-
Notifications
You must be signed in to change notification settings - Fork 1
/
main.py
193 lines (156 loc) · 7.25 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
import torch, pickle, time, os
import numpy as np
from options import parse_args
from torch.utils.data import DataLoader
from model import HGMN, BPRLoss
from data_utils import prepare_dgl_graph, MyDataset
from utils import load_data, load_model, save_model, fix_random_seed_as
from tqdm import tqdm
class Model():
def __init__(self, args):
self.args = args
self.device = torch.device('cuda' if args.cuda and torch.cuda.is_available() else 'cpu')
self.dataset = load_data(args.data_path)
val_neg_data = load_data(args.val_neg_path)
test_neg_data = load_data(args.test_neg_path)
trainset = MyDataset(self.dataset, 'train')
valset = MyDataset(self.dataset, 'val', val_neg_data)
testset = MyDataset(self.dataset, 'test', test_neg_data)
self.trainloader = DataLoader(
dataset=trainset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers
)
self.valloader = DataLoader(
dataset=valset,
batch_size=args.test_batch_size * 101,
shuffle=False,
num_workers=args.num_workers
)
self.testloader = DataLoader(
dataset=testset,
batch_size=args.test_batch_size * 101,
shuffle=False,
num_workers=args.num_workers
)
self.graph = prepare_dgl_graph(args, self.dataset).to(self.device)
self.model = HGMN(args, self.dataset['userCount'], self.dataset['itemCount'], self.dataset['categoryCount'])
self.model = self.model.to(self.device)
self.criterion = BPRLoss(args.reg)
self.optimizer = torch.optim.Adam([
{'params': self.model.parameters()}
], lr=args.lr)
self.scheduler = torch.optim.lr_scheduler.StepLR(
self.optimizer,
step_size=args.decay_step,
gamma=args.decay
)
if args.checkpoint:
load_model(self.model, args.checkpoint, self.optimizer)
def train(self):
args = self.args
best_hr, best_ndcg, best_epoch, wait = 0, 0, 0, 0
start_time = time.time()
for self.epoch in range(1, args.n_epoch + 1):
epoch_losses = self.train_one_epoch(self.trainloader, self.graph)
print('epoch {} done! elapsed {:.2f}.s, epoch_losses {}'.format(
self.epoch, time.time() - start_time, epoch_losses
), flush=True)
hr, ndcg = self.validate(self.testloader, self.graph)
cur_best = hr + ndcg > best_hr + best_ndcg
if cur_best:
best_hr, best_ndcg, best_epoch = hr, ndcg, self.epoch
wait = 0
else:
wait += 1
print('+ epoch {} tested, elapsed {:.2f}s, N@{}: {:.4f}, R@{}: {:.4f}'.format(
self.epoch, time.time() - start_time, args.topk, ndcg, args.topk, hr
), flush=True)
if args.model_dir and cur_best:
desc = f'{args.dataset}_hid_{args.n_hid}_layer_{args.n_layers}_mem_{args.mem_size}_' + \
f'lr_{args.lr}_reg_{args.reg}_decay_{args.decay}_step_{args.decay_step}_batch_{args.batch_size}'
perf = '' # f'N/R_{ndcg:.4f}/{hr:.4f}'
fname = f'{args.desc}_{desc}_{perf}.pth'
save_model(self.model, os.path.join(args.model_dir, fname), self.optimizer)
if wait >= args.patience:
print(f'Early stop at epoch {self.epoch}, best epoch {best_epoch}')
break
print(f'Best N@{args.topk} {best_ndcg:.4f}, R@{args.topk} {best_hr:.4f}', flush=True)
def train_one_epoch(self, dataloader, graph):
self.model.train()
epoch_losses = [0] * 2
dataloader.dataset.neg_sample()
tqdm_dataloader = tqdm(dataloader)
for iteration, batch in enumerate(tqdm_dataloader):
user_idx, pos_idx, neg_idx = batch
rep, user_pool = self.model(graph)
user = rep[user_idx] + user_pool[user_idx]
pos = rep[self.model.n_user + pos_idx]
neg = rep[self.model.n_user + neg_idx]
pos_preds = self.model.predict(user, pos)
neg_preds = self.model.predict(user, neg)
loss, losses = self.criterion(pos_preds, neg_preds, user, pos, neg)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
epoch_losses = [x + y for x, y in zip(epoch_losses, losses)]
tqdm_dataloader.set_description('Epoch {}, loss: {:.4f}'.format(self.epoch, loss.item()))
if self.scheduler is not None:
self.scheduler.step()
epoch_losses = [sum(epoch_losses)] + epoch_losses
return epoch_losses
def calc_hr_and_ndcg(self, preds, topk):
preds = preds.reshape(-1, 101)
labels = torch.zeros_like(preds)
labels[:, 0] = 1
_, indices = preds.topk(topk)
hits = labels.gather(1, indices)
hrs = hits.sum(1).tolist()
weights = 1 / torch.log2(torch.arange(2, 2 + topk).float()).to(hits.device)
ndcgs = (hits * weights).sum(1).tolist()
return hrs, ndcgs
def validate(self, dataloader, graph):
self.model.eval()
hrs, ndcgs = [], []
with torch.no_grad():
tqdm_dataloader = tqdm(dataloader)
for iteration, batch in enumerate(tqdm_dataloader, start=1):
user_idx, item_idx = batch
rep, user_pool = self.model(graph)
user = rep[user_idx] + user_pool[user_idx]
item = rep[self.model.n_user + item_idx]
preds = self.model.predict(user, item)
preds_hrs, preds_ndcgs = self.calc_hr_and_ndcg(preds, self.args.topk)
hrs += preds_hrs
ndcgs += preds_ndcgs
return np.mean(hrs), np.mean(ndcgs)
def test(self):
load_model(self.model, args.checkpoint)
self.model.eval()
with torch.no_grad():
rep, user_pool = self.model(self.graph)
""" Save embeddings """
user_emb = (rep[:self.model.n_user] + user_pool).cpu().numpy()
item_emb = rep[self.model.n_user: self.model.n_user + self.model.n_item].cpu().numpy()
with open(f'HGMN-{self.args.dataset}-embeds.pkl', 'wb') as f:
pickle.dump({'user_embed': user_emb, 'item_embed': item_emb}, f)
""" Save results """
tqdm_dataloader = tqdm(self.testloader)
uids, hrs, ndcgs = [], [], []
for iteration, batch in enumerate(tqdm_dataloader, start=1):
user_idx, item_idx = batch
user = rep[user_idx] + user_pool[user_idx]
item = rep[self.model.n_user + item_idx]
preds = self.model.predict(user, item)
preds_hrs, preds_ndcgs = self.calc_hr_and_ndcg(preds, self.args.topk)
hrs += preds_hrs
ndcgs += preds_ndcgs
uids += user_idx[::101].tolist()
with open(f'HGMN-{self.args.dataset}-test.pkl', 'wb') as f:
pickle.dump({uid: (hr, ndcg) for uid, hr, ndcg in zip(uids, hrs, ndcgs)}, f)
if __name__ == "__main__":
args = parse_args()
fix_random_seed_as(args.seed)
app = Model(args)
app.train()