-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathmain_qm9.py
151 lines (123 loc) · 6.43 KB
/
main_qm9.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
import argparse
import os
import numpy as np
import torch
from torch_geometric.datasets import QM9
from torch_geometric.loader import DataLoader
import pytorch_lightning as pl
from lightning_wrappers.callbacks import EMA, EpochTimer
from lightning_wrappers.qm9 import PONITA_QM9
# TODO: do we need this?
import torch.multiprocessing
torch.multiprocessing.set_sharing_strategy('file_system')
# ------------------------ Start of the main experiment script
if __name__ == "__main__":
parser = argparse.ArgumentParser()
# ------------------------ Input arguments
# Run parameters
parser.add_argument('--epochs', type=int, default=1000,
help='number of epochs')
parser.add_argument('--warmup', type=int, default=10,
help='number of epochs')
parser.add_argument('--batch_size', type=int, default=96,
help='Batch size. Does not scale with number of gpus.')
parser.add_argument('--lr', type=float, default=5e-4,
help='learning rate')
parser.add_argument('--weight_decay', type=float, default=1e-10,
help='weight decay')
parser.add_argument('--log', type=eval, default=True,
help='logging flag')
parser.add_argument('--enable_progress_bar', type=eval, default=True,
help='enable progress bar')
parser.add_argument('--num_workers', type=int, default=0,
help='Num workers in dataloader')
parser.add_argument('--seed', type=int, default=0,
help='Random seed')
# Train settings
parser.add_argument('--train_augm', type=eval, default=True,
help='whether or not to use random rotations during training')
# Test settings
parser.add_argument('--repeats', type=int, default=5,
help='number of repeated forward passes at test-time')
# QM9 Dataset
parser.add_argument('--root', type=str, default="datasets/qm9",
help='Data set location')
parser.add_argument('--target', type=str, default="alpha",
help='MD17 target')
# Graph connectivity settings
parser.add_argument('--radius', type=eval, default=1000.,
help='radius for the radius graph construction in front of the force loss')
parser.add_argument('--loop', type=eval, default=True,
help='enable self interactions')
# PONTA model settings
parser.add_argument('--num_ori', type=int, default=-1,
help='num elements of spherical grid')
parser.add_argument('--hidden_dim', type=int, default=128,
help='internal feature dimension')
parser.add_argument('--basis_dim', type=int, default=256,
help='number of basis functions')
parser.add_argument('--degree', type=int, default=3,
help='degree of the polynomial embedding')
parser.add_argument('--layers', type=int, default=5,
help='Number of message passing layers')
parser.add_argument('--widening_factor', type=int, default=4,
help='Number of message passing layers')
parser.add_argument('--layer_scale', type=float, default=0,
help='Initial layer scale factor in ConvNextBlock, 0 means do not use layer scale')
parser.add_argument('--multiple_readouts', type=eval, default=False,
help='Whether or not to readout after every layer')
# Parallel computing stuff
parser.add_argument('-g', '--gpus', default=1, type=int,
help='number of gpus to use (assumes all are on one node)')
# Arg parser
args = parser.parse_args()
# ------------------------ Device settings
if args.gpus > 0:
accelerator = "gpu"
devices = args.gpus
else:
accelerator = "cpu"
devices = "auto"
if args.num_workers == -1:
args.num_workers = os.cpu_count()
# ------------------------ Dataset
# Load the dataset and set the dataset specific settings
dataset = QM9(root=args.root)
# Create train, val, test split (same random seed and splits as DimeNet)
random_state = np.random.RandomState(seed=42)
perm = torch.from_numpy(random_state.permutation(np.arange(130831)))
train_idx, val_idx, test_idx = perm[:110000], perm[110000:120000], perm[120000:]
datasets = {'train': dataset[train_idx], 'val': dataset[val_idx], 'test': dataset[test_idx]}
# Select the right target
targets = ['mu', 'alpha', 'homo', 'lumo', 'gap', 'r2', 'zpve', 'U0',
'U', 'H', 'G', 'Cv', 'U0_atom', 'U_atom', 'H_atom', 'G_atom', 'A', 'B', 'C']
idx = torch.tensor([0, 1, 2, 3, 4, 5, 6, 12, 13, 14, 15, 11, 12, 13, 14, 15]) # We will automatically replace U0 -> U0_atom etc.
dataset.data.y = dataset.data.y[:, idx]
dataset.data.y = dataset.data.y[:, targets.index(args.target)]
# Make the dataloaders
dataloaders = {
split: DataLoader(dataset, batch_size=args.batch_size, shuffle=(split == 'train'), num_workers=args.num_workers)
for split, dataset in datasets.items()}
# ------------------------ Load and initialize the model
model = PONITA_QM9(args)
model.set_dataset_statistics(datasets['train'])
# ------------------------ Weights and Biases logger
if args.log:
logger = pl.loggers.WandbLogger(project="PONITA-QM9", name=args.target.replace(" ", "_"), config=args, save_dir='logs')
else:
logger = None
# ------------------------ Set up the trainer
# Seed
pl.seed_everything(args.seed, workers=True)
# Pytorch lightning call backs
callbacks = [EMA(0.99),
pl.callbacks.ModelCheckpoint(monitor='valid MAE', mode = 'min'),
EpochTimer()]
if args.log: callbacks.append(pl.callbacks.LearningRateMonitor(logging_interval='epoch'))
# Initialize the trainer
trainer = pl.Trainer(logger=logger, max_epochs=args.epochs, callbacks=callbacks, inference_mode=False, # Important for force computation via backprop
gradient_clip_val=0.5, accelerator=accelerator, devices=devices, enable_progress_bar=args.enable_progress_bar)
# Do the training
trainer.fit(model, dataloaders['train'], dataloaders['val'])
# And test
trainer.test(model, dataloaders['test'], ckpt_path = "best")