-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_qm9.py
215 lines (179 loc) · 8.74 KB
/
train_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
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
"""
script to train on QM9 targets.
"""
import torch
import torch.nn as nn
from torch import Tensor
from datasets.QM9Dataset import QM9, conversion
from models.input_encoder import QM9InputEncoder, EmbeddingEncoder
import train_utils
from interfaces.pl_model_interface import PlGNNTestonValModule
from interfaces.pl_data_interface import PlPyGDataTestonValModule
from lightning.pytorch import seed_everything
from lightning.pytorch import Trainer
from lightning.pytorch.loggers import WandbLogger
from lightning.pytorch.callbacks import ModelCheckpoint, LearningRateMonitor, Timer
from lightning.pytorch.callbacks.progress import TQDMProgressBar
import wandb
from torchmetrics import MeanAbsoluteError
from torchmetrics.functional.regression.mae import _mean_absolute_error_compute
import torch_geometric.transforms as T
from torch_geometric.data import Data
class InputTransform(object):
"""QM9 input feature transformation. Concatenate x and z together.
"""
def __init__(self):
super().__init__()
def __call__(self,
data: Data) -> Data:
x = data.x
z = data.z
data.x = torch.cat([z.unsqueeze(-1), x], dim=-1)
data.edge_attr = torch.where(data.edge_attr == 1)[-1]
return data
class MeanAbsoluteErrorQM9(MeanAbsoluteError):
def __init__(self,
std,
conversion,
**kwargs):
super().__init__(**kwargs)
self.std = std
self.conversion = conversion
def compute(self) -> Tensor:
return (_mean_absolute_error_compute(self.sum_abs_error, self.total) * self.std) / self.conversion
def main():
parser = train_utils.args_setup()
parser.add_argument('--dataset_name', type=str, default="QM9", help='Name of dataset.')
parser.add_argument('--task', type=int, default=11, choices=list(range(19)), help='Train target.')
parser.add_argument('--search', action="store_true", help="If true, run all first 12 targets.")
args = parser.parse_args()
args = train_utils.update_args(args, add_task=False)
path, pre_transform, follow_batch = train_utils.data_setup(args)
if args.search:
for target in range(12):
args.task = target
dataset = QM9(path,
pre_transform=T.Compose([InputTransform(), pre_transform]),
transform=train_utils.PostTransform(args.wo_node_feature,
args.wo_edge_feature,
args.task
))
dataset = dataset.shuffle()
tenprecent = int(len(dataset) * 0.1)
mean = dataset.data.y[tenprecent:].mean(dim=0)
std = dataset.data.y[tenprecent:].std(dim=0)
dataset.data.y = (dataset.data.y - mean) / std
train_dataset = dataset[2 * tenprecent:]
test_dataset = dataset[:tenprecent]
val_dataset = dataset[tenprecent:2 * tenprecent]
logger = WandbLogger(name=f'target_{str(args.task + 1)}',
project=args.exp_name,
save_dir=args.save_dir,
offline=args.offline)
logger.log_hyperparams(args)
timer = Timer(duration=dict(weeks=4))
# Set random seed
seed = train_utils.get_seed(args.seed)
seed_everything(seed)
datamodule = PlPyGDataTestonValModule(train_dataset=train_dataset,
val_dataset=val_dataset,
test_dataset=test_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers,
follow_batch=follow_batch)
loss_cri = nn.MSELoss()
evaluator = MeanAbsoluteErrorQM9(std[args.task].item(), conversion[args.task].item())
init_encoder = QM9InputEncoder(args.hidden_channels)
edge_encoder = EmbeddingEncoder(4, args.inner_channels)
modelmodule = PlGNNTestonValModule(loss_criterion=loss_cri,
evaluator=evaluator,
args=args,
init_encoder=init_encoder,
edge_encoder=edge_encoder)
trainer = Trainer(
accelerator="auto",
devices="auto",
max_epochs=args.num_epochs,
enable_checkpointing=True,
enable_progress_bar=True,
logger=logger,
callbacks=[
TQDMProgressBar(refresh_rate=20),
ModelCheckpoint(monitor="val/metric", mode="min"),
LearningRateMonitor(logging_interval="epoch"),
timer
]
)
trainer.fit(modelmodule, datamodule=datamodule)
val_result, test_result = trainer.test(modelmodule, datamodule=datamodule, ckpt_path="best")
results = {"final/best_val_metric": val_result["val/metric"],
"final/best_test_metric": test_result["test/metric"],
"final/avg_train_time_epoch": timer.time_elapsed("train") / args.num_epochs,
}
logger.log_metrics(results)
wandb.finish()
else:
dataset = QM9(path,
pre_transform=T.Compose([InputTransform(), pre_transform]),
transform=train_utils.PostTransform(args.wo_node_feature,
args.wo_edge_feature,
args.task
))
dataset = dataset.shuffle()
tenprecent = int(len(dataset) * 0.1)
mean = dataset.data.y[tenprecent:].mean(dim=0)
std = dataset.data.y[tenprecent:].std(dim=0)
dataset.data.y = (dataset.data.y - mean) / std
train_dataset = dataset[2 * tenprecent:]
test_dataset = dataset[:tenprecent]
val_dataset = dataset[tenprecent:2 * tenprecent]
logger = WandbLogger(name=f'target_{str(args.task + 1)}',
project=args.exp_name,
save_dir=args.save_dir,
offline=args.offline)
logger.log_hyperparams(args)
timer = Timer(duration=dict(weeks=4))
# Set random seed
seed = train_utils.get_seed(args.seed)
seed_everything(seed)
datamodule = PlPyGDataTestonValModule(train_dataset=train_dataset,
val_dataset=val_dataset,
test_dataset=test_dataset,
batch_size=args.batch_size,
num_workers=args.num_workers,
follow_batch=follow_batch)
loss_cri = nn.MSELoss()
evaluator = MeanAbsoluteErrorQM9(std[args.task].item(), conversion[args.task].item())
args.mode = "min"
init_encoder = QM9InputEncoder(args.hidden_channels)
edge_encoder = EmbeddingEncoder(4, args.inner_channels)
modelmodule = PlGNNTestonValModule(loss_criterion=loss_cri,
evaluator=evaluator,
args=args,
init_encoder=init_encoder,
edge_encoder=edge_encoder)
trainer = Trainer(
accelerator="auto",
devices="auto",
max_epochs=args.num_epochs,
enable_checkpointing=True,
enable_progress_bar=True,
logger=logger,
callbacks=[
TQDMProgressBar(refresh_rate=5),
ModelCheckpoint(monitor="val/metric", mode=args.mode),
LearningRateMonitor(logging_interval="epoch"),
timer
]
)
trainer.fit(modelmodule, datamodule=datamodule)
val_result, test_result = trainer.test(modelmodule, datamodule=datamodule, ckpt_path="best")
results = {"final/best_val_metric": val_result["val/metric"],
"final/best_test_metric": test_result["test/metric"],
"final/avg_train_time_epoch": timer.time_elapsed("train") / args.num_epochs,
}
logger.log_metrics(results)
wandb.finish()
return
if __name__ == "__main__":
main()