File size: 14,510 Bytes
5a87d8d |
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 |
import sys
sys.path.append("./BranchSBM")
import torch
import pytorch_lightning as pl
from branchsbm.ema import EMA
import itertools
from utils import wasserstein_distance, plot_lidar
import matplotlib.pyplot as plt
class BranchInterpolantTrain(pl.LightningModule):
def __init__(
self,
flow_matcher,
args,
skipped_time_points: list = None,
ot_sampler=None,
state_cost=None,
data_manifold_metric=None,
):
super().__init__()
self.save_hyperparameters()
self.args = args
self.flow_matcher = flow_matcher
# list of geopath nets
self.geopath_nets = flow_matcher.geopath_nets
self.branches = len(self.geopath_nets)
self.metric_clusters = args.metric_clusters
self.ot_sampler = ot_sampler
self.skipped_time_points = skipped_time_points if skipped_time_points else []
self.optimizer_name = args.geopath_optimizer
self.lr = args.geopath_lr
self.weight_decay = args.geopath_weight_decay
self.args = args
self.multiply_validation = 4
self.first_loss = None
self.timesteps = None
self.computing_reference_loss = False
# updates
self.state_cost = state_cost
self.data_manifold_metric = data_manifold_metric
self.whiten = args.whiten
def forward(self, x0, x1, t, branch_idx):
# return specific branch interpolant
return self.geopath_nets[branch_idx](x0, x1, t)
def on_train_start(self):
self.first_loss = self.compute_initial_loss()
print("first loss")
print(self.first_loss)
# to edit
def compute_initial_loss(self):
# Set all GeoPath networks to eval mode
for net in self.geopath_nets:
net.train(mode=False)
total_loss = 0
total_count = 0
with torch.enable_grad():
self.t_val = []
for i in range(
self.trainer.datamodule.num_timesteps - len(self.skipped_time_points)
):
self.t_val.append(
torch.rand(
self.trainer.datamodule.batch_size * self.multiply_validation,
requires_grad=True,
)
)
self.computing_reference_loss = True
with torch.no_grad():
old_alpha = self.flow_matcher.alpha
self.flow_matcher.alpha = 0
for batch in self.trainer.datamodule.train_dataloader():
self.timesteps = torch.linspace(
0.0, 1.0, len(batch[0]["train_samples"][0])
)
loss = self._compute_loss(
batch[0]["train_samples"][0],
batch[0]["metric_samples"][0],
)
print("initial loss")
print(loss)
total_loss += loss.item()
total_count += 1
self.flow_matcher.alpha = old_alpha
self.computing_reference_loss = False
# Set all GeoPath networks back to training mode
for net in self.geopath_nets:
net.train(mode=True)
return total_loss / total_count if total_count > 0 else 1.0
def _compute_loss(self, main_batch, metric_samples_batch=None):
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 == 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)
]
"""samples0, samples1, samples2 = (
metric_samples_batch[0],
metric_samples_batch[1],
metric_samples_batch[2]
)"""
assert len(x1s_list) == self.branches, "Mismatch between x1s_list and expected branches"
# compute sum of velocities for each branch
loss = 0
velocities = []
for branch_idx in range(self.branches):
ts, xts, uts = self._process_flow(x0s, x1s_list[branch_idx], branch_idx)
for i in range(len(ts)):
# calculate kinetic and potential energy of the predicted interpolant
if self.args.manifold:
start_samples, end_samples = branch_sample_pairs[branch_idx]
samples = torch.cat([start_samples, end_samples], dim=0)
#print("metric sample shape")
#print(samples.shape)
vel, _, _ = self.data_manifold_metric.calculate_velocity(
xts[i], uts[i], samples, i
)
else:
vel = torch.sqrt((uts[i]**2).sum(dim =-1) + self.state_cost(xts[i]))
#vel = (uts[i]**2).sum(dim =-1)
velocities.append(vel)
loss = torch.mean(torch.cat(velocities) ** 2)
self.log(
"BranchPathNet/mean_velocity_geopath",
loss,
on_step=False,
on_epoch=True,
prog_bar=True,
)
return loss
def _process_flow(self, x0s, x1s, branch_idx):
ts, xts, uts = [], [], []
t_start = self.timesteps[0]
i_start = 0
for i, (x0, x1) in enumerate(zip(x0s, x1s)):
x0, x1 = torch.squeeze(x0), torch.squeeze(x1)
if self.trainer.validating or self.computing_reference_loss:
repeat_tuple = (self.multiply_validation, 1) + (1,) * (
len(x0.shape) - 2
)
x0 = x0.repeat(repeat_tuple)
x1 = x1.repeat(repeat_tuple)
if self.ot_sampler is not None:
x0, x1 = self.ot_sampler.sample_plan(
x0,
x1,
replace=True,
)
if self.skipped_time_points and i + 1 >= self.skipped_time_points[0]:
t_start_next = self.timesteps[i + 2]
else:
t_start_next = self.timesteps[i + 1]
t = None
if self.trainer.validating or self.computing_reference_loss:
t = self.t_val[i]
t, xt, ut = self.flow_matcher.sample_location_and_conditional_flow(
x0, x1, t_start, t_start_next, branch_idx, training_geopath_net=True, t=t
)
ts.append(t)
xts.append(xt)
uts.append(ut)
t_start = t_start_next
return ts, xts, uts
def training_step(self, batch, batch_idx):
if self.args.data_type in ["scrna", "tahoe"]:
main_batch = batch[0]["train_samples"][0]
metric_batch = batch[0]["metric_samples"][0]
else:
main_batch = batch["train_samples"][0]
metric_batch = batch["metric_samples"][0]
tangential_velocity_loss = self._compute_loss(main_batch, metric_batch)
if self.first_loss:
tangential_velocity_loss = tangential_velocity_loss / self.first_loss
self.log(
"BranchPathNet/mean_geopath_geopath",
(self.flow_matcher.geopath_net_output.abs().mean()),
on_step=False,
on_epoch=True,
prog_bar=True,
)
self.log(
"BranchPathNet/train_loss_geopath",
tangential_velocity_loss,
on_step=True,
on_epoch=True,
prog_bar=True,
logger=True,
)
return tangential_velocity_loss
def validation_step(self, batch, batch_idx):
if self.args.data_type in ["scrna", "tahoe"]:
main_batch = batch[0]["val_samples"][0]
metric_batch = batch[0]["metric_samples"][0]
else:
main_batch = batch["val_samples"][0]
metric_batch = batch["metric_samples"][0]
tangential_velocity_loss = self._compute_loss(main_batch, metric_batch)
if self.first_loss:
tangential_velocity_loss = tangential_velocity_loss / self.first_loss
self.log(
"BranchPathNet/val_loss_geopath",
tangential_velocity_loss,
on_step=False,
on_epoch=True,
prog_bar=True,
logger=True,
)
return tangential_velocity_loss
def test_step(self, batch, batch_idx):
main_batch = batch["test_samples"][0]
metric_batch = batch["metric_samples"][0]
x0 = main_batch["x0"][0] # [B, D]
cloud_points = main_batch["dataset"][0] # full dataset, [N, D]
x0 = x0.to(self.device)
cloud_points = cloud_points.to(self.device)
t_vals = [0.25, 0.5, 0.75]
t_labels = ["t=1/4", "t=1/2", "t=3/4"]
colors = {
"x0": "#4D176C",
"t=1/4": "#5C3B9D",
"t=1/2": "#6172B9",
"t=3/4": "#AC4E51",
"x1": "#771F4F",
}
# Unwhiten cloud points if needed
if self.whiten:
cloud_points = torch.tensor(
self.trainer.datamodule.scaler.inverse_transform(cloud_points.cpu().numpy())
)
for i in range(self.branches):
geopath = self.geopath_nets[i]
x1_key = f"x1_{i + 1}"
if x1_key not in main_batch:
print(f"Skipping branch {i + 1}: no final distribution {x1_key}")
continue
x1 = main_batch[x1_key][0].to(self.device)
print(x1.shape)
print(x0.shape)
interpolated_points = []
with torch.no_grad():
for t_scalar in t_vals:
t_tensor = torch.full((x0.shape[0], 1), t_scalar, device=self.device) # [B, 1]
xt = geopath(x0, x1, t_tensor).cpu() # [B, D]
if self.whiten:
xt = torch.tensor(
self.trainer.datamodule.scaler.inverse_transform(xt.numpy())
)
interpolated_points.append(xt)
if self.whiten:
x0_plot = torch.tensor(
self.trainer.datamodule.scaler.inverse_transform(x0.cpu().numpy())
)
x1_plot = torch.tensor(
self.trainer.datamodule.scaler.inverse_transform(x1.cpu().numpy())
)
else:
x0_plot = x0.cpu()
x1_plot = x1.cpu()
# Plot
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)
# Initial x₀
ax.scatter(
x0_plot[:, 0], x0_plot[:, 1], x0_plot[:, 2],
s=15, alpha=1.0, color=colors["x0"], label="x₀", depthshade=True,
edgecolors="white",
linewidths=0.3
)
# Interpolated points
for xt, t_label in zip(interpolated_points, t_labels):
ax.scatter(
xt[:, 0], xt[:, 1], xt[:, 2],
s=15, alpha=1.0, color=colors[t_label], label=t_label, depthshade=True,
edgecolors="white",
linewidths=0.3
)
# Final x₁
ax.scatter(
x1_plot[:, 0], x1_plot[:, 1], x1_plot[:, 2],
s=15, alpha=1.0, color=colors["x1"], label="x₁", depthshade=True,
edgecolors="white",
linewidths=0.3
)
ax.legend()
save_path = f"/raid/st512/branchsbm/figures/{self.args.data_type}/lidar_geopath_branch_{i+1}.png"
plt.savefig(save_path, dpi=300)
plt.close()
def optimizer_step(self, *args, **kwargs):
super().optimizer_step(*args, **kwargs)
if isinstance(self.geopath_nets, EMA):
self.geopath_nets.update_ema()
def configure_optimizers(self):
if self.optimizer_name == "adam":
"""optimizer = torch.optim.Adam(
self.geopath_nets.parameters(),
lr=self.lr,
)"""
optimizer = torch.optim.Adam(
itertools.chain(*[net.parameters() for net in self.geopath_nets]), lr=self.lr
)
elif self.optimizer_name == "adamw":
"""optimizer = torch.optim.AdamW(
self.geopath_nets.parameters(),
lr=self.lr,
weight_decay=self.weight_decay,
)"""
optimizer = torch.optim.AdamW(
itertools.chain(*[net.parameters() for net in self.geopath_nets]), lr=self.lr
)
return optimizer
|