-
Notifications
You must be signed in to change notification settings - Fork 0
/
datamodule.py
159 lines (122 loc) · 4.48 KB
/
datamodule.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
import numpy as np
import pandas as pd
import pytorch_lightning as pl
import torch
from torch.utils.data import Dataset
from torch import nn, optim
from utils import recall_score, f1_score, CosineScheduler
class lstm_dataset(Dataset):
def __init__(self, X, P, y) -> None:
super().__init__()
self.X = torch.tensor(X)
self.P = torch.tensor(P)
self.y = torch.tensor(y)
def __len__(self):
return len(self.X)
def __getitem__(self, idx):
return self.X[idx], self.P[idx], self.y[idx]
class bert_dataset(Dataset):
def __init__(self, X, mask, y) -> None:
super().__init__()
if type(X) != torch.Tensor:
X = torch.tensor(X)
if type(mask) != torch.Tensor:
mask = torch.tensor(mask)
if type(y) != torch.Tensor:
y = torch.tensor(y)
self.X = X
self.mask = mask
self.y = y
def __len__(self):
return len(self.X)
def __getitem__(self, idx):
return self.X[idx], self.mask[idx], self.y[idx]
class Net(pl.LightningModule):
def __init__(self, model, lr, epoch=500, warmup_t=10, crf=False, use_scheduler=True):
super().__init__()
print("Making Model...")
self.model = model
self.criterion = nn.CrossEntropyLoss()
self.optimizer = optim.Adam(
self.model.parameters(),
lr=lr,
betas=(0.9, 0.999),
amsgrad=True,
)
self.crf = crf
self.use_scheduler = use_scheduler
self.scheduler = {
"scheduler": CosineScheduler(
self.optimizer,
t_initial=epoch - warmup_t,
lr_min=1e-9,
warmup_t=warmup_t,
warmup_lr_init=1e-6,
warmup_prefix=True,
),
"interval": "epoch",
}
def forward(self, x, sub):
output = self.model(x, sub)
return output
def predict(self, x, sub):
if self.crf:
output = self.model.decode(x, sub)
else:
output = self.forward(x)
return output
def loss_fn(self, pred, label):
pred = pred.reshape(-1, pred.shape[-1])
label = label.view(-1).to(torch.int64)
loss = self.criterion(pred, label)
return loss
def training_step(self, batch, batch_idx):
input, sub_input, label = batch
if self.crf:
loss = self.model.forward(input, sub_input, label)
loss = torch.sum(-loss)
else:
pred = self.forward(input, sub_input)
loss = self.loss_fn(pred, label)
self.log("loss", loss)
return {"loss": loss}
def validation_step(self, batch, batch_idx):
input, sub_input, label = batch
batch_size = input.shape[0]
label = label.to("cpu")
pred = self.predict(input, sub_input).to("cpu")
pred = pred.squeeze(-1)
acc = (pred == label).sum().item() / batch_size
pred = pred.view(-1)
label = label.view(-1)
f1 = f1_score(label, pred).tolist()
return {"acc": acc, "f1": f1}
def validation_epoch_end(self, outputs):
ave_acc = torch.tensor([x["acc"] for x in outputs]).to(torch.float).mean()
ave_f1 = torch.tensor([x["f1"] for x in outputs]).to(torch.float).mean()
self.log("acc", ave_acc)
self.log("f1", ave_f1)
self.log("lr", self.optimizer.param_groups[0]["lr"])
return {"acc": ave_acc}
def test_step(self, batch, batch_idx):
input, sub_input, label = batch
batch_size = input.shape[0]
label = label.to("cpu")
pred = self.predict(input, sub_input).to("cpu")
pred = pred.squeeze(-1)
acc = (pred == label).sum().item() / batch_size
pred = pred.view(-1)
label = label.view(-1)
recall = recall_score(label, pred, average="macro").tolist()
f1 = f1_score(label, pred).tolist()
self.log("test_acc", acc)
self.log("test_recall", recall)
self.log("test_f1", f1)
def test_epoch_end(self, outputs) -> None:
return super().test_epoch_end(outputs)
def configure_optimizers(self):
if self.use_scheduler:
lr_scheduler = self.scheduler
return [self.optimizer], [lr_scheduler]
else:
return [self.optimizer]