-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathfinetune_dynamic.py
137 lines (109 loc) · 5.65 KB
/
finetune_dynamic.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
import sys
sys.path.append('./')
from os import path
from utils.parse_args import args
from utils.dataloader import EdgeListData
from tqdm import tqdm
# import dgl
import random
import numpy as np
import torch
from utils.logger import Logger, log_exceptions
from modules.dynamicGNN.trainer_roland import Trainer
import importlib
import setproctitle
import pandas as pd
import datetime
setproctitle.setproctitle('GraphPro')
modules_class = 'modules.dynamicGNN.'
def import_pretrained_model():
module = importlib.import_module('modules.LightGCN')
return getattr(module, 'LightGCN')
def import_finetune_model():
module = importlib.import_module(modules_class + args.f_model)
return getattr(module, args.f_model)
def init_seed(seed):
random.seed(seed)
np.random.seed(seed)
# dgl.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
init_seed(args.seed)
logger = Logger(args)
pretrain_data = path.join(args.data_path, "pretrain.txt")
pretrain_val_data = path.join(args.data_path, "pretrain_val.txt")
finetune_data = path.join(args.data_path, "fine_tune.txt")
test_data_num = 8 if args.data_path.split('/')[-1] == 'amazon' else 4
logger.log(f"test_data_num: {test_data_num}")
test_datas = [path.join(args.data_path, f"test_{i}.txt") for i in range(1, test_data_num+1)]
all_data = [pretrain_data, pretrain_val_data, finetune_data, *test_datas]
recalls, ndcgs = [], []
if args.f_model == 'roland':
@log_exceptions
def run():
pretrain_dataset = EdgeListData(pretrain_data, pretrain_val_data)
# LightGCN
pretrain_model = import_pretrained_model()(pretrain_dataset, phase='vanilla').to(args.device)
pretrain_model.load_state_dict(torch.load(args.pre_model_path), strict=False)
pretrain_model.eval()
meta_model_sd = pretrain_model.state_dict()
for num_stage in range(1,len(test_datas)+1):
test_data_idx = num_stage + 2
ft_data_idx = test_data_idx - 1
finetune_dataset = EdgeListData(train_file=all_data[ft_data_idx], test_file=all_data[test_data_idx], phase='finetune', pre_dataset=pretrain_dataset, has_time=True, user_hist_files=all_data[:ft_data_idx])
if num_stage == 1:
model = import_finetune_model()(finetune_dataset, pretrain_model, meta_model=pretrain_model).to(args.device)
if num_stage > 1:
# update model for next stage
model = import_finetune_model()(finetune_dataset, meta_model=updated_model).to(args.device)
print(model)
logger.info(f"ROLAND Learning Stage {num_stage}, test data: {all_data[test_data_idx]}, incremental train data: {all_data[ft_data_idx]}")
trainer = Trainer(finetune_dataset, logger)
best_perform = trainer.train_finetune(model)
recalls.append(best_perform['recall'][0])
ndcgs.append(best_perform['ndcg'][0])
# update meta model
# reload the best model
model.load_state_dict(torch.load(trainer.save_path))
model.meta_model = None
updated_model = model.update_meta_model(model, meta_model_sd)
meta_model_sd = updated_model.state_dict()
# update exp time for saving new model
args.exp_time = datetime.datetime.now().strftime('%b-%d-%Y_%H-%M-%S')
logger.info(f"recalls: {recalls} \n ndcgs: {ndcgs} \n avg. recall: {np.round(np.mean(recalls), 4)}, avg. ndcg: {np.round(np.mean(ndcgs), 4)}")
run()
elif args.f_model in ['evolveGCN_H', 'evolveGCN_O']:
@log_exceptions
def run():
pretrain_dataset = EdgeListData(pretrain_data, pretrain_val_data)
# LightGCN
pretrain_model = import_pretrained_model()(pretrain_dataset, phase='vanilla').to(args.device)
pretrain_model.load_state_dict(torch.load(args.pre_model_path), strict=False)
pretrain_model.eval()
for num_stage in range(1,len(test_datas)+1):
test_data_idx = num_stage + 2
ft_data_idx = test_data_idx - 1
finetune_dataset = EdgeListData(train_file=all_data[ft_data_idx], test_file=all_data[test_data_idx], phase='finetune', pre_dataset=pretrain_dataset, has_time=True, user_hist_files=all_data[:ft_data_idx])
if num_stage == 1:
last_emb = torch.concat(pretrain_model.generate(), dim=0)
model = import_finetune_model()(finetune_dataset, pretrain_model, last_emb).to(args.device)
if num_stage > 1:
# update model for next stage
model = import_finetune_model()(finetune_dataset, last_emb=last_emb).to(args.device)
print(model)
logger.info(f"EvolveGCN Learning Stage {num_stage}, test data: {all_data[test_data_idx]}, incremental train data: {all_data[ft_data_idx]}")
trainer = Trainer(finetune_dataset, logger)
best_perform = trainer.train_finetune(model)
recalls.append(best_perform['recall'][0])
ndcgs.append(best_perform['ndcg'][0])
if args.f_model == 'evolveGCN_H':
last_emb = torch.concat(model.generate(), dim=0)
elif args.f_model == 'evolveGCN_O':
last_emb = torch.concat([model.user_embedding, model.item_embedding], dim=0)
# update exp time for saving new model
args.exp_time = datetime.datetime.now().strftime('%b-%d-%Y_%H-%M-%S')
logger.info(f"recalls: {recalls} \n ndcgs: {ndcgs} \n avg. recall: {np.round(np.mean(recalls), 4)}, avg. ndcg: {np.round(np.mean(ndcgs), 4)}")
run()