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
| class MyDataModule(pl.LightningDataModule):
def __init__(self, train_indices, val_indices, test_indices, batch_size):
super().__init__()
self.batch_size = batch_size
self.train_indices = train_indices
self.val_indices = val_indices
self.test_indices = test_indices
def train_dataloader(self):
self.train_dataset = CREembeddingDataset(data_df.iloc[self.train_indices], embeddings, target='class')
return DataLoader(self.train_dataset, batch_size=self.batch_size)
def val_dataloader(self):
self.val_dataset = CREembeddingDataset(data_df.iloc[self.val_indices], embeddings, target='class')
return DataLoader(self.val_dataset, batch_size=self.batch_size)
def test_dataloader(self):
self.test_dataset = CREembeddingDataset(data_df.iloc[self.test_indices], embeddings, target='class')
return DataLoader(self.test_dataset, batch_size=self.batch_size)
class MyModule(pl.LightningModule):
def __init__(self, foldnum, model, num_epochs, batch_size=batch_size, class_weights=class_weights):
"""
foldnum: 第几折
model: 模型
batch_size: 批次大小
class_weights: 类别权重
"""
super().__init__()
self.mymodel = model
self.batch_size = batch_size
self.class_weights = class_weights
self.loss_fn = nn.CrossEntropyLoss(weight=torch.tensor(class_weights, dtype=torch.bfloat16))
self.f1 = f1_score
self.init_metric_target(foldnum, num_epochs)
def init_metric_target(self, foldnum, num_epochs):
self.val_outputs = []
self.val_targets = []
self.test_outputs = []
self.test_targets = []
self.running_loss = 0
self.val_loss = 0
self.test_loss = 0
self.epoch = 0
self.fold = foldnum
self.num_epochs = num_epochs
def forward(self, x):
return self.mymodel(x)
def configure_optimizers(self):
optimizer = optim.AdamW(self.mymodel.parameters(), lr=0.0001, weight_decay=1e-3)
warm_up_iter = 5
T_max = 50
lr_max = 1e-4
lr_min = 0
lambda0 = lambda cur_iter: (lr_max - lr_min)/warm_up_iter * cur_iter / 0.0001 if cur_iter <= warm_up_iter else \
(lr_min + 0.5*(lr_max-lr_min)*(1.0+math.cos((cur_iter-warm_up_iter)/(T_max-warm_up_iter)*math.pi))) / 0.0001
scheduler = LambdaLR(optimizer, lr_lambda=lambda0)
return {"optimizer": optimizer, "lr_scheduler": scheduler}
def training_step(self, batch, batch_idx):
embeddings, targets = batch
outputs = self(embeddings)
loss = self.loss_fn(outputs, targets)
self.running_loss += loss.item()
if batch_idx % 2 == 1:
avg_train_batch_loss = self.running_loss / (batch_idx + 1)
logging.info(f'Fold {self.fold+1}, Epoch {self.epoch + 1}/{self.num_epochs}, Batch {batch_idx + 1}, Train Loss(CrossEntropy): {avg_train_batch_loss:.4f}')
return loss
def validation_step(self, batch, batch_idx):
embeddings, targets = batch
outputs = self(embeddings)
loss = self.loss_fn(outputs, targets)
self.val_loss += loss.item()
_, preds = torch.max(outputs, 1)
_, decode_target = torch.max(targets, 1)
self.val_outputs.extend(preds.cpu().numpy())
self.val_targets.extend(decode_target.cpu().numpy())
return loss
def test_step(self, batch, batch_idx):
embeddings, targets = batch
outputs = self(embeddings)
loss = self.loss_fn(outputs, targets)
self.test_loss += loss.item()
_, preds = torch.max(outputs, 1)
_, decode_target = torch.max(targets, 1)
self.test_outputs.extend(preds.cpu().numpy())
self.test_targets.extend(decode_target.cpu().numpy())
return loss
def on_train_epoch_start(self):
self.running_loss = 0
current_lr = self.configure_optimizers()["optimizer"].param_groups[0]['lr']
logging.info(f'Epoch {self.epoch + 1}/{self.num_epochs}, Learning Rate: {current_lr:.5f}')
def on_train_epoch_end(self):
avg_train_loss = self.running_loss / len(self.train_dataloader())
logging.info(f'Fold {self.fold+1}, Epoch {self.epoch+1}/{self.num_epochs}, Train Loss(CrossEntropy): {avg_train_loss:.4f}')
self.log('train_loss', avg_train_loss, prog_bar=True)
def on_validation_epoch_start(self):
self.val_loss = 0
self.val_outputs = []
self.val_targets = []
def on_validation_epoch_end(self):
val_f1 = self.f1(self.val_targets, self.val_outputs, average='macro')
avg_val_loss = self.val_loss / len(self.val_targets)
logging.info(f'Fold {self.fold+1}, Epoch {self.epoch+1}/{self.num_epochs}, Val Loss(CrossEntropy): {avg_val_loss:.4f}, Val F1: {val_f1:.4f}')
self.log('val_f1', val_f1, prog_bar=True)
self.log('val_loss', avg_val_loss, prog_bar=True)
def on_test_epoch_start(self):
self.test_loss = 0
self.test_outputs = []
self.test_targets = []
def on_test_epoch_end(self):
test_f1 = self.f1(self.test_targets, self.test_outputs, average='macro')
avg_test_loss = self.test_loss / len(self.test_targets)
logging.info(f'Fold {self.fold+1}, Epoch {self.epoch+1}/{self.num_epochs}, Test Loss(CrossEntropy): {avg_test_loss:.4f}, Val F1: {test_f1:.4f}')
self.log('test_f1', test_f1, prog_bar=True)
self.log('test_loss', avg_test_loss, prog_bar=True)
self.epoch += 1
|
