-
Notifications
You must be signed in to change notification settings - Fork 2
/
small_model_dg.py
172 lines (150 loc) · 7.3 KB
/
small_model_dg.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
from utils.data_reading import read_jsonl_file, load_data_for_expert
import os
import torch
from torch.utils.data import Dataset, DataLoader
from transformers import BertForSequenceClassification, BertTokenizer, AdamW, get_linear_schedule_with_warmup, RobertaForSequenceClassification, RobertaTokenizer
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, f1_score
from tqdm import tqdm
data_path = "/home/liuhui/unify/data"
dataset_name_C = "Constraint"
dataset_name_G = "GOSSIPCOP"
dataset_name_P = "POLITIFACT"
# dataload
# "ID": ID, "MESSAGE": MESSAGE, "EVIDENCE": EVIDENCE, "label": label
# dataset = {"test": testset, "train": trainset, "val": valset}
C_dataset, _ = load_data_for_expert(os.path.join(data_path, dataset_name_C), dataset_name_C, mode="binary", gq_file=None, sq_file=None, evo_file=None, evo_flag=False)
G_dataset, _ = load_data_for_expert(os.path.join(data_path, dataset_name_G), dataset_name_G, mode="binary", gq_file=None, sq_file=None, evo_file=None, evo_flag=False)
P_dataset, _ = load_data_for_expert(os.path.join(data_path, dataset_name_P), dataset_name_P, mode="binary", gq_file=None, sq_file=None, evo_file=None, evo_flag=False)
# C，G -> P
# source_train_dataset = C_dataset["train"] + G_dataset["train"]
# source_val_dataset = C_dataset["val"] + G_dataset["val"]
# target_dataset = P_dataset["test"]
# # C，P -> G
# source_train_dataset = C_dataset["train"] + P_dataset["train"]
# source_val_dataset = C_dataset["val"] + P_dataset["val"]
# target_dataset = G_dataset["test"]
#
# # G，P -> C
source_train_dataset = G_dataset["train"] + P_dataset["train"]
source_val_dataset = G_dataset["val"] + P_dataset["val"]
target_dataset = C_dataset["test"]
# Create DataLoader instances
batch_size = 32 # Set your desired batch size
class CustomDataset(Dataset):
def __init__(self, data):
self.data = data
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
sample = self.data[idx]
return {
'MESSAGE': sample['MESSAGE'],
'label': sample['label']
}
# Create custom datasets
source_train_dataset = CustomDataset(source_train_dataset)
source_val_dataset = CustomDataset(source_val_dataset )
target_dataset = CustomDataset(target_dataset )
# Create DataLoader instances
source_train_loader = DataLoader(source_train_dataset, batch_size=batch_size, shuffle=True)
source_val_loader = DataLoader(source_val_dataset, batch_size=batch_size, shuffle=False)
target_loader = DataLoader(target_dataset, batch_size=batch_size, shuffle=False)
# lode model
model_name = 'roberta-base'
tokenizer = RobertaTokenizer.from_pretrained(model_name)
model = RobertaForSequenceClassification.from_pretrained(model_name, num_labels=2)
# model_name = 'bert-base-uncased'
# tokenizer = BertTokenizer.from_pretrained(model_name)
# model = BertForSequenceClassification.from_pretrained(model_name, num_labels=2) # Assuming binary classification
model.to("cuda")
# Training loop
epochs = 8 # You may need to adjust the number of epochs
best_val_loss = float('inf')
optimizer = AdamW(model.parameters(), lr=2e-5) # You may need to adjust the learning rate
total_steps = len(source_train_loader) * epochs
scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=0, num_training_steps=total_steps)
for epoch in range(epochs):
model.train()
total_loss = 0.0
predictions_list = []
labels_list = []
best_val_accuracy = 0.0
best_model_state_dict = None
for batch in tqdm(source_train_loader, desc=f'Epoch {epoch + 1}/{epochs}'):
# Prepare input data
inputs = tokenizer(batch['MESSAGE'], padding=True, truncation=True, return_tensors="pt", max_length=512)
inputs = {key: value.to("cuda") for key, value in inputs.items()}
labels_str = batch['label']
labels = torch.tensor([1 if label.lower() == "true" else 0 for label in labels_str]).to(
"cuda") # Convert "true" to 1, "false" to 0
# Forward pass
outputs = model(**inputs, labels=labels)
loss = outputs.loss
# Backward pass
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
total_loss += loss.item()
# Save predictions and labels for later evaluation
logits = outputs.logits
predictions = torch.argmax(logits, dim=1)
predictions_list.extend(predictions.cpu().numpy())
labels_list.extend(labels.cpu().numpy())
del inputs, labels
avg_loss = total_loss / len(source_train_loader)
accuracy = accuracy_score(labels_list, predictions_list)
f1 = f1_score(labels_list, predictions_list)
print(f'Epoch {epoch + 1}/{epochs} - Avg Loss: {avg_loss:.4f} - Accuracy: {accuracy:.4f} - F1 Score: {f1:.4f}')
# Validation
model.eval()
val_loss = 0.
val_predictions_list = []
val_labels_list = []
with torch.no_grad():
for val_batch in source_val_loader:
val_inputs = tokenizer(val_batch['MESSAGE'], padding=True, truncation=True, return_tensors="pt")
val_inputs = {key: value.to("cuda") for key, value in val_inputs.items()}
val_labels_str = val_batch['label']
val_labels = torch.tensor([1 if label.lower() == "true" else 0 for label in val_labels_str]).to("cuda") # Convert "true" to 1, "false" to 0
val_outputs = model(**val_inputs, labels=val_labels)
val_logits = val_outputs.logits
val_predictions = torch.argmax(val_logits, dim=1)
val_predictions_list.extend(val_predictions.cpu().numpy())
val_labels_list.extend(val_labels.cpu().numpy())
del val_inputs, val_labels
accuracy_val = accuracy_score(val_labels_list, val_predictions_list)
f1_val = f1_score(val_labels_list, val_predictions_list)
# Save the best model based on validation accuracy
if accuracy_val > best_val_accuracy:
best_val_accuracy = accuracy_val
best_model_state_dict = model.state_dict()
print(f'Epoch {epoch + 1}/{epochs} - Val Accuracy: {accuracy_val:.4f} - Val F1 Score: {f1_val:.4f}')
# Load the best model for testing
del model
best_model = RobertaForSequenceClassification.from_pretrained(model_name, num_labels=2)
best_model.load_state_dict(best_model_state_dict)
best_model.to("cuda")
# Testing on target dataset
best_model.eval()
test_predictions = []
test_labels = []
with torch.no_grad():
for batch in target_loader:
inputs = tokenizer(batch['MESSAGE'], padding=True, truncation=True, return_tensors="pt")
inputs = {key: value.to("cuda") for key, value in inputs.items()} # Move inputs to the same device as labels
labels_str = batch['label']
labels = torch.tensor([1 if label.lower() == "true" else 0 for label in labels_str]).to(
"cuda") # Convert "true" to 1, "false" to 0
outputs = best_model(**inputs, labels=labels)
logits = outputs.logits
predictions = torch.argmax(logits, dim=1)
test_predictions.extend(predictions.cpu().numpy())
test_labels.extend(labels.cpu().numpy())
del inputs, labels
# Evaluate test results (you can use appropriate metrics)
accuracy_test = accuracy_score(test_labels, test_predictions)
f1_test = f1_score(test_labels, test_predictions)
print(
f'Best Model - Test Accuracy: {accuracy_test:.4f} - Test F1 Score: {f1_test:.4f}')