-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.py
168 lines (136 loc) · 6.93 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
# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
# SPDX-License-Identifier: Apache-2.0
import os
import time
import torch
import argparse
import pickle
from model import SASRec
from utils import *
import torch.nn as nn
from numpy.linalg import inv
import matplotlib.pyplot as plt
with open('user_item_all.pkl', 'rb') as f:
user_item = pickle.load(f)
with open('item_emb.pkl', 'rb') as f:
item_emb = pickle.load(f)
with open('item_llm.pkl', 'rb') as f:
item_llm = pickle.load(f)
with open('item_dist.pkl', 'rb') as f:
item_dist = pickle.load(f)
# with open('item_cov.pkl', 'rb') as f:
# item_cov = pickle.load(f)
item_llm = np.stack(item_llm)
print("Data loaded!")
class params:
def __init__(self):
self.num_epochs = 20
self.usernum = len(user_item)
self.maxlen = 100
self.batch_size = 64
self.train = True
self.train_rate = 0.8
self.itemnum = 209171 # not consequtive numbers, 62423 in total
self.item_emb_dim = item_emb[1][0].shape[0]
self.l2_emb = 0
self.layer = 2
self.num_heads = 4
self.num_blocks = 2
self.dropout_rate = 0.5
self.hidden_units = 32
self.lr = 0.001
# self.rho = 0.0001
self.rho =0.01
self.k = 71
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
args = params()
num_batch = len(user_item) // args.batch_size # tail? + ((len(user_train) % args.batch_size) != 0)
sampler = WarpSampler(user_item, item_emb, args.usernum, args.itemnum, args.item_emb_dim, args.train, args.train_rate, batch_size=args.batch_size, maxlen=args.maxlen, n_workers=3)
print("Sampler made!")
model = SASRec(args.usernum, args.itemnum, args).to(args.device) # no ReLU activation in original SASRec implementation?
model = model.float()
for name, param in model.named_parameters():
try:
torch.nn.init.xavier_normal_(param.data)
except:
pass # just ignore those failed init layers
model.train() # enable model training
epoch_start_idx = 1
bce_criterion = torch.nn.BCEWithLogitsLoss() # torch.nn.BCELoss()
adam_optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, betas=(0.9, 0.98))
T = 0.0
t0 = time.time()
loss_best = 10000000000
loss_rec = 0
loss_rcd = []
loss_reg_rcd = []
loss_pos_rcd = []
loss_neg_rcd = []
for epoch in range(epoch_start_idx, args.num_epochs + 1):
for step in range(num_batch): # tqdm(range(num_batch), total=num_batch, ncols=70, leave=False, unit='b'):
_, seq, pos, seq_emb, pos_emb, neg_emb = sampler.next_batch() # tuples to ndarray
seq, pos, seq_emb, pos_emb, neg_emb = np.array(seq), np.array(pos), np.array(seq_emb), np.array(pos_emb), np.array(neg_emb)
pos_logits, neg_logits, seqs_out = model(seq, seq_emb, pos_emb, neg_emb)
pos_labels, neg_labels = torch.ones(pos_logits.shape, device=args.device), torch.zeros(neg_logits.shape, device=args.device)
# regularization
# z
llm_emb_z = model.item_emb(torch.from_numpy(item_llm).to(args.device))
## item sequence
seq = seq.flatten()
idx = np.where(seq != 0)
seq = seq[idx].tolist()
## Trained embedding
seqs_out = seqs_out.reshape(-1, args.hidden_units)
seqs_out = seqs_out[idx, :].squeeze() # len(seq), hidden_units
## Trained embedding of nearest items
seq_z = torch.stack([llm_emb_z[item_dist[item][:args.k], :] for item in seq]).reshape(len(seq), args.k, -1) # len(seq), k, hidden_units
seq_z = torch.sum((seqs_out.unsqueeze(1) - seq_z)**2, dim = -1) # len(seq), k
seq_z = seq_z[:, 1:]
# x
## LLM embedding
seq_emb = seq_emb.reshape(-1, 384)
seq_emb = seq_emb[idx, :]
## Cov
cov = np.stack([np.std(item_llm[item_dist[item][:args.k], :], axis = 0) for item in seq]) # len(seq), 384
cov = 1 / (cov + 1e-3)
cov = np.repeat(cov[:, np.newaxis, :], args.k, axis = 1) # len(seq), k, 384
## LLM embedding of nearest items
seq_x = np.stack([item_llm[item_dist[item][:args.k], :] for item in seq]).reshape(len(seq), args.k, -1) # len(seq), k, 384
seq_x = np.sum(((seq_emb.squeeze()[:, np.newaxis, :] - seq_x)**2) * cov, axis = -1) # len(seq), k
seq_x = torch.exp(-0.5*torch.from_numpy(seq_x)).to(args.device)
seq_x = seq_x[:, 1:]
loss_reg = torch.sum(seq_x * seq_z)
adam_optimizer.zero_grad()
indices = np.where(pos != 0)
loss_pos = bce_criterion(pos_logits[indices], pos_labels[indices])
loss_neg = bce_criterion(neg_logits[indices], neg_labels[indices])
loss = loss_pos + loss_neg + args.rho * loss_reg
loss_rcd.append(loss.detach().cpu().numpy().item())
loss_reg_rcd.append(loss_reg.detach().cpu().numpy().item())
loss_pos_rcd.append(loss_pos.detach().cpu().numpy().item())
loss_neg_rcd.append(loss_neg.detach().cpu().numpy().item())
for param in model.item_emb.parameters():
loss += args.l2_emb * torch.norm(param)
loss_rec += loss.item()
loss.backward()
adam_optimizer.step()
print("loss in epoch {} iteration {}: {} loss_reg={} loss_pos={} loss_neg={}".format(epoch, step, loss.item(), loss_reg.item(), loss_pos.item(), loss_neg.item())) # expected 0.4~0.6 after init few epochs
loss_rec /= num_batch
if loss_rec < loss_best:
folder = "./model_save"
fname = 'SASRec.epoch={}.lr={}.layer={}.head={}.hidden={}.maxlen={}.rho={}.pth'.format(args.num_epochs, args.lr, args.num_blocks, args.num_heads, args.hidden_units, args.maxlen, args.rho)
# fname = fname.format(args.num_epochs, args.lr, args.num_blocks, args.num_heads, args.hidden_units, args.maxlen)
torch.save(model.state_dict(), os.path.join(folder, fname))
loss_rec = 0
plt.figure()
plt.plot(loss_rcd)
plt.savefig('./model_save/loss.epoch={}.lr={}.layer={}.head={}.hidden={}.maxlen={}.rho={}.png'.format(args.num_epochs, args.lr, args.num_blocks, args.num_heads, args.hidden_units, args.maxlen, args.rho))
plt.figure()
plt.plot(loss_pos_rcd)
plt.savefig('./model_save/loss_pos.epoch={}.lr={}.layer={}.head={}.hidden={}.maxlen={}.rho={}.png'.format(args.num_epochs, args.lr, args.num_blocks, args.num_heads, args.hidden_units, args.maxlen, args.rho))
plt.figure()
plt.plot(loss_neg_rcd)
plt.savefig('./model_save/loss_neg.epoch={}.lr={}.layer={}.head={}.hidden={}.maxlen={}.rho={}.png'.format(args.num_epochs, args.lr, args.num_blocks, args.num_heads, args.hidden_units, args.maxlen, args.rho))
plt.figure()
plt.plot(loss_reg_rcd)
plt.savefig('./model_save/loss_reg.epoch={}.lr={}.layer={}.head={}.hidden={}.maxlen={}.rho={}.png'.format(args.num_epochs, args.lr, args.num_blocks, args.num_heads, args.hidden_units, args.maxlen, args.rho))