SMASH/train.py at main · hly1998/SMASH · 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
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
import argparse
import os
import torch
from utils import *
import torch.optim as optim
from model import *
from scipy import sparse
from scipy.sparse import csc_matrix
import numpy as np


def train(args):
    if not args.gpunum:
        parser.error("Need to provide the GPU number.")
    if not args.dataset:
        parser.error("Need to provide the dataset.")
    os.environ["CUDA_VISIBLE_DEVICES"] = args.gpunum
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    dataset, data_fmt = args.dataset.split('.')
    if dataset in ['reuters', 'tmc', 'rcv1', 'edudata', 'edudata20']:
        single_label_flag = False
    else:
        single_label_flag = True
    if single_label_flag:
        train_set = SingleLabelTextDataset('dataset/{}'.format(dataset),
                                           subset='train',
                                           bow_format=data_fmt,
                                           download=True)
        test_set = SingleLabelTextDataset('dataset/{}'.format(dataset),
                                          subset='test',
                                          bow_format=data_fmt,
                                          download=True)
    else:
        train_set = MultiLabelTextDataset('dataset/{}'.format(dataset),
                                          subset='train',
                                          bow_format=data_fmt,
                                          download=True)
        test_set = MultiLabelTextDataset('dataset/{}'.format(dataset),
                                         subset='test',
                                         bow_format=data_fmt,
                                         download=True)

    train_loader = torch.utils.data.DataLoader(dataset=train_set,
                                               batch_size=args.batch_size,
                                               shuffle=True)
    test_loader = torch.utils.data.DataLoader(dataset=test_set,
                                              batch_size=args.batch_size,
                                              shuffle=True)
    setup_seed(2023)
    num_bits = args.nbits
    n_sample = args.n_sample
    top_k = args.top_k
    num_features = train_set[0][0].size(0)
    best_precision = 0
    best_precision_epoch = 0
    time_max = int(len(train_set) / args.batch_size)
    model = SMASH(dataset,
                  num_features,
                  num_bits,
                  dropoutProb=0.1,
                  time_max=time_max,
                  em_alpha=args.em_alpha,
                  sigma=args.sigma,
                  device=device,
                  em_length=len(train_set))
    model.to(device)
    num_epochs = args.num_epochs
    optimizer = optim.Adam(model.parameters(), lr=args.lr)
    if dataset == 'ng20':
        scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
                                                    step_size=5e3,
                                                    gamma=0.96)
    L_max = 0
    L_t_minus_1 = 0
    transfrom_flag = False
    transfrom_count = 0
    for epoch in range(num_epochs):
        total_loss = []
        reconstr_loss = []
        propagation_loss = []
        balance_loss = []
        model.train()
        for _, (xb, idxs, yb) in enumerate(train_loader):
            # print(L_max, L_t_minus_1)
            xb = xb.to(device)
            yb = yb.to(device)
            loss_change_term = np.abs(L_max - L_t_minus_1) / (L_max + 0.00001)
            logprob_w, logprob_w_noise, z, z_noise, long_z, noise_long_z_e, em_out = model(xb, idxs, epoch, loss_change_term, n_sample)
            # 计算epsilon
            # 这里这个值是手动设置，和长bit code长度一致
            noise_z_reshape = noise_long_z_e.reshape(xb.shape[0], 128, n_sample)
            mult = torch.matmul(long_z.unsqueeze(1), noise_z_reshape).squeeze(1)
            epsilon = torch.nn.functional.softmax(mult, dim=1)
            # print(logprob_w_noise.shape, epsilon.shape)

            rec_loss = compute_reconstr_loss(logprob_w, xb) + compute_reconstr_noise_loss(logprob_w_noise, xb, epsilon)
            pro_loss = relevance_propagation_v1(z, long_z) + relevance_propagation_v1(z_noise, noise_long_z_e)
            loss = rec_loss + args.lsc_weight * pro_loss
            if transfrom_flag:
                ba_loss = code_balance_global_v2(num_bits,
                                                    em_out,
                                                    z,
                                                    alpha=args.bb_weight,
                                                    beta=args.bd_weight)
                loss = loss + ba_loss
            # if loss.item() > L_max:
            #     L_max = loss.item()
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            scheduler.step()
            L_t_minus_1 = loss.item()
            total_loss.append(loss.item())
            reconstr_loss.append(rec_loss.item())
            propagation_loss.append(pro_loss.item())
            if transfrom_flag:
                balance_loss.append(ba_loss.item())
            else:
                balance_loss.append(0)
        model.eval()
        if np.mean(total_loss) > L_max:
            L_max = np.mean(total_loss)
        with torch.no_grad():
            train_b, test_b, train_y, test_y = model.get_binary_code(
                train_loader, test_loader)
            precision = precisionK(train_b, test_b, train_y, test_y, num_bits)
        if precision < best_precision and epoch > 10:
            transfrom_count = transfrom_count + 1
        else:
            transfrom_count = 0
        if transfrom_count == 5:
            transfrom_flag = True
        if precision > best_precision:
            best_precision = precision
            best_precision_epoch = epoch + 1
            state = {
                'epoch': epoch,
                'model': model.state_dict(),
                'best_recall': best_precision,
            }
            save_file = os.path.join(
                './checkpoint',
                '{},{},{},{},{}_best.pth'.format(args.dataset, args.nbits,
                                                    args.bb_weight,
                                                    args.bd_weight,
                                                    args.lsc_weight))
            print('saving the best model!')
            torch.save(state, save_file)
        print(
            'total_loss:{:.4f} reconstr_loss:{:.4f} propagation_loss:{:.4f} balance_loss:{:.4f} MAX_Loss: {:.4f} Best Precision:({}){:.4f} Now Precision:({}){:.4f}'
            .format(np.mean(total_loss), np.mean(reconstr_loss),
                    np.mean(propagation_loss),
                    np.mean(balance_loss), L_max, best_precision_epoch,
                    best_precision, epoch + 1, precision))
    return best_precision, best_precision_epoch


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument("-g",
                        "--gpunum",
                        help="GPU number to train the model.")
    parser.add_argument("-d", "--dataset", help="Name of the dataset.")
    parser.add_argument("-b",
                        "--nbits",
                        help="Number of bits of the embedded vector.",
                        type=int)
    parser.add_argument("--dropout",
                        help="Dropout probability (0 means no dropout)",
                        default=0.1,
                        type=float)
    parser.add_argument("--num_epochs", default=30, type=int)
    parser.add_argument("--raduis_r", default=1, type=int)
    parser.add_argument("--top_k", default=100, type=int)
    parser.add_argument("--model", type=str)
    parser.add_argument("--batch_size", default=128, type=int)
    parser.add_argument("--lr", default=0.001, type=float)
    parser.add_argument("--lsc_weight", default=0.5, type=float)
    parser.add_argument("--bb_weight", default=1, type=float)
    parser.add_argument("--bd_weight", default=0.01, type=float)
    parser.add_argument("--em_alpha", default=0.5, type=float)
    parser.add_argument("--sigma", default=0.6, type=float)
    parser.add_argument("--n_sample", default=3, type=int)
    args = parser.parse_args()
    train(args)