MM4Rec/inference.py at main · idssplab/MM4Rec · 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
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
import glob,pickle,joblib,xlsxwriter
import seaborn as sns
from tqdm import tqdm
from dateutil import parser
import argparse
from sklearn.preprocessing import LabelEncoder
from typing import Optional, Any, Union, Callable,Tuple
import copy,math,itertools,random,zipfile
import uuid,json,time
from collections import OrderedDict
import torch
import torch.nn.functional as F
import torch.optim as optim
import torch.nn as nn
from torch.autograd import Variable
from torch.utils.data import Dataset,DataLoader
from collections import Counter
from transformer import *
from Dataset import *
from utils import *
from model import *
from trainer import *
print(torch.__version__)
train_on_gpu = torch.cuda.is_available()
if not train_on_gpu:
    print('CUDA is not available.  Training on CPU ...')
else:
    print('CUDA is available!  Training on GPU ...')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


parser = argparse.ArgumentParser()
# basic parameters
parser.add_argument('--seed',type=int,default=100)
parser.add_argument('--learning_rate',type=float,default=3e-4,help='Specify learning rate for optimizer. (default: 3e-4)')
parser.add_argument('--weight_decay',type=float,default=1e-3,help='Specify weight_decay for optimizer. (default: 1e-3)')
parser.add_argument('--dropout',type=float,default=0.5)
parser.add_argument('--num_epoch',type=int,default=60,help='Number of training epochs. (default: 20)')
parser.add_argument('--batch_size',type=int,default=128,help='Batch size for data loaders. (default: 32)' )
parser.add_argument('--pad_id',type=int,default=999999,help='padding_id (default: 999999)')
parser.add_argument('--add_PE',default=True,help='positional encoding (default:true)')
parser.add_argument('--d_model',type=int,default=128)
parser.add_argument('--nhead',type=int,default=16)
parser.add_argument('--num_encoder_layers', type=int,default=2)
parser.add_argument( '--num_expert',type=int,default=3)
parser.add_argument('--ff_dim',type=int, default=128 )
parser.add_argument('--llm_size',type=int, default=4096 )
parser.add_argument('--ckpt_path',type=str, default='../ckpt/AviviD_DatasetA/best_ckpt_1fe94')
parser.add_argument('--dataset',type=str, default='AviviD_DatasetA' )
parser.add_argument('--model_name',type=str, default='MM4Rec' )


if __name__ == "__main__":
    args = parser.parse_args()

    # read the testing dataset from 2 source and scenario

    args.dataset = 'AviviD_DatasetA'

    click_seq_train_1= load('../data/'+str(args.dataset)+'/Push_Ads/click_seq_train_v.pkl')
    click_seq_train_2= load('../data/'+str(args.dataset)+'/Browse_Ads/click_seq_train_v.pkl')
    click_seq_train_3= load('../data/'+str(args.dataset)+'/Push_News/click_seq_train_v.pkl')
    click_seq_train_4= load('../data/'+str(args.dataset)+'/Browse_News/click_seq_train_v.pkl')

    click_seq_val_1= load('../data/'+str(args.dataset)+'/Push_Ads/click_seq_val_v.pkl')
    click_seq_val_2= load('../data/'+str(args.dataset)+'/Browse_Ads/click_seq_val_v.pkl')
    click_seq_val_3= load('../data/'+str(args.dataset)+'/Push_News/click_seq_val_v.pkl')
    click_seq_val_4= load('../data/'+str(args.dataset)+'/Browse_News/click_seq_val_v.pkl')

    click_seq_test_1= load('../data/'+str(args.dataset)+'/Push_Ads/click_seq_test_v.pkl')
    click_seq_test_2= load('../data/'+str(args.dataset)+'/Browse_Ads/click_seq_test_v.pkl')
    click_seq_test_3= load('../data/'+str(args.dataset)+'/Push_News/click_seq_test_v.pkl')
    click_seq_test_4= load('../data/'+str(args.dataset)+'/Browse_News/click_seq_test_v.pkl')


    click_seq_train=pd.concat([click_seq_train_1,click_seq_train_2,click_seq_train_3,click_seq_train_4],axis=0).reset_index(drop=True)
    click_seq_val =pd.concat([click_seq_val_1,click_seq_val_2,click_seq_val_3,click_seq_val_4],axis=0).reset_index(drop=True)
    click_seq_test=pd.concat([click_seq_test_1,click_seq_test_2,click_seq_test_3,click_seq_test_4],axis=0).reset_index(drop=True)
    print(click_seq_test.shape)


     # create ads/news feature matrix
    banner_id_data=load('../data/'+str(args.dataset)+'/banner_id_data.pkl')
    news_article_data=load('../data/'+str(args.dataset)+'/news_article_data.pkl')

    # padding operation
    banner_id_data['section_encode']=21
    news_article_data['charge_mode_encode']=3
    news_article_data['unit_price_encode']=21

    ads_feature_matrix=torch.from_numpy(banner_id_data[['charge_mode_encode','unit_price_encode','section_encode','banner_id_encode']].values)
    ads_feature_matrix=torch.cat([torch.Tensor(banner_id_data.title_emb_taide7b),ads_feature_matrix],1)
    ads_feature_matrix=torch.cat([ads_feature_matrix,torch.cat([torch.tensor([0]*4096),torch.tensor([4,22,22,999999])]).unsqueeze(0)])
    news_feature_matrix=torch.from_numpy(news_article_data[['charge_mode_encode','unit_price_encode','section_encode','url_encode']].values)
    news_feature_matrix=torch.cat([torch.Tensor(news_article_data.title_emb_taide7b),news_feature_matrix],1)
    news_feature_matrix=torch.cat([news_feature_matrix,torch.cat([torch.tensor([0]*4096),torch.tensor([4,22,22,999999])]).unsqueeze(0)])

    # set seed
    my_seed = args.seed #100
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False
    np.random.seed(my_seed)
    torch.manual_seed(my_seed)

    # get model
    def get_model(model, model_params):
        models = {
            'MM4Rec':MM4Rec,
            }
        return models.get(model)(**model_params)

    # model & ckpt_path
    model_name = args.model_name #'MM4Rec'
    ckpt_path = args.ckpt_path

    # training parameters
    weight_decay =  args.weight_decay    #1e-3
    learning_rate = args.learning_rate #3e-4
    num_epoch =   args.num_epoch        #60
    batch_size = args.batch_size      #2048
    seq_len = 20
    pad_id=args.pad_id                #999999

    if (args.dataset == 'AviviD_DatasetA'):
      n_section =23
    else:
      n_section=52

    # model parameters
    model_params = {
          'pad_id':pad_id,
          'd_model':args.d_model, #128 ,
          'nhead' : args.nhead, #16,
          'num_encoder_layers' :args.num_encoder_layers ,#2 ,
          'num_expert':args.num_expert,#3,
          'ff_dim':args.ff_dim,#128,
          'llm_size':args.llm_size,#4096,
          'dropout' : args.dropout,#0.5,
          'add_PE':args.add_PE#True
          }

    model = get_model(model_name,model_params).to(device)
    loss_fn = torch.nn.BCELoss()
    optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate , weight_decay=args.weight_decay)
    train_set  = MyDataset(click_seq_train ,ads_feature_matrix,news_feature_matrix)
    val_set    = MyDataset(click_seq_val   ,ads_feature_matrix,news_feature_matrix)
    test_set   = MyDataset(click_seq_test  ,ads_feature_matrix,news_feature_matrix)
    train_loader  = DataLoader(train_set , batch_size=args.batch_size        ,collate_fn=train_set.collate_fn,shuffle=True)
    val_loader    = DataLoader(val_set   , batch_size=args.batch_size        ,collate_fn=val_set.collate_fn,shuffle=True)
    test_loader   = DataLoader(test_set  , batch_size=1                      ,collate_fn=test_set.collate_fn,shuffle=False)
    total_steps = len(train_loader)

    opt1 = Optimization(model=model, loss_fn=loss_fn, optimizer=optimizer)
    metrics5_test, metrics10_test, metrics20_test = opt1.evaluate(test_loader,ckpt_path = args.ckpt_path)