SVIP/train_labeling_model.py at main · ASTRAL-Group/SVIP · GitHub

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
import torch
from torch.utils.data import DataLoader, random_split
from transformers import DataCollatorWithPadding, AutoTokenizer
import torch.nn as nn
import torch.optim as optim
from tqdm import tqdm
import numpy as np
import argparse
import yaml
from models import *
from mydatasets import *
from utils import *

class ContrastiveLoss(nn.Module):
    def __init__(self, margin=1.0, alpha=0.5):
        super(ContrastiveLoss, self).__init__()
        self.margin = margin
        self.alpha = alpha

    def forward(self, outputs, encode_sentences):
        N = len(outputs)

        sentence_loss = 0.0
        for i in range(N):
            sentence_loss += distance_consistency_loss(outputs[i], encode_sentences)
        sentence_loss /= N

        secret_distance = 0.0
        for i in range(N):
            for j in range(i + 1, N):
                secret_distance += distance_margin_loss(outputs[i], outputs[j], self.margin)
        secret_loss = secret_distance / (N * (N - 1) / 2)

        loss = self.alpha * sentence_loss + (1 - self.alpha) * secret_loss
        return loss, sentence_loss, secret_loss

def load_config(config_path):
    with open(config_path, 'r') as file:
        config = yaml.safe_load(file)
    return config

def train_model(model, train_loader, secret_batch_size, secret_dim, optimizer, scheduler, criterion, device):
    model.train()
    total_loss = 0.0
    total_sentence_loss = 0.0
    total_secret_loss = 0.0
    for batch in tqdm(train_loader):
        optimizer.zero_grad()
        sentence_batch = batch['input_ids'].to(device)
        attention_mask = batch['attention_mask'].to(device)
        outputs = []
        for n in range(secret_batch_size):
            secret_tensor = torch.randint(0, 2, (secret_dim,)).float().to(device)
            secret_batch = secret_tensor.unsqueeze(0).expand(sentence_batch.shape[0], secret_dim)

            output, encode_sentence = model(sentence_batch, attention_mask, secret_batch)
            outputs.append(output)

        loss, sentence_loss, secret_loss = criterion(outputs, encode_sentence)

        loss.backward()
        optimizer.step()

        total_loss += loss.item()
        total_sentence_loss += sentence_loss.item()
        total_secret_loss += secret_loss.item()

    print(f"Training Loss: {total_loss / len(train_loader)}")
    print(f"Training Sentence Loss: {total_sentence_loss / len(train_loader)}")
    print(f"Training Secret Loss: {total_secret_loss / len(train_loader)}")

def test_model(model, test_loader, secret_batch_size, secret_dim, criterion, device):
    model.eval()
    total_loss = 0.0
    total_sentence_loss = 0.0
    total_secret_loss = 0.0
    with torch.no_grad():
        for batch in tqdm(test_loader):
            sentence_batch = batch['input_ids'].to(device)
            attention_mask = batch['attention_mask'].to(device)
            outputs = []
            for n in range(secret_batch_size):
                secret_tensor = torch.randint(0, 2, (secret_dim,)).float().to(device)
                secret_batch = secret_tensor.unsqueeze(0).expand(sentence_batch.shape[0], secret_dim)

                output, encode_sentence = model(sentence_batch, attention_mask, secret_batch)
                outputs.append(output)

            loss, sentence_loss, secret_loss = criterion(outputs, encode_sentence)
            total_loss += loss.item()
            total_sentence_loss += sentence_loss.item()
            total_secret_loss += secret_loss.item()

    print(f"Test Loss: {total_loss / len(test_loader)}")
    print(f"Test Sentence Loss: {total_sentence_loss / len(test_loader)}")
    print(f"Test Secret Loss: {total_secret_loss / len(test_loader)}")

def distance_between_two_secrets(model, test_loader, secret_dim, threshold, device, secret_n=8):
    model.eval()
    diffs = []
    over_threshold_count = 0
    total_count = 0

    with torch.no_grad():
        for batch in tqdm(test_loader):
            sentence_batch = batch['input_ids'].to(device)
            attention_mask = batch['attention_mask'].to(device)
            for _ in range(secret_n):

                secret_batch1 = torch.randint(0, 2, (sentence_batch.shape[0], secret_dim)).float().to(device)
                secret_batch2 = torch.randint(0, 2, (sentence_batch.shape[0], secret_dim)).float().to(device)


                output1, _ = model(sentence_batch, attention_mask, secret_batch1)
                output2, _ = model(sentence_batch, attention_mask, secret_batch2)

                diff = F.pairwise_distance(output1, output2, p=2)
                diffs.append(diff)

                over_threshold_count += (diff > threshold).sum().item()
                total_count += len(diff)

    diffs_tensor = torch.cat(diffs, dim=0)

    avg_diffs = diffs_tensor.mean().item()
    print(f"Average difference between outputs with different secrets: {avg_diffs}")
    print(f"Percentage of samples with difference greater than {threshold}: {over_threshold_count / total_count * 100:.4f}%")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Training labeling model")
    parser.add_argument('--config_path', type=str, required=True, help='Path to config file')
    parser.add_argument('--dataset_path', type=str, required=True, help='Path to dataset')
    args = parser.parse_args()

    config = load_config(args.config_path)

    torch.manual_seed(config["seed"])
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"learnable_y_{config['secret_dim']}_m{config['margin']}")

    tokenizer = AutoTokenizer.from_pretrained(config["sentence_encode_model_name"])
    data_collator = DataCollatorWithPadding(tokenizer)

    dataset = TextDataset(args.dataset_path)
    subset_length = len(dataset) // 10
    dataset = torch.utils.data.Subset(dataset, range(subset_length))

    test_size = int(config["test_split_ratio"] * len(dataset))
    train_size = len(dataset) - test_size
    train_dataset, test_dataset = random_split(dataset, [train_size, test_size])

    train_loader = DataLoader(train_dataset, batch_size=config["batch_size"], shuffle=True, num_workers=12, collate_fn=data_collator)
    test_loader = DataLoader(test_dataset, batch_size=config["batch_size"], shuffle=False, num_workers=12, collate_fn=data_collator)

    model = LearnableYModel(secret_dim=config["secret_dim"], output_dim=config["output_dim"],
                            sentence_model_name=config["sentence_encode_model_name"], sentence_embed_dim=config["sentence_embed_dim"],
                            output_range_max=config["output_range_max"]).to(device)
    criterion = ContrastiveLoss(margin=config["margin"], alpha=config["alpha"])
    optimizer = optim.AdamW(model.parameters(), lr=config["learning_rate"], weight_decay=1e-2)
    scheduler = None

    # Training loop
    for epoch in range(config["num_epochs"]):
        print(f"Epoch :{epoch}")
        train_model(model, train_loader, config["secret_batch_size"], config["secret_dim"], optimizer, scheduler, criterion, device)
        test_model(model, test_loader, config["secret_batch_size"], config["secret_dim"], criterion, device)
        distance_between_two_secrets(model, test_loader, config["secret_dim"], config["threshold"], device)
        torch.save(model.state_dict(), f'./models/Ymodel.pth')