File size: 11,016 Bytes
5a87d8d a03ffb8 5a87d8d a03ffb8 5a87d8d a03ffb8 5a87d8d a03ffb8 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 |
import sys
sys.path.append("./BranchSBM")
import os
import sys
import argparse
import copy
from pytorch_lightning import Trainer
from pytorch_lightning.loggers import WandbLogger
import wandb
import hydra
from omegaconf import DictConfig, OmegaConf
from torchcfm.optimal_transport import OTPlanSampler
from branchsbm.branchsbm import BranchSBM
from branchsbm.branch_flow_net_train import FlowNetTrainCell, FlowNetTrainLidar
from branchsbm.branch_interpolant_train import BranchInterpolantTrain
from branchsbm.branch_growth_net_train import GrowthNetTrain, GrowthNetTrainCell, GrowthNetTrainLidar
from dataloaders.trajectory_data import TemporalDataModule
from dataloaders.mouse_data import WeightedBranchedCellDataModule
from dataloaders.three_branch_data import ThreeBranchTahoeDataModule
from dataloaders.clonidine_v2_data import ClonidineV2DataModule
from dataloaders.clonidine_single_branch import ClonidineSingleBranchDataModule
from dataloaders.trametinib_single import TrametinibSingleBranchDataModule
from dataloaders.lidar_data import WeightedBranchedLidarDataModule
from dataloaders.lidar_data_single import LidarSingleDataModule
from networks.flow_mlp import VelocityNet
from networks.growth_mlp import GrowthNet
from networks.interpolant_mlp import GeoPathMLP
from utils import set_seed
from train.parsers import parse_args
from branchsbm.ema import EMA
from train.train_utils import (
load_config,
merge_config,
generate_group_string,
dataset_name2datapath,
create_callbacks,
)
from state_costs.metric_factory import DataManifoldMetric
import torch.nn as nn
def main(args: argparse.Namespace, seed: int, t_exclude: int) -> None:
set_seed(seed)
branches = args.branches
skipped_time_points = [t_exclude] if t_exclude else []
### DATAMODULES ###
if args.data_name == "lidar":
datamodule = WeightedBranchedLidarDataModule(args=args)
elif args.data_name == "lidarsingle":
datamodule = LidarSingleDataModule(args=args)
elif args.data_name == "mouse":
datamodule = WeightedBranchedCellDataModule(args=args)
elif args.data_name in ["clonidine50D", "clonidine100D", "clonidine150D"]:
datamodule = ClonidineV2DataModule(args=args)
elif args.data_name == "clonidine50Dsingle":
datamodule = ClonidineSingleBranchDataModule(args=args)
elif args.data_name == "trametinib":
datamodule = ThreeBranchTahoeDataModule(args=args)
elif args.data_name == "trametinibsingle":
datamodule = TrametinibSingleBranchDataModule(args=args)
flow_nets = nn.ModuleList()
geopath_nets = nn.ModuleList()
growth_nets = nn.ModuleList()
##### initialize branched flow and growth networks #####
for i in range(branches):
flow_net = VelocityNet(
dim=args.dim,
hidden_dims=args.hidden_dims_flow,
activation=args.activation_flow,
batch_norm=False,
)
geopath_net = GeoPathMLP(
input_dim=args.dim,
hidden_dims=args.hidden_dims_geopath,
time_geopath=args.time_geopath,
activation=args.activation_geopath,
batch_norm=False,
)
if i == 0:
growth_net = GrowthNet(
dim=args.dim,
hidden_dims=args.hidden_dims_growth,
activation=args.activation_growth,
batch_norm=False,
negative=True
)
else:
growth_net = GrowthNet(
dim=args.dim,
hidden_dims=args.hidden_dims_growth,
activation=args.activation_growth,
batch_norm=False,
negative=False
)
if args.ema_decay is not None:
flow_net = EMA(model=flow_net, decay=args.ema_decay)
geopath_net = EMA(model=geopath_net, decay=args.ema_decay)
growth_net = EMA(model=growth_net, decay=args.ema_decay)
flow_nets.append(flow_net)
geopath_nets.append(geopath_net)
growth_nets.append(growth_net)
ot_sampler = (
OTPlanSampler(method=args.optimal_transport_method)
if args.optimal_transport_method != "None"
else None
)
wandb.init(
project=f"branchsbm-{args.data_name}-{branches}-branches",
group=args.group_name,
config=vars(args),
dir=args.working_dir,
)
flow_matcher_base = BranchSBM(
geopath_nets=geopath_nets,
sigma=args.sigma,
alpha=int(args.branchsbm),
)
##### STAGE 1: Training of Geodesic Interpolants Beginning #####
geopath_callbacks = create_callbacks(
args, phase="geopath", data_type=args.data_type, run_id=wandb.run.id
)
# define state cost
data_manifold_metric = DataManifoldMetric(
args=args,
skipped_time_points=skipped_time_points,
datamodule=datamodule,
)
geopath_model = BranchInterpolantTrain(
flow_matcher=flow_matcher_base,
skipped_time_points=skipped_time_points,
ot_sampler=ot_sampler,
args=args,
data_manifold_metric=data_manifold_metric
)
wandb_logger = WandbLogger()
trainer = Trainer(
max_epochs=args.epochs,
callbacks=geopath_callbacks,
accelerator=args.accelerator,
logger=wandb_logger,
num_sanity_val_steps=0,
default_root_dir=args.working_dir,
gradient_clip_val=(1.0 if args.data_type == "image" else None),
)
if args.load_geopath_model_ckpt:
best_model_path = args.load_geopath_model_ckpt
else:
trainer.fit(
geopath_model,
datamodule=datamodule,
)
best_model_path = geopath_callbacks[0].best_model_path
geopath_model = BranchInterpolantTrain.load_from_checkpoint(best_model_path)
flow_matcher_base.geopath_nets = geopath_model.geopath_nets
##### STAGE 1: Training of Geodesic Interpolants End #####
##### STAGE 2: Flow Matching Beginning #####
flow_callbacks = create_callbacks(
args,
phase="flow",
data_type=args.data_type,
run_id=wandb.run.id,
datamodule=datamodule,
)
if args.data_type == "lidar":
FlowNetTrain = FlowNetTrainLidar
else:
FlowNetTrain = FlowNetTrainCell
flow_train = FlowNetTrain(
flow_matcher=flow_matcher_base,
flow_nets=flow_nets,
ot_sampler=ot_sampler,
skipped_time_points=skipped_time_points,
args=args,
)
wandb_logger = WandbLogger()
trainer = Trainer(
max_epochs=args.epochs,
callbacks=flow_callbacks,
check_val_every_n_epoch=args.check_val_every_n_epoch,
accelerator=args.accelerator,
logger=wandb_logger,
default_root_dir=args.working_dir,
gradient_clip_val=(1.0 if args.data_type == "image" else None),
num_sanity_val_steps=(0 if args.data_type == "image" else None),
)
trainer.fit(
flow_train, datamodule=datamodule, ckpt_path=args.resume_flow_model_ckpt
)
if args.data_type == "lidar":
trainer.test(flow_train, datamodule=datamodule)
##### STAGE 2: Flow Matching End #####
##### STAGE 3: Training Growth Networks Beginning ####
flow_nets = flow_train.flow_nets
growth_callbacks = create_callbacks(
args,
phase="growth",
data_type=args.data_type,
run_id=wandb.run.id,
datamodule=datamodule,
)
if args.data_type == "lidar":
GrowthNetTrain = GrowthNetTrainLidar
else:
GrowthNetTrain = GrowthNetTrainCell
growth_train = GrowthNetTrain(
flow_nets = flow_nets,
growth_nets = growth_nets,
ot_sampler=ot_sampler,
skipped_time_points=skipped_time_points,
args=args,
data_manifold_metric=data_manifold_metric,
joint = False
)
wandb_logger = WandbLogger()
trainer = Trainer(
max_epochs=args.epochs,
callbacks=growth_callbacks,
check_val_every_n_epoch=args.check_val_every_n_epoch,
accelerator=args.accelerator,
logger=wandb_logger,
default_root_dir=args.working_dir,
gradient_clip_val=(1.0 if args.data_type == "image" else None),
num_sanity_val_steps=(0 if args.data_type == "image" else None),
)
trainer.fit(
growth_train, datamodule=datamodule, ckpt_path=None
)
trainer.test(growth_train, datamodule=datamodule)
##### STAGE 3: Training Growth Networks End ####
##### STAGE 4: Joint Training Beginning ####
growth_nets = growth_train.growth_nets
joint_callbacks = create_callbacks(
args,
phase="joint",
data_type=args.data_type,
run_id=wandb.run.id,
datamodule=datamodule,
)
if args.data_type == "lidar":
GrowthNetTrain = GrowthNetTrainLidar
else:
GrowthNetTrain = GrowthNetTrainCell
joint_train = GrowthNetTrain(
flow_nets = flow_nets,
growth_nets = growth_nets,
ot_sampler=ot_sampler,
skipped_time_points=skipped_time_points,
args=args,
data_manifold_metric=data_manifold_metric,
joint = True
)
wandb_logger = WandbLogger()
trainer = Trainer(
max_epochs=args.epochs,
callbacks=joint_callbacks,
check_val_every_n_epoch=args.check_val_every_n_epoch,
accelerator=args.accelerator,
logger=wandb_logger,
default_root_dir=args.working_dir,
gradient_clip_val=(1.0 if args.data_type == "image" else None),
num_sanity_val_steps=(0 if args.data_type == "image" else None),
)
trainer.fit(
joint_train, datamodule=datamodule, ckpt_path=None
)
trainer.test(joint_train, datamodule=datamodule)
##### STAGE 4: Joint Training End ####
wandb.finish()
if __name__ == "__main__":
args = parse_args()
updated_args = copy.deepcopy(args)
if args.config_path:
config = load_config(args.config_path)
updated_args = merge_config(updated_args, config)
updated_args.group_name = generate_group_string()
updated_args.data_path = dataset_name2datapath(
updated_args.data_name, updated_args.working_dir
)
for seed in updated_args.seeds:
if updated_args.t_exclude:
for i, t_exclude in enumerate(updated_args.t_exclude):
updated_args.t_exclude_current = t_exclude
updated_args.seed_current = seed
updated_args.gamma_current = updated_args.gammas[i]
main(updated_args, seed=seed, t_exclude=t_exclude)
else:
updated_args.seed_current = seed
updated_args.gamma_current = updated_args.gammas[0]
main(updated_args, seed=seed, t_exclude=None) |