TAGSAM/dataset.py at master · SundayVHan/TAGSAM · 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
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
import os
import pickle

import numpy as np
import torch
from torch_geometric.loader import NeighborSampler
from tqdm import tqdm

from model import TextEncoder


class GraphDataset:
    def __init__(self, args):
        self.args = args

        data = torch.load(f'./data/{args.dataset_name}.pt', weights_only=False, map_location="cpu")
        self.node_f = data['node_f']
        self.edge_index = data['edge_index']
        self.text_list = data['text_list']
        self.label_list = data['label_list']
        self.all_labels = data['all_labels']
        self.labels_desc = data['labels_desc']

        self.text_embeds = None
        self.all_labels_embeds = None
        self.process_text2emb()

        self.test_labels = None
        self.test_samples = None
        self.test_split()

        self.tasks = None
        self.process_subgraph()

    def __len__(self):
        return len(self.node_f)

    @torch.no_grad()
    def process_text2emb(self):
        cache_file = os.path.join(self.args.buffer_save_dir, "cache.pt")
        if os.path.exists(cache_file):
            with open(cache_file, 'rb') as f:
                cache_data = pickle.load(f)
                self.text_embeds = cache_data.get('text_embeds')
                self.all_labels_embeds = cache_data.get('all_labels_embeds')
                return

        print("Start to process text to embeddings...")
        all_text_embeds = []
        all_labels_embeds = []
        batch_size = 64
        text_encoder = TextEncoder(self.args).to(self.args.device)

        for i in tqdm(range(0, len(self.text_list), batch_size)):
            batch_texts = self.text_list[i:i + batch_size]
            batch_embeds = text_encoder(batch_texts).detach().cpu()

            all_text_embeds.append(batch_embeds)
            del batch_embeds
            torch.cuda.empty_cache()

        labels_with_desc = [f"{label} {desc}" for label, desc in zip(self.all_labels, self.labels_desc)]
        for i in tqdm(range(0, len(self.all_labels), batch_size)):
            batch_texts = labels_with_desc[i:i + batch_size]
            batch_embeds = text_encoder(batch_texts).detach().cpu()

            all_labels_embeds.append(batch_embeds)
            del batch_embeds
            torch.cuda.empty_cache()

        self.text_embeds = torch.cat(all_text_embeds, dim=0).to("cpu")
        self.all_labels_embeds = torch.cat(all_labels_embeds, dim=0).to("cpu")

        cache_data = {
            'text_embeds': self.text_embeds,
            'all_labels_embeds': self.all_labels_embeds
        }
        with open(cache_file, 'wb') as f:
            pickle.dump(cache_data, f)

    def test_split(self):
        filtered_labels = [label for label in self.all_labels if label != "nan"]
        filtered_labels = np.array(filtered_labels)
        num_labels = len(filtered_labels)
        n_way = 5
        num_groups = num_labels // n_way

        label_to_all_index = {label: idx for idx, label in enumerate(self.all_labels)}

        label_to_list_index = {}
        for idx, label in enumerate(self.label_list):
            if label not in label_to_list_index:
                label_to_list_index[label] = []
            label_to_list_index[label].append(idx)

        label_to_list_index = {label: np.array(indices) for label, indices in label_to_list_index.items()}

        labels = []
        samples = []
        for i in range(num_groups):
            test_labels = filtered_labels[i * n_way: (i + 1) * n_way]
            test_labels_idx = [label_to_all_index[label] for label in test_labels]
            labels.append(test_labels_idx)

            test_samples_idx = np.concatenate([label_to_list_index[label] for label in test_labels])
            samples.append(test_samples_idx)

        self.test_labels = labels
        self.test_samples = samples

    def process_subgraph(self):
        cache_file = os.path.join(self.args.buffer_save_dir, f"cache_subgraph_{self.args.batch_size_test}.pt")
        if os.path.exists(cache_file):
            with open(cache_file, 'rb') as f:
                cache_data = pickle.load(f)
                self.tasks = cache_data.get('tasks')
                return

        print("Start to process test subgraph...")

        tasks = []

        for labels_idx, samples_idx in tqdm(zip(self.test_labels, self.test_samples), total=len(self.test_labels)):
            samples_idx = torch.tensor(samples_idx)
            if self.args.batch_size_test == -1:
                sampler = NeighborSampler(self.edge_index, node_idx=samples_idx,
                                          sizes=self.args.sample_size, batch_size=len(samples_idx),
                                          shuffle=True, num_workers=16)
            else:
                sampler = NeighborSampler(self.edge_index, node_idx=samples_idx,
                                          sizes=self.args.sample_size, batch_size=self.args.batch_size_test,
                                          shuffle=True, num_workers=16)

            task = []
            for batch_size, n_id, adjs in sampler:
                task.append((labels_idx, batch_size, n_id, adjs))

            tasks.append(task)

        self.tasks = tasks

        cache_data = {
            'tasks': self.tasks
        }
        with open(cache_file, 'wb') as f:
            pickle.dump(cache_data, f)


class SynGraphDataset(GraphDataset):
    def __init__(self, syn_graph, syn_text, args):
        self.args = args

        self.node_f = syn_graph.detach().to(args.device).requires_grad_(True)
        self.edge_index = torch.stack([torch.arange(len(self)), torch.arange(len(self))], dim=0)
        self.graph_encoder_lr = torch.tensor(args.graph_encoder_lr).to(args.device).requires_grad_(True)
        self.text_encoder_lr = torch.tensor(args.text_encoder_lr).to(args.device).requires_grad_(True)
        self.text_embeds = syn_text.to(args.device).requires_grad_(False)

        self.optimizer = torch.optim.SGD([
            {'params': self.node_f, 'lr': args.syn_lr, "momentum": 0.5},
            {'params': self.graph_encoder_lr, 'lr': args.syn_lr_lr, "momentum": 0.5},
            {'params': self.text_encoder_lr, 'lr': args.syn_lr_lr, "momentum": 0.5}
        ])
        self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer, step_size=200, gamma=0.95)

    def __len__(self):
        return len(self.node_f)

    def set_eval_model(self):
        self.node_f.requires_grad_(False)
        self.graph_encoder_lr.requires_grad_(False)
        self.text_encoder_lr.requires_grad_(False)

    def set_train_model(self):
        self.node_f.requires_grad_(True)
        self.graph_encoder_lr.requires_grad_(True)
        self.text_encoder_lr.requires_grad_(True)

    def compute_grad(self, loss):
        grad = torch.autograd.grad(loss, [self.node_f, self.graph_encoder_lr, self.text_encoder_lr], allow_unused=True)
        self.node_f.grad = grad[0]
        self.graph_encoder_lr.grad = grad[1]
        self.text_encoder_lr.grad = grad[2]
        torch.nn.utils.clip_grad_norm_(self.node_f, 5)

    def step(self):
        self.optimizer.step()
        self.scheduler.step()

    def zero_grad(self):
        self.optimizer.zero_grad()