File size: 27,631 Bytes
b2cb4a0 | 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 | #
# For licensing see accompanying LICENSE file.
# Copyright (c) 2025 Apple Inc. Licensed under MIT License.
#
import os
import shutil
import copy
import numpy as np
from pathlib import Path
from einops import repeat
import torch
import torch.nn.functional as F
from torch.optim.optimizer import Optimizer
import torch.distributed as dist
from torch.distributed.fsdp import FullyShardedDataParallel
from torch.optim.swa_utils import AveragedModel
from torch.nn.utils import clip_grad_norm_
import lightning
import lightning.pytorch as pl
from fairscale.nn.data_parallel import FullyShardedDataParallel as FSDP
from fairscale.nn.wrap import enable_wrap, wrap
from onescience.utils.simplefold.esm_utils import _af2_to_esm, esm_registry
from onescience.datapipes.boltz_data_pipeline.types import Record, Structure
from onescience.utils.simplefold.boltz_utils import (
weighted_rigid_align,
center_random_augmentation,
process_structure,
save_structure
)
def logit_normal_sample(n=1, m=0.0, s=1.0):
# Logit-Normal Sampling from https://arxiv.org/pdf/2403.03206.pdf
u = torch.randn(n) * s + m
t = 1 / (1 + torch.exp(-u))
return t
def lddt_dist(dmat_predicted, dmat_true, mask, cutoff=15.0, per_atom=False):
# NOTE: the mask is a pairwise mask which should have the identity elements already masked out
# Compute mask over distances
dists_to_score = (dmat_true < cutoff).float() * mask
dist_l1 = torch.abs(dmat_true - dmat_predicted)
score = 0.25 * (
(dist_l1 < 0.5).float()
+ (dist_l1 < 1.0).float()
+ (dist_l1 < 2.0).float()
+ (dist_l1 < 4.0).float()
)
# Normalize over the appropriate axes.
if per_atom:
mask_no_match = torch.sum(dists_to_score, dim=-1) != 0
norm = 1.0 / (1e-10 + torch.sum(dists_to_score, dim=-1))
score = norm * (1e-10 + torch.sum(dists_to_score * score, dim=-1))
return score, mask_no_match.float()
else:
norm = 1.0 / (1e-10 + torch.sum(dists_to_score, dim=(-2, -1)))
score = norm * (1e-10 + torch.sum(dists_to_score * score, dim=(-2, -1)))
total = torch.sum(dists_to_score, dim=(-1, -2))
return score, total
class SimpleFold(pl.LightningModule):
def __init__(
self,
architecture,
processor,
loss,
path,
sampler,
optimizer=None,
scheduler=None,
plddt_module=None,
ema_decay=0.999,
esm_model="esm2_3B",
aa_bolt_link=None,
use_rigid_align=True,
smooth_lddt_loss_weight=1.0,
lddt_cutoff=15.0,
clip_grad_norm_val=None,
lddt_weight_schedule=False,
plddt_training=False,
sample_dir='artifacts/',
):
super().__init__()
self.save_hyperparameters(logger=False)
self.model = architecture
self.model_ema = AveragedModel(
self.model,
multi_avg_fn=torch.optim.swa_utils.get_ema_multi_avg_fn(
self.hparams.ema_decay
),
use_buffers=True,
)
self.model_ema.eval()
self.loss = loss
self.path = path
self.sampler = sampler
self.use_rigid_align = use_rigid_align
self.lddt_cutoff = lddt_cutoff
self.smooth_lddt_loss_weight = smooth_lddt_loss_weight
self.use_smooth_lddt_loss = smooth_lddt_loss_weight > 0.0
self.lddt_weight_schedule = lddt_weight_schedule
self.plddt_training = plddt_training
self.sample_dir = sample_dir
self.aa_bolt_link = aa_bolt_link
self.nval_steps = 0
try:
self.t_eps = self.sampler.t_eps
except AttributeError:
self.t_eps = 0.0
self.use_esm = esm_model is not None
if self.use_esm:
self.esm_model, self.esm_dict = esm_registry[esm_model]()
self.esm_model.eval()
self.af2_to_esm = _af2_to_esm(self.esm_dict)
print(f"Using ESM model: {esm_model}")
else:
self.esm_model = None
self.esm_dict = None
self.af2_to_esm = None
self.plddt_module = plddt_module
if self.plddt_training:
assert self.plddt_module is not None, "PLDDT module must be provided for PLDDT training"
self.model.eval()
def register(self, name, tensor):
self.register_buffer(name, tensor.type(torch.float32))
def loss_masking(self, loss, atom_mask):
loss_mask = repeat(atom_mask, "b s -> b s d", d=loss.shape[-1])
loss *= loss_mask
denom = torch.sum(atom_mask, -1, keepdim=True)
denom = denom.unsqueeze(-1)
loss = torch.sum(loss, dim=1, keepdim=True) / denom
return loss
def smooth_lddt_loss(
self,
pred_coords,
true_coords,
# is_nucleotide,
coords_mask,
t,
):
"""Compute weighted alignment.
Parameters
----------
pred_coords: torch.Tensor
The predicted atom coordinates
true_coords: torch.Tensor
The ground truth atom coordinates
coords_mask: torch.Tensor
The atoms mask
"""
B, N, _ = true_coords.shape
true_dists = torch.cdist(true_coords, true_coords)
mask = (true_dists < self.lddt_cutoff).float()
mask = mask * (1 - torch.eye(pred_coords.shape[1], device=pred_coords.device))
mask = mask * (coords_mask.unsqueeze(-1) * coords_mask.unsqueeze(-2))
# Compute distances between all pairs of atoms
pred_dists = torch.cdist(pred_coords, pred_coords)
dist_diff = torch.abs(true_dists - pred_dists)
# Compute epsilon values
eps = (
(
(
F.sigmoid(0.5 - dist_diff)
+ F.sigmoid(1.0 - dist_diff)
+ F.sigmoid(2.0 - dist_diff)
+ F.sigmoid(4.0 - dist_diff)
)
/ 4.0
)
.view(B, N, N)
.mean(dim=0)
)
# Calculate masked averaging
num = (eps * mask).sum(dim=(-1, -2))
den = mask.sum(dim=(-1, -2)).clamp(min=1)
lddt = num / den
if self.lddt_weight_schedule:
t_weight = 1 + 8 * torch.relu(t - 0.5)
lddt = (1.0 - lddt) * t_weight
return lddt.mean()
else:
return (1.0 - lddt.mean()) * self.smooth_lddt_loss_weight
def plddt_loss(
self,
pred_lddt,
pred_atom_coords,
true_atom_coords,
true_coords_resolved_mask,
feats,
# multiplicity=1,
):
"""Compute plddt loss.
Parameters
----------
pred_lddt: torch.Tensor
The plddt logits
pred_atom_coords: torch.Tensor
The predicted atom coordinates
true_atom_coords: torch.Tensor
The atom coordinates after symmetry correction
true_coords_resolved_mask: torch.Tensor
The resolved mask after symmetry correction
feats: Dict[str, torch.Tensor]
Dictionary containing the model input
Returns
-------
torch.Tensor
Plddt loss
"""
# extract necessary features
atom_mask = true_coords_resolved_mask
R_set_to_rep_atom = feats["r_set_to_rep_atom"].float()
# R_set_to_rep_atom = R_set_to_rep_atom.repeat_interleave(multiplicity, 0).float()
token_type = feats["mol_type"]
# token_type = token_type.repeat_interleave(multiplicity, 0)
# is_nucleotide_token = (token_type == const.chain_type_ids["DNA"]).float() + (
# token_type == const.chain_type_ids["RNA"]
# ).float()
B = true_atom_coords.shape[0]
# atom_to_token = feats["atom_to_token"].float()
# atom_to_token = atom_to_token.repeat_interleave(multiplicity, 0)
token_to_rep_atom = feats["token_to_rep_atom"].float()
# token_to_rep_atom = token_to_rep_atom.repeat_interleave(multiplicity, 0)
true_token_coords = torch.bmm(token_to_rep_atom, true_atom_coords)
pred_token_coords = torch.bmm(token_to_rep_atom, pred_atom_coords)
# compute true lddt
true_d = torch.cdist(
true_token_coords,
torch.bmm(R_set_to_rep_atom, true_atom_coords),
)
pred_d = torch.cdist(
pred_token_coords,
torch.bmm(R_set_to_rep_atom, pred_atom_coords),
)
# compute mask
pair_mask = atom_mask.unsqueeze(-1) * atom_mask.unsqueeze(-2)
pair_mask = (
pair_mask
* (1 - torch.eye(pair_mask.shape[1], device=pair_mask.device))[None, :, :]
)
pair_mask = torch.einsum("bnm,bkm->bnk", pair_mask, R_set_to_rep_atom)
pair_mask = torch.bmm(token_to_rep_atom, pair_mask)
atom_mask = torch.bmm(token_to_rep_atom, atom_mask.unsqueeze(-1).float())
# is_nucleotide_R_element = torch.bmm(
# R_set_to_rep_atom, torch.bmm(atom_to_token, is_nucleotide_token.unsqueeze(-1))
# ).squeeze(-1)
# cutoff = 15 + 15 * is_nucleotide_R_element.reshape(B, 1, -1).repeat(
# 1, true_d.shape[1], 1
# )
# compute lddt
target_lddt, mask_no_match = lddt_dist(
pred_d, true_d, pair_mask, cutoff=15.0, per_atom=True
)
# compute loss
num_bins = pred_lddt.shape[-1]
bin_index = torch.floor(target_lddt * num_bins).long()
bin_index = torch.clamp(bin_index, max=(num_bins - 1))
lddt_one_hot = F.one_hot(bin_index, num_classes=num_bins)
errors = -1 * torch.sum(
lddt_one_hot * F.log_softmax(pred_lddt, dim=-1),
dim=-1,
)
atom_mask = atom_mask.squeeze(-1)
loss = torch.sum(errors * atom_mask * mask_no_match, dim=-1) / (
1e-7 + torch.sum(atom_mask * mask_no_match, dim=-1)
)
# Average over the batch dimension
loss = torch.mean(loss)
self.log(
"loss/plddt",
loss.item(),
on_epoch=True,
logger=True,
prog_bar=True,
rank_zero_only=True,
)
return loss
def plddt_train_step(self, batch, batch_idx):
with torch.no_grad():
batch = self.processor.preprocess_training(
batch,
esm_model=self.esm_model,
esm_dict=self.esm_dict,
af2_to_esm=self.af2_to_esm,
)
noise = torch.randn_like(batch['coords']).to(self.device)
out_dict = self.sampler.sample(
self.model_ema.module.forward, self.path,
noise, batch
)
# out_dict = self.processor.postprocess(out_dict, batch)
# denoised_coords = center_of_mass_norm(
# out_dict["denoised_coords"], batch['atom_pad_mask']
# )
# true_coords = center_of_mass_norm(
# batch['coords'], batch['atom_pad_mask']
# )
denoised_coords = center_random_augmentation(
out_dict["denoised_coords"],
batch['atom_pad_mask'],
augmentation=False,
centering=True,
)
true_coords = center_random_augmentation(
batch['coords'],
batch['atom_pad_mask'],
augmentation=False,
centering=True,
)
out_dict["denoised_coords"] = denoised_coords * self.processor.scale
out_dict["coords"] = true_coords * self.processor.scale
out_dict["true_coords_resolved_mask"] = batch["atom_resolved_mask"]
t = torch.ones(batch['coords'].shape[0], device=self.device)
out_feat = self.model(denoised_coords, t, batch) # use unscaled coords
# Compute plddt loss
plddt_out_dict = self.plddt_module(
out_feat["latent"].detach(),
batch,
)
plddt_loss = self.plddt_loss(
plddt_out_dict["plddt_logits"],
out_dict["denoised_coords"],
out_dict["coords"],
out_dict["true_coords_resolved_mask"],
batch,
)
return plddt_loss
def flow_matching_train_step(self, batch, batch_idx):
batch = self.processor.preprocess_training(
batch,
esm_model=self.esm_model,
esm_dict=self.esm_dict,
af2_to_esm=self.af2_to_esm,
)
# timestep resampling
t_size = batch['coords'].shape[0]
t = 0.98 * logit_normal_sample(n=t_size, m=0.8, s=1.7) + 0.02 * torch.rand(t_size)
t = t.to(self.device)
t = t * (1 - 2 * self.t_eps) + self.t_eps
noise = torch.randn_like(batch['coords']).to(self.device)
_, y_t, v_t = self.path.interpolant(t, noise, batch["coords"])
out_dict = self.model(y_t, t, batch)
resolved_atom_mask = batch["atom_resolved_mask"].float()
align_weights = y_t.new_ones(y_t.shape[:2])
if self.use_rigid_align:
with torch.no_grad(), torch.autocast("cuda", enabled=False):
v_t = out_dict['predict_velocity'].detach().float()
denoised_coords = y_t + v_t * (1.0 - t[:, None, None])
coords = batch["coords"].detach().float()
coords_aligned = weighted_rigid_align(
coords,
denoised_coords.detach().float(),
align_weights.detach().float(),
mask=resolved_atom_mask.detach().float(),
)
_, _, v_t_aligned = self.path.interpolant(t, noise, coords_aligned)
target = v_t_aligned
else:
target = v_t
loss = F.mse_loss(out_dict['predict_velocity'], target, reduction='none')
loss_mask = resolved_atom_mask * align_weights
loss = self.loss_masking(loss, loss_mask)
loss = loss.mean()
self.log(
"loss/mse",
loss.item(),
on_epoch=True,
logger=True,
prog_bar=True,
rank_zero_only=True,
)
if self.use_smooth_lddt_loss:
# one-step Euler to get denoised coordinates
denoised_coords = y_t + \
out_dict['predict_velocity'] * (1.0 - t[:, None, None])
# rescale coordinates to angstroms
# denoised_coords = center_of_mass_norm(denoised_coords, batch['atom_pad_mask'])
# true_coords = center_of_mass_norm(batch['coords'], batch['atom_pad_mask'])
denoised_coords = center_random_augmentation(
denoised_coords,
batch['atom_pad_mask'],
augmentation=False,
centering=True,
)
true_coords = center_random_augmentation(
batch['coords'],
batch['atom_pad_mask'],
augmentation=False,
centering=True,
)
denoised_coords = denoised_coords * self.processor.scale
true_coords = true_coords * self.processor.scale
smooth_lddt_loss = self.smooth_lddt_loss(
denoised_coords,
true_coords,
resolved_atom_mask,
t,
)
loss += smooth_lddt_loss
self.log(
"loss/smooth_lddt",
smooth_lddt_loss.item(),
on_epoch=True,
logger=True,
prog_bar=True,
rank_zero_only=True,
)
self.log(
"loss/loss",
loss.item(),
on_epoch=True,
logger=True,
prog_bar=True,
rank_zero_only=True,
)
self.log(
"trainer/global_step",
self.global_step,
on_epoch=False,
logger=True,
prog_bar=False,
rank_zero_only=True,
)
self.global_training_step = self.trainer.global_step
self.epoch = self.trainer.current_epoch
self.world_size = self.trainer.world_size
return loss
def training_step(self, batch, batch_idx):
if self.plddt_training:
return self.plddt_train_step(batch, batch_idx)
else:
return self.flow_matching_train_step(batch, batch_idx)
@torch.no_grad()
def validation_step(self, batch, batch_idx):
# we skip validation step in training mode
return
def on_train_start(self):
global_seed = os.environ.get("PL_GLOBAL_SEED", 42)
pl.seed_everything(int(global_seed) + self.trainer.global_rank, True)
return
@torch.no_grad()
def predict_step(self, batch, batch_idx):
with torch.autocast(device_type='cuda', dtype=torch.float32):
batch = self.processor.preprocess_inference(
batch,
esm_model=self.esm_model,
esm_dict=self.esm_dict,
af2_to_esm=self.af2_to_esm,
)
num_repeats = batch.get("num_repeats", torch.tensor(1, device=self.device)).item()
multiplicity = batch["mol_type"].shape[0]
num_iter = np.ceil(num_repeats / multiplicity).astype(int)
print(f"Generating {num_repeats} samples with num_iter: {num_iter}, multiplicity: {multiplicity}")
# num_repeats is the total number of samples to generate for one protein
# multiplicity is the number of samples to generate at once
curr_idx = 0
for i in range(num_iter):
batch_in = copy.deepcopy(batch)
noise = torch.randn_like(batch_in['coords']).to(self.device)
out_dict = self.sampler.sample(
self.model_ema.module.forward, self.path,
noise, batch_in
)
if self.plddt_module is not None:
denoised_coords = center_random_augmentation(
out_dict["denoised_coords"],
batch['atom_pad_mask'],
augmentation=False,
centering=True,
)
t = torch.ones(batch['coords'].shape[0], device=self.device)
out_feat = self.model(denoised_coords, t, batch) # use unscaled coords
plddt_out_dict = self.plddt_module(
out_feat["latent"],
batch_in,
)
plddts = plddt_out_dict["plddt"] * 100.0
else:
plddts = None
out_dict = self.processor.postprocess(out_dict, batch_in)
record = Record(**batch_in['record'][0])
gt_coord = out_dict['coords']
sampled_coord = out_dict['denoised_coords']
pad_mask = batch_in['atom_pad_mask']
if num_repeats - curr_idx < multiplicity:
curr_num_copies = num_repeats - curr_idx
gt_coord = gt_coord[:curr_num_copies]
sampled_coord = sampled_coord[:curr_num_copies]
pad_mask = pad_mask[:curr_num_copies]
plddts = plddts[:curr_num_copies] if plddts is not None else None
else:
curr_num_copies = multiplicity
data_dir = self.trainer.datamodule.predict_dataloader().dataset.target_dir
sample_dir = self.sample_dir
path = Path(data_dir) / "structures" / f"{record.id}.npz"
structure: Structure = Structure.load(path)
for j in range(curr_num_copies):
file_id = j + curr_idx
try:
sampled_structure = copy.deepcopy(structure)
sampled_structure = process_structure(
sampled_structure, sampled_coord[j], pad_mask[j], record
)
# save mmcif structure
sampled_struct_dir = Path(sample_dir)
outname = f"{record.id}_sampled_{str(file_id)}"
save_structure(
sampled_structure, sampled_struct_dir, outname,
plddts=plddts[j] if plddts is not None else None,
output_format="mmcif",
)
# save pdb structure
save_structure(
sampled_structure, sampled_struct_dir, outname,
plddts=plddts[j] if plddts is not None else None,
output_format="pdb",
)
except:
print(f"Error processing {record.id}")
continue
curr_idx += curr_num_copies
self.world_size = self.trainer.world_size
return
def on_predict_epoch_end(self):
dist.barrier() # wait for all processes to finish
return
def reset_esm(self, esm_model: str):
self.esm_model, self.esm_dict = esm_registry[esm_model]()
self.esm_model.eval()
self.af2_to_esm = _af2_to_esm(self.esm_dict)
self.esm_model = self.esm_model.to(self.device)
self.af2_to_esm = self.af2_to_esm.to(self.device)
print(f"Successfully reset ESM model {esm_model}")
def setup(self, stage: str) -> None:
"""Lightning hook that is called at the beginning of fit (train + validate), validate,
test, or predict.
This is a good hook when you need to build models dynamically or adjust something about
them. This hook is called on every process when using DDP.
:param stage: Either `"fit"`, `"validate"`, `"test"`, or `"predict"`.
"""
self.processor = self.hparams.processor(device=self.device)
if self.use_esm:
if stage == "fit" and not isinstance(
self.trainer.strategy, lightning.pytorch.strategies.fsdp.FSDPStrategy
):
# initialize the model with FSDP wrapper
fsdp_params = dict(
mixed_precision=True,
flatten_parameters=True,
state_dict_device=torch.device("cpu"), # reduce GPU mem usage
cpu_offload=True, # enable cpu offloading
fp32_reduce_scatter=True, # use fp32 reduce scatter
)
with enable_wrap(wrapper_cls=FSDP, **fsdp_params):
self.esm_model.eval()
# Wrap each layer in FSDP separately
for name, child in self.esm_model.named_children():
if name == "layers":
for layer_name, layer in child.named_children():
wrapped_layer = wrap(layer)
setattr(child, layer_name, wrapped_layer)
self.esm_model = wrap(self.esm_model)
self.af2_to_esm = self.af2_to_esm.to(self.device)
else:
self.esm_model = self.esm_model.to(self.device)
self.af2_to_esm = self.af2_to_esm.to(self.device)
if stage == "fit":
self.training_gpus = self.trainer.world_size
self.hparams["training_gpus"] = self.training_gpus
batch = next(iter(self.trainer.datamodule.train_dataloader()))
batch = self.processor.preprocess_training(
batch,
self.esm_model,
self.esm_dict,
self.af2_to_esm,
)
y = batch["coords"]
t = torch.zeros((y.shape[0])).cuda()
def on_train_batch_end(self, outputs, batch, batch_idx):
optimizer = self.optimizers()
self.log(
"trainer/lr",
optimizer.param_groups[0]["lr"],
on_epoch=True,
logger=True,
prog_bar=True,
rank_zero_only=True,
)
def on_before_optimizer_step(self, optimizer: Optimizer) -> None:
if isinstance(
self.trainer.strategy, lightning.pytorch.strategies.fsdp.FSDPStrategy
):
with FullyShardedDataParallel.summon_full_params(
self.trainer.strategy.model, with_grads=True
):
clip_grad_norm_(
self.trainer.strategy.model.model.parameters(),
self.hparams.clip_grad_norm_val,
norm_type=2.0,
error_if_nonfinite=True,
)
else:
clip_grad_norm_(
self.model.parameters(),
self.hparams.clip_grad_norm_val,
norm_type=2.0,
error_if_nonfinite=True,
)
return
def on_before_zero_grad(self, optimizer: Optimizer) -> None:
# if self.eval_ema:
if isinstance(
self.trainer.strategy, lightning.pytorch.strategies.fsdp.FSDPStrategy
):
with FullyShardedDataParallel.summon_full_params(
self.trainer.strategy.model
):
self.trainer.strategy.model.model_ema.update_parameters(
self.trainer.strategy.model.model
)
else:
self.model_ema.update_parameters(self.model)
return
def on_save_checkpoint(self, checkpoint) -> None:
if not isinstance(
self.trainer.strategy, lightning.pytorch.strategies.fsdp.FSDPStrategy
):
layers_to_delete = []
for k in checkpoint["state_dict"].keys():
if k.startswith("esm_model._fsdp_wrapped_module"):
layers_to_delete.append(k)
for k in layers_to_delete:
del checkpoint["state_dict"][k]
else:
try:
del checkpoint["hyper_parameters"]
except:
print("No hyper_parameters in checkpoint")
return super().on_save_checkpoint(checkpoint)
def on_load_checkpoint(self, checkpoint) -> None:
if not isinstance(
self.trainer.strategy, lightning.pytorch.strategies.fsdp.FSDPStrategy
):
self.training_gpus = checkpoint["hyper_parameters"]["training_gpus"]
self.fwd_flops = checkpoint["hyper_parameters"]["fwd_flops"]
if checkpoint["loops"] is not None:
self.trainer.fit_loop.load_state_dict(checkpoint["loops"]["fit_loop"])
self.trainer.validate_loop.load_state_dict(
checkpoint["loops"]["validate_loop"]
)
return super().on_load_checkpoint(checkpoint)
def configure_optimizers(self):
optimizer = self.hparams.optimizer(
params=self.trainer.model.parameters()
)
if self.hparams.scheduler is not None:
scheduler = self.hparams.scheduler(optimizer=optimizer)
return {
"optimizer": optimizer,
"lr_scheduler": {
"scheduler": scheduler,
"interval": "step",
"frequency": 1,
},
}
return {"optimizer": optimizer}
|