File size: 43,188 Bytes
b55bace | 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 | import os
import sys
import torch
import wandb
import matplotlib.pyplot as plt
import pytorch_lightning as pl
from torch.optim import AdamW
from torchmetrics.functional import mean_squared_error
from torchdyn.core import NeuralODE
import numpy as np
import lpips
from .networks.utils import flow_model_torch_wrapper
from .utils import plot_lidar
from .ema import EMA
from torchdiffeq import odeint as odeint2
from .losses.energy_loss import EnergySolver, ReconsLoss
class GrowthNetTrain(pl.LightningModule):
def __init__(
self,
flow_nets,
growth_nets,
skipped_time_points=None,
ot_sampler=None,
args=None,
state_cost=None,
data_manifold_metric=None,
joint = False
):
super().__init__()
#self.save_hyperparameters()
self.flow_nets = flow_nets
if not joint:
for param in self.flow_nets.parameters():
param.requires_grad = False
self.growth_nets = growth_nets # list of growth networks for each branch
self.ot_sampler = ot_sampler
self.skipped_time_points = skipped_time_points
self.optimizer_name = args.growth_optimizer
self.lr = args.growth_lr
self.weight_decay = args.growth_weight_decay
self.whiten = args.whiten
self.working_dir = args.working_dir
self.args = args
#branching
self.state_cost = state_cost
self.data_manifold_metric = data_manifold_metric
self.branches = len(growth_nets)
self.metric_clusters = args.metric_clusters
self.recons_loss = ReconsLoss()
# loss weights
self.lambda_energy = args.lambda_energy
self.lambda_mass = args.lambda_mass
self.lambda_match = args.lambda_match
self.lambda_recons = args.lambda_recons
self.joint = joint
def forward(self, t, xt, branch_idx):
# output growth rate given branch_idx
return self.growth_nets[branch_idx](t, xt)
def _compute_loss(self, main_batch, metric_samples_batch=None, validation=False):
x0s = main_batch["x0"][0]
w0s = main_batch["x0"][1]
x1s_list = []
w1s_list = []
if self.branches > 1:
for i in range(self.branches):
x1s_list.append([main_batch[f"x1_{i+1}"][0]])
w1s_list.append([main_batch[f"x1_{i+1}"][1]])
else:
x1s_list.append([main_batch["x1"][0]])
w1s_list.append([main_batch["x1"][1]])
if self.args.manifold:
#changed
if self.metric_clusters == 7 and self.branches == 6:
# Weinreb 6-branch scenario: cluster 0 (root) → clusters 1-6 (6 branches)
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[2]), # x0 → x1_2 (branch 2)
(metric_samples_batch[0], metric_samples_batch[3]), # x0 → x1_3 (branch 3)
(metric_samples_batch[0], metric_samples_batch[4]), # x0 → x1_4 (branch 4)
(metric_samples_batch[0], metric_samples_batch[5]), # x0 → x1_5 (branch 5)
(metric_samples_batch[0], metric_samples_batch[6]), # x0 → x1_6 (branch 6)
]
elif self.metric_clusters == 4:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[2]), # x0 → x1_2 (branch 2)
(metric_samples_batch[0], metric_samples_batch[3]),
]
elif self.metric_clusters == 3:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[2]), # x0 → x1_2 (branch 2)
]
elif self.metric_clusters == 2 and self.branches == 2:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_2 (branch 2)
]
elif self.metric_clusters == 2:
# For any number of branches with 2 metric clusters (initial vs remaining)
# All branches use the same metric cluster pair
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]) # x0 → all branches
] * self.branches
else:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
]
batch_size = x0s.shape[0]
assert len(x1s_list) == self.branches, "Mismatch between x1s_list and expected branches"
energy_loss = [0.] * self.branches
mass_loss = 0.
neg_weight_penalty = 0.
match_loss = [0.] * self.branches
recons_loss = [0.] * self.branches
dtype = x0s[0].dtype
#w0s = torch.zeros((batch_size, 1), dtype=dtype)
m0s = torch.zeros_like(w0s, dtype=dtype)
start_state = (x0s, w0s, m0s)
xt = [x0s.clone() for _ in range(self.branches)]
w0_branch = torch.zeros_like(w0s, dtype=dtype)
w0_branches = []
w0_branches.append(w0s)
for _ in range(self.branches - 1):
w0_branches.append(w0_branch)
#w0_branches = [w0_branch.clone() for _ in range(self.branches - 1)]
wt = w0_branches
mt = [m0s.clone() for _ in range(self.branches)]
# loop through timesteps
for step_idx, (s, t) in enumerate(zip(self.timesteps[:-1], self.timesteps[1:])):
time = torch.Tensor([s, t])
total_w_t = 0
# loop through branches
for i in range(self.branches):
if self.args.manifold:
start_samples, end_samples = branch_sample_pairs[i]
samples = torch.cat([start_samples, end_samples], dim=0)
else:
samples = None
# initialize weight and energy
start_state = (xt[i], wt[i], mt[i])
# loop over timesteps
xt_next, wt_next, mt_next = self.take_step(time, start_state, i, samples, timestep_idx=step_idx)
# placeholders for next state
xt_last = xt_next[-1]
wt_last = wt_next[-1]
mt_last = mt_next[-1]
total_w_t += wt_last
energy_loss[i] += (mt_last - mt[i])
neg_weight_penalty += torch.relu(-wt_last).sum()
# update branch state
xt[i] = xt_last.clone().detach()
wt[i] = wt_last.clone().detach()
mt[i] = mt_last.clone().detach()
# calculate mass loss from all branches
target = torch.ones_like(total_w_t)
mass_loss += mean_squared_error(total_w_t, target)
# calculate loss that matches final weights
for i in range(self.branches):
match_loss[i] = mean_squared_error(wt[i], w1s_list[i][0])
# compute reconstruction loss
recons_loss[i] = self.recons_loss(xt[i], x1s_list[i][0])
# average across time steps (loop runs len(timesteps)-1 times)
mass_loss = mass_loss / max(len(self.timesteps) - 1, 1)
# Weighted mean across branches (inversely weighted by cluster size)
# Get cluster sizes from datamodule if available
if hasattr(self.trainer, 'datamodule') and hasattr(self.trainer.datamodule, 'cluster_sizes'):
cluster_sizes = self.trainer.datamodule.cluster_sizes
max_size = max(cluster_sizes)
# Inverse weighting: smaller clusters get higher weight
branch_weights = torch.tensor([max_size / size for size in cluster_sizes],
dtype=energy_loss[0].dtype, device=energy_loss[0].device)
# Normalize weights to sum to num_branches for fair comparison
branch_weights = branch_weights * self.branches / branch_weights.sum()
energy_loss = torch.mean(torch.stack([e.mean() for e in energy_loss]) * branch_weights)
match_loss = torch.mean(torch.stack(match_loss) * branch_weights)
recons_loss = torch.mean(torch.stack(recons_loss) * branch_weights)
else:
# Fallback to uniform weighting
energy_loss = torch.mean(torch.stack([e.mean() for e in energy_loss]))
match_loss = torch.mean(torch.stack(match_loss))
recons_loss = torch.mean(torch.stack(recons_loss))
loss = (self.lambda_energy * energy_loss) + (self.lambda_mass * (mass_loss + neg_weight_penalty)) + (self.lambda_match * match_loss) \
+ (self.lambda_recons * recons_loss)
if self.joint:
if validation:
self.log("JointTrain/val_mass_loss", mass_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/val_neg_penalty_loss", neg_weight_penalty, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/val_match_loss", match_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/val_energy_loss", energy_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/val_recons_loss", recons_loss, on_step=False, on_epoch=True, prog_bar=True)
else:
self.log("JointTrain/train_mass_loss", mass_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/train_neg_penalty_loss", neg_weight_penalty, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/train_match_loss", match_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/train_energy_loss", energy_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/train_recons_loss", recons_loss, on_step=False, on_epoch=True, prog_bar=True)
else:
if validation:
self.log("GrowthNet/val_mass_loss", mass_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/val_neg_penalty_loss", neg_weight_penalty, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/val_match_loss", match_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/val_energy_loss", energy_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/val_recons_loss", recons_loss, on_step=False, on_epoch=True, prog_bar=True)
else:
self.log("GrowthNet/train_mass_loss", mass_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/train_neg_penalty_loss", neg_weight_penalty, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/train_match_loss", match_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/train_energy_loss", energy_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/train_recons_loss", recons_loss, on_step=False, on_epoch=True, prog_bar=True)
return loss
def take_step(self, t, start_state, branch_idx, samples=None, timestep_idx=0):
flow_net = self.flow_nets[branch_idx]
growth_net = self.growth_nets[branch_idx]
x_t, w_t, m_t = odeint2(EnergySolver(flow_net, growth_net, self.state_cost, self.data_manifold_metric, samples, timestep_idx), start_state, t, options=dict(step_size=0.1),method='euler')
return x_t, w_t, m_t
def training_step(self, batch, batch_idx):
if isinstance(batch, (list, tuple)):
batch = batch[0]
if isinstance(batch, dict) and "train_samples" in batch:
main_batch = batch["train_samples"]
metric_batch = batch["metric_samples"]
if isinstance(main_batch, tuple):
main_batch = main_batch[0]
if isinstance(metric_batch, tuple):
metric_batch = metric_batch[0]
else:
# Fallback
main_batch = batch.get("train_samples", batch)
metric_batch = batch.get("metric_samples", [])
self.timesteps = torch.linspace(0.0, 1.0, len(main_batch["x0"])).tolist()
loss = self._compute_loss(main_batch, metric_batch, validation=False)
if self.joint:
self.log(
"JointTrain/train_loss",
loss,
on_step=False,
on_epoch=True,
prog_bar=True,
logger=True,
)
else:
self.log(
"GrowthNet/train_loss",
loss,
on_step=False,
on_epoch=True,
prog_bar=True,
logger=True,
)
return loss
def validation_step(self, batch, batch_idx):
if isinstance(batch, (list, tuple)):
batch = batch[0]
if isinstance(batch, dict) and "val_samples" in batch:
main_batch = batch["val_samples"]
metric_batch = batch["metric_samples"]
if isinstance(main_batch, tuple):
main_batch = main_batch[0]
if isinstance(metric_batch, tuple):
metric_batch = metric_batch[0]
else:
# Fallback
main_batch = batch.get("val_samples", batch)
metric_batch = batch.get("metric_samples", [])
self.timesteps = torch.linspace(0.0, 1.0, len(main_batch["x0"])).tolist()
val_loss = self._compute_loss(main_batch, metric_batch, validation=True)
if self.joint:
self.log(
"JointTrain/val_loss",
val_loss,
on_step=False,
on_epoch=True,
prog_bar=True,
logger=True,
)
else:
self.log(
"GrowthNet/val_loss",
val_loss,
on_step=False,
on_epoch=True,
prog_bar=True,
logger=True,
)
return val_loss
def optimizer_step(self, *args, **kwargs):
super().optimizer_step(*args, **kwargs)
for net in self.growth_nets:
if isinstance(net, EMA):
net.update_ema()
if self.joint:
for net in self.flow_nets:
if isinstance(net, EMA):
net.update_ema()
def configure_optimizers(self):
params = []
for net in self.growth_nets:
params += list(net.parameters())
if self.joint:
for net in self.flow_nets:
params += list(net.parameters())
if self.optimizer_name == "adamw":
optimizer = AdamW(
params,
lr=self.lr,
weight_decay=self.weight_decay,
)
elif self.optimizer_name == "adam":
optimizer = torch.optim.Adam(
params,
lr=self.lr,
)
return optimizer
@torch.no_grad()
def get_mass_and_position(self, main_batch, metric_samples_batch=None):
if isinstance(main_batch, dict):
main_batch = main_batch
else:
main_batch = main_batch[0]
x0s = main_batch["x0"][0]
w0s = main_batch["x0"][1]
if self.args.manifold:
if self.metric_clusters == 4:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[2]), # x0 → x1_2 (branch 2)
(metric_samples_batch[0], metric_samples_batch[3]),
]
elif self.metric_clusters == 3:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[2]), # x0 → x1_2 (branch 2)
]
elif self.metric_clusters == 2 and self.branches == 2:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_2 (branch 2)
]
elif self.metric_clusters == 2:
# For any number of branches with 2 metric clusters (initial vs remaining)
# All branches use the same metric cluster pair
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]) # x0 → all branches
] * self.branches
else:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
]
batch_size = x0s.shape[0]
dtype = x0s[0].dtype
m0s = torch.zeros_like(w0s, dtype=dtype)
xt = [x0s.clone() for _ in range(self.branches)]
w0_branch = torch.zeros_like(w0s, dtype=dtype)
w0_branches = []
w0_branches.append(w0s)
for _ in range(self.branches - 1):
w0_branches.append(w0_branch)
wt = w0_branches
mt = [m0s.clone() for _ in range(self.branches)]
time_points = []
mass_over_time = [[] for _ in range(self.branches)]
energy_over_time = [[] for _ in range(self.branches)]
# record per-sample weights at each time for each branch (to allow OT with per-sample masses)
weights_over_time = [[] for _ in range(self.branches)]
all_trajs = [[] for _ in range(self.branches)]
t_span = torch.linspace(0, 1, 101)
for step_idx, (s, t) in enumerate(zip(t_span[:-1], t_span[1:])):
time_points.append(t.item())
time = torch.Tensor([s, t])
for i in range(self.branches):
if self.args.manifold:
start_samples, end_samples = branch_sample_pairs[i]
samples = torch.cat([start_samples, end_samples], dim=0)
else:
samples = None
start_state = (xt[i], wt[i], mt[i])
xt_next, wt_next, mt_next = self.take_step(time, start_state, i, samples, timestep_idx=step_idx)
xt[i] = xt_next[-1].clone().detach()
wt[i] = wt_next[-1].clone().detach()
mt[i] = mt_next[-1].clone().detach()
all_trajs[i].append(xt[i].clone().detach())
mass_over_time[i].append(wt[i].mean().item())
energy_over_time[i].append(mt[i].mean().item())
# store per-sample weights (clone to detach from graph)
try:
weights_over_time[i].append(wt[i].clone().detach())
except Exception:
# fallback: store mean as singleton tensor
weights_over_time[i].append(torch.tensor(wt[i].mean().item()).unsqueeze(0))
return time_points, xt, all_trajs, mass_over_time, energy_over_time, weights_over_time
@torch.no_grad()
def _plot_mass_and_energy(self, main_batch, metric_samples_batch=None, save_dir=None):
x0s = main_batch["x0"][0]
w0s = main_batch["x0"][1]
if self.args.manifold:
if self.metric_clusters == 7 and self.branches == 6:
# Weinreb 6-branch scenario: cluster 0 (root) → clusters 1-6 (6 branches)
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[2]), # x0 → x1_2 (branch 2)
(metric_samples_batch[0], metric_samples_batch[3]), # x0 → x1_3 (branch 3)
(metric_samples_batch[0], metric_samples_batch[4]), # x0 → x1_4 (branch 4)
(metric_samples_batch[0], metric_samples_batch[5]), # x0 → x1_5 (branch 5)
(metric_samples_batch[0], metric_samples_batch[6]), # x0 → x1_6 (branch 6)
]
elif self.metric_clusters == 4:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[2]), # x0 → x1_2 (branch 2)
(metric_samples_batch[0], metric_samples_batch[3]),
]
elif self.metric_clusters == 3:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[2]), # x0 → x1_2 (branch 2)
]
elif self.metric_clusters == 2 and self.branches == 2:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_2 (branch 2)
]
else:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1]), # x0 → x1_1 (branch 1)
]
batch_size = x0s.shape[0]
dtype = x0s[0].dtype
m0s = torch.zeros_like(w0s, dtype=dtype)
xt = [x0s.clone() for _ in range(self.branches)]
w0_branch = torch.zeros_like(w0s, dtype=dtype)
w0_branches = []
w0_branches.append(w0s)
for _ in range(self.branches - 1):
w0_branches.append(w0_branch)
wt = w0_branches
mt = [m0s.clone() for _ in range(self.branches)]
time_points = []
mass_over_time = [[] for _ in range(self.branches)]
energy_over_time = [[] for _ in range(self.branches)]
t_span = torch.linspace(0, 1, 101)
for step_idx, (s, t) in enumerate(zip(t_span[:-1], t_span[1:])):
time_points.append(t.item())
time = torch.Tensor([s, t])
for i in range(self.branches):
if self.args.manifold:
start_samples, end_samples = branch_sample_pairs[i]
samples = torch.cat([start_samples, end_samples], dim=0)
else:
samples = None
start_state = (xt[i], wt[i], mt[i])
xt_next, wt_next, mt_next = self.take_step(time, start_state, i, samples, timestep_idx=step_idx)
xt[i] = xt_next[-1].clone().detach()
wt[i] = wt_next[-1].clone().detach()
mt[i] = mt_next[-1].clone().detach()
mass_over_time[i].append(wt[i].mean().item())
energy_over_time[i].append(mt[i].mean().item())
if save_dir is None:
run_name = self.args.run_name if hasattr(self.args, 'run_name') and self.args.run_name else self.args.data_name
save_dir = os.path.join(self.args.working_dir, 'results', run_name, 'figures')
os.makedirs(save_dir, exist_ok=True)
# Use tab10 colormap to get visually distinct colors
if self.args.branches == 3:
branch_colors = ['#9793F8', '#50B2D7', '#D577FF'] # tuple of RGBs
else:
branch_colors = ['#50B2D7', '#D577FF'] # tuple of RGBs
# --- Plot Mass ---
plt.figure(figsize=(8, 5))
for i in range(self.branches):
color = branch_colors[i]
plt.plot(time_points, mass_over_time[i], color=color, linewidth=2.5, label=f"Mass Branch {i}")
plt.xlabel("Time")
plt.ylabel("Mass")
plt.title("Mass Evolution per Branch")
plt.legend()
plt.grid(True)
if self.joint:
mass_path = os.path.join(save_dir, f"{self.args.data_name}_joint_mass.png")
else:
mass_path = os.path.join(save_dir, f"{self.args.data_name}_growth_mass.png")
plt.savefig(mass_path, dpi=300, bbox_inches="tight")
plt.close()
# --- Plot Energy ---
plt.figure(figsize=(8, 5))
for i in range(self.branches):
color = branch_colors[i]
plt.plot(time_points, energy_over_time[i], color=color, linewidth=2.5, label=f"Energy Branch {i}")
plt.xlabel("Time")
plt.ylabel("Energy")
plt.title("Energy Evolution per Branch")
plt.legend()
plt.grid(True)
if self.joint:
energy_path = os.path.join(save_dir, f"{self.args.data_name}_joint_energy.png")
else:
energy_path = os.path.join(save_dir, f"{self.args.data_name}_growth_energy.png")
plt.savefig(energy_path, dpi=300, bbox_inches="tight")
plt.close()
class GrowthNetTrainLidar(GrowthNetTrain):
def test_step(self, batch, batch_idx):
# Handle both tuple and dict batch formats from CombinedLoader
if isinstance(batch, dict):
main_batch = batch["test_samples"][0]
metric_batch = batch["metric_samples"][0]
else:
# batch is a tuple: (test_samples, metric_samples)
main_batch = batch[0][0]
metric_batch = batch[1][0]
self._plot_mass_and_energy(main_batch, metric_batch)
x0 = main_batch["x0"][0] # [B, D]
cloud_points = main_batch["dataset"][0] # full dataset, [N, D]
t_span = torch.linspace(0, 1, 101)
all_trajs = []
for i, flow_net in enumerate(self.flow_nets):
node = NeuralODE(
flow_model_torch_wrapper(flow_net),
solver="euler",
sensitivity="adjoint",
)
with torch.no_grad():
traj = node.trajectory(x0, t_span).cpu() # [T, B, D]
if self.whiten:
traj_shape = traj.shape
traj = traj.reshape(-1, 3)
traj = self.trainer.datamodule.scaler.inverse_transform(
traj.cpu().detach().numpy()
).reshape(traj_shape)
traj = torch.tensor(traj)
traj = torch.transpose(traj, 0, 1) # [B, T, D]
all_trajs.append(traj)
# Inverse-transform the point cloud once
if self.whiten:
cloud_points = torch.tensor(
self.trainer.datamodule.scaler.inverse_transform(
cloud_points.cpu().detach().numpy()
)
)
# ===== Plot all trajectories together =====
fig = plt.figure(figsize=(6, 5))
ax = fig.add_subplot(111, projection="3d", computed_zorder=False)
ax.view_init(elev=30, azim=-115, roll=0)
for i, traj in enumerate(all_trajs):
plot_lidar(ax, cloud_points, xs=traj, branch_idx=i)
run_name = self.args.run_name if hasattr(self.args, 'run_name') and self.args.run_name else self.args.data_name
results_dir = os.path.join(self.args.working_dir, 'results', run_name)
lidar_fig_dir = os.path.join(results_dir, 'figures')
os.makedirs(lidar_fig_dir, exist_ok=True)
if self.joint:
plt.savefig(os.path.join(lidar_fig_dir, 'joint_lidar_all_branches.png'), dpi=300)
else:
plt.savefig(os.path.join(lidar_fig_dir, 'growth_lidar_all_branches.png'), dpi=300)
plt.close()
# ===== Plot each trajectory separately =====
for i, traj in enumerate(all_trajs):
fig = plt.figure(figsize=(6, 5))
ax = fig.add_subplot(111, projection="3d", computed_zorder=False)
ax.view_init(elev=30, azim=-115, roll=0)
plot_lidar(ax, cloud_points, xs=traj, branch_idx=i)
if self.joint:
plt.savefig(os.path.join(lidar_fig_dir, f'joint_lidar_branch_{i + 1}.png'), dpi=300)
else:
plt.savefig(os.path.join(lidar_fig_dir, f'growth_lidar_branch_{i + 1}.png'), dpi=300)
plt.close()
class GrowthNetTrainCell(GrowthNetTrain):
def test_step(self, batch, batch_idx):
if self.args.data_type in ["scrna", "tahoe"]:
main_batch = batch[0]["test_samples"][0]
metric_batch = batch[0]["metric_samples"][0]
else:
main_batch = batch["test_samples"][0]
metric_batch = batch["metric_samples"][0]
self._plot_mass_and_energy(main_batch, metric_batch)
class SequentialGrowthNetTrain(pl.LightningModule):
"""
Sequential growth network training for multi-timepoint data.
Learns growth rates for transitions between consecutive timepoints.
"""
def __init__(
self,
flow_nets,
growth_nets,
skipped_time_points=None,
ot_sampler=None,
args=None,
data_manifold_metric=None,
joint=False
):
super().__init__()
self.flow_nets = flow_nets
if not joint:
for param in self.flow_nets.parameters():
param.requires_grad = False
self.growth_nets = growth_nets
self.ot_sampler = ot_sampler
self.skipped_time_points = skipped_time_points
self.optimizer_name = args.growth_optimizer
self.lr = args.growth_lr
self.weight_decay = args.growth_weight_decay
self.whiten = args.whiten
self.working_dir = args.working_dir
self.args = args
self.data_manifold_metric = data_manifold_metric
self.branches = len(growth_nets)
self.metric_clusters = args.metric_clusters
self.recons_loss = ReconsLoss()
# loss weights
self.lambda_energy = args.lambda_energy
self.lambda_mass = args.lambda_mass
self.lambda_match = args.lambda_match
self.lambda_recons = args.lambda_recons
self.joint = joint
self.num_timepoints = None
self.timepoint_keys = None
def forward(self, t, xt, branch_idx):
return self.growth_nets[branch_idx](t, xt)
def setup(self, stage=None):
"""Initialize timepoint keys before training/validation starts."""
if self.timepoint_keys is None:
timepoint_data = self.trainer.datamodule.get_timepoint_data()
self.timepoint_keys = [k for k in sorted(timepoint_data.keys())
if not any(x in k for x in ['_', 'time_labels'])]
self.num_timepoints = len(self.timepoint_keys)
print(f"Training sequential growth for {self.num_timepoints} timepoints: {self.timepoint_keys}")
def _compute_loss(self, main_batch, metric_samples_batch=None, validation=False):
"""Compute loss for sequential growth between timepoints."""
x0s = main_batch["x0"][0]
w0s = main_batch["x0"][1]
# Setup metric sample pairs
if self.args.manifold:
if self.metric_clusters == 2:
branch_sample_pairs = [
(metric_samples_batch[0], metric_samples_batch[1])
] * self.branches
else:
branch_sample_pairs = []
for b in range(self.branches):
if b + 1 < len(metric_samples_batch):
branch_sample_pairs.append(
(metric_samples_batch[0], metric_samples_batch[b + 1])
)
else:
branch_sample_pairs.append(
(metric_samples_batch[0], metric_samples_batch[1])
)
total_loss = 0
total_energy_loss = 0
total_mass_loss = 0
total_match_loss = 0
total_recons_loss = 0
num_transitions = 0
# Process each consecutive timepoint transition
for i in range(len(self.timepoint_keys) - 1):
t_curr_key = self.timepoint_keys[i]
t_next_key = self.timepoint_keys[i + 1]
batch_curr_key = f"x{t_curr_key.replace('t', '').replace('final', '1')}"
x_curr = main_batch[batch_curr_key][0]
w_curr = main_batch[batch_curr_key][1]
if i == len(self.timepoint_keys) - 2:
# Final transition to branches
# Get cluster size weights if available
if hasattr(self.trainer, 'datamodule') and hasattr(self.trainer.datamodule, 'cluster_sizes'):
cluster_sizes = self.trainer.datamodule.cluster_sizes
max_size = max(cluster_sizes)
# Inverse weighting: smaller clusters get higher weight
branch_weights = [max_size / size for size in cluster_sizes]
else:
branch_weights = [1.0] * self.branches
for b in range(self.branches):
x_next = main_batch[f"x1_{b+1}"][0]
w_next = main_batch[f"x1_{b+1}"][1]
# Compute growth-based loss for this transition
loss, energy_l, mass_l, match_l, recons_l = self._compute_transition_loss(
x_curr, w_curr, x_next, w_next, b, i,
branch_sample_pairs[b] if self.args.manifold else None
)
# Apply branch weight
total_loss += loss * branch_weights[b]
total_energy_loss += energy_l * branch_weights[b]
total_mass_loss += mass_l * branch_weights[b]
total_match_loss += match_l * branch_weights[b]
total_recons_loss += recons_l * branch_weights[b]
num_transitions += 1
else:
# Regular consecutive timepoints
batch_next_key = f"x{t_next_key.replace('t', '').replace('final', '1')}"
x_next = main_batch[batch_next_key][0]
w_next = main_batch[batch_next_key][1]
for b in range(self.branches):
loss, energy_l, mass_l, match_l, recons_l = self._compute_transition_loss(
x_curr, w_curr, x_next, w_next, b, i,
branch_sample_pairs[b] if self.args.manifold else None
)
total_loss += loss
total_energy_loss += energy_l
total_mass_loss += mass_l
total_match_loss += match_l
total_recons_loss += recons_l
num_transitions += 1
# Average losses
avg_energy_loss = total_energy_loss / num_transitions if num_transitions > 0 else total_energy_loss
avg_mass_loss = total_mass_loss / num_transitions if num_transitions > 0 else total_mass_loss
avg_match_loss = total_match_loss / num_transitions if num_transitions > 0 else total_match_loss
avg_recons_loss = total_recons_loss / num_transitions if num_transitions > 0 else total_recons_loss
# Log individual components
if self.joint:
if validation:
self.log("JointTrain/val_energy_loss", avg_energy_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/val_mass_loss", avg_mass_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/val_match_loss", avg_match_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/val_recons_loss", avg_recons_loss, on_step=False, on_epoch=True, prog_bar=True)
else:
self.log("JointTrain/train_energy_loss", avg_energy_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/train_mass_loss", avg_mass_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/train_match_loss", avg_match_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("JointTrain/train_recons_loss", avg_recons_loss, on_step=False, on_epoch=True, prog_bar=True)
else:
if validation:
self.log("GrowthNet/val_energy_loss", avg_energy_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/val_mass_loss", avg_mass_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/val_match_loss", avg_match_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/val_recons_loss", avg_recons_loss, on_step=False, on_epoch=True, prog_bar=True)
else:
self.log("GrowthNet/train_energy_loss", avg_energy_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/train_mass_loss", avg_mass_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/train_match_loss", avg_match_loss, on_step=False, on_epoch=True, prog_bar=True)
self.log("GrowthNet/train_recons_loss", avg_recons_loss, on_step=False, on_epoch=True, prog_bar=True)
return total_loss
def _compute_transition_loss(self, x0, w0, x1, w1, branch_idx, transition_idx, metric_pair):
"""Compute loss for a single timepoint transition."""
if self.ot_sampler is not None:
x0, x1 = self.ot_sampler.sample_plan(x0, x1, replace=True)
# Simulate trajectory using flow network
t_span = torch.linspace(0, 1, 10, device=x0.device)
flow_model = flow_model_torch_wrapper(self.flow_nets[branch_idx])
node = NeuralODE(flow_model, solver="euler", sensitivity="adjoint")
with torch.no_grad():
traj = node.trajectory(x0, t_span)
# Compute energy and mass losses
energy_loss = 0
mass_loss = 0
neg_weight_penalty = 0
for t_idx in range(len(t_span)):
t = t_span[t_idx]
xt = traj[t_idx]
# Growth rate
growth = self.growth_nets[branch_idx](t.unsqueeze(0).expand(xt.shape[0]), xt)
# Energy loss
if self.args.manifold and metric_pair is not None:
start_samples, end_samples = metric_pair
samples = torch.cat([start_samples, end_samples], dim=0)
_, kinetic, potential = self.data_manifold_metric.calculate_velocity(
xt, torch.zeros_like(xt), samples, transition_idx
)
energy = kinetic + potential
else:
energy = (growth ** 2).sum(dim=-1)
energy_loss += energy.mean()
# Mass conservation
growth_sum = growth.sum(dim=-1, keepdim=True) # Keep dimension for proper broadcasting
wt = w0 * torch.exp(growth_sum)
mass = wt.sum()
mass_loss += (mass - w1.sum()).abs()
neg_weight_penalty += torch.relu(-wt).sum()
# Match and reconstruction losses (computed at final time)
xt_final = traj[-1]
match_loss = mean_squared_error(wt, w1)
recons_loss = self.recons_loss(xt_final, x1)
total_loss = (
self.lambda_energy * energy_loss +
self.lambda_mass * (mass_loss + neg_weight_penalty) +
self.lambda_match * match_loss +
self.lambda_recons * recons_loss
)
return total_loss, energy_loss, mass_loss + neg_weight_penalty, match_loss, recons_loss
def training_step(self, batch, batch_idx):
if isinstance(batch, (list, tuple)):
batch = batch[0]
main_batch = batch["train_samples"]
metric_batch = batch["metric_samples"]
if isinstance(main_batch, tuple):
main_batch = main_batch[0]
if isinstance(metric_batch, tuple):
metric_batch = metric_batch[0]
loss = self._compute_loss(main_batch, metric_batch)
if self.joint:
self.log(
"JointTrain/train_loss",
loss,
on_step=False,
on_epoch=True,
prog_bar=True,
logger=True,
)
else:
self.log(
"GrowthNet/train_loss",
loss,
on_step=False,
on_epoch=True,
prog_bar=True,
logger=True,
)
return loss
def validation_step(self, batch, batch_idx):
if isinstance(batch, (list, tuple)):
batch = batch[0]
main_batch = batch["val_samples"]
metric_batch = batch["metric_samples"]
if isinstance(main_batch, tuple):
main_batch = main_batch[0]
if isinstance(metric_batch, tuple):
metric_batch = metric_batch[0]
loss = self._compute_loss(main_batch, metric_batch, validation=True)
if self.joint:
self.log(
"JointTrain/val_loss",
loss,
on_step=False,
on_epoch=True,
prog_bar=True,
logger=True,
)
else:
self.log(
"GrowthNet/val_loss",
loss,
on_step=False,
on_epoch=True,
prog_bar=True,
logger=True,
)
return loss
def configure_optimizers(self):
import itertools
params = list(itertools.chain(*[net.parameters() for net in self.growth_nets]))
if self.joint:
params += list(itertools.chain(*[net.parameters() for net in self.flow_nets]))
if self.optimizer_name == "adam":
optimizer = torch.optim.Adam(params, lr=self.lr)
elif self.optimizer_name == "adamw":
optimizer = torch.optim.AdamW(
params,
lr=self.lr,
weight_decay=self.weight_decay,
)
return optimizer |