File size: 9,818 Bytes
11c72a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260

import torch
from torch import nn
import torch.nn.functional as F


class DETM(nn.Module):
    """
        The Dynamic Embedded Topic Model. 2019

        Adji B. Dieng, Francisco J. R. Ruiz, David M. Blei
    """
    def __init__(self, vocab_size, num_times, train_size, train_time_wordfreq,
                 num_topics=50, train_WE=True, pretrained_WE=None, en_units=800,
                 eta_hidden_size=200, rho_size=300, enc_drop=0.0, eta_nlayers=3,
                 eta_dropout=0.0, delta=0.005, theta_act='relu', device='cpu'):
        super().__init__()

        ## define hyperparameters
        self.num_topics = num_topics
        self.num_times = num_times
        self.vocab_size = vocab_size
        self.eta_hidden_size = eta_hidden_size
        self.rho_size = rho_size
        self.enc_drop = enc_drop
        self.eta_nlayers = eta_nlayers
        self.t_drop = nn.Dropout(enc_drop)
        self.eta_dropout = eta_dropout
        self.delta = delta
        self.train_WE = train_WE
        self.train_size = train_size
        self.rnn_inp = train_time_wordfreq
        self.device = device

        self.theta_act = self.get_activation(theta_act)

        ## define the word embedding matrix \rho
        if self.train_WE:
            self.rho = nn.Linear(self.rho_size, self.vocab_size, bias=False)
        else:
            rho = nn.Embedding(pretrained_WE.size())
            rho.weight.data = torch.from_numpy(pretrained_WE)
            self.rho = rho.weight.data.clone().float().to(self.device)

        ## define the variational parameters for the topic embeddings over time (alpha) ... alpha is K x T x L
        self.mu_q_alpha = nn.Parameter(torch.randn(self.num_topics, self.num_times, self.rho_size))
        self.logsigma_q_alpha = nn.Parameter(torch.randn(self.num_topics, self.num_times, self.rho_size))

        ## define variational distribution for \theta_{1:D} via amortizartion... theta is K x D
        self.q_theta = nn.Sequential(
            nn.Linear(self.vocab_size + self.num_topics, en_units), 
            self.theta_act,
            nn.Linear(en_units, en_units),
            self.theta_act,
        )
        self.mu_q_theta = nn.Linear(en_units, self.num_topics, bias=True)
        self.logsigma_q_theta = nn.Linear(en_units, self.num_topics, bias=True)

        ## define variational distribution for \eta via amortizartion... eta is K x T
        self.q_eta_map = nn.Linear(self.vocab_size, self.eta_hidden_size)
        self.q_eta = nn.LSTM(self.eta_hidden_size, self.eta_hidden_size, self.eta_nlayers, dropout=self.eta_dropout)
        self.mu_q_eta = nn.Linear(self.eta_hidden_size + self.num_topics, self.num_topics, bias=True)
        self.logsigma_q_eta = nn.Linear(self.eta_hidden_size + self.num_topics, self.num_topics, bias=True)

        self.decoder_bn = nn.BatchNorm1d(vocab_size)
        self.decoder_bn.weight.requires_grad = False

    def get_activation(self, act):
        activations = {
            'tanh': nn.Tanh(),
            'relu': nn.ReLU(),
            'softplus': nn.Softplus(),
            'rrelu': nn.RReLU(),
            'leakyrelu': nn.LeakyReLU(),
            'elu': nn.ELU(),
            'selu': nn.SELU(),
            'glu': nn.GLU(),
        }

        if act in activations:
            act = activations[act]
        else:
            print('Defaulting to tanh activations...')
            act = nn.Tanh()
        return act

    def reparameterize(self, mu, logvar):
        """Returns a sample from a Gaussian distribution via reparameterization.
        """
        if self.training:
            std = torch.exp(0.5 * logvar)
            eps = torch.randn_like(std)
            return eps.mul_(std).add_(mu)
        else:
            return mu

    def get_kl(self, q_mu, q_logsigma, p_mu=None, p_logsigma=None):
        """Returns KL( N(q_mu, q_logsigma) || N(p_mu, p_logsigma) ).
        """
        if p_mu is not None and p_logsigma is not None:
            sigma_q_sq = torch.exp(q_logsigma)
            sigma_p_sq = torch.exp(p_logsigma)
            kl = ( sigma_q_sq + (q_mu - p_mu)**2 ) / ( sigma_p_sq + 1e-6 )
            kl = kl - 1 + p_logsigma - q_logsigma
            kl = 0.5 * torch.sum(kl, dim=-1)
        else:
            kl = -0.5 * torch.sum(1 + q_logsigma - q_mu.pow(2) - q_logsigma.exp(), dim=-1)
        return kl

    def get_alpha(self): ## mean field
        alphas = torch.zeros(self.num_times, self.num_topics, self.rho_size).to(self.device)
        kl_alpha = []

        alphas[0] = self.reparameterize(self.mu_q_alpha[:, 0, :], self.logsigma_q_alpha[:, 0, :])

        # TODO: why logsigma_p_0 is zero?
        p_mu_0 = torch.zeros(self.num_topics, self.rho_size).to(self.device)
        logsigma_p_0 = torch.zeros(self.num_topics, self.rho_size).to(self.device)
        kl_0 = self.get_kl(self.mu_q_alpha[:, 0, :], self.logsigma_q_alpha[:, 0, :], p_mu_0, logsigma_p_0)
        kl_alpha.append(kl_0)
        for t in range(1, self.num_times):
            alphas[t] = self.reparameterize(self.mu_q_alpha[:, t, :], self.logsigma_q_alpha[:, t, :]) 

            p_mu_t = alphas[t - 1]
            logsigma_p_t = torch.log(self.delta * torch.ones(self.num_topics, self.rho_size).to(self.device))
            kl_t = self.get_kl(self.mu_q_alpha[:, t, :], self.logsigma_q_alpha[:, t, :], p_mu_t, logsigma_p_t)
            kl_alpha.append(kl_t)
        kl_alpha = torch.stack(kl_alpha).sum()
        return alphas, kl_alpha.sum()

    def get_eta(self, rnn_inp): ## structured amortized inference
        inp = self.q_eta_map(rnn_inp).unsqueeze(1)
        hidden = self.init_hidden()
        output, _ = self.q_eta(inp, hidden)
        output = output.squeeze()

        etas = torch.zeros(self.num_times, self.num_topics).to(self.device)
        kl_eta = []

        inp_0 = torch.cat([output[0], torch.zeros(self.num_topics,).to(self.device)], dim=0)
        mu_0 = self.mu_q_eta(inp_0)
        logsigma_0 = self.logsigma_q_eta(inp_0)
        etas[0] = self.reparameterize(mu_0, logsigma_0)

        p_mu_0 = torch.zeros(self.num_topics,).to(self.device)
        logsigma_p_0 = torch.zeros(self.num_topics,).to(self.device)
        kl_0 = self.get_kl(mu_0, logsigma_0, p_mu_0, logsigma_p_0)
        kl_eta.append(kl_0)

        for t in range(1, self.num_times):
            inp_t = torch.cat([output[t], etas[t-1]], dim=0)
            mu_t = self.mu_q_eta(inp_t)
            logsigma_t = self.logsigma_q_eta(inp_t)
            etas[t] = self.reparameterize(mu_t, logsigma_t)

            p_mu_t = etas[t-1]
            logsigma_p_t = torch.log(self.delta * torch.ones(self.num_topics,).to(self.device))
            kl_t = self.get_kl(mu_t, logsigma_t, p_mu_t, logsigma_p_t)
            kl_eta.append(kl_t)
        kl_eta = torch.stack(kl_eta).sum()

        return etas, kl_eta

    def get_theta(self, bows, times, eta=None): ## amortized inference
        """Returns the topic proportions.
        """

        normalized_bows = bows / bows.sum(1, keepdims=True)

        if eta is None and self.training is False:
            eta, kl_eta = self.get_eta(self.rnn_inp)

        eta_td = eta[times]
        inp = torch.cat([normalized_bows, eta_td], dim=1)
        q_theta = self.q_theta(inp)
        if self.enc_drop > 0:
            q_theta = self.t_drop(q_theta)
        mu_theta = self.mu_q_theta(q_theta)
        logsigma_theta = self.logsigma_q_theta(q_theta)
        z = self.reparameterize(mu_theta, logsigma_theta)
        theta = F.softmax(z, dim=-1)
        kl_theta = self.get_kl(mu_theta, logsigma_theta, eta_td, torch.zeros(self.num_topics).to(self.device))

        if self.training:
            return theta, kl_theta
        else:
            return theta

    @property
    def word_embeddings(self):
        return self.rho.weight

    @property
    def topic_embeddings(self):
        alpha, _ = self.get_alpha()
        return alpha

    def get_beta(self, alpha=None):
        """Returns the topic matrix \beta of shape T x K x V
        """

        if alpha is None and self.training is False:
            alpha, kl_alpha = self.get_alpha()

        if self.train_WE:
            logit = self.rho(alpha.view(alpha.size(0) * alpha.size(1), self.rho_size))
        else:
            tmp = alpha.view(alpha.size(0) * alpha.size(1), self.rho_size)
            logit = torch.mm(tmp, self.rho.permute(1, 0))
        logit = logit.view(alpha.size(0), alpha.size(1), -1)

        beta = F.softmax(logit, dim=-1)

        return beta

    def get_NLL(self, theta, beta, bows):
        theta = theta.unsqueeze(1)
        loglik = torch.bmm(theta, beta).squeeze(1)
        loglik = torch.log(loglik + 1e-12)
        nll = -loglik * bows
        nll = nll.sum(-1)
        return nll

    def forward(self, bows, times):
        bsz = bows.size(0)
        coeff = self.train_size / bsz
        eta, kl_eta = self.get_eta(self.rnn_inp)
        theta, kl_theta = self.get_theta(bows, times, eta)
        kl_theta = kl_theta.sum() * coeff

        alpha, kl_alpha = self.get_alpha()
        beta = self.get_beta(alpha)

        beta = beta[times]
        # beta = beta[times.type('torch.LongTensor')]
        nll = self.get_NLL(theta, beta, bows)
        nll = nll.sum() * coeff

        loss = nll + kl_eta + kl_theta

        rst_dict = {
            'loss': loss,
            'nll': nll,
            'kl_eta': kl_eta,
            'kl_theta': kl_theta
        }

        loss += kl_alpha
        rst_dict['kl_alpha'] = kl_alpha

        return rst_dict

    def init_hidden(self):
        """Initializes the first hidden state of the RNN used as inference network for \\eta.
        """
        weight = next(self.parameters())
        nlayers = self.eta_nlayers
        nhid = self.eta_hidden_size
        return (weight.new_zeros(nlayers, 1, nhid), weight.new_zeros(nlayers, 1, nhid))