MolGEN/train-Transition1x-equivariant.py at main · tuoping/MolGEN · GitHub

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
from mdgen.parsing import parse_train_args
args = parse_train_args()
from mdgen.logger import get_logger
logger = get_logger(__name__)

import torch, os
from mdgen.dataset import BucketBatchSampler
from mdgen.equivariant_wrapper import EquivariantMDGenWrapper
from pytorch_lightning.callbacks import ModelCheckpoint, ModelSummary
import pytorch_lightning as pl

class ResetLrCallback(pl.Callback):
    def __init__(self, new_lr: float):
        self.new_lr = new_lr

    # runs right after checkpoint restore, before the first batch
    def on_train_epoch_start(self, trainer, pl_module):
        for optimizer in trainer.optimizers:
            for pg in optimizer.param_groups:
                pg["lr"] = self.new_lr
        ## (optional) reset schedulers if you wish
        # scheduler = pl_module.lr_schedulers()
        # scheduler.base_lrs = [self.new_lr]
        # scheduler.last_epoch = 1499  # starts fresh


torch.set_float32_matmul_precision('medium')
from torch.utils.data import ConcatDataset
from torch.utils.data import Subset

train_dataset = torch.load(os.path.join("data/Transition1x", "tps_masked_train-fragmented_cutoffx1.5.pt"), weights_only=False)
#train_dataset = ConcatDataset([
#                                torch.load(os.path.join("data/Transition1x", "tps_masked_train-fragmented_cutoffx1.5.pt"), weights_only=False),
#                                torch.load(os.path.join("data/RGD1", "tps_masked_train.pt"), weights_only=False),
#                          ])
trainsampler = BucketBatchSampler(train_dataset, batch_size=args.batch_size)

if args.overfit:
    val_dataset = train_dataset
    valsampler = trainsampler
else:
    val_dataset = torch.load(os.path.join("data/Transition1x", "tps_masked_val-fragmented_cutoffx1.5.pt"), weights_only=False)
    valsampler = BucketBatchSampler(val_dataset, batch_size=args.batch_size, drop_last=True)

train_loader = torch.utils.data.DataLoader(
    train_dataset,
    batch_sampler=trainsampler,
    num_workers=args.num_workers,
)

val_loader = torch.utils.data.DataLoader(
    val_dataset,
    batch_sampler=valsampler,
    num_workers=args.num_workers,
)

model = EquivariantMDGenWrapper(args)


if args.weight_loss_var_x0 > 0:
    callbacks_fn = [
        # ResetLrCallback(args.lr),
        ModelCheckpoint(
            dirpath=os.environ["MODEL_DIR"],
            save_top_k=1,
            monitor="val_loss_gen",
            every_n_epochs=args.ckpt_freq,
            save_last=True,
        ),
        ModelCheckpoint(
            dirpath=os.environ["MODEL_DIR"],
            save_top_k=1,
            monitor="val_meanRMSD_Kabsch",
            every_n_epochs=args.ckpt_freq,
        ),
        ModelCheckpoint(
            dirpath=os.environ["MODEL_DIR"],
            save_top_k=1,
            monitor="val_loss_var",
            every_n_epochs=args.ckpt_freq,
        ),
        ModelSummary(max_depth=2),

    ]
else:
    callbacks_fn = [
        # ResetLrCallback(args.lr),
        ModelCheckpoint(
            dirpath=os.environ["MODEL_DIR"],
            save_top_k=1,
            monitor="val_loss_gen",
            every_n_epochs=args.ckpt_freq,
            save_last=True,
        ),
        ModelCheckpoint(
            dirpath=os.environ["MODEL_DIR"],
            filename="{epoch:03d}-{step:07d}-{val_meanRMSD_Kabsch:.4f}",
            save_top_k=1,
            monitor="val_meanRMSD_Kabsch",
            every_n_epochs=args.ckpt_freq,
        ),
        ModelSummary(max_depth=2),

    ]

trainer = pl.Trainer(
    accelerator="gpu" if torch.cuda.is_available() else 'auto',
    max_epochs=args.epochs,
    limit_train_batches=args.train_batches or 1.0,
    limit_val_batches=0.0 if args.no_validate else (args.val_batches or 1.0),
    num_sanity_val_steps=1,
    precision=args.precision,
    enable_progress_bar=not args.wandb or os.getlogin() == 'hstark',
    gradient_clip_val=args.grad_clip,
    default_root_dir=os.environ["MODEL_DIR"],
    callbacks=callbacks_fn,
    accumulate_grad_batches=args.accumulate_grad,
    val_check_interval=args.val_freq,
    check_val_every_n_epoch=args.val_epoch_freq,
    logger=False
)

# torch.manual_seed(137)
# np.random.seed(137)


if args.validate:
    trainer.validate(model, val_loader, ckpt_path=args.ckpt)
    # trainer.validate(model, val_loader)
else:
    trainer.fit(model, train_loader, val_loader, ckpt_path=args.ckpt)
    # trainer.fit(model, train_loader, val_loader)