-
Notifications
You must be signed in to change notification settings - Fork 179
/
train_DFOLD_dynamics.py
1583 lines (1341 loc) · 70.2 KB
/
train_DFOLD_dynamics.py
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
"""Pytorch script for training SE(3) protein diffusion.
To run:
> python experiments/train_se3_diffusion.py
Without Wandb,
> python experiments/train_se3_diffusion.py experiment.use_wandb=False
To modify config options with the command line,
> python experiments/train_se3_diffusion.py experiment.batch_size=32
"""
import os
import torch
import GPUtil
import time
import tree
import numpy as np
import wandb
import copy
import hydra
import logging
import copy
import random
import pandas as pd
from collections import defaultdict,deque
from datetime import datetime
from omegaconf import DictConfig,OmegaConf
from torch.nn import DataParallel as DP
from torch.nn.parallel import DistributedDataParallel as DDP
import torch.distributed as dist
from torch.utils import data
from torch.utils.tensorboard import SummaryWriter
from Bio.SVDSuperimposer import SVDSuperimposer
import gc
import matplotlib.pyplot as plt
from hydra.core.hydra_config import HydraConfig
from src.analysis import utils as au
from src.analysis import metrics
from src.data import Dfold_data_loader_dynamic,se3_diffuser,all_atom
from src.data import utils as du
from src.model import Dfold_network_dynamic
from src.experiments import utils as eu
from openfold.utils.loss import lddt, lddt_ca,torsion_angle_loss
from openfold.np import residue_constants#
from openfold.utils.superimposition import superimpose
from openfold.utils.validation_metrics import gdt_ts,gdt_ha,drmsd
from openfold.utils.lr_schedulers import AlphaFoldLRScheduler
from openfold.utils import rigid_utils as ru
from src.toolbox.rot_trans_error import average_quaternion_distances,average_translation_distances
import MDAnalysis as mda
from MDAnalysis.analysis import rms,align,rdf,contacts
from scipy.stats import pearsonr
import pickle
from tqdm import tqdm
import mdtraj as md
def format_func(value, tick_number):
return f'{value:.1f}'
from matplotlib.ticker import FuncFormatter
formatter = FuncFormatter(format_func)
#https://github.com/scipy/scipy/blob/main/scipy/spatial/transform/_rotation.pyx
def rmsdalign(a, b, weights=None): # alignes B to A # [*, N, 3]
B = a.shape[:-2]
N = a.shape[-2]
if weights == None:
weights = a.new_ones(*B, N)
weights = weights.unsqueeze(-1)
a_mean = (a * weights).sum(-2, keepdims=True) / weights.sum(-2, keepdims=True)
a = a - a_mean
b_mean = (b * weights).sum(-2, keepdims=True) / weights.sum(-2, keepdims=True)
b = b - b_mean
B = torch.einsum('...ji,...jk->...ik', weights * a, b)
u, s, vh = torch.linalg.svd(B)
# Correct improper rotation if necessary (as in Kabsch algorithm)
'''
if torch.linalg.det(u @ vh) < 0:
s[-1] = -s[-1]
u[:, -1] = -u[:, -1]
'''
sgn = torch.sign(torch.linalg.det(u @ vh))
s[...,-1] *= sgn
u[...,:,-1] *= sgn.unsqueeze(-1)
C = u @ vh # c rotates B to A
return b @ C.mT + a_mean
@torch.no_grad()
def compute_validation_metrics_all(gt_pos, out_pos,gt_mask,superimposition_metrics=False):
metrics = {}
gt_coords =gt_pos# batch["atom37_pos"]
pred_coords = out_pos#outputs["atom37_pos"]
all_atom_mask = gt_mask
gt_coords = gt_coords#torch.from_numpy(gt_coords)
pred_coords = torch.from_numpy(pred_coords)
all_atom_mask = all_atom_mask#torch.from_numpy(all_atom_mask)
# print(gt_coords.shape,pred_coords.shape, all_atom_mask[..., None].shape)
# This is super janky for superimposition. Fix later
gt_coords_masked = gt_coords * all_atom_mask[..., None]
pred_coords_masked = pred_coords * all_atom_mask[..., None]
ca_pos = residue_constants.atom_order["CA"]
gt_coords_masked_ca = gt_coords_masked[..., ca_pos, :] # [11,N,3]
pred_coords_masked_ca = pred_coords_masked[..., ca_pos, :]# [11,N,3]
all_atom_mask_ca = all_atom_mask[..., ca_pos]
#[11,N]
drmsd_ca_score = drmsd(
pred_coords_masked_ca,
gt_coords_masked_ca,
mask=all_atom_mask_ca, # still required here to compute n
)
metrics["drmsd_ca"] = drmsd_ca_score
frame_time,Nseq,_,_ = gt_coords_masked.shape
gt_coords_masked = gt_coords_masked.reshape([frame_time,-1,3])
pred_coords_masked = pred_coords_masked.reshape([frame_time,-1,3])
diff = gt_coords_masked - pred_coords_masked # [F,N*37,3]
# print(diff.shape,all_atom_mask.shape,all_atom_mask_ca.shape)
# print(diff[1,:10,0],all_atom_mask.reshape([4,-1])[1,:10])
# torch.Size([4, 2516(68*37), 3]) torch.Size([4, 68, 37]) torch.Size([4, 68])
#xit()
# diff torch.Size([4, 37N, 3])
#all_atom_mask [F,N,37]
metrics["rmsd_all"] = torch.sqrt(torch.sum(diff**2,axis=(-1,-2))/(torch.sum(all_atom_mask, dim=(-1, -2)) + 1e-4))
diff = gt_coords_masked_ca - pred_coords_masked_ca # [F,N,3]
#all_atom_mask_ca [F,N]
metrics["rmsd_ca"] = torch.sqrt(torch.sum(diff**2,axis=(-1,-2))/(torch.sum(all_atom_mask_ca, dim=-1) + 1e-4))
# print('='*100)
superimposed_pred, alignment_rmsd = superimpose(
gt_coords_masked_ca, pred_coords_masked_ca,
)
metrics["rmsd_ca_aligned"] = alignment_rmsd
return metrics
@torch.no_grad()
def plot_curve_merged(metric_merged,save_path,row_num=2,col_num=5,suffer_fix=None):
total_width = col_num * 2
total_height = row_num * 2
fig, axes = plt.subplots(row_num, col_num,figsize=(total_width, total_height),dpi=300)
# 输出每一行的信息
for key in metric_merged.keys():
data = metric_merged[key]
for index, row in data.iterrows():
name = row['pdb_name']
col_id = index
if col_num == 1:
axes[0].plot(row['rmsd_all'],label=key, marker='o', linestyle='-')
axes[1].plot(row['rmsd_ca_aligned'],label=key, marker='o', linestyle='-')
axes[2].plot(row['rmsd_ca'],label=key, marker='o', linestyle='-')
axes[0].set_title(name+' | RMSE')
axes[1].set_title(name+' | RMSD_ca_a')
axes[2].set_title(name+' | RMSE_ca')
else:
axes[0, col_id].plot(row['rmsd_all'],label=key, marker='o', linestyle='-')
axes[1, col_id].plot(row['rmsd_ca_aligned'],label=key, marker='o', linestyle='-')
axes[2, col_id].plot(row['rmsd_ca'],label=key, marker='o', linestyle='-')
axes[0, col_id].set_title(name+' | RMSE')
axes[1, col_id].set_title(name+' | RMSD_ca_a')
axes[2, col_id].set_title(name+' | RMSE_ca')
# 在每个子图上显示图例
plt.suptitle('RSME over Atoms')
plt.tight_layout()
plt.legend()
# plt.axis('off')
if suffer_fix is not None:
plt.savefig(f'{save_path}/rmse_rmsd_{suffer_fix}.png')
else:
plt.savefig(f'{save_path}/rmse_rmsd.png')
return fig
# rot_trans_error_dict = {'name':[],"ave_rot":[],"ave_trans":[],"first_rot":[],"first_trans":[]}
@torch.no_grad()
def plot_rot_trans_curve(error_dict,save_path,frame_step=1):
rows,cols = 2,len(error_dict['name'])
fig, axes = plt.subplots(rows, cols, figsize=(cols*2,rows*2),dpi=300)
# /cpfs01/projects-HDD/cfff-6f3a36a0cd1e_HDD/public/protein/workspace/chengkaihui/code/DFOLDv2/plot/4ue8_B_simulation_1_frame_step_1_wj.pickle
motion_pkl_path = f'/cpfs01/projects-HDD/cfff-6f3a36a0cd1e_HDD/public/protein/workspace/chengkaihui/code/DFOLDv2/plot/4ue8_B_simulation_1_frame_step_{frame_step}_wj.pickle'# 1a62_A_ 4ue8_B
# motion_pkl_path = f'/cpfs01/projects-HDD/cfff-6f3a36a0cd1e_HDD/public/protein/workspace/chengkaihui/code/DFOLDv2/plot/4ue8_b_38_{frame_step}_atlas.pickle'
try:# atlas_average_motion_traj1
with open(motion_pkl_path, 'rb') as handle:
loaded_data = pickle.load(handle)
except (EOFError, FileNotFoundError, pickle.UnpicklingError) as e:
# 在这里捕获可能的异常并跳过处理
loaded_data=None
print(f'======> motion step {frame_step}:',loaded_data)
print('======> error dict:',error_dict)
for idx,name in enumerate(error_dict['name']):
if cols==1:
axes[0].plot(error_dict['ave_rot'][idx],label='Pred', marker='o', linestyle='-')
axes[0].plot(error_dict['first_rot'][idx],label='RefAsPred', marker='o', linestyle='-')
# x = np.arange(1, len(error_dict['first_rot'][idx]))
# axes[0].plot(x,error_dict['time_rot_dif'][idx][1:],label='RM', marker='o', linestyle='-')
axes[1].plot(error_dict['ave_trans'][idx],label='Pred', marker='o', linestyle='-')
axes[1].plot(error_dict['first_trans'][idx],label='RefAsPred', marker='o', linestyle='-')
# axes[1].plot(x,error_dict['time_trans_dif'][idx][1:],label='RM', marker='o', linestyle='-')
# plot percent compare with traj motion
if loaded_data is not None:
rots_traj_motion = loaded_data[name]['ARC_rot']
rots_traj_motion = np.array([rots_traj_motion]* len(error_dict['ave_rot'][idx]))
axes[0].plot(rots_traj_motion,label='TrajMotion', marker='o', linestyle='-')
percent_rots = error_dict['ave_rot'][idx]/rots_traj_motion
x = np.arange(0, len(percent_rots))
for i in range(len(x)):
axes[0].annotate(f'{percent_rots[i]:.2f}',
xy=(x[i], error_dict['ave_rot'][idx][i]),
xytext=(2, 0), # points vertical offset
textcoords="offset points",
ha='left', va='center',fontsize=8)
trans_traj_motion = loaded_data[name]['ARC_trans_MSE']
trans_traj_motion = np.array([trans_traj_motion]* len(error_dict['ave_trans'][idx]))
axes[1].plot(trans_traj_motion,label='TrajMotion', marker='o', linestyle='-')
percent_trans = error_dict['ave_trans'][idx]/trans_traj_motion
x = np.arange(0, len(percent_trans))
for i in range(len(x)):
axes[1].annotate(f'{percent_trans[i]:.2f}',
xy=(x[i], error_dict['ave_trans'][idx][i]),
xytext=(2, 0), # points vertical offset
textcoords="offset points",
ha='left', va='center',fontsize=8)
axes[0].set_title(name)
axes[1].yaxis.set_major_formatter(formatter)
axes[0].set_ylabel('Rotation/°')
axes[1].set_ylabel('Translation/Å')
else:
axes[0,idx].plot(error_dict['ave_rot'][idx],label='Pred', marker='o', linestyle='-')
axes[0,idx].plot(error_dict['first_rot'][idx],label='RefAsPred', marker='o', linestyle='-')
# x = np.arange(1, len(error_dict['first_rot'][idx]))
# axes[0,idx].plot(x,error_dict['time_rot_dif'][idx][1:],label='RM', marker='o', linestyle='-')
axes[1,idx].plot(error_dict['ave_trans'][idx],label='Pred', marker='o', linestyle='-')
axes[1,idx].plot(error_dict['first_trans'][idx],label='RefAsPred', marker='o', linestyle='-')
# axes[1,idx].plot(x,error_dict['time_trans_dif'][idx][1:],label='RM', marker='o', linestyle='-')
if loaded_data is not None:
rots_traj_motion = loaded_data[name]['ARC_rot']
rots_traj_motion = np.array([rots_traj_motion]* len(error_dict['ave_rot'][idx]))
axes[0,idx].plot(rots_traj_motion,label='TrajMotion', marker='o', linestyle='-')
percent_rots = error_dict['ave_rot'][idx]/rots_traj_motion
x = np.arange(0, len(percent_rots))
for i in range(len(x)):
axes[0,idx].annotate(f'{percent_rots[i]:.2f}',
xy=(x[i], error_dict['ave_rot'][idx][i]),
xytext=(2, 0), # points vertical offset
textcoords="offset points",
ha='left', va='center',fontsize=8)
trans_traj_motion = loaded_data[name]['ARC_trans_MSE']
trans_traj_motion = np.array([trans_traj_motion]* len(error_dict['ave_trans'][idx]))
axes[1,idx].plot(trans_traj_motion,label='TrajMotion', marker='o', linestyle='-')
percent_trans = error_dict['ave_trans'][idx]/trans_traj_motion
x = np.arange(0, len(percent_trans))
for i in range(len(x)):
axes[1,idx].annotate(f'{percent_trans[i]:.2f}',
xy=(x[i], error_dict['ave_trans'][idx][i]),
xytext=(2, 0), # points vertical offset
textcoords="offset points",
ha='left', va='center',fontsize=8)
axes[0, idx].set_title(name)
axes[1,idx].yaxis.set_major_formatter(formatter)
if idx==0:
axes[0,idx].set_ylabel('Rotation/°')
axes[1,idx].set_ylabel('Translation/Å')
plt.tight_layout()
plt.legend()
plt.savefig(f'{save_path}/rot_trans_error.png')
return fig
@torch.no_grad()
def plot_curve(metric_dict,save_path,row_num=2,col_num=5,suffer_fix=None):
fig, axes = plt.subplots(row_num, col_num, figsize=(16, 12),dpi=300)
data = metric_dict
# 输出每一行的信息
for index, row in data.iterrows():
name = row['pdb_name']
row_id = index//col_num
col_id = index% col_num
axes[row_id, col_id].plot(row['rmsd_all'],label='RMSE')
axes[row_id, col_id].plot(row['rmsd_ca_aligned'],label='RMSD_ca_a')
axes[row_id, col_id].plot(row['rmsd_ca'],label='RMSE_ca')
axes[row_id, col_id].set_title(name)
# 在每个子图上显示图例
plt.suptitle('RSME over Atoms')
plt.tight_layout()
plt.legend()
# plt.axis('off')
if suffer_fix is not None:
plt.savefig(f'{save_path}/rmse_rmsd_{suffer_fix}.png')
else:
plt.savefig(f'{save_path}/rmse_rmsd.png')
return fig
@torch.no_grad()
def calculate_rmsf(pdb_file, reference_select="protein and name CA"):
u = mda.Universe(pdb_file)
atoms = u.select_atoms(reference_select)
aligner = align.AlignTraj(u, atoms, select=reference_select, in_memory=True).run()
atoms = u.select_atoms(reference_select)
rmsf_analysis = rms.RMSF(atoms).run()
return rmsf_analysis.rmsf
class Experiment:
def __init__(
self,
*,
conf: DictConfig,
):
"""Initialize experiment.
Args:
exp_cfg: Experiment configuration.
"""
self._log = logging.getLogger(__name__)
self._available_gpus = ''.join([str(x) for x in GPUtil.getAvailable(order='memory', limit = 8)])
# Configs
self._conf = conf
self._exp_conf = conf.experiment
if HydraConfig.initialized() and 'num' in HydraConfig.get().job:
self._exp_conf.name = (f'{self._exp_conf.name}_{HydraConfig.get().job.num}')
self._diff_conf = conf.diffuser
self._model_conf = conf.model
self._data_conf = conf.data
self._use_tensorboard = self._exp_conf.use_tensorboard
self._use_ddp = self._exp_conf.use_ddp
self.dt_string = datetime.now().strftime("%dD_%mM_%YY_%Hh_%Mm_%Ss")
# 1. initialize ddp info if in ddp mode
# 2. silent rest of logger when use ddp mode
# 3. silent wandb logger
# 4. unset checkpoint path if rank is not 0 to avoid saving checkpoints and evaluation
if self._use_ddp :
torch.cuda.set_device(int(os.environ["LOCAL_RANK"]))
dist.init_process_group(backend='nccl')
self.ddp_info = eu.get_ddp_info()
if self.ddp_info['rank'] not in [0,-1]:
self._log.addHandler(logging.NullHandler())
self._log.setLevel("ERROR")
self._use_tensorboard = False
self._exp_conf.ckpt_dir = None
self.trained_epochs = 0
self.trained_steps = 0
# Initialize experiment objects
self._diffuser = se3_diffuser.SE3Diffuser(self._diff_conf)
self._model = Dfold_network_dynamic.FullScoreNetwork(self._model_conf, self.diffuser)
if conf.experiment.warm_start:
ckpt_path = conf.experiment.warm_start
self.load_pretrianed_model(ckpt_path=ckpt_path)
# print(next(self._model.parameters()).device)
num_parameters = sum(p.numel() for p in self._model.parameters())
if self._conf.model.ipa.temporal and self._conf.model.ipa.frozen_spatial:
self._log.info('Frozen model and only train temporal module')
# only train motion module
for param in self._model.parameters():
param.requires_grad = False
for name, param in self._model.named_parameters():
if 'temporal' in name: # 'frame'
param.requires_grad = True
# 冻结/解冻后计算总参数数量（应与初始值相同）
trainable_num_parameters = sum(p.numel() for p in self._model.parameters() if p.requires_grad)
self._exp_conf.num_parameters = num_parameters
self._exp_conf.trainable_num_parameters = num_parameters
self._log.info(f'Number of model parameters {num_parameters}, trainable parameters:{trainable_num_parameters}')
self._optimizer = torch.optim.Adam(self._model.parameters(), lr=self._exp_conf.learning_rate,amsgrad=True)#,weight_decay=1e-3
self._init_log()
self._init_best_eval()
if not self.conf.experiment.training:
seed = 0
else:
seed = dist.get_rank()
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
random.seed(seed)
np.random.seed(seed)
def _init_best_eval(self):
self.best_trained_steps = 0
self.best_trained_epoch = 0
self.best_rmsd_ca = 10000
self.best_rmsd_all = 10000
self.best_drmsd = 10000
self.best_rmsd_ca_aligned = 10000
self.best_rot_error=1000
self.best_trans_error = 1000
self.best_ref_rot_error = 1000
self.best_ref_trans_error = 1000
def _init_log(self):
if self._exp_conf.ckpt_dir is not None:
# Set-up checkpoint location
ckpt_dir = os.path.join(
self._exp_conf.ckpt_dir,
self._exp_conf.name,
self.dt_string )
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir, exist_ok=True)
self._exp_conf.ckpt_dir = ckpt_dir
self._log.info(f'Checkpoints saved to: {ckpt_dir}')
else:
self._log.info('Checkpoint not being saved.')
if self._exp_conf.eval_dir is not None :
eval_dir = os.path.join(
self._exp_conf.eval_dir,
self._exp_conf.name,
self.dt_string )
self._exp_conf.eval_dir = eval_dir
self._log.info(f'Evaluation saved to: {eval_dir}')
else:
self._exp_conf.eval_dir = os.devnull
self._log.info(f'Evaluation will not be saved.')
# self._aux_data_history = deque(maxlen=100)
def load_pretrianed_model(self, ckpt_path):
try:
self._log.info(f'Loading checkpoint from {ckpt_path}')
ckpt_pkl = torch.load(ckpt_path, map_location='cpu')
if ckpt_pkl is not None and 'model' in ckpt_pkl:
ckpt_model = ckpt_pkl['model']
# if 'epoch' in ckpt_pkl:
# self.trained_epochs = ckpt_pkl['epoch']
# if 'step' in ckpt_pkl:
# self.trained_steps = ckpt_pkl['step']
if ckpt_model is not None:
ckpt_model = {k.replace('module.', ''): v for k, v in ckpt_model.items()}
model_state_dict = self._model.state_dict()
# pretrained_dict = {k: v for k, v in ckpt_model.items() if k in model_state_dict}
pretrained_dict = {k: v for k, v in ckpt_model.items() if k in model_state_dict and v.shape == model_state_dict[k].shape}
model_state_dict.update(pretrained_dict)
self._model.load_state_dict(model_state_dict)
self._log.info(f'Warm starting from: {ckpt_path}')
del ckpt_pkl,ckpt_model,pretrained_dict,model_state_dict
gc.collect()
return True
else:
self._log.error("Checkpoint model is None.")
return False
else:
self._log.error("Checkpoint or model not found in checkpoint file.")
return False
except Exception as e:
self._log.error(f"Error loading checkpoint: {e}")
return False
@property
def diffuser(self):
return self._diffuser
@property
def model(self):
return self._model
@property
def conf(self):
return self._conf
def create_dataset(self):
if self._data_conf.is_extrapolation:
train_dataset = Dfold_data_loader_dynamic.PdbDatasetExtrapolation(
data_conf=self._data_conf,
diffuser=self._diffuser,
is_training=True
)
valid_dataset = Dfold_data_loader_dynamic.PdbDatasetExtrapolation(
data_conf=self._data_conf,
diffuser=self._diffuser,
is_training=False
)
else:
# Datasets
train_dataset = Dfold_data_loader_dynamic.PdbDataset(
data_conf=self._data_conf,
diffuser=self._diffuser,
is_training=True
)
valid_dataset = Dfold_data_loader_dynamic.PdbDataset(
data_conf=self._data_conf,
diffuser=self._diffuser,
is_training=False
)
# Loaders
num_workers = self._exp_conf.num_loader_workers
persistent_workers = True if num_workers > 0 else False
prefetch_factor=2
prefetch_factor = 2 if num_workers == 0 else prefetch_factor
sampler = torch.utils.data.distributed.DistributedSampler(train_dataset, num_replicas=dist.get_world_size(), rank=dist.get_rank())
train_loader = data.DataLoader(
train_dataset,
batch_size=self._exp_conf.batch_size if not self._exp_conf.use_ddp else self._exp_conf.batch_size // self.ddp_info['world_size'],
shuffle=False,
num_workers=num_workers,
prefetch_factor=prefetch_factor,
persistent_workers=persistent_workers,
drop_last=False,
sampler=sampler,
multiprocessing_context='fork' if num_workers != 0 else None,
timeout=60000,
)
valid_loader = data.DataLoader(
valid_dataset,
batch_size=self._exp_conf.eval_batch_size,
shuffle=False,
num_workers=num_workers,
prefetch_factor=prefetch_factor,
persistent_workers=persistent_workers,
drop_last=False,
multiprocessing_context='fork' if num_workers != 0 else None,
)
return train_loader, valid_loader
def init_tensorboard(self):
self._log.info('Initializing TensorBoard.')
conf_dict = OmegaConf.to_container(self._conf, resolve=True)
# Initialize TensorBoard SummaryWriter
tensorboard_log_dir = os.path.join(self._exp_conf.tensorboard_dir,self._exp_conf.name, self.dt_string ,self._exp_conf.name)
os.makedirs(tensorboard_log_dir, exist_ok=True)
self.writer = SummaryWriter(log_dir=tensorboard_log_dir)
# Log configuration details
self.writer.add_text('Config', str(conf_dict))
# You can also log other initial details if needed
self._exp_conf.run_id = 'unique_run_id' # You may need to generate this appropriately
self._log.info(f'TensorBoard: run_id={self._exp_conf.run_id}, log_dir={tensorboard_log_dir}')
def start_training(self, return_logs=False):
# Set environment variables for which GPUs to use.
if HydraConfig.initialized() and 'num' in HydraConfig.get().job:
replica_id = int(HydraConfig.get().job.num)
else:
replica_id = 0
if self._use_tensorboard and replica_id == 0:
self.init_tensorboard()
assert(not self._exp_conf.use_ddp or self._exp_conf.use_gpu)
# GPU mode
if torch.cuda.is_available() and self._exp_conf.use_gpu:
# single GPU mode
if self._exp_conf.num_gpus==1 :
gpu_id = self._available_gpus[replica_id]
device = f"cuda:{gpu_id}"
self._model = self.model.to(device)
self._log.info(f"Using device: {device}")
#muti gpu mode
elif self._exp_conf.num_gpus > 1:
device_ids = [f"cuda:{i}" for i in self._available_gpus[:self._exp_conf.num_gpus]]
#DDP mode
if self._use_ddp :
device = torch.device("cuda",self.ddp_info['local_rank'])
model = self.model.to(device)
self._model = DDP(model, device_ids=[self.ddp_info['local_rank']], output_device=self.ddp_info['local_rank'],find_unused_parameters=True)
self._log.info(f"Multi-GPU training on GPUs in DDP mode, node_id : {self.ddp_info['node_id']}, devices: {device_ids}")
#DP mode
else:
if len(self._available_gpus) < self._exp_conf.num_gpus:
raise ValueError(f"require {self._exp_conf.num_gpus} GPUs, but only {len(self._available_gpus)} GPUs available ")
self._log.info(f"Multi-GPU training on GPUs in DP mode: {device_ids}")
gpu_id = self._available_gpus[replica_id]
device = f"cuda:{gpu_id}"
self._model = DP(self._model, device_ids=device_ids)
self._model = self.model.to(device)
else:
device = 'cpu'
self._model = self.model.to(device)
self._log.info(f"Using device: {device}")
# if self.conf.experiment.warm_start:
# for state in self._optimizer.state.values():
# for k, v in state.items():
# if torch.is_tensor(v):
# state[k] = v.to(device)
self._model.train()
(train_loader,valid_loader) = self.create_dataset()
logs = []
# torch.cuda.empty_cache()
for epoch in range(self.trained_epochs, self._exp_conf.num_epoch):
self.trained_epochs = epoch
train_loader.sampler.set_epoch(epoch)
epoch_log = self.train_epoch(
train_loader,
valid_loader,
device,
return_logs=return_logs
)
# self._schedule.step()
if return_logs:
logs.append(epoch_log)
self._log.info('Done')
return logs
def update_fn(self, data):
"""Updates the state using some data and returns metrics."""
self._optimizer.zero_grad()
loss, aux_data = self.loss_fn(data)
loss.backward()
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
self._optimizer.step()
return loss, aux_data
def train_epoch(self, train_loader, valid_loader, device,return_logs=False):
log_lossses = defaultdict(list)
global_logs = []
log_time = time.time()
step_time = time.time()
for train_feats in train_loader:
self.model.train()
train_feats = tree.map_structure(lambda x: x.to(device), train_feats)
# TODO flatten the dim of batch and frame_time
for k in train_feats.keys():
v = train_feats[k]
if len(v.shape)>1:
reshaped_tensor = torch.flatten(v, start_dim=0, end_dim=1)#torch.reshape(v, (tensor.size(0), -1)) # 在这里修改形状，这里示例将张量展平为二维
train_feats[k] = reshaped_tensor
loss, aux_data = self.update_fn(train_feats)
if return_logs:
global_logs.append(loss)
for k,v in aux_data.items():
log_lossses[k].append(du.move_to_np(v))
self.trained_steps += 1
if self.trained_steps == 1 or self.trained_steps % self._exp_conf.log_freq == 0:
elapsed_time = time.time() - log_time
log_time = time.time()
step_per_sec = self._exp_conf.log_freq / elapsed_time
rolling_losses = tree.map_structure(np.mean, log_lossses)
loss_log = ' '.join([
f'{k}={v[0]:.4f}'
for k,v in rolling_losses.items() if 'batch' not in k
])
self._log.info(f'Epoch[{self.trained_epochs}/{self._exp_conf.num_epoch}] trained_steps: [{self.trained_steps}]: {loss_log}, steps/sec={step_per_sec:.5f}')
log_lossses = defaultdict(list)
# Take checkpoint
if False:
ckpt_path = os.path.join(self._exp_conf.ckpt_dir, f'step_{self.trained_steps}.pth')
du.write_checkpoint(
ckpt_path,
copy.deepcopy(self.model.state_dict()),
self._conf,
copy.deepcopy(self._optimizer.state_dict()),
self.trained_epochs,
self.trained_steps,
logger=self._log,
use_torch=True
)
# Run evaluation
self._log.info(f'Running evaluation of {ckpt_path}')
start_time = time.time()
eval_dir = os.path.join(self._exp_conf.eval_dir, f'step_{self.trained_steps}')
os.makedirs(eval_dir, exist_ok=True)
ckpt_metrics,curve_fig,curve_fig_aligned,error_fig,model_ckpt_update,rot_trans_error_mean = self.eval_fn(
eval_dir, valid_loader, device,
noise_scale=self._exp_conf.noise_scale
)
eval_time = time.time() - start_time
self._log.info(f'Finished evaluation in {eval_time:.2f}s')
else:
ckpt_metrics = None
eval_time = None
model_ckpt_update=False
# Remote log to Wandb.
if self._exp_conf.ckpt_dir is not None and self.trained_epochs!=0 and self.trained_epochs % self._exp_conf.ckpt_freq ==0:
old_best_rmsd = self.best_rmsd_all
start_time = time.time()
eval_dir = os.path.join(self._exp_conf.eval_dir, f'step_{self.trained_steps}')
os.makedirs(eval_dir, exist_ok=True)
ckpt_path = os.path.join(
self._exp_conf.ckpt_dir, f'step_{self.trained_steps}.pth')
du.write_checkpoint(
ckpt_path,
copy.deepcopy(self.model.state_dict()),
self._conf,
copy.deepcopy(self._optimizer.state_dict()),
self.trained_epochs,
self.trained_steps,
logger=self._log,
use_torch=True
)
eval_time = time.time() - start_time
self._log.info(f'Finished evaluation in {eval_time:.2f}s')
if return_logs:
return global_logs
def eval_fn(self, eval_dir, valid_loader, device, min_t=None, num_t=None, noise_scale=1.0,is_training=True):
# initial some metrics and base save path
# diff_first_metric_list=[]
metric_list = []
metric_all_list = []
metric_aligned_list = []
metric_aligned_all_list = []
first_frame_all_list = []
save_name_list = []
rot_trans_error_dict = {'name':[],"ave_rot":[],"ave_trans":[],'all_atom_mae':[],'all_atom_mse':[],'all_atom_rmsd':[]}
sample_root_path = os.path.join(eval_dir,'sample')
gt_root_path = os.path.join(eval_dir,'gt')
sample_frame_root_path = os.path.join(eval_dir,'frame','sample')
gt_frame_root_path = os.path.join(eval_dir,'frame','gt')
prot_traj_root_path = os.path.join(eval_dir,'traj')
rigids_path = os.path.join(eval_dir,'rigids')
if not os.path.exists(rigids_path):
os.makedirs(rigids_path,exist_ok=True)
if not is_training:
pred_numpy_path = os.path.join(eval_dir,'pred_npz')
if not os.path.exists(pred_numpy_path):
os.makedirs(pred_numpy_path,exist_ok=True)
if not os.path.exists(prot_traj_root_path):
os.makedirs(prot_traj_root_path,exist_ok=True)
if not os.path.exists(sample_root_path):
os.makedirs(sample_root_path,exist_ok=True)
if not os.path.exists(gt_root_path):
os.makedirs(gt_root_path,exist_ok=True)
if not os.path.exists(sample_frame_root_path):
os.makedirs(sample_frame_root_path,exist_ok=True)
if not os.path.exists(gt_frame_root_path):
os.makedirs(gt_frame_root_path,exist_ok=True)
# ergodic the validation
idx = 0
for valid_feats, pdb_names in valid_loader:
idx = idx + 1
# initialize input data
sample_length = valid_feats['aatype'].shape[-1]
frame_time = valid_feats['aatype'].shape[1]
res_mask = np.ones((frame_time,sample_length))
fixed_mask = np.zeros_like(res_mask)
res_idx = torch.arange(1, sample_length+1).unsqueeze(0).repeat(frame_time,1)
ref_sample = self.diffuser.sample_ref(
n_samples=sample_length*frame_time,
as_tensor_7=True,
)
ref_sample = tree.map_structure(lambda x: x[None].to(device), ref_sample)
init_feats = {
'res_mask': res_mask[None],
'seq_idx': res_idx[None],
'fixed_mask': fixed_mask[None],
#'torsion_angles_sin_cos': np.zeros((sample_length, 7, 2))[None],
'torsion_angles_sin_cos':valid_feats['torsion_angles_sin_cos'],
'torsion_angles_mask':valid_feats['torsion_angles_mask'],
'sc_ca_t': np.zeros((frame_time,sample_length, 3))[None],
'node_repr':valid_feats['node_repr'],
'edge_repr':valid_feats['edge_repr'],
'aatype':valid_feats['aatype'],
**ref_sample,
'rigids_0':valid_feats['rigids_0'], #TODO
'atom37_pos':valid_feats['atom37_pos'],
'atom37_mask':valid_feats['atom37_mask'],
'force':valid_feats['force'],
'vel':valid_feats['vel']
# 'rigids_t': diff_rigids_t[None].to(device) # rigids_t based on gt
}
# TODO here
# fasta_aatype = du.move_to_np(valid_feats['aatype'])[0] # remove the batch(1,...) to (...),conver form [1,N] to [N,]
init_feats = tree.map_structure(lambda x: x if torch.is_tensor(x) else torch.tensor(x), init_feats)
init_feats = tree.map_structure(lambda x: x.to(device), init_feats)
# TODO flatten the dim of batch and frame_time
for k in init_feats.keys():
v = init_feats[k]
if len(v.shape)>1:
reshaped_tensor = torch.flatten(v, start_dim=0, end_dim=1)#torch.reshape(v, (tensor.size(0), -1)) # 在这里修改形状，这里示例将张量展平为二维
init_feats[k] = reshaped_tensor
# start inference
# start_time = time.time()
sample_out = self.inference_fn(
init_feats,
num_t=num_t,
min_t=min_t,
aux_traj=True,
noise_scale=noise_scale,
)
# 设置第一帧
# align_sample_rigids_list = []
sample_rigids = sample_out['rigid_traj'][0]
# save the predication
save_name = pdb_names[0].split('.')[0]
sample_path = os.path.join(sample_root_path, f'{save_name}.pdb')
gt_path = os.path.join(gt_root_path, f'{save_name}_gt.pdb')
diffuse_mask = np.ones(sample_length)
b_factors = np.tile((diffuse_mask * 100)[:, None], (1, 37))
save_name_list.append(save_name)
# diff_first_metric_list.append({k: compare_to_first_dict[k].cpu().numpy() if torch.is_tensor(compare_to_first_dict[k]) else compare_to_first_dict[k] for k in compare_to_first_dict})
all_atom37_pos = valid_feats['atom37_pos'][0].cpu().numpy()
# caculate the rotation and translation error
ave_quat, ave_trans, ref_ave_quat, ref_ave_trans = self._calc_rot_trans_error(sample_out['rigid_traj'][0],gt_rigids=init_feats['rigids_0'].cpu().numpy())
rot_trans_error_dict['name'].append(save_name)
rot_trans_error_dict['ave_rot'].append(ave_quat[-1:])
rot_trans_error_dict['ave_trans'].append(ave_trans[-1:])
atom_mask = valid_feats['atom37_mask'][0].unsqueeze(-1).expand(-1,-1,-1,3).cpu().numpy()
atom_mask[-1, -1, -1] = 0.0
atom_gt = valid_feats['atom37_pos'][0].cpu().numpy()
atom_mae = (np.abs(atom_gt - sample_out['prot_traj'][0]) * atom_mask).sum(axis=(-1,-2,-3)) / atom_mask.sum(axis=(-1,-2,-3))
rot_trans_error_dict['all_atom_mae'].append(atom_mae[-1:])
atom_all_gt = atom_gt[-1].reshape(atom_gt.shape[1]*atom_gt.shape[2], 3)
atom_all_pred = sample_out['prot_traj'][0, -1].reshape(atom_gt.shape[1]*atom_gt.shape[2], 3)
atom_all_mask = atom_mask[-1].reshape(atom_gt.shape[1]*atom_gt.shape[2], 3)[:, 0]
atom_all_index = np.where(atom_all_mask>0)[0]
sup = SVDSuperimposer()
sup.set(atom_all_pred[atom_all_index], atom_all_gt[atom_all_index])
sup.run()
sup.get_transformed()
rmsd = sup.get_rms()
rot_trans_error_dict['all_atom_rmsd'].append(np.array([rmsd]))
atom_mse = ((valid_feats['atom37_pos'][0].cpu().numpy() - sample_out['prot_traj'][0])**2 * atom_mask).sum(axis=(-1,-2,-3)) / atom_mask.sum(axis=(-1,-2,-3))
rot_trans_error_dict['all_atom_mse'].append(atom_mse[-1:])
rot_trans_error_dict['ave_rot'] = np.concatenate(rot_trans_error_dict['ave_rot'], axis=0)
rot_trans_error_dict['ave_trans'] = np.concatenate(rot_trans_error_dict['ave_trans'], axis=0)
rot_trans_error_dict['all_atom_mae'] = np.concatenate(rot_trans_error_dict['all_atom_mae'], axis=0)
rot_trans_error_dict['all_atom_rmsd'] = np.concatenate(rot_trans_error_dict['all_atom_rmsd'], axis=0)
mask = rot_trans_error_dict['ave_trans'] < 10000.0
print('ave_rot:', rot_trans_error_dict['ave_rot'].sum() / mask.sum())
print('ave_trans:', rot_trans_error_dict['ave_trans'].sum() / mask.sum())
print('ave_atom_mae:', rot_trans_error_dict['all_atom_mae'].sum() / mask.sum())
print('ave_atom_rmsd:', rot_trans_error_dict['all_atom_rmsd'].sum() / mask.sum())
#print('ave_atom_rmsd_median:', np.median(rot_trans_error_dict['all_atom_rmsd']))
# should reture eval_dict
return rot_trans_error_dict
def eval_extension(self, eval_dir, valid_loader, device, min_t=None, num_t=None, noise_scale=1.0,is_training=True):
# ergodic the validation
length = 200
atom_traj = []
rigid_traj = []
for valid_feats, pdb_names in valid_loader:
for j in range(length):
if j % 100 == 0:
print(f'Finish Setp {j}',end='\r')
sample_length = valid_feats['aatype'].shape[-1]
frame_time = valid_feats['aatype'].shape[1]
res_mask = np.ones((frame_time,sample_length))
fixed_mask = np.zeros_like(res_mask)
res_idx = torch.arange(1, sample_length+1).unsqueeze(0).repeat(frame_time,1)
ref_sample = self.diffuser.sample_ref(
n_samples=sample_length*frame_time,
as_tensor_7=True,
)
ref_sample = tree.map_structure(lambda x: x[None].to(device), ref_sample)
init_feats = {
'res_mask': res_mask[None],
'seq_idx': res_idx[None],
'fixed_mask': fixed_mask[None],
'torsion_angles_sin_cos': np.zeros((sample_length, 7, 2))[None],
'sc_ca_t': np.zeros((frame_time,sample_length, 3))[None],
'node_repr':valid_feats['node_repr'],
'edge_repr':valid_feats['edge_repr'],
'aatype':valid_feats['aatype'],
**ref_sample,
'rigids_0':valid_feats['rigids_0'] #TODO
# 'rigids_t': diff_rigids_t[None].to(device) # rigids_t based on gt
}
# TODO here
# fasta_aatype = du.move_to_np(valid_feats['aatype'])[0] # remove the batch(1,...) to (...),conver form [1,N] to [N,]
init_feats = tree.map_structure(lambda x: x if torch.is_tensor(x) else torch.tensor(x), init_feats)
init_feats = tree.map_structure(lambda x: x.to(device), init_feats)
# TODO flatten the dim of batch and frame_time
for k in init_feats.keys():
v = init_feats[k]
if len(v.shape)>1:
reshaped_tensor = torch.flatten(v, start_dim=0, end_dim=1)#torch.reshape(v, (tensor.size(0), -1)) # 在这里修改形状，这里示例将张量展平为二维
init_feats[k] = reshaped_tensor
# start inference
# start_time = time.time()
sample_out = self.inference_fn(
init_feats,
num_t=num_t,
min_t=min_t,
aux_traj=True,
noise_scale=noise_scale,
)
atom_pred = sample_out['prot_traj'][0]
rigid_pred = sample_out['rigid_traj'][0]
atom_traj.append(atom_pred[-1:]) # concate the last frame
rigid_traj.append(rigid_pred[-1:])
valid_feats['rigids_0'] = torch.from_numpy(np.concatenate([rigid_pred[1:], rigid_pred[-1:]], axis=0)).unsqueeze(0).to(valid_feats['rigids_0'].device).to(valid_feats['rigids_0'].dtype)
atom_traj = np.concatenate(atom_traj, axis=0)
rigid_traj = np.concatenate(rigid_traj, axis=0)
save_path = os.path.join(eval_dir,'extension.npz')
np.savez_compressed(save_path, atom_traj=atom_traj, rigid_traj=rigid_traj, aatype=valid_feats['aatype'])