-
Notifications
You must be signed in to change notification settings - Fork 0
/
pretrain_bert.py
241 lines (206 loc) · 9.39 KB
/
pretrain_bert.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
import argparse
from util import str2bool, load_config
from pathlib import Path
import os
import wandb
import torch
from torch import _pin_memory, logit, nn
import pytorch_lightning as pl
from pytorch_lightning.loggers import WandbLogger
from pytorch_lightning.strategies import DDPStrategy
import torch
import torchmetrics
from torchmetrics import Accuracy, Recall, Precision
import numpy as np
from data.dataset import CodeFile
from transformers import get_linear_schedule_with_warmup
from transformers import RobertaTokenizer
from tokenizers import Tokenizer
# from models.bert import BERTforPreTraining
from models.rankfromsets import InnerProduct
from models.lstm import LSTM_model
from typing import Optional
import warnings
from dotenv import load_dotenv
import pickle
warnings.filterwarnings("ignore")
class PreTrainingModule(pl.LightningModule):
def __init__(self, model, run_d):
super().__init__()
self.model = model
self.run_d = run_d
self.cross_loss = nn.BCEWithLogitsLoss()
self.learning_rate = self.run_d["learning_rate"]
self.accuracy = torchmetrics.Accuracy(average = "micro", num_classes = 256)
self.precision_metric = torchmetrics.Precision(average = "macro", num_classes = 256)
self.recall_metric = torchmetrics.Recall(average = "macro", num_classes = 256)
self.save_hyperparameters()
self.val_predictions = []
self.val_labels = []
self.counter_variable = 1
def training_step(self, batch, batch_idx):
metrics_labels = batch.pop("metrics_labels")
logits = self.model(**batch)
loss = self.cross_loss(logits, batch["labels"].float())
self.log("train_loss", loss)
return loss
def validation_step(self, batch, batch_idx):
metrics_labels = batch.pop("metrics_labels")
logits = self.model(**batch)
preds = torch.argmax(logits, dim=1)
self.accuracy.update(preds, metrics_labels)
self.precision_metric.update(preds, metrics_labels)
self.recall_metric.update(preds, metrics_labels)
preds = np.argmax(logits.detach().cpu().numpy(), axis=1)
self.val_predictions.extend(preds)
def test_step(self, batch, batch_idx):
metrics_labels = batch.pop("metrics_labels")
logits = self.model(**batch)
preds = torch.argmax(logits, dim=1)
self.accuracy.update(preds, metrics_labels)
self.precision_metric.update(preds, metrics_labels)
self.recall_metric.update(preds, metrics_labels)
preds = np.argmax(logits.detach().cpu().numpy(), axis=1)
self.val_predictions.extend(preds)
def configure_optimizers(self):
# set up optimizer
if self.run_d["optimizer"] == "adam":
optimizer = torch.optim.Adam(
self.model.parameters(),
lr=self.run_d["learning_rate"],
betas=eval(self.run_d["betas"]),
weight_decay=self.run_d["weight_decay"],
)
elif self.run_d["optimizer"] == "adamW":
optimizer = torch.optim.AdamW(
self.model.parameters(), lr=self.run_d["learning_rate"], weight_decay=self.run_d["weight_decay"]
)
else:
raise NotImplementedError
# set up scheduler
if self.run_d["scheduler"] == "step":
scheduler = torch.optim.lr_scheduler.StepLR(
optimizer, self.run_d["scheduler_period"], gamma=self.run_d["scheduler_ratio"]
)
elif self.run_d["scheduler"] == "plateau":
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, mode="min", patience=self.run_d["scheduler_period"], factor=self.run_d["scheduler_ratio"]
)
elif self.run_d["scheduler"] == "linear_with_warmup":
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=self.num_training_steps * 0.1, num_training_steps=self.num_training_steps
)
else:
return optimizer
return [optimizer], [{"scheduler": scheduler, "interval": "step"}]
def training_epoch_end(self, outputs):
if self.global_step >= 0.9 * self.num_training_steps:
self.model.max_len = 512
if self.trainer.is_global_zero and self.run_d["save_state"]:
self.save_model()
def validation_epoch_end(self, outputs):
with open(f'val_preds_{self.counter_variable}.pkl', 'wb') as f:
pickle.dump(self.val_predictions, f)
self.val_predictions = []
self.counter_variable += 1
self.log("Validation Accuracy", self.accuracy.compute())
self.accuracy.reset()
self.log("Validation Precision", self.precision_metric.compute())
self.precision_metric.reset()
self.log("Validation Recall", self.recall_metric.compute())
self.recall_metric.reset()
def test_epoch_end(self, outputs):
self.log("Test Accuracy", self.accuracy.compute())
self.accuracy.reset()
self.log("Test Precision", self.precision_metric.compute())
self.precision_metric.reset()
self.log("Test Recall", self.recall_metric.compute())
self.recall_metric.reset()
def save_model(self):
mr_fp = os.path.join(wandb.run.dir, "bert-pre-train.pt")
torch.save(self.model.state_dict(), mr_fp)
@property
def num_training_steps(self) -> int:
"""Total training steps inferred from datamodule and devices."""
if self.trainer.max_steps:
return self.trainer.max_steps
limit_batches = self.trainer.limit_train_batches
batches = len(self.train_dataloader())
batches = min(batches, limit_batches) if isinstance(limit_batches, int) else int(limit_batches * batches)
num_devices = max(1, self.trainer.num_gpus, self.trainer.num_processes)
if self.trainer.tpu_cores:
num_devices = max(num_devices, self.trainer.tpu_cores)
effective_accum = self.trainer.accumulate_grad_batches * num_devices
return (batches // effective_accum) * self.trainer.max_epochs
class TrainingDataModule(pl.LightningDataModule):
def __init__(self, data_d, run_d, collate):
super().__init__()
self.data_d = data_d
self.run_d = run_d
self.collate = collate
def setup(self, stage: Optional[str] = None) -> None:
self.train_data = CodeFile(self.data_d["train_path"])
self.val_data = CodeFile(self.data_d["val_path"])
self.test_data = CodeFile(self.data_d["test_path"])
def train_dataloader(self) -> torch.utils.data.DataLoader:
return torch.utils.data.DataLoader(
self.train_data,
batch_size=self.run_d["batch_size"],
shuffle=True,
collate_fn=self.collate,
pin_memory=True,
num_workers=self.run_d["num_workers"],
)
def val_dataloader(self) -> torch.utils.data.DataLoader:
return torch.utils.data.DataLoader(
self.val_data,
batch_size = self.run_d["eval_batch_size"],
shuffle=False,
collate_fn=self.collate,
pin_memory=True,
num_workers=self.run_d["num_workers"]
)
def test_dataloader(self) -> torch.utils.data.DataLoader:
return torch.utils.data.DataLoader(
self.test_data,
batch_size = self.run_d["batch_size"],
shuffle=False,
collate_fn=self.collate,
pin_memory=True,
num_workers=self.run_d["num_workers"]
)
if __name__ == "__main__":
load_dotenv()
parser = argparse.ArgumentParser(description="Train")
parser.add_argument(
"-t", "--template_fp", type=str, default="config/template.yaml", help="path to template config file"
)
parser.add_argument("-c", "--custom_fp", type=str, required=False, help="path to custom config file")
parser.add_argument("-l", "--log_dir", type=str, default="./", help="path to log/output directory")
flags = parser.parse_args()
project_name, run_name, data_d, model_d, run_d = load_config(flags.template_fp, flags.custom_fp)
Path(flags.log_dir).mkdir(parents=True, exist_ok=True)
wandb_logger = WandbLogger(project=project_name, name=run_name, save_dir=flags.log_dir)
trainer = pl.Trainer(
max_epochs=run_d["num_epochs"],
gpus=torch.cuda.device_count(),
logger=wandb_logger,
strategy=DDPStrategy(find_unused_parameters=False),
precision=16,
profiler="simple",
default_root_dir=flags.log_dir,
)
tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
if model_d["model_name"] == "rankfromsets":
model_dict = {"n_labels":model_d["classes"],"n_attributes":len(tokenizer),"emb_size":model_d["emb_size"],"sparse":False,"mode":model_d["mode"]}
model = InnerProduct(**model_dict)
elif model_d["model_name"] == "lstm":
model_dict = {"num_words":len(tokenizer),"emb_size":model_d["emb_size"],"outputs":model_d["n_labels"],"dropout":model_d["dropout"]}
model = LSTM_model(**model_dict)
data_module = TrainingDataModule(data_d, run_d, model.collate_fn)
if "load_state" in run_d and run_d["load_state"]:
pre_training_module = PreTrainingModule.load_from_checkpoint(run_d["chkpt_path"])
else:
pre_training_module = PreTrainingModule(model, run_d)
trainer.fit(pre_training_module, data_module)
trainer.test(pre_training_module, data_module)