to(device)
netG.load_state_dict(torch.load(f"{self.c.path}/Gen.pt"))
netD.load_state_dict(torch.load(f"{self.c.path}/Disc.pt"))
noise = torch.load(f"{self.c.path}/noise.pt")
netG.eval()
with torch.no_grad():
# delta is an int that dictates how much of the centre of the seed is random
if delta is None:
if min(noise.shape[2:]) < 10:
mask_noise = False
else:
delta = torch.tensor(noise.shape[2:]) - 10
mask_noise = True
elif delta == "rand":
mask_noise = False
plot_noise = make_noise(
noise.detach().clone(), device, mask_noise=mask_noise, delta=delta
)
img = netG(plot_noise).detach()
f = update_pixmap_rect(
self.training_imgs, img, self.c, save_path=save_path, border=border
)
if save_path:
axs = f.axes
f.savefig(f"{save_path}_border.png", transparent=True)
for ax in axs:
ax.patches = []
f.savefig(f"{save_path}.png", transparent=True)
return img
|