-
Notifications
You must be signed in to change notification settings - Fork 24
/
multitask_train.py
309 lines (272 loc) · 9.63 KB
/
multitask_train.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
import argparse
import logging
import os
import pdb
import sys
import tempfile
import traceback
from functools import partial
from typing import Optional
import torch
from pyprojroot import here as project_root
sys.path.insert(0, str(project_root()))
from fs_mol.data import (
NUM_EDGE_TYPES,
NUM_NODE_FEATURES,
DataFold,
FSMolDataset,
FSMolTaskSample,
)
from fs_mol.data.multitask import (
FSMolMultitaskBatch,
MultitaskTaskSampleBatchIterable,
get_multitask_inference_batcher,
)
from fs_mol.models.abstract_torch_fsmol_model import (
AbstractTorchFSMolModel,
MetricType,
eval_model_by_finetuning_on_task,
train_loop,
create_optimizer,
save_model,
)
from fs_mol.models.gnn_multitask import (
GNNMultitaskConfig,
GNNMultitaskModel,
GNNConfig,
create_model,
)
from fs_mol.utils.cli_utils import add_train_cli_args, set_up_train_run, str2bool
from fs_mol.utils.metrics import (
avg_metrics_over_tasks,
BinaryEvalMetrics,
)
from fs_mol.utils.test_utils import eval_model
SMALL_NUMBER = 1e-7
logger = logging.getLogger(__name__)
def validate_by_finetuning_on_tasks(
model: AbstractTorchFSMolModel[FSMolMultitaskBatch],
dataset: FSMolDataset,
learning_rate: float,
task_specific_learning_rate: float,
batch_size: int = 128,
metric_to_use: MetricType = "avg_precision",
seed: int = 0,
aml_run=None,
) -> float:
with tempfile.TemporaryDirectory() as tempdir:
# First, store the current state of the model, so that we can just load it back in
# repeatedly as starting point during finetuning:
current_model_path = os.path.join(tempdir, "cur_model.pt")
save_model(current_model_path, model)
# Move model off GPU to make space for validation model:
model_device = model.device
model = model.to(torch.device("cpu"))
def test_model_fn(
task_sample: FSMolTaskSample, temp_out_folder: str, seed: int
) -> BinaryEvalMetrics:
return eval_model_by_finetuning_on_task(
current_model_path,
model_cls=GNNMultitaskModel,
task_sample=task_sample,
temp_out_folder=temp_out_folder,
batcher=get_multitask_inference_batcher(max_num_graphs=batch_size),
learning_rate=learning_rate,
task_specific_learning_rate=task_specific_learning_rate,
metric_to_use=metric_to_use,
seed=seed,
quiet=True,
device=model_device,
)
task_to_results = eval_model(
test_model_fn=test_model_fn,
dataset=dataset,
train_set_sample_sizes=[16, 128],
num_samples=3,
valid_size_or_ratio=0.2,
test_size_or_ratio=512,
fold=DataFold.VALIDATION,
seed=seed,
)
mean_metrics = avg_metrics_over_tasks(task_to_results)
if aml_run is not None:
for metric_name, (metric_mean, _) in mean_metrics.items():
aml_run.log(f"valid_task_test_{metric_name}", float(metric_mean))
model = model.to(model_device)
return mean_metrics[metric_to_use][0]
def add_model_arguments(parser: argparse.ArgumentParser):
# GNN parameters.
parser.add_argument(
"--gnn_type",
type=str,
default="PNA",
choices=["MultiHeadAttention", "MultiAggr", "PNA", "Plain"],
help="Type of GNN architecture to use.",
)
parser.add_argument(
"--num_gnn_layers", type=int, default=10, help="Number of GNN layers to use."
)
parser.add_argument(
"--node_embed_dim", type=int, default=128, help="Size of GNN node representations."
)
parser.add_argument(
"--num_heads",
type=int,
default=4,
help="Number of heads used in each GNN message propagation step. Relevant in MultiHeadAttention.",
)
parser.add_argument(
"--per_head_dim",
type=int,
default=64,
help="Size of message representation in each attention head.",
)
parser.add_argument(
"--intermediate_dim",
type=int,
default=1024,
help="Size of intermediate representation used in BOOM layer. Set to 0 to deactivate BOOM layer.",
)
parser.add_argument("--message_function_depth", type=int, default=1)
parser.add_argument(
"--readout_type",
type=str,
default="combined",
choices=["sum", "min", "max", "mean", "weighted_sum", "weighted_mean", "combined"],
help="Readout used to summarise atoms into a molecule",
)
parser.add_argument(
"--readout_use_all_states",
type=str2bool,
default=True,
help="Indicates if all intermediate GNN activations or only the final ones should be used when computing a graph-level representation.",
)
parser.add_argument("--num_tail_layers", type=int, default=2)
def make_model_from_args(
num_tasks: int, args: argparse.Namespace, device: Optional[torch.device] = None
):
model_config = GNNMultitaskConfig(
num_tasks=num_tasks,
node_feature_dim=NUM_NODE_FEATURES,
gnn_config=GNNConfig(
type=args.gnn_type,
hidden_dim=args.node_embed_dim,
num_edge_types=NUM_EDGE_TYPES,
num_heads=args.num_heads,
per_head_dim=args.per_head_dim,
intermediate_dim=args.intermediate_dim,
message_function_depth=args.message_function_depth,
num_layers=args.num_gnn_layers,
),
num_outputs=1,
readout_type=args.readout_type,
readout_use_only_last_timestep=not args.readout_use_all_states,
num_tail_layers=args.num_tail_layers,
)
model = create_model(model_config, device=device)
return model
def add_train_loop_arguments(parser: argparse.ArgumentParser):
parser.add_argument("--batch_size", type=int, default=256)
parser.add_argument("--num_epochs", type=int, default=100)
parser.add_argument("--patience", type=int, default=10)
parser.add_argument(
"--learning-rate",
type=float,
default=0.00005,
help="Learning rate for shared model components.",
)
parser.add_argument(
"--metric-to-use",
type=str,
choices=[
"acc",
"balanced_acc",
"f1",
"prec",
"recall",
"roc_auc",
"avg_precision",
"kappa",
],
default="avg_precision",
help="Metric to evaluate on validation data.",
)
def main():
parser = argparse.ArgumentParser(
description="Train a Multitask GNN model.",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
add_train_cli_args(parser)
add_model_arguments(parser)
# Training parameters:
add_train_loop_arguments(parser)
parser.add_argument(
"--task-specific-lr",
type=float,
default=0.0001,
help="Learning rate for shared model components. By default, 10x core learning rate.",
)
parser.add_argument(
"--finetune-lr-scale",
type=float,
default=1.0,
help="Scaling factor for LRs used in finetuning eval.",
)
args = parser.parse_args()
out_dir, fsmol_dataset, aml_run = set_up_train_run("Multitask", args, torch=True)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = make_model_from_args(
num_tasks=fsmol_dataset.get_num_fold_tasks(DataFold.TRAIN), args=args, device=device
)
logger.info(f"\tNum parameters {sum(p.numel() for p in model.parameters())}")
logger.info(f"\tDevice: {device}")
logger.info(f"\tModel:\n{model}")
train_task_name_to_id = {
name: i for i, name in enumerate(fsmol_dataset.get_task_names(data_fold=DataFold.TRAIN))
}
if args.task_specific_lr is not None:
task_specific_lr = args.task_specific_lr
else:
task_specific_lr = 10 * args.learning_rate
optimizer, lr_scheduler = create_optimizer(
model,
lr=args.learning_rate,
task_specific_lr=task_specific_lr,
warmup_steps=100,
task_specific_warmup_steps=100,
)
# Validation is done by finetuning on a bunch of tasks:
valid_fn = partial(
validate_by_finetuning_on_tasks,
dataset=fsmol_dataset,
learning_rate=args.finetune_lr_scale * args.learning_rate,
task_specific_learning_rate=args.finetune_lr_scale * task_specific_lr,
batch_size=args.batch_size,
metric_to_use=args.metric_to_use,
seed=args.seed,
aml_run=aml_run,
)
_, best_model_state = train_loop(
model=model,
optimizer=optimizer,
lr_scheduler=lr_scheduler,
train_data=MultitaskTaskSampleBatchIterable(
fsmol_dataset,
data_fold=DataFold.TRAIN,
task_name_to_id=train_task_name_to_id,
max_num_graphs=args.batch_size,
),
valid_fn=valid_fn,
metric_to_use=args.metric_to_use,
max_num_epochs=args.num_epochs,
patience=args.patience,
aml_run=aml_run,
)
torch.save(best_model_state, os.path.join(out_dir, "best_model.pt"))
if __name__ == "__main__":
try:
main()
except Exception:
_, value, tb = sys.exc_info()
traceback.print_exc()
pdb.post_mortem(tb)