-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
156 lines (127 loc) · 5.62 KB
/
model.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
import pytorch_lightning as pl
import torchmetrics
from pytorch_lightning.callbacks import ModelCheckpoint, EarlyStopping
from pytorch_lightning.loggers import TensorBoardLogger
from transformers import AutoModel
from transformers import (
get_cosine_schedule_with_warmup,
get_linear_schedule_with_warmup,
)
from typing import Any, Tuple, Iterable, Dict
import torch.nn as nn
import torch
class RuleTakerModel(pl.LightningModule):
def __init__(
self,
plm: str,
n_classes: int,
n_training_steps=None,
n_warmup_steps=None,
lr=None,
):
super().__init__()
self.plm = plm
self.encoder = AutoModel.from_pretrained(plm, return_dict=True)
self.classifier = nn.Linear(self.encoder.config.hidden_size, n_classes)
self.dropout = nn.Dropout(self.encoder.config.hidden_dropout_prob)
self.lr = lr
# self.warmup_steps = warmup_steps
self.n_training_steps = n_training_steps
self.n_warmup_steps = n_warmup_steps
self.criterion = nn.CrossEntropyLoss()
self.train_metrics = torchmetrics.Accuracy()
self.val_metrics = torchmetrics.Accuracy()
# self.train_metrics = torchmetrics.Accuracy()
# self.val_metrics = torchmetrics.Accuracy()
def forward(self, input_ids, attention_mask, label=None):
output = self.encoder(input_ids, attention_mask=attention_mask)
output = self.classifier(output.pooler_output)
label_logits = self.dropout(output)
output_dic = {}
loss = 0
output_dic["label_logits"] = label_logits
output_dic["label_probs"] = nn.functional.softmax(label_logits, dim=1)
output_dic["answer_index"] = label_logits.argmax(1)
if label is not None:
loss = self.criterion(label_logits, label)
is_correct = output_dic["answer_index"] == label
output_dic["is_correct"] = is_correct
return loss, output_dic
# def log_metrics(self, split, pred, label) -> None:
# self.metrics[split](pred, label)
# # self.metrics[split](pred, label)
# self.log(f"Acc_{split}", self.metrics[split], on_step=False, on_epoch=True)
def training_step(self, batch, batch_idx):
input_ids = batch["token_ids"]
attention_mask = batch["attention_mask"]
label = batch["label"]
loss, outputs = self(input_ids, attention_mask, label)
self.log("loss_train", loss, on_step=True, on_epoch=True)
return {"loss": loss, "predictions": outputs["answer_index"], "label": label}
def training_step_end(self, step_output):
split = "train"
self.train_metrics(step_output["predictions"], step_output["label"])
self.log(f"Acc_{split}", self.train_metrics, on_step=False, on_epoch=True)
# self.log_metrics("train", step_output["predictions"], step_output["label"])
def validation_step_end(self, step_output):
# self.log_metrics("val", step_output["predictions"], step_output["label"])
split = "val"
self.val_metrics(step_output["predictions"], step_output["label"])
self.log(f"Acc_{split}", self.val_metrics, on_step=False, on_epoch=True)
# def test_step(self, batch, batch_idx):
# return self.val_test_step("test", batch, batch_idx)
def validation_step(self, batch, batch_idx):
return self.val_test_step("val", batch, batch_idx)
def val_test_step(self, split, batch, batch_idx):
input_ids = batch["token_ids"]
attention_mask = batch["attention_mask"]
label = batch["label"]
loss, outputs = self(input_ids, attention_mask, label)
# self.metrics["val"](outputs["answer_index"], label)
# self.log_metrics("val", outputs["answer_index"], label)
# acc = self.metrics
self.log("loss_val", loss, on_step=True, on_epoch=True)
# self.log("performance", {"loss": loss, "acc": acc})
# self.log(f"{split}_loss={loss}\tAcc={acc}", prog_bar=True, logger=True, sync_dist=True)
return {"loss": loss, "predictions": outputs["answer_index"], "label": label}
def validation_epoch_end(self, outputs: Iterable[Any]) -> None:
return self.val_test_epoch_end("val", outputs)
# def test_epoch_end(self, outputs: Iterable[Any]) -> None:
# return self.val_test_epoch_end("test", outputs)
def training_epoch_end(self, outputs):
self.train_metrics.reset()
def val_test_epoch_end(self, split: str, outputs: Iterable[Any]) -> None:
val = self.val_metrics.compute()
self.log("val_acc_epoch", val)
self.val_metrics.reset()
print("val acc ", val)
def configure_optimizers(self) -> Dict[str, Any]:
assert self.trainer is not None
max_steps = (
self.trainer.max_epochs
* len(self.trainer.datamodule.train_dataloader())
// self.trainer.accumulate_grad_batches
)
return get_optimizers(
self.parameters(), self.lr, self.n_warmup_steps, max_steps,
)
def get_optimizers(
parameters: Iterable[torch.nn.parameter.Parameter],
lr: float,
num_warmup_steps: int,
num_training_steps: int,
) -> Dict[str, Any]:
"""
Get an AdamW optimizer with linear learning rate warmup and cosine decay.
"""
optimizer = torch.optim.AdamW(parameters, lr=lr)
# scheduler = get_cosine_schedule_with_warmup(
scheduler = get_linear_schedule_with_warmup(
optimizer,
num_warmup_steps=num_warmup_steps,
num_training_steps=num_training_steps,
)
return {
"optimizer": optimizer,
"lr_scheduler": {"scheduler": scheduler, "interval": "step",},
}