-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain.py
379 lines (351 loc) · 18.2 KB
/
main.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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
"""
To be merged later;
"""
# import wandb
import optuna
from helper.args import *
from helper.data_utils import *
from helper.hete_utils import generate_graph as generate_graph_syn
from helper.hete_utils import generate_base_features
from helper.hete_utils import new_edge_addition, save_pyg_data, local_hetero_edge_addition, hetero_edge_addition
from backward.main_backward_new import main_backward
from forward.main_forward_new import main_forward
from transfer.main_distill_backup import main_transfer
import torch_geometric
from datetime import datetime
from pathlib import Path
from copy import deepcopy
import itertools
import os.path as osp
import logging
import torch
logging.basicConfig(level=logging.INFO)
def test_acc(res, data):
return (res[0].reshape(-1).cuda()[data.test_mask] == data.y.cuda()[data.test_mask]).sum() / torch.sum(data.test_mask)
def test_acc_res(res, data):
return (res[0].reshape(-1).cuda()[data.test_mask] == data.y.cuda()[data.test_mask]).sum() / torch.sum(data.test_mask)
def main():
seeds = [1, 3, 5, 7, 11] # , 13, 17, 19, 23, 29
args = generate_args()
if args.generate_graph:
generate_graph(args)
elif args.expmode == 'inductive':
inductive_run(args, seeds)
elif args.expmode == 'transductive':
fix_param_run(args, seeds)
def generate_graph(args):
if args.generate_mode == 'sitao':
with open(args.generation_config, 'r') as f:
exec(f.read())
generation_config = locals()['generation']
generate_graph_syn(
generation_config['num_class'],
generation_config['num_node_total'],
generation_config['degree_intra'],
generation_config['num_graph'],
generation_config['graph_type'],
generation_config['edge_homos'],
generation_config['base_dir']
)
_, data = get_dataset(args)
for i in range(generation_config['num_graph']):
base_features = generate_base_features(data, generation_config['num_node_total'],
'feature',
generation_config['num_node_total'] // generation_config['num_class'],
generation_config['num_class'])
data_dir = f"{generation_config['base_dir']}/features/{args.dataset}"
Path(data_dir).mkdir(parents=True, exist_ok=True)
torch.save(base_features, f"{data_dir}/{args.dataset}_{i}.pt")
else:
_, data = get_dataset(args)
if args.is_new and args.dataset in ["Cora", "CiteSeer", "PubMed", "photo", "cs"]:
data = torch.load(f"/mnt/home/haitaoma/Graph-smooth/code/CPF/dataset/{args.dataset}.pt")
data.edge_index = data.edge_index.to(torch.int64)
if args.dataset == 'Cora':
orig_data = deepcopy(data)
for i in range(10):
num_classes = orig_data.y.max().item() + 1
mask = 'train'
if mask == 'train':
ratio = 1.0
unit = 200
else:
ratio = 0.05
unit = 100
mode = 'tohete'
orig_mask = mask
if i == 0:
orig_data = local_hetero_edge_addition(orig_data, ratio = ratio, edges=0, dataset = args.dataset, mask=mask, mode = mode)
else:
orig_data = local_hetero_edge_addition(orig_data, ratio = ratio, edges=unit, node_label_cat=orig_data.original_index, dataset = args.dataset, mask=mask, mode = mode)
save_pyg_data(orig_data, i * unit, ratio, args.dataset, orig_mask)
elif args.dataset == 'PubMed':
orig_data = deepcopy(data)
for i in range(10):
mask = 'train'
if mask == 'train':
ratio = 1.0
unit = 300
else:
ratio = 0.05
unit = 300
mode = 'tohete'
orig_mask = mask
if i == 0:
orig_data = local_hetero_edge_addition(orig_data, ratio = ratio, edges=0, dataset = args.dataset, mask=mask, mode = mode)
else:
orig_data = local_hetero_edge_addition(orig_data, ratio = ratio, edges=unit, node_label_cat=orig_data.original_index, dataset = args.dataset, mask=mask, mode = mode)
save_pyg_data(orig_data, i * unit, ratio, args.dataset, orig_mask)
else:
orig_data = deepcopy(data)
for i in range(10):
mask = 'test'
mode = 'tohomo'
orig_mask = mask
if mask == 'train':
ratio = 1.0
edge_unit = 1500
else:
ratio = 0.15
edge_unit = 100
if i == 0:
orig_data = local_hetero_edge_addition(orig_data, ratio = ratio, edges=0, dataset = args.dataset, mask=mask, mode = mode)
else:
orig_data = local_hetero_edge_addition(orig_data, ratio=ratio, edges=edge_unit, node_label_cat=orig_data.original_index, dataset = args.dataset, mask = mask, mode=mode)
mask = torch.zeros(orig_data.x.shape[0], dtype=torch.bool)
total_sampled_node_idxs = torch.LongTensor(list(itertools.chain(*orig_data.original_index.values())))
mask[total_sampled_node_idxs] = True
save_pyg_data(orig_data, i * edge_unit, ratio, args.dataset, orig_mask)
def fix_param_run(args, seeds):
dataset, odata = get_dataset(args)
acc_list = []
odata.edge_index, _ = torch_geometric.utils.remove_self_loops(torch_geometric.utils.to_undirected(odata.edge_index))
split_seeds = np.array(range(10))
best_acc_across_split = -1
all_preds = []
all_res = []
# ipdb.set_trace()
for split_idx in range(args.num_split):
split_seed = split_seeds[split_idx].item()
data, _ = get_split(args, dataset, odata, split_seed)
data.edge_index, _ = torch_geometric.utils.remove_self_loops(torch_geometric.utils.to_undirected(data.edge_index))
data = data.cuda()
vars(args)["num_node"] = data.x.shape[0]
vars(args)["num_feat"] = data.x.shape[1]
vars(args)["num_class"] = max(data.y).item() + 1
vars(args)["split_seed"] = split_seed
for random_seed in seeds:
args.random_seed = random_seed
set_seed_config(random_seed)
vars(args)["split_seed"] = split_seed
vars(args)["num_node"] = data.x.shape[0]
vars(args)["num_feat"] = data.x.shape[1]
vars(args)["num_class"] = max(data.y).item() + 1
if args.teacher != 'No':
acc_teacher, acc_student, out_student, pred_student, _ = main_transfer(data, args)
acc = acc_student if args.is_distill else acc_teacher
all_preds.append(pred_student)
acc_list.append(acc)
#acc_list.append(acc)
#all_preds.append(pred)
all_res.append(out_student)
res = out_student
pred = pred_student
else:
if args.train_schema == "forward":
acc, res, pred, _, best_model = main_forward(data, args)
elif args.train_schema == "backward":
acc, res, pred, _ = main_backward(data, args)
best_model = None
acc_list.append(acc)
all_preds.append(pred)
all_res.append(res)
if best_acc_across_split < acc:
best_acc_across_split = acc
acc = torch.sum(all_res[-1][data.test_mask].argmax(-1) == data.y[data.test_mask]).item() / torch.sum(data.test_mask).item()
acc_mean, acc_std = np.mean(acc_list), np.std(acc_list)
logging.info(f'acc: {acc_mean}')
logging.info(f'acc_std: {acc_std}')
logging.info(f'acc_log: {acc_list}')
if args.best_run:
if not args.is_synthetic:
logging.info(f"Best run: {args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-{args.expmode}")
save_best(args, osp.join(args.best_path, f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-args.pkl"))
save_best(all_preds, osp.join(args.best_path, f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-pred.pkl"))
save_best(all_res, osp.join(args.best_path, f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-res.pkl"))
else:
logging.info(f"Best run: {args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-{args.expmode}")
save_best(args, osp.join(args.best_path, f"{args.algo_name}-{args.change}-args.pkl"))
save_best(all_preds, osp.join(args.best_path, f"{args.algo_name}-{args.change}-pred.pkl"))
save_best(all_res, osp.join(args.best_path, f"{args.algo_name}-{args.change}-res.pkl"))
else:
save_best(args, osp.join(args.save_path, f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-args.pkl"))
save_best(all_preds, osp.join(args.save_path, f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-pred.pkl"))
save_best(all_res, osp.join(args.save_path, f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-res.pkl"))
# ipdb.set_trace()
print(args)
return acc_mean
# return best_model
def inductive_run(args, seeds):
dataset, odata = get_dataset(args)
acc_list = []
odata.edge_index, _ = torch_geometric.utils.remove_self_loops(torch_geometric.utils.to_undirected(odata.edge_index))
split_seeds = np.array(range(10))
all_preds = []
all_res = []
ind_list = []
for split_idx in range(args.num_split):
split_seed = split_seeds[split_idx].item()
data, _ = get_split(args, dataset, odata, split_seed)
vars(args)["num_node"] = data.x.shape[0]
vars(args)["num_feat"] = data.x.shape[1]
vars(args)["num_class"] = max(data.y).item() + 1
data.edge_index, _ = torch_geometric.utils.remove_self_loops(torch_geometric.utils.to_undirected(data.edge_index))
data = data.cuda()
vars(args)["split_seed"] = split_seed
indices = graph_split(data.train_idx, data.val_idx, data.test_idx, args.inductive_rate, split_seed)
obs_idx_train, obs_idx_val, obs_idx_test, idx_obs, idx_test_ind = indices
data.obs_idx_train = obs_idx_train
data.obs_idx_val = obs_idx_val
data.obs_idx_test = obs_idx_test
data.idx_obs = idx_obs.sort().values
data.idx_test_ind = idx_test_ind
for random_seed in seeds:
args.random_seed = random_seed
set_seed_config(random_seed)
vars(args)["split_seed"] = split_seed
vars(args)["num_node"] = data.x.shape[0]
vars(args)["num_feat"] = data.x.shape[1]
vars(args)["num_class"] = max(data.y).item() + 1
if args.teacher != 'No':
acc_teacher, acc_student, out_student, pred_student, ind_test_acc = main_transfer(data, args)
acc = acc_student if args.is_distill else acc_teacher
all_preds.append(pred_student)
acc_list.append(acc)
all_res.append(out_student)
ind_list.append(ind_test_acc)
continue
if args.train_schema == "forward":
acc, res, pred, ind_acc, best_model = main_forward(data, args)
elif args.train_schema == "backward":
acc, res, pred, ind_acc = main_backward(data, args)
acc_list.append(acc)
all_res.append(res)
all_preds.append(pred)
ind_list.append(ind_acc)
acc_mean, acc_std = np.mean(acc_list), np.std(acc_list)
logging.info(f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}")
logging.info(f"test_acc: {acc_mean}, test_acc_std: {acc_std}")
# ipdb.set_trace()
ind_acc_mean, ind_acc_std = np.mean(ind_list), np.std(ind_list)
logging.info(f"ind_test_acc: {ind_acc_mean}, ind_test_acc_std: {ind_acc_std}")
idx = {'obs_idx_train':data.obs_idx_train, 'obs_idx_val':data.obs_idx_val, 'obs_idx_test':data.obs_idx_test, 'idx_test_ind':data.idx_test_ind}
if args.best_run:
logging.info(f"Best run: {args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-{args.expmode}")
save_best(args, osp.join(args.best_path, "inductive", f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-args.pkl"))
save_best(all_preds, osp.join(args.best_path, "inductive", f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-pred.pkl"))
save_best(all_res, osp.join(args.best_path, "inductive", f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-res.pkl"))
save_best(idx, osp.join(args.best_path, "inductive", f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-mask.pkl"))
else:
save_best(args, osp.join(args.save_path, "inductive", f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-args.pkl"))
save_best(all_preds, osp.join(args.save_path, "inductive", f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-pred.pkl"))
save_best(all_res, osp.join(args.save_path, "inductive", f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-res.pkl"))
save_best(idx, osp.join(args.save_path, "inductive", f"{args.exp_name}-{args.algo_name}-{args.dataset}-{args.num_split}-{args.is_new}-mask.pkl"))
print(args)
def sweep_run(trial, args, seeds, param_f):
if args.dataset in ['Cora', 'CiteSeer', 'PubMed'] or 'Cora' in args.dataset:
mode = 'normal'
elif args.dataset == 'arxiv' or args.dataset == 'IGB_tiny':
mode = 'arxiv'
else:
mode = 'hete'
params = param_f(trial, args.algo_name, mode)
dataset, odata = get_dataset(args)
args = update_parameter(args, params)
acc_list = []
odata.edge_index, _ = torch_geometric.utils.remove_self_loops(torch_geometric.utils.to_undirected(odata.edge_index))
split_seeds = np.array(range(10))
best_acc_across_split = -1
ind_list = []
seeds = [1]
for split_idx in range(args.sweep_split):
split_seed = split_seeds[split_idx].item()
data, _ = get_split(args, dataset, odata, split_seed)
data.edge_index, _ = torch_geometric.utils.remove_self_loops(torch_geometric.utils.to_undirected(data.edge_index))
data = data.cuda()
vars(args)["split_seed"] = split_seed
vars(args)["num_node"] = data.x.shape[0]
vars(args)["num_feat"] = data.x.shape[1]
vars(args)["num_class"] = max(data.y).item() + 1
if args.expmode == 'inductive':
indices = graph_split(data.train_idx, data.val_idx, data.test_idx, args.inductive_rate, split_seed)
obs_idx_train, obs_idx_val, obs_idx_test, idx_obs, idx_test_ind = indices
data.obs_idx_train = obs_idx_train
data.obs_idx_val = obs_idx_val
data.obs_idx_test = obs_idx_test
data.idx_obs = idx_obs.sort().values
data.idx_test_ind = idx_test_ind
for random_seed in seeds:
set_seed_config(random_seed)
args.random_seed = random_seed
set_seed_config(random_seed)
vars(args)["split_seed"] = split_seed
vars(args)["num_node"] = data.x.shape[0]
vars(args)["num_feat"] = data.x.shape[1]
vars(args)["num_class"] = max(data.y).item() + 1
if args.teacher != 'No':
acc_teacher, acc_student, out_student, pred_student, ind_acc = main_transfer(data, args)
acc = acc_student if args.is_distill else acc_teacher
acc_list.append(acc)
ind_list.append(ind_acc)
continue
if args.train_schema == "forward":
acc, res, pred, ind_acc, best_model = main_forward(data, args)
elif args.train_schema == "backward":
acc, res, pred, ind_acc = main_backward(data, args)
acc_list.append(acc)
ind_list.append(ind_acc)
if best_acc_across_split < acc:
best_acc_across_split = acc
acc_mean, acc_std = np.mean(acc_list), np.std(acc_list)
logging.info(f"test_acc: {acc_mean}, test_acc_std: {acc_std}")
ind_acc_mean, ind_acc_std = np.mean(ind_list), np.std(ind_list)
logging.info(f"ind_test_acc: {ind_acc_mean}, ind_test_acc_std: {ind_acc_std}")
if args.expmode == 'transductive':
return acc_mean
else:
return ind_acc_mean
def gen(dataset):
params = {}
params["single"] = False
args = generate_args()
args = update_parameter(args, params)
args = check_parameter(args)
args = generate_prefix(args)
vars(args)['dataset'] = dataset
vars(args)['is_new'] = 1
vars(args)['is_fix'] = 1
vars(args)['expmode'] = 'transductive'
_, data = get_dataset(args)
if args.is_new and args.dataset in ["Cora", "CiteSeer", "PubMed", "photo", "cs"]:
data = torch.load(f"/mnt/home/haitaoma/Graph-smooth/code/CPF/dataset/{args.dataset}.pt")
data.edge_index = data.edge_index.to(torch.int64)
orig_data = deepcopy(data)
total_edges = data.edge_index.shape[1] // 2
for i in range(11):
mask = torch.ones(orig_data.x.shape[0], dtype=torch.bool)
if i == 0:
K = 0
orig_data = hetero_edge_addition(orig_data, K = 0, mode = 'tohete')
else:
K = int(i * 2 / 10 * total_edges)
if dataset == 'Cora':
orig_data = hetero_edge_addition(orig_data, K = int(2 / 10 * total_edges), mode = 'tohete')
elif dataset == 'Squirrel':
orig_data = hetero_edge_addition(orig_data, K = int(2 / 10 * total_edges), mode = 'tohomo')
filename = osp.join("/mnt/home/haitaoma/Graph-smooth/code/synthetic_data/full", f"full-syn-{dataset}-{K}.pt")
torch.save(orig_data, filename)
if __name__ == '__main__':
main()
#gen('Cora')
# gen('Squirrel')