File size: 39,117 Bytes
3c8ff2e |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 |
"""
Main script for image reconstruction experiments
Author: Patrick Ebel (github/PatrickTUM), based on the scripts of
Vivien Sainte Fare Garnot (github/VSainteuf)
License: MIT
"""
import os
import sys
import time
import json
import random
import pprint
import argparse
import numpy as np
from tqdm import tqdm
from matplotlib import pyplot as plt
dirname = os.path.dirname(os.path.abspath(__file__))
sys.path.append(os.path.dirname(dirname))
from parse_args import create_parser
from data.dataLoader import SEN12MSCR
from src.model_utils import (
get_model,
save_model,
freeze_layers,
load_model,
load_checkpoint,
)
from src.learning.metrics import img_metrics, avg_img_metrics
import torch
import torchnet as tnt
from torch.utils.tensorboard import SummaryWriter
from src import utils, losses
from src.learning.weight_init import weight_init
S2_BANDS = 13
parser = create_parser(mode="train")
config = utils.str2list(
parser.parse_args(), list_args=["encoder_widths", "decoder_widths", "out_conv"]
)
if config.model in ["unet", "utae"]:
assert len(config.encoder_widths) == len(config.decoder_widths)
config.loss = "l2"
if config.model == "unet":
# train U-Net from scratch
config.pretrain = True
config.trained_checkp = ""
if config.pretrain: # pre-training is on a single time point
config.input_t = config.n_head = 1
config.sample_type = "pretrain"
if config.model == "unet":
config.batch_size = 32
config.positional_encoding = False
if config.loss in ["GNLL", "MGNLL"]:
# for univariate losses, default to univariate mode (batched across channels)
if config.loss in ["GNLL"]:
config.covmode = "uni"
if config.covmode == "iso":
config.out_conv[-1] += 1
elif config.covmode in ["uni", "diag"]:
config.out_conv[-1] += S2_BANDS
config.var_nonLinearity = "softplus"
# grab the PID so we can look it up in the logged config for server-side process management
config.pid = os.getpid()
# import & re-load a previous configuration, e.g. to resume training
if config.resume_from:
load_conf = os.path.join(config.res_dir, config.experiment_name, "conf.json")
if config.experiment_name != config.trained_checkp.split("/")[-2]:
raise ValueError("Mismatch of loaded config file and checkpoints")
with open(load_conf, "rt") as f:
t_args = argparse.Namespace()
# do not overwrite the following flags by their respective values in the config file
no_overwrite = [
"pid",
"num_workers",
"root1",
"root2",
"root3",
"resume_from",
"trained_checkp",
"epochs",
"encoder_widths",
"decoder_widths",
"lr",
]
conf_dict = {
key: val for key, val in json.load(f).items() if key not in no_overwrite
}
for key, val in vars(config).items():
if key in no_overwrite:
conf_dict[key] = val
t_args.__dict__.update(conf_dict)
config = parser.parse_args(namespace=t_args)
config = utils.str2list(
config, list_args=["encoder_widths", "decoder_widths", "out_conv"]
)
# resume at a specified epoch and update optimizer accordingly
if config.resume_at >= 0:
config.lr = config.lr * config.gamma**config.resume_at
# fix all RNG seeds,
# throw the whole bunch at 'em
def seed_packages(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
os.environ["PYTHONHASHSEED"] = str(seed)
# torch.use_deterministic_algorithms(True, warn_only=True)
torch.backends.cudnn.benchmark = False
def seed_worker(worker_id):
worker_seed = torch.initial_seed() % 2**32
np.random.seed(worker_seed)
random.seed(worker_seed)
# seed everything
seed_packages(config.rdm_seed)
# seed generators for train & val/test dataloaders
f, g = torch.Generator(), torch.Generator()
f.manual_seed(config.rdm_seed + 0) # note: this may get re-seeded each epoch
g.manual_seed(config.rdm_seed) # keep this one fixed
if __name__ == "__main__":
pprint.pprint(config)
# instantiate tensorboard logger
writer = SummaryWriter(
os.path.join(os.path.dirname(config.res_dir), "logs", config.experiment_name)
)
def plot_img(imgs, mod, plot_dir, file_id=None):
if not os.path.exists(plot_dir):
os.makedirs(plot_dir)
try:
imgs = imgs.cpu().numpy()
for tdx, img in enumerate(imgs): # iterate over temporal dimension
time = "" if imgs.shape[0] == 1 else f"_t-{tdx}"
if mod in ["pred", "in", "target", "s2"]:
rgb = [3, 2, 1] if img.shape[0] == S2_BANDS else [5, 4, 3]
img, val_min, val_max = img[rgb, ...], 0, 1
elif mod == "s1":
img, val_min, val_max = img[[0], ...], 0, 1
elif mod == "mask":
img, val_min, val_max = img[[0], ...], 0, 1
elif mod == "err":
img, val_min, val_max = img[[0], ...], 0, 0.01
elif mod == "var":
img, val_min, val_max = img[[0], ...], 0, 0.000025
else:
raise NotImplementedError
if file_id is not None: # export into file name
img = img.clip(
val_min, val_max
) # note: this only removes outliers, vmin/vmax below do the global rescaling (else doing instance-wise min/max scaling)
plt.imsave(
os.path.join(plot_dir, f"img-{file_id}_{mod}{time}.png"),
np.moveaxis(img, 0, -1).squeeze(),
dpi=100,
cmap="gray",
vmin=val_min,
vmax=val_max,
)
except:
if isinstance(imgs, plt.Figure): # the passed argument is a pre-rendered figure
plt.savefig(os.path.join(plot_dir, f"img-{file_id}_{mod}.png"), dpi=100)
else:
raise NotImplementedError
def export(arrs, mod, export_dir, file_id=None):
if not os.path.exists(export_dir):
os.makedirs(export_dir)
for tdx, arr in enumerate(arrs): # iterate over temporal dimension
num = "" if arrs.shape[0] == 1 else f"_t-{tdx}"
np.save(os.path.join(export_dir, f"img-{file_id}_{mod}{num}.npy"), arr.cpu())
def prepare_data(batch, device, config):
if config.pretrain:
return prepare_data_mono(batch, device, config)
else:
return prepare_data_multi(batch, device, config)
def prepare_data_mono(batch, device, config):
x = batch["input"]["S2"].to(device).unsqueeze(1)
if config.use_sar:
x = torch.cat((batch["input"]["S1"].to(device).unsqueeze(1), x), dim=2)
m = batch["input"]["masks"].to(device).unsqueeze(1)
y = batch["target"]["S2"].to(device).unsqueeze(1)
return x, y, m
def prepare_data_multi(batch, device, config):
in_S2 = recursive_todevice(batch["input"]["S2"], device)
in_S2_td = recursive_todevice(batch["input"]["S2 TD"], device)
if config.batch_size > 1:
in_S2_td = torch.stack((in_S2_td)).T
in_m = torch.stack(recursive_todevice(batch["input"]["masks"], device)).swapaxes(
0, 1
)
target_S2 = recursive_todevice(batch["target"]["S2"], device)
y = torch.cat(target_S2, dim=0).unsqueeze(1)
if config.use_sar:
in_S1 = recursive_todevice(batch["input"]["S1"], device)
in_S1_td = recursive_todevice(batch["input"]["S1 TD"], device)
if config.batch_size > 1:
in_S1_td = torch.stack((in_S1_td)).T
x = torch.cat((torch.stack(in_S1, dim=1), torch.stack(in_S2, dim=1)), dim=2)
dates = (
torch.stack((torch.tensor(in_S1_td), torch.tensor(in_S2_td)))
.float()
.mean(dim=0)
.to(device)
)
else:
x = torch.stack(in_S2, dim=1)
dates = torch.tensor(in_S2_td).float().to(device)
return x, y, in_m, dates
def log_aleatoric(writer, config, mode, step, var, name, img_meter=None):
# if var is of shape [B x 1 x C x C x H x W] then it's a covariance tensor
if len(var.shape) > 5:
covar = var
# get [B x 1 x C x H x W] variance tensor
var = var.diagonal(dim1=2, dim2=3).moveaxis(-1, 2)
# compute spatial-average to visualize patch-wise covariance matrices
patch_covmat = covar.mean(dim=-1).mean(dim=-1).squeeze(dim=1)
for bdx, img in enumerate(patch_covmat): # iterate over [B x C x C] covmats
img = img.detach().numpy()
max_abs = max(abs(img.min()), abs(img.max()))
scale_rel_left, scale_rel_right = -max_abs, +max_abs
fig = continuous_matshow(img, min=scale_rel_left, max=scale_rel_right)
writer.add_figure(f"Img/{mode}/patch covmat relative {bdx}", fig, step)
scale_center0_absolute = (
1 / 4 * 1**2
) # assuming covmat has been rescaled already, this is an upper bound
fig = continuous_matshow(
img, min=-scale_center0_absolute, max=scale_center0_absolute
)
writer.add_figure(f"Img/{mode}/patch covmat absolute {bdx}", fig, step)
# aleatoric uncertainty: comput during train, val and test
# note: the quantile statistics are computed solely over the variances (and would be much different if involving covariances, e.g. in the isotopic case)
avg_var = torch.mean(
var, dim=2, keepdim=True
) # avg over bands, note: this only considers variances (else diag COV's avg would be tiny)
q50 = (
avg_var[:, 0, ...].view(avg_var.shape[0], -1).median(dim=-1)[0].detach().clone()
)
q75 = (
avg_var[:, 0, ...]
.view(avg_var.shape[0], -1)
.quantile(0.75, dim=-1)
.detach()
.clone()
)
q50, q75 = q50[0], q75[0] # take batch's first item as a summary
binning = 256 # see: https://pytorch.org/docs/stable/tensorboard.html#torch.utils.tensorboard.writer.SummaryWriter.add_histogram
if config.loss in ["GNLL", "MGNLL"]:
writer.add_image(
f"Img/{mode}/{name}aleatoric [0,1]",
avg_var[0, 0, ...].clip(0, 1),
step,
dataformats="CHW",
) # map image to [0, 1]
writer.add_image(
f"Img/{mode}/{name}aleatoric [0,q75]",
avg_var[0, 0, ...].clip(0.0, q75) / q75,
step,
dataformats="CHW",
) # map image to [0, q75]
writer.add_histogram(
f"Hist/{mode}/{name}aleatoric",
avg_var[0, 0, ...].flatten().clip(0, 1),
step,
bins=binning,
max_bins=binning,
)
else:
raise NotImplementedError
writer.add_scalar(f"{mode}/{name}aleatoric median all", q50, step)
writer.add_scalar(f"{mode}/{name}aleatoric q75 all", q75, step)
if img_meter is not None:
writer.add_scalar(f"{mode}/{name}UCE SE", img_meter.value()["UCE SE"], step)
writer.add_scalar(f"{mode}/{name}AUCE SE", img_meter.value()["AUCE SE"], step)
def log_train(writer, config, model, step, x, out, y, in_m, name="", var=None):
# logged loss is before rescaling by learning rate
_, loss = model.criterion, model.loss_G.cpu()
if name != "":
name = f"model_{name}/"
writer.add_scalar(f"train/{name}{config.loss}", loss, step)
writer.add_scalar(f"train/{name}total", loss, step)
# use add_images for batch-wise adding across temporal dimension
if config.use_sar:
writer.add_image(
f"Img/train/{name}in_s1", x[0, :, [0], ...], step, dataformats="NCHW"
)
writer.add_image(
f"Img/train/{name}in_s2", x[0, :, [5, 4, 3], ...], step, dataformats="NCHW"
)
else:
writer.add_image(
f"Img/train/{name}in_s2", x[0, :, [3, 2, 1], ...], step, dataformats="NCHW"
)
writer.add_image(
f"Img/train/{name}out", out[0, 0, [3, 2, 1], ...], step, dataformats="CHW"
)
writer.add_image(
f"Img/train/{name}y", y[0, 0, [3, 2, 1], ...], step, dataformats="CHW"
)
writer.add_image(
f"Img/train/{name}m", in_m[0, :, None, ...], step, dataformats="NCHW"
)
# analyse cloud coverage
# covered at ALL time points (AND) or covered at ANY time points (OR)
# and_m, or_m = torch.prod(in_m[0,:, ...], dim=0, keepdim=True), torch.sum(in_m[0,:, ...], dim=0, keepdim=True).clip(0,1)
and_m, or_m = torch.prod(in_m, dim=1, keepdim=True), torch.sum(
in_m, dim=1, keepdim=True
).clip(0, 1)
writer.add_scalar(f"train/{name}OR m %", or_m.float().mean(), step)
writer.add_scalar(f"train/{name}AND m %", and_m.float().mean(), step)
writer.add_image(f"Img/train/{name}AND m", and_m, step, dataformats="NCHW")
writer.add_image(f"Img/train/{name}OR m", or_m, step, dataformats="NCHW")
and_m_gray = in_m.float().mean(axis=1).cpu()
for bdx, img in enumerate(and_m_gray):
fig = discrete_matshow(img, n_colors=config.input_t)
writer.add_figure(f"Img/train/temp overlay m {bdx}", fig, step)
if var is not None:
# log aleatoric uncertainty statistics, excluding computation of ECE
log_aleatoric(writer, config, "train", step, var, name, img_meter=None)
def discrete_matshow(data, n_colors=5, min=0, max=1):
fig, ax = plt.subplots()
# get discrete colormap
cmap = plt.get_cmap("gray", n_colors + 1)
ax.matshow(data, cmap=cmap, vmin=min, vmax=max)
ax.axis("off")
fig.tight_layout()
return fig
def continuous_matshow(data, min=0, max=1):
fig, ax = plt.subplots()
# get discrete colormap
cmap = plt.get_cmap("seismic")
ax.matshow(data, cmap=cmap, vmin=min, vmax=max)
ax.axis("off")
# optionally: provide a colorbar and tick at integers
# cax = plt.colorbar(mat, ticks=np.arange(min, max + 1))
return fig
def iterate(model, data_loader, config, writer, mode="train", epoch=None, device=None):
if len(data_loader) == 0:
raise ValueError("Received data loader with zero samples!")
# loss meter, needs 1 meter per scalar (see https://tnt.readthedocs.io/en/latest/_modules/torchnet/meter/averagevaluemeter.html);
loss_meter = tnt.meter.AverageValueMeter()
img_meter = avg_img_metrics()
# collect sample-averaged uncertainties and errors
errs, errs_se, errs_ae, vars_aleatoric = [], [], [], []
t_start = time.time()
for i, batch in enumerate(tqdm(data_loader)):
step = (epoch - 1) * len(data_loader) + i
if config.sample_type == "cloudy_cloudfree":
x, y, in_m, dates = prepare_data(batch, device, config)
elif config.sample_type == "pretrain":
x, y, in_m = prepare_data(batch, device, config)
dates = None
else:
raise NotImplementedError
inputs = {"A": x, "B": y, "dates": dates, "masks": in_m}
if mode != "train": # val or test
with torch.no_grad():
# compute single-model mean and variance predictions
model.set_input(inputs)
model.forward()
model.get_loss_G()
model.rescale()
out = model.fake_B
if hasattr(model.netG, "variance") and model.netG.variance is not None:
var = model.netG.variance
model.netG.variance = None
else:
var = out[:, :, S2_BANDS:, ...]
out = out[:, :, :S2_BANDS, ...]
batch_size = y.size()[0]
for bdx in range(batch_size):
# only compute statistics on variance estimates if using e.g. NLL loss or combinations thereof
if config.loss in ["GNLL", "MGNLL"]:
# if the variance variable is of shape [B x 1 x C x C x H x W] then it's a covariance tensor
if len(var.shape) > 5:
covar = var
# get [B x 1 x C x H x W] variance tensor
var = var.diagonal(dim1=2, dim2=3).moveaxis(-1, 2)
extended_metrics = img_metrics(y[bdx], out[bdx], var=var[bdx])
vars_aleatoric.append(extended_metrics["mean var"])
errs.append(extended_metrics["error"])
errs_se.append(extended_metrics["mean se"])
errs_ae.append(extended_metrics["mean ae"])
else:
extended_metrics = img_metrics(y[bdx], out[bdx])
img_meter.add(extended_metrics)
idx = i * batch_size + bdx # plot and export every k-th item
if config.plot_every > 0 and idx % config.plot_every == 0:
plot_dir = os.path.join(
config.res_dir,
config.experiment_name,
"plots",
f"epoch_{epoch}",
f"{mode}",
)
plot_img(x[bdx], "in", plot_dir, file_id=idx)
plot_img(out[bdx], "pred", plot_dir, file_id=idx)
plot_img(y[bdx], "target", plot_dir, file_id=idx)
plot_img(
((out[bdx] - y[bdx]) ** 2).mean(1, keepdims=True),
"err",
plot_dir,
file_id=idx,
)
plot_img(
discrete_matshow(
in_m.float().mean(axis=1).cpu()[bdx],
n_colors=config.input_t,
),
"mask",
plot_dir,
file_id=idx,
)
if var is not None:
plot_img(
var.mean(2, keepdims=True)[bdx],
"var",
plot_dir,
file_id=idx,
)
if config.export_every > 0 and idx % config.export_every == 0:
export_dir = os.path.join(
config.res_dir,
config.experiment_name,
"export",
f"epoch_{epoch}",
f"{mode}",
)
export(out[bdx], "pred", export_dir, file_id=idx)
export(y[bdx], "target", export_dir, file_id=idx)
if var is not None:
try:
export(covar[bdx], "covar", export_dir, file_id=idx)
except:
export(var[bdx], "var", export_dir, file_id=idx)
else: # training
# compute single-model mean and variance predictions
model.set_input(inputs)
model.optimize_parameters() # not using model.forward() directly
out = model.fake_B.detach().cpu()
# read variance predictions stored on generator
if hasattr(model.netG, "variance") and model.netG.variance is not None:
var = model.netG.variance.cpu()
else:
var = out[:, :, S2_BANDS:, ...]
out = out[:, :, :S2_BANDS, ...]
if config.plot_every > 0:
plot_out = out.detach().clone()
batch_size = y.size()[0]
for bdx in range(batch_size):
idx = i * batch_size + bdx # plot and export every k-th item
if idx % config.plot_every == 0:
plot_dir = os.path.join(
config.res_dir,
config.experiment_name,
"plots",
f"epoch_{epoch}",
f"{mode}",
)
plot_img(x[bdx], "in", plot_dir, file_id=i)
plot_img(plot_out[bdx], "pred", plot_dir, file_id=i)
plot_img(y[bdx], "target", plot_dir, file_id=i)
if mode == "train":
# periodically log stats
if step % config.display_step == 0:
out, x, y, in_m = out.cpu(), x.cpu(), y.cpu(), in_m.cpu()
if config.loss in ["GNLL", "MGNLL"]:
var = var.cpu()
log_train(writer, config, model, step, x, out, y, in_m, var=var)
else:
log_train(writer, config, model, step, x, out, y, in_m)
# log the loss, computed via model.backward_G() at train time & via model.get_loss_G() at val/test time
loss_meter.add(model.loss_G.item())
# after each batch, close any leftover figures
plt.close("all")
# --- end of epoch ---
# after each epoch, log the loss metrics
t_end = time.time()
total_time = t_end - t_start
print("Epoch time : {:.1f}s".format(total_time))
metrics = {f"{mode}_epoch_time": total_time}
# log the loss, only computed within model.backward_G() at train time
metrics[f"{mode}_loss"] = loss_meter.value()[0]
if mode == "train": # after each epoch, update lr acc. to scheduler
current_lr = model.optimizer_G.state_dict()["param_groups"][0]["lr"]
writer.add_scalar("Etc/train/lr", current_lr, step)
model.scheduler_G.step()
if mode == "test" or mode == "val":
# log the metrics
# log image metrics
for key, val in img_meter.value().items():
writer.add_scalar(f"{mode}/{key}", val, step)
# any loss is currently only computed within model.backward_G() at train time
writer.add_scalar(f"{mode}/loss", metrics[f"{mode}_loss"], step)
# use add_images for batch-wise adding across temporal dimension
if config.use_sar:
writer.add_image(
f"Img/{mode}/in_s1", x[0, :, [0], ...], step, dataformats="NCHW"
)
writer.add_image(
f"Img/{mode}/in_s2", x[0, :, [5, 4, 3], ...], step, dataformats="NCHW"
)
else:
writer.add_image(
f"Img/{mode}/in_s2", x[0, :, [3, 2, 1], ...], step, dataformats="NCHW"
)
writer.add_image(
f"Img/{mode}/out", out[0, 0, [3, 2, 1], ...], step, dataformats="CHW"
)
writer.add_image(
f"Img/{mode}/y", y[0, 0, [3, 2, 1], ...], step, dataformats="CHW"
)
writer.add_image(
f"Img/{mode}/m", in_m[0, :, None, ...], step, dataformats="NCHW"
)
# compute Expected Calibration Error (ECE)
if config.loss in ["GNLL", "MGNLL"]:
sorted_errors_se = compute_ece(
vars_aleatoric, errs_se, len(data_loader.dataset), percent=5
)
sorted_errors = {"se_sortAleatoric": sorted_errors_se}
plot_discard(
sorted_errors["se_sortAleatoric"], config, mode, step, is_se=True
)
# compute ECE
uce_l2, auce_l2 = compute_uce_auce(
vars_aleatoric,
errs,
len(data_loader.dataset),
percent=5,
l2=True,
mode=mode,
step=step,
)
# no need for a running mean here
img_meter.value()["UCE SE"] = uce_l2.cpu().numpy().item()
img_meter.value()["AUCE SE"] = auce_l2.cpu().numpy().item()
if config.loss in ["GNLL", "MGNLL"]:
log_aleatoric(writer, config, mode, step, var, f"model/", img_meter)
return metrics, img_meter.value()
else:
return metrics
def plot_discard(sorted_errors, config, mode, step, is_se=True):
metric = "SE" if is_se else "AE"
fig, ax = plt.subplots()
x_axis = np.arange(0.0, 1.0, 0.05)
ax.scatter(
x_axis,
sorted_errors,
c="b",
alpha=1.0,
marker=r".",
label=f"{metric}, sorted by uncertainty",
)
# fit a linear regressor with slope b and intercept a
sorted_errors[np.isnan(sorted_errors)] = np.nanmean(sorted_errors)
b, a = np.polyfit(x_axis, sorted_errors, deg=1)
x_seq = np.linspace(0, 1.0, num=1000)
ax.plot(
x_seq,
a + b * x_seq,
c="k",
lw=1.5,
alpha=0.75,
label=f"linear fit, {round(a, 3)} + {round(b, 3)} * x",
)
plt.xlabel("Fraction of samples, sorted ascendingly by uncertainty")
plt.ylabel("Error")
plt.legend(loc="upper left")
plt.grid()
fig.tight_layout()
writer.add_figure(f"Img/{mode}/discard_uncertain", fig, step)
if mode == "test": # export the final test split plots for print
path_to = os.path.join(config.res_dir, config.experiment_name)
print(f"Logging discard plots to path {path_to}")
fig.savefig(
os.path.join(path_to, f"plot_{mode}_{metric}_discard.png"),
bbox_inches="tight",
dpi=int(1e3),
)
fig.savefig(
os.path.join(path_to, f"plot_{mode}_{metric}_discard.pdf"),
bbox_inches="tight",
dpi=int(1e3),
)
def compute_ece(vars, errors, n_samples, percent=5):
# rank sample-averaged uncertainties ascendingly, and errors accordingly
_, vars_indices = torch.sort(torch.Tensor(vars))
errors = torch.Tensor(errors)
errs_sort = errors[vars_indices]
# incrementally remove 5% of errors, ranked by highest uncertainty
bins = torch.linspace(0, n_samples, 100 // percent + 1, dtype=int)[1:]
# get uncertainty-sorted cumulative errors, i.e. at x-tick 65% we report the average error for the 65% most certain predictions
sorted_errors = np.array(
[torch.nanmean(errs_sort[:rdx]).cpu().numpy() for rdx in bins]
)
return sorted_errors
binarize = lambda arg, n_bins, floor=0, ceil=1: np.digitize(
arg, bins=np.linspace(floor, ceil, num=n_bins)[1:]
)
def compute_uce_auce(var, errors, n_samples, percent=5, l2=True, mode="val", step=0):
n_bins = 100 // percent
var, errors = torch.Tensor(var), torch.Tensor(errors)
# metric: IN: standard deviation & error
# OUT: either root mean variance & root mean squared error or mean standard deviation & mean absolute error
metric = (
lambda arg: torch.sqrt(torch.mean(arg**2))
if l2
else torch.mean(torch.abs(arg))
)
m_str = "L2" if l2 else "L1"
# group uncertainty values into n_bins
var_idx = torch.Tensor(binarize(var, n_bins, floor=var.min(), ceil=var.max()))
# compute bin-wise statistics, defaults to nan if no data contained in bin
bk_var, bk_err = torch.empty(n_bins), torch.empty(n_bins)
for bin_idx in range(n_bins): # for each of the n_bins ...
bk_var[bin_idx] = metric(
var[var_idx == bin_idx].sqrt()
) # note: taking the sqrt to wrap into metric function,
bk_err[bin_idx] = metric(
errors[var_idx == bin_idx]
) # apply same metric function on error
calib_err = torch.abs(
bk_err - bk_var
) # calibration error: discrepancy of error vs uncertainty
bk_weight = (
torch.histogram(var_idx, n_bins)[0] / n_samples
) # fraction of total data per bin, for bin-weighting
uce = torch.nansum(bk_weight * calib_err) # calc. weighted UCE,
auce = torch.nanmean(calib_err) # calc. unweighted AUCE
# plot bin-wise error versus bin-wise uncertainty
fig, ax = plt.subplots()
x_min, x_max = bk_var[~bk_var.isnan()].min(), bk_var[~bk_var.isnan()].max()
y_min, y_max = 0, bk_err[~bk_err.isnan()].max()
x_axis = np.linspace(x_min, x_max, num=n_bins)
ax.plot(x_axis, x_axis) # diagonal reference line
ax.bar(
x_axis,
bk_err,
width=x_axis[1] - x_axis[0],
alpha=0.75,
edgecolor="k",
color="gray",
)
plt.xlim(x_min, x_max)
plt.ylim(y_min, y_max)
plt.xlabel("Uncertainty")
plt.ylabel(f"{m_str} Error")
plt.legend(loc="upper left")
plt.grid()
fig.tight_layout()
writer.add_figure(f"Img/{mode}/err_vs_var_{m_str}", fig, step)
return uce, auce
def recursive_todevice(x, device):
if isinstance(x, torch.Tensor):
return x.to(device)
elif isinstance(x, dict):
return {k: recursive_todevice(v, device) for k, v in x.items()}
else:
return [recursive_todevice(c, device) for c in x]
def prepare_output(config):
os.makedirs(os.path.join(config.res_dir, config.experiment_name), exist_ok=True)
def checkpoint(log, config):
with open(
os.path.join(config.res_dir, config.experiment_name, "trainlog.json"), "w"
) as outfile:
json.dump(log, outfile, indent=4)
def save_results(metrics, path, split="test"):
with open(os.path.join(path, f"{split}_metrics.json"), "w") as outfile:
json.dump(metrics, outfile, indent=4)
# check for file of pre-computed statistics, e.g. indices or cloud coverage
def import_from_path(split, config):
if os.path.exists(
os.path.join(os.path.dirname(os.getcwd()), "util", "precomputed")
):
import_path = os.path.join(
os.path.dirname(os.getcwd()),
"util",
"precomputed",
f"generic_{config.input_t}_{split}_{config.region}_s2cloudless_mask.npy",
)
else:
import_path = os.path.join(
config.precomputed,
f"generic_{config.input_t}_{split}_{config.region}_s2cloudless_mask.npy",
)
import_data_path = import_path if os.path.isfile(import_path) else None
return import_data_path
def main(config):
prepare_output(config)
device = torch.device(config.device)
# define data sets
if config.pretrain: # pretrain / training on mono-temporal data
dt_train = SEN12MSCR(
os.path.expanduser(config.root3),
split="train",
region=config.region,
sample_type=config.sample_type,
)
dt_val = SEN12MSCR(
os.path.expanduser(config.root3),
split="test",
region=config.region,
sample_type=config.sample_type,
)
dt_test = SEN12MSCR(
os.path.expanduser(config.root3),
split="val",
region=config.region,
sample_type=config.sample_type,
)
else:
pass
# wrap to allow for subsampling, e.g. for test runs etc
dt_train = torch.utils.data.Subset(
dt_train,
range(
0,
min(
config.max_samples_count,
len(dt_train),
int(len(dt_train) * config.max_samples_frac),
),
),
)
dt_val = torch.utils.data.Subset(
dt_val,
range(
0,
min(
config.max_samples_count,
len(dt_val),
int(len(dt_train) * config.max_samples_frac),
),
),
)
dt_test = torch.utils.data.Subset(
dt_test,
range(
0,
min(
config.max_samples_count,
len(dt_test),
int(len(dt_train) * config.max_samples_frac),
),
),
)
# instantiate dataloaders, note: worker_init_fn is needed to get reproducible random samples across runs if vary_samples=True
train_loader = torch.utils.data.DataLoader(
dt_train,
batch_size=config.batch_size,
shuffle=True,
worker_init_fn=seed_worker,
generator=f,
num_workers=config.num_workers,
pin_memory=True,
persistent_workers=True,
)
val_loader = torch.utils.data.DataLoader(
dt_val,
batch_size=config.batch_size,
shuffle=False,
worker_init_fn=seed_worker,
generator=g,
num_workers=config.num_workers,
pin_memory=True,
persistent_workers=True,
)
test_loader = torch.utils.data.DataLoader(
dt_test,
batch_size=config.batch_size,
shuffle=False,
worker_init_fn=seed_worker,
generator=g,
num_workers=config.num_workers,
pin_memory=True,
persistent_workers=True,
)
print("Train {}, Val {}, Test {}".format(len(dt_train), len(dt_val), len(dt_test)))
# model definition
# (compiled model hangs up in validation step on some systems, retry in the future for pytorch > 2.0)
model = get_model(config) # torch.compile(get_model(config))
# set model properties
model.len_epoch = len(train_loader)
config.N_params = utils.get_ntrainparams(model)
print("\n\nTrainable layers:")
for name, p in model.named_parameters():
if p.requires_grad:
print(f"\t{name}")
model = model.to(device)
# do random weight initialization
print("\nInitializing weights randomly.")
model.netG.apply(weight_init)
if config.trained_checkp and len(config.trained_checkp) > 0:
# load weights from the indicated checkpoint
print(f"Loading weights from (pre-)trained checkpoint {config.trained_checkp}")
load_model(
config,
model,
train_out_layer=False,
load_out_partly=config.model in ["uncrtaints"],
)
with open(
os.path.join(config.res_dir, config.experiment_name, "conf.json"), "w"
) as file:
file.write(json.dumps(vars(config), indent=4))
print(f"TOTAL TRAINABLE PARAMETERS: {config.N_params}\n")
print(model)
# Optimizer and Loss
model.criterion = losses.get_loss(config)
# track best loss, checkpoint at best validation performance
is_better, best_loss = lambda new, prev: new <= prev, float("inf")
# Training loop
trainlog = {}
# resume training at scheduler's latest epoch, != 0 if --resume_from
begin_at = (
config.resume_at
if config.resume_at >= 0
else model.scheduler_G.state_dict()["last_epoch"]
)
for epoch in range(begin_at + 1, config.epochs + 1):
print("\nEPOCH {}/{}".format(epoch, config.epochs))
# put all networks in training mode again
model.train()
model.netG.train()
# unfreeze all layers after specified epoch
if epoch > config.unfreeze_after and hasattr(model, "frozen") and model.frozen:
print("Unfreezing all network layers")
model.frozen = False
freeze_layers(model.netG, grad=True)
# re-seed train generator for each epoch anew, depending on seed choice plus current epoch number
# ~ else, dataloader provides same samples no matter what epoch training starts/resumes from
# ~ note: only re-seed train split dataloader (if config.vary_samples), but keep all others consistent
# ~ if desiring different runs, then the seeds must at least be config.epochs numbers apart
if config.vary_samples:
# condition dataloader samples on current epoch count
f.manual_seed(config.rdm_seed + epoch)
train_loader = torch.utils.data.DataLoader(
dt_train,
batch_size=config.batch_size,
shuffle=True,
worker_init_fn=seed_worker,
generator=f,
num_workers=config.num_workers,
)
train_metrics = iterate(
model,
data_loader=train_loader,
config=config,
writer=writer,
mode="train",
epoch=epoch,
device=device,
)
# do regular validation steps at the end of each training epoch
if epoch % config.val_every == 0 and epoch > config.val_after:
print("Validation . . . ")
model.eval()
model.netG.eval()
val_metrics, val_img_metrics = iterate(
model,
data_loader=val_loader,
config=config,
writer=writer,
mode="val",
epoch=epoch,
device=device,
)
# use the training loss for validation
print("Using training loss as validation loss")
if "val_loss" in val_metrics:
val_loss = val_metrics["val_loss"]
else:
val_loss = val_metrics["val_loss_ensembleAverage"]
print(f"Validation Loss {val_loss}")
print(f"validation image metrics: {val_img_metrics}")
save_results(
val_img_metrics,
os.path.join(config.res_dir, config.experiment_name),
split=f"test_epoch_{epoch}",
)
print(
f"\nLogged validation epoch {epoch} metrics to path {os.path.join(config.res_dir, config.experiment_name)}"
)
# checkpoint best model
trainlog[epoch] = {**train_metrics, **val_metrics}
checkpoint(trainlog, config)
if is_better(val_loss, best_loss):
best_loss = val_loss
save_model(config, epoch, model, "model")
else:
trainlog[epoch] = {**train_metrics}
checkpoint(trainlog, config)
# always checkpoint the current epoch's model
save_model(config, epoch, model, f"model_epoch_{epoch}")
print(f"Completed current epoch of experiment {config.experiment_name}.")
# following training, test on hold-out data
print("Testing best epoch . . .")
load_checkpoint(config, config.res_dir, model, "model")
model.eval()
model.netG.eval()
test_metrics, test_img_metrics = iterate(
model,
data_loader=test_loader,
config=config,
writer=writer,
mode="test",
epoch=epoch,
device=device,
)
if "test_loss" in test_metrics:
test_loss = test_metrics["test_loss"]
else:
test_loss = test_metrics["test_loss_ensembleAverage"]
print(f"Test Loss {test_loss}")
print(f"\nTest image metrics: {test_img_metrics}")
save_results(
test_img_metrics,
os.path.join(config.res_dir, config.experiment_name),
split="test",
)
print(
f"\nLogged test metrics to path {os.path.join(config.res_dir, config.experiment_name)}"
)
# close tensorboard logging
writer.close()
print(f"Finished training experiment {config.experiment_name}.")
if __name__ == "__main__":
main(config)
exit()
|