-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
173 lines (145 loc) · 7.44 KB
/
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
from sklearn.metrics import accuracy_score, f1_score, classification_report
import torch
from torch.utils.data import DataLoader
from torch.optim import AdamW
from tqdm import tqdm
import os
import yaml
import wandb
from prompt import prompt_generator, PromptDataset
from utils import read_config, read_files
from baseline import Baseline
class Train(Baseline):
def __init__(self, config):
super().__init__(config)
self.batch_size = config["batch_size"]
self.num_epochs = config["num_epochs"]
self.lr = config["lr"]
self.use_all_corpus = config["use_all_corpus"]
self.run_name = config["run_name"]
self.load_checkpoint = config["load_checkpoint"]
self.checkpoint_load_path = config["checkpoint_load_path"]
self.use_wandb = config["use_wandb"]
self.train_loader = None
self.val_loader = None
self.test_loader = None
self.optimizer = None
self.criterion = None
self.prompt_format = None
if self.load_checkpoint:
state_dict = torch.load(self.checkpoint_load_path)
self.model.load_state_dict(state_dict)
print(f"Checkpoint loaded from {self.checkpoint_load_path}")
def ready_for_train(self):
print("Reading files ...")
train_contexts, val_contexts, dictionary = read_files(self.config)
print("Generating prompt ...")
if self.use_all_corpus:
train_data = prompt_generator(train_contexts, dictionary, use_all=True, split=False)
# length: 3,390,121
val_data, test_data = prompt_generator(val_contexts, dictionary, use_all=True, split=True, test_size=0.5)
# length: 374,927 / 2 each
else:
train_data, val_data, test_data = prompt_generator(val_contexts, dictionary, use_all=False, split=True, test_size=0.3)
# length: 262,448 | 56,239 | 56,240
self.prompt_format = test_data["inputs"][0]
print("Making datasets ...")
train_dataset = PromptDataset(train_data, tokenizer=self.tokenizer)
val_dataset = PromptDataset(val_data, tokenizer=self.tokenizer)
test_dataset = PromptDataset(test_data, tokenizer=self.tokenizer)
print("Converting to dataloader ...")
self.train_loader = DataLoader(train_dataset, batch_size=self.batch_size, shuffle=True)
self.val_loader = DataLoader(val_dataset, batch_size=self.batch_size, shuffle=True)
self.test_loader = DataLoader(test_dataset, batch_size=self.batch_size, shuffle=True)
self.optimizer = AdamW(self.model.parameters(), lr=self.lr)
self.criterion = torch.nn.CrossEntropyLoss()
print("Ready for training !")
print(f"Train dataloader length: {len(self.train_loader)}")
print(f"Validation dataloader length: {len(self.val_loader)}")
print(f"Test dataloader length: {len(self.test_loader)}")
def train(self):
device = self.device
print(f"Using Device: {device}")
self.model.to(device)
checkpoint_dir = os.path.join(self.run_name, f"checkpoints")
if not os.path.isdir(checkpoint_dir):
os.makedirs(checkpoint_dir)
config_path = os.path.join(self.run_name, f"config.yaml")
with open(config_path, 'w', encoding="utf-8") as f:
self.config["prompt_format"] = self.prompt_format
yaml.dump(self.config, f, allow_unicode=True)
print(f"Config saved at {config_path}")
num_steps = len(self.train_loader)
for epoch in range(self.num_epochs):
self.model.train()
for step, (input_ids, attention_masks, labels) in tqdm(enumerate(self.train_loader), desc=f"Epoch {epoch+1} train"):
input_ids, attention_masks, labels = input_ids.to(device), attention_masks.to(device), labels.to(device)
outputs = self.model(input_ids=input_ids, attention_mask=attention_masks)[0]
loss = self.criterion(outputs, labels)
if step % 50 == 0:
if self.use_wandb == True:
wandb.log({"train_loss": loss.item()}, step= epoch*num_steps + step)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
checkpoint_path = os.path.join(checkpoint_dir, f"epoch_{epoch+1}.pt")
torch.save(self.model.state_dict(), checkpoint_path)
print(f"Checkpoint saved at {checkpoint_path}")
self.model.eval()
total_preds = []
total_labels = []
with torch.no_grad():
for input_ids, attention_masks, labels in tqdm(self.val_loader, desc=f"Epoch {epoch+1} validation"):
input_ids, attention_masks, labels = input_ids.to(device), attention_masks.to(device), labels.to(device)
outputs = self.model(input_ids=input_ids, attention_mask=attention_masks)[0]
preds = torch.argmax(outputs, dim=1)
total_preds.extend(preds.cpu().numpy())
total_labels.extend(labels.cpu().numpy())
print(f"Epoch {epoch+1} validation result:")
print(classification_report(y_true=total_labels, y_pred=total_preds, digits=4))
acc = accuracy_score(y_true=total_labels, y_pred=total_preds)
f1_weighted = f1_score(y_true=total_labels, y_pred=total_preds, average="weighted")
if self.use_wandb == True:
wandb.log({"val_acc": acc, "val_f1_weigthed": f1_weighted})
def evaluation(self):
device = self.device
self.model.eval()
self.model.to(device)
total_preds = []
total_labels = []
with torch.no_grad():
for input_ids, attention_masks, labels in tqdm(self.test_loader, desc=f"Evalutation ..."):
input_ids, attention_masks, labels = input_ids.to(device), attention_masks.to(device), labels.to(device)
outputs = self.model(input_ids=input_ids, attention_mask=attention_masks)[0]
preds = torch.argmax(outputs, dim=1)
total_preds.extend(preds.cpu().numpy())
total_labels.extend(labels.cpu().numpy())
print(f"Evaluation result:")
report = classification_report(y_true=total_labels, y_pred=total_preds, digits=4)
print(report)
acc = accuracy_score(y_true=total_labels, y_pred=total_preds)
f1_weighted = f1_score(y_true=total_labels, y_pred=total_preds, average="weighted")
if self.use_wandb == True:
wandb.log({"test_acc": acc, "test_f1_weigthed": f1_weighted})
output_dir = os.path.join(self.run_name)
if not os.path.isdir(output_dir):
os.mkdir(output_dir)
report_path = os.path.join(output_dir, "evaluation_report.txt")
with open(report_path, "w") as f:
f.write(report)
print(f"Evaluation result saved at {report_path}")
if __name__ == "__main__":
config = read_config()
train = Train(config)
if train.use_wandb == True:
wandb.init(project="NLP-WSD-KOR", config=config)
wandb.config["prompt_format"] = train.prompt_format
wandb.run.name = config["run_name"]
wandb.define_metric("train_loss", summary="min")
wandb.define_metric("val_acc", summary="max")
wandb.define_metric("val_f1_weighted", summary="max")
train.ready_for_train()
train.train()
train.evaluation()
if train.use_wandb == True:
wandb.finish()