File size: 7,540 Bytes
569596a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from collections import namedtuple

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()
import tensorflow_probability as tfp
tfd = tfp.distributions
import numpy as np

from tf_utils import dense_layer, shape


LSTMAttentionCellState = namedtuple(
    'LSTMAttentionCellState',
    ['h1', 'c1', 'h2', 'c2', 'h3', 'c3', 'alpha', 'beta', 'kappa', 'w', 'phi']
)


class LSTMAttentionCell(tf.nn.rnn_cell.RNNCell):

    def __init__(

        self,

        lstm_size,

        num_attn_mixture_components,

        attention_values,

        attention_values_lengths,

        num_output_mixture_components,

        bias,

        reuse=None,

    ):
        self.reuse = reuse
        self.lstm_size = lstm_size
        self.num_attn_mixture_components = num_attn_mixture_components
        self.attention_values = attention_values
        self.attention_values_lengths = attention_values_lengths
        self.window_size = shape(self.attention_values, 2)
        self.char_len = tf.shape(attention_values)[1]
        self.batch_size = tf.shape(attention_values)[0]
        self.num_output_mixture_components = num_output_mixture_components
        self.output_units = 6*self.num_output_mixture_components + 1
        self.bias = bias

    @property
    def state_size(self):
        return LSTMAttentionCellState(
            self.lstm_size,
            self.lstm_size,
            self.lstm_size,
            self.lstm_size,
            self.lstm_size,
            self.lstm_size,
            self.num_attn_mixture_components,
            self.num_attn_mixture_components,
            self.num_attn_mixture_components,
            self.window_size,
            self.char_len,
        )

    @property
    def output_size(self):
        return self.lstm_size

    def zero_state(self, batch_size, dtype):
        return LSTMAttentionCellState(
            tf.zeros([batch_size, self.lstm_size]),
            tf.zeros([batch_size, self.lstm_size]),
            tf.zeros([batch_size, self.lstm_size]),
            tf.zeros([batch_size, self.lstm_size]),
            tf.zeros([batch_size, self.lstm_size]),
            tf.zeros([batch_size, self.lstm_size]),
            tf.zeros([batch_size, self.num_attn_mixture_components]),
            tf.zeros([batch_size, self.num_attn_mixture_components]),
            tf.zeros([batch_size, self.num_attn_mixture_components]),
            tf.zeros([batch_size, self.window_size]),
            tf.zeros([batch_size, self.char_len]),
        )

    def __call__(self, inputs, state, scope=None):
        with tf.variable_scope(scope or type(self).__name__, reuse=tf.AUTO_REUSE):

            # lstm 1
            s1_in = tf.concat([state.w, inputs], axis=1)
            cell1 = tf.compat.v1.nn.rnn_cell.LSTMCell(self.lstm_size)
            s1_out, s1_state = cell1(s1_in, state=(state.c1, state.h1))

            # attention
            attention_inputs = tf.concat([state.w, inputs, s1_out], axis=1)
            attention_params = dense_layer(attention_inputs, 3*self.num_attn_mixture_components, scope='attention')
            alpha, beta, kappa = tf.split(tf.nn.softplus(attention_params), 3, axis=1)
            kappa = state.kappa + kappa / 25.0
            beta = tf.clip_by_value(beta, .01, np.inf)

            kappa_flat, alpha_flat, beta_flat = kappa, alpha, beta
            kappa, alpha, beta = tf.expand_dims(kappa, 2), tf.expand_dims(alpha, 2), tf.expand_dims(beta, 2)

            enum = tf.reshape(tf.range(self.char_len), (1, 1, self.char_len))
            u = tf.cast(tf.tile(enum, (self.batch_size, self.num_attn_mixture_components, 1)), tf.float32)
            phi_flat = tf.reduce_sum(alpha*tf.exp(-tf.square(kappa - u) / beta), axis=1)

            phi = tf.expand_dims(phi_flat, 2)
            sequence_mask = tf.cast(tf.sequence_mask(self.attention_values_lengths, maxlen=self.char_len), tf.float32)
            sequence_mask = tf.expand_dims(sequence_mask, 2)
            w = tf.reduce_sum(phi*self.attention_values*sequence_mask, axis=1)

            # lstm 2
            s2_in = tf.concat([inputs, s1_out, w], axis=1)
            cell2 = tf.compat.v1.nn.rnn_cell.LSTMCell(self.lstm_size)
            s2_out, s2_state = cell2(s2_in, state=(state.c2, state.h2))

            # lstm 3
            s3_in = tf.concat([inputs, s2_out, w], axis=1)
            cell3 = tf.compat.v1.nn.rnn_cell.LSTMCell(self.lstm_size)
            s3_out, s3_state = cell3(s3_in, state=(state.c3, state.h3))

            new_state = LSTMAttentionCellState(
                s1_state.h,
                s1_state.c,
                s2_state.h,
                s2_state.c,
                s3_state.h,
                s3_state.c,
                alpha_flat,
                beta_flat,
                kappa_flat,
                w,
                phi_flat,
            )

            return s3_out, new_state

    def output_function(self, state):
        params = dense_layer(state.h3, self.output_units, scope='gmm', reuse=tf.AUTO_REUSE)
        pis, mus, sigmas, rhos, es = self._parse_parameters(params)
        mu1, mu2 = tf.split(mus, 2, axis=1)
        mus = tf.stack([mu1, mu2], axis=2)
        sigma1, sigma2 = tf.split(sigmas, 2, axis=1)

        covar_matrix = [tf.square(sigma1), rhos*sigma1*sigma2,
                        rhos*sigma1*sigma2, tf.square(sigma2)]
        covar_matrix = tf.stack(covar_matrix, axis=2)
        covar_matrix = tf.reshape(covar_matrix, (self.batch_size, self.num_output_mixture_components, 2, 2))

        mvn = tfd.MultivariateNormalFullCovariance(loc=mus, covariance_matrix=covar_matrix)
        b = tfd.Bernoulli(probs=es)
        c = tfd.Categorical(probs=pis)

        sampled_e = b.sample()
        sampled_coords = mvn.sample()
        sampled_idx = c.sample()

        idx = tf.stack([tf.range(self.batch_size), sampled_idx], axis=1)
        coords = tf.gather_nd(sampled_coords, idx)
        return tf.concat([coords, tf.cast(sampled_e, tf.float32)], axis=1)

    def termination_condition(self, state):
        char_idx = tf.cast(tf.argmax(state.phi, axis=1), tf.int32)
        final_char = char_idx >= self.attention_values_lengths - 1
        past_final_char = char_idx >= self.attention_values_lengths
        output = self.output_function(state)
        es = tf.cast(output[:, 2], tf.int32)
        is_eos = tf.equal(es, tf.ones_like(es))
        return tf.logical_or(tf.logical_and(final_char, is_eos), past_final_char)

    def _parse_parameters(self, gmm_params, eps=1e-8, sigma_eps=1e-4):
        pis, sigmas, rhos, mus, es = tf.split(
            gmm_params,
            [
                1*self.num_output_mixture_components,
                2*self.num_output_mixture_components,
                1*self.num_output_mixture_components,
                2*self.num_output_mixture_components,
                1
            ],
            axis=-1
        )
        pis = pis*(1 + tf.expand_dims(self.bias, 1))
        sigmas = sigmas - tf.expand_dims(self.bias, 1)

        pis = tf.nn.softmax(pis, axis=-1)
        pis = tf.where(pis < .01, tf.zeros_like(pis), pis)
        sigmas = tf.clip_by_value(tf.exp(sigmas), sigma_eps, np.inf)
        rhos = tf.clip_by_value(tf.tanh(rhos), eps - 1.0, 1.0 - eps)
        es = tf.clip_by_value(tf.nn.sigmoid(es), eps, 1.0 - eps)
        es = tf.where(es < .01, tf.zeros_like(es), es)

        return pis, mus, sigmas, rhos, es