DDTCDR/engine.py at master · lpworld/DDTCDR · 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
import torch
from torch.autograd import Variable
from mlp import MLP
from utils import save_checkpoint, use_optimizer, use_cuda
from sklearn.metrics import mean_squared_error, mean_absolute_error
import os
import numpy as np

class Engine(object):
    """Meta Engine for training & evaluating NCF model

    Note: Subclass should implement self.model !
    """

    def __init__(self, config):
        self.config = config  # model configuration
        self.modelA = MLP(config)
        self.modelB = MLP(config)
        if config['use_cuda'] is True:
            use_cuda(True, config['device_id'])
            self.modelA.cuda()
            self.modelB.cuda()
        print(self.modelA)
        if config['pretrain']:
            self.model.load_pretrain_weights()
        self.optA = use_optimizer(self.modelA, config)
        self.optB = use_optimizer(self.modelB, config)
        self.crit = torch.nn.MSELoss()
        self.alpha = config['alpha']

    def train_single_batch(self, book_user_embeddings, book_item_embeddings, book_rating,
                          movie_user_embeddings, movie_item_embeddings, movie_rating):
        self.optA.zero_grad()
        self.optB.zero_grad()
        book_ratings_pred1 = self.modelA(book_user_embeddings, book_item_embeddings)
        lossA1 = self.crit(book_ratings_pred1.squeeze(1), book_rating)
        book_ratings_pred2 = self.modelB(book_user_embeddings, book_item_embeddings, dual=True)
        lossA2 = self.crit(book_ratings_pred2.squeeze(1), book_rating)
        movie_ratings_pred1 = self.modelB(movie_user_embeddings, movie_item_embeddings)
        lossB1 = self.crit(movie_ratings_pred1.squeeze(1), movie_rating)
        movie_ratings_pred2 = self.modelA(movie_user_embeddings, movie_item_embeddings, dual=True)
        lossB2 = self.crit(movie_ratings_pred2.squeeze(1), movie_rating)
        lossA = (1-self.alpha)*lossA1 + self.alpha*Variable(lossA2.data, requires_grad=False)
        lossB = (1-self.alpha)*lossB1 + self.alpha*Variable(lossB2.data, requires_grad=False)
        lossA.backward(retain_graph=True)
        lossB.backward(retain_graph=True)
        orth_loss_A, orth_loss_B = torch.zeros(1), torch.zeros(1)
        reg = 1e-6
        for name, param in self.modelA.bridge.named_parameters():
            if 'bias' not in name:
                param_flat = param.view(param.shape[0], -1)
                sym = torch.mm(param_flat, torch.t(param_flat))
                sym -= torch.eye(param_flat.shape[0])
                orth_loss_A = orth_loss_A + (reg * sym.abs().sum())
        orth_loss_A.backward()
        for name, param in self.modelB.bridge.named_parameters():
            if 'bias' not in name:
                param_flat = param.view(param.shape[0], -1)
                sym = torch.mm(param_flat, torch.t(param_flat))
                sym -= torch.eye(param_flat.shape[0])
                orth_loss_B = orth_loss_B + (reg * sym.abs().sum())
        orth_loss_B.backward()
        self.optA.step()
        self.optB.step()
        if self.config['use_cuda'] is True:
            lossA = lossA.data.cpu().numpy()[0]
            lossB = lossB.data.cpu().numpy()[0]
            orth_loss_A = orth_loss_A.data.cpu().numpy()[0]
            orth_loss_B = orth_loss_B.data.cpu().numpy()[0]
        else:
            lossA = lossA.data.numpy()
            lossB = lossB.data.numpy()
            orth_loss_A = orth_loss_A.data.numpy()
            orth_loss_B = orth_loss_B.data.numpy()
        return lossA + lossB + orth_loss_A + orth_loss_B

    def train_an_epoch(self, train_book_loader, train_movie_loader, epoch_id):
        self.modelA.train()
        self.modelB.train()
        total_loss = 0
        for book_batch, movie_batch in zip(train_book_loader, train_movie_loader):
            assert isinstance(book_batch[0], torch.LongTensor)
            book_rating, book_user_embeddings, book_item_embeddings = Variable(book_batch[2]), Variable(book_batch[3]), Variable(book_batch[4])
            movie_rating, movie_user_embeddings, movie_item_embeddings = Variable(movie_batch[2]), Variable(movie_batch[3]), Variable(movie_batch[4])
            book_rating = book_rating.float()
            movie_rating = movie_rating.float()
            if self.config['use_cuda'] is True:
                book_rating = book_rating.cuda()
                movie_rating = movie_rating.cuda()
                book_user_embeddings = book_user_embeddings.cuda()
                book_item_embeddings = book_item_embeddings.cuda()
                movie_user_embeddings = movie_user_embeddings.cuda()
                movie_item_embeddings = movie_item_embeddings.cuda()
            loss = self.train_single_batch(book_user_embeddings, book_item_embeddings, book_rating,
                                           movie_user_embeddings, movie_item_embeddings, movie_rating)
            total_loss += loss

    def evaluate(self, evaluate_book_data, evaluate_movie_data, epoch_id):
        self.modelA.eval()
        self.modelB.eval()
        book_user, book_item, book_user_embeddings, book_item_embeddings, \
            book_golden = evaluate_book_data[0], evaluate_book_data[1], \
                Variable(evaluate_book_data[2]), Variable(evaluate_book_data[3]), evaluate_book_data[4]
        movie_user, movie_item, movie_user_embeddings, movie_item_embeddings, \
            movie_golden = evaluate_movie_data[0], evaluate_movie_data[1], \
                Variable(evaluate_movie_data[2]), Variable(evaluate_movie_data[3]), evaluate_movie_data[4]
        if self.config['use_cuda'] is True:
            book_user_embeddings = book_user_embeddings.cuda()
            book_item_embeddings = book_item_embeddings.cuda()
            movie_user_embeddings = movie_user_embeddings.cuda()
            movie_item_embeddings = movie_item_embeddings.cuda()
        book_scores = self.modelA(book_user_embeddings, book_item_embeddings)
        book_scores = book_scores.detach().numpy()
        movie_scores = self.modelB(movie_user_embeddings, movie_item_embeddings)
        movie_scores = movie_scores.detach().numpy()
        book_MSE = mean_squared_error(book_golden, book_scores)
        book_MAE = mean_absolute_error(book_golden, book_scores)
        movie_MSE = mean_squared_error(movie_golden, movie_scores)
        movie_MAE = mean_absolute_error(movie_golden, movie_scores)

        unique_book_user = list(set(book_user))
        unique_movie_user = list(set(movie_user))
        book_recommend, movie_recommend = [], []
        book_precision, movie_precision, book_recall, movie_recall = [], [], [], []
        for index in range(len(book_user)):
            book_recommend.append((book_user[index],book_item[index],book_golden[index],book_scores[index]))
        for index in range(len(movie_user)):
            movie_recommend.append((movie_user[index],movie_item[index],movie_golden[index],movie_scores[index]))
        for user in unique_book_user:
            user_ratings = [x for x in book_recommend if x[0]==user]
            user_ratings.sort(key=lambda x:x[3], reverse=True)
            n_rel = sum((true_r >= 0.5) for (_, _, true_r, _) in user_ratings)
            n_rec_k = sum((est >= 0.5) for (_, _, _, est) in user_ratings)
            n_rel_and_rec_k = sum(((true_r >= 0.5) and (est >= 0.5))
                            for (_, _, true_r, est) in user_ratings)
            book_precision.append(n_rel_and_rec_k / n_rec_k if n_rec_k!=0 else 1)
            book_recall.append(n_rel_and_rec_k / n_rel if n_rel!=0 else 1)
        book_precision = np.mean(book_precision)
        book_recall = np.mean(book_recall)
        for user in unique_movie_user:
            user_ratings = [x for x in movie_recommend if x[0]==user]
            user_ratings.sort(key=lambda x:x[3], reverse=True)
            n_rel = sum((true_r >= 0.5) for (_, _, true_r, _) in user_ratings)
            n_rec_k = sum((est >= 0.5) for (_, _, _, est) in user_ratings)
            n_rel_and_rec_k = sum(((true_r >= 0.5) and (est >= 0.5))
                              for (_, _, true_r, est) in user_ratings)
            movie_precision.append(n_rel_and_rec_k / n_rec_k if n_rec_k!=0 else 0)
            movie_recall.append(n_rel_and_rec_k / n_rel if n_rel!=0 else 1)
        movie_precision = np.mean(movie_precision)
        movie_recall = np.mean(movie_recall)

        print('[Book Evluating Epoch {}] MSE = {:.4f}, MAE = {:.4f}, Precision = {}, Recall = {}'.format(epoch_id, book_MSE, book_MAE, book_precision, book_recall))
        print('[Movie Evluating Epoch {}] MSE = {:.4f}, MAE = {:.4f}, Precision = {}, Recall = {}'.format(epoch_id, movie_MSE, movie_MAE, movie_precision, movie_recall))

        return book_MSE, book_MAE, movie_MSE, movie_MAE

    def save(self, dirname, filename):
        with open(os.path.join(dirname, filename)+'A', 'wb') as f:
            torch.save(self.modelA.state_dict(), f)
        with open(os.path.join(dirname, filename)+'B', 'wb') as f:
            torch.save(self.modelB.state_dict(), f)