LEAM/model.py at master · guoyinwang/LEAM · 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
import tensorflow as tf
import numpy as np
from tensorflow.contrib import layers

def embedding(features, opt, prefix='', is_reuse=None):
    """Customized function to transform batched x into embeddings."""
    # Convert indexes of words into embeddings.
    with tf.variable_scope(prefix + 'embed', reuse=is_reuse):
        if opt.fix_emb:
            assert (hasattr(opt, 'W_emb'))
            assert (np.shape(np.array(opt.W_emb)) == (opt.n_words, opt.embed_size))
            W = tf.get_variable('W', initializer=opt.W_emb, trainable=True)
            print("initialize word embedding finished")
        else:
            weightInit = tf.random_uniform_initializer(-0.001, 0.001)
            W = tf.get_variable('W', [opt.n_words, opt.embed_size], initializer=weightInit)
    if hasattr(opt, 'relu_w') and opt.relu_w:
        W = tf.nn.relu(W)

    word_vectors = tf.nn.embedding_lookup(W, features)

    return word_vectors, W

def embedding_class(features, opt, prefix='', is_reuse=None):
    """Customized function to transform batched y into embeddings."""
    # Convert indexes of words into embeddings.
    with tf.variable_scope(prefix + 'embed', reuse=is_reuse):
        if opt.fix_emb:
            assert (hasattr(opt, 'W_class_emb'))
            W = tf.get_variable('W_class', initializer=opt.W_class_emb, trainable=True)
            print("initialize class embedding finished")
        else:
            weightInit = tf.random_uniform_initializer(-0.001, 0.001)
            W = tf.get_variable('W_class', [opt.num_class, opt.embed_size], initializer=weightInit)
    if hasattr(opt, 'relu_w') and opt.relu_w:
        W = tf.nn.relu(W)
    word_vectors = tf.nn.embedding_lookup(W, features)

    return word_vectors, W

def att_emb_ngram_encoder_maxout(x_emb, x_mask, W_class, W_class_tran, opt):
    x_mask = tf.expand_dims(x_mask, axis=-1) # b * s * 1
    x_emb_0 = tf.squeeze(x_emb,) # b * s * e
    x_emb_1 = tf.multiply(x_emb_0, x_mask) # b * s * e

    x_emb_norm = tf.nn.l2_normalize(x_emb_1, dim=2) # b * s * e
    W_class_norm = tf.nn.l2_normalize(W_class_tran, dim = 0) # e * c
    G = tf.contrib.keras.backend.dot(x_emb_norm, W_class_norm) # b * s * c
    x_full_emb = x_emb_0
    Att_v = tf.contrib.layers.conv2d(G, num_outputs=opt.num_class,kernel_size=[opt.ngram], padding='SAME',activation_fn=tf.nn.relu) #b * s *  c

    Att_v = tf.reduce_max(Att_v, axis=-1, keep_dims=True)
    Att_v_max = partial_softmax(Att_v, x_mask, 1, 'Att_v_max') # b * s * 1

    x_att = tf.multiply(x_full_emb, Att_v_max)
    H_enc = tf.reduce_sum(x_att, axis=1)
    return H_enc

def att_emb_ngram_encoder_cnn(x_emb, x_mask, W_class, W_class_tran, opt):
    x_mask = tf.expand_dims(x_mask, axis=-1) # b * s * 1
    x_emb_0 = tf.squeeze(x_emb,) # b * s * e
    x_emb_1 = tf.multiply(x_emb_0, x_mask) # b * s * e

    H = tf.contrib.layers.conv2d(x_emb_0, num_outputs=opt.embed_size,kernel_size=[10], padding='SAME',activation_fn=tf.nn.relu) #b * s *  c


    G = tf.contrib.keras.backend.dot(H, W_class_tran) # b * s * c
    Att_v_max = partial_softmax(G, x_mask, 1, 'Att_v_max') # b * s * c

    x_att = tf.contrib.keras.backend.batch_dot(tf.transpose(H,[0,2,1]), Att_v_max)
    H_enc = tf.squeeze(x_att)
    return H_enc


def aver_emb_encoder(x_emb, x_mask):
    """ compute the average over all word embeddings """
    x_mask = tf.expand_dims(x_mask, axis=-1)
    # x_mask = tf.expand_dims(x_mask, axis=-1)  # batch L 1 1

    x_sum = tf.multiply(x_emb, x_mask)  # batch L emb
    H_enc_0 = tf.reduce_sum(x_sum, axis=1, keep_dims=True)  # batch 1 emb
    H_enc = tf.squeeze(H_enc_0, [1,])  # batch emb
    x_mask_sum = tf.reduce_sum(x_mask, axis=1, keep_dims=True)  # batch 1 1
    x_mask_sum = tf.squeeze(x_mask_sum, [2,])  # batch 1

    #pdb.set_trace()

    H_enc = H_enc / x_mask_sum  # batch emb

    return H_enc

def gru_encoder(X_emb, opt, prefix = '', is_reuse=None, res=None):
    with tf.variable_scope(prefix + 'gru_encoder', reuse=True):
        cell_fw = tf.contrib.rnn.GRUCell(opt.n_hid)
        cell_bw = tf.contrib.rnn.GRUCell(opt.n_hid)
    with tf.variable_scope(prefix + 'gru_encoder', reuse=is_reuse):
        weightInit = tf.random_uniform_initializer(-0.001, 0.001)

        packed_output, state = tf.nn.bidirectional_dynamic_rnn(cell_fw, cell_bw, X_emb , dtype = tf.float32)
        h_fw = state[0]
        h_bw = state[1]

        hidden = tf.concat((h_fw, h_bw), 1)

        hidden = tf.nn.l2_normalize(hidden, 1)
    return hidden, res

def discriminator_1layer(H, opt, dropout, prefix='', num_outputs=1, is_reuse=None):
    # last layer must be linear
    H = tf.squeeze(H)
    biasInit = tf.constant_initializer(0.001, dtype=tf.float32)
    H_dis = layers.fully_connected(tf.nn.dropout(H, keep_prob=dropout), num_outputs=opt.H_dis,
                                   biases_initializer=biasInit, activation_fn=tf.nn.relu, scope=prefix + 'dis_1',
                                   reuse=is_reuse)
    return H_dis


def discriminator_0layer(H, opt, dropout, prefix='', num_outputs=1, is_reuse=None):
    H = tf.squeeze(H)
    biasInit = tf.constant_initializer(0.001, dtype=tf.float32)
    logits = layers.linear(tf.nn.dropout(H, keep_prob=dropout), num_outputs=num_outputs, biases_initializer=biasInit,
                           scope=prefix + 'dis', reuse=is_reuse)
    return logits

def linear_layer(x, output_dim):
    input_dim = x.get_shape().as_list()[1]
    thres = np.sqrt(6.0 / (input_dim + output_dim))
    W = tf.get_variable("W", [input_dim, output_dim], scope=prefix + '_W',
                        initializer=tf.random_uniform_initializer(minval=-thres, maxval=thres))
    b = tf.get_variable("b", [output_dim], scope=prefix + '_b', initializer=tf.constant_initializer(0.0))
    return tf.matmul(x, W) + b


def discriminator_2layer(H, opt, dropout, prefix='', num_outputs=1, is_reuse=None):
    # last layer must be linear
    biasInit = tf.constant_initializer(0.001, dtype=tf.float32)
    H_dis = layers.fully_connected(tf.nn.dropout(H, keep_prob=dropout), num_outputs=opt.H_dis,
                                   biases_initializer=biasInit, activation_fn=tf.nn.relu, scope=prefix + 'dis_1',
                                   reuse=is_reuse)
    logits = layers.linear(tf.nn.dropout(H_dis, keep_prob=dropout), num_outputs=num_outputs,
                           biases_initializer=biasInit, scope=prefix + 'dis_2', reuse=is_reuse)
    return logits

def discriminator_3layer(H, opt, dropout, prefix='', num_outputs=1, is_reuse=None):
    # last layer must be linear
    biasInit = tf.constant_initializer(0.001, dtype=tf.float32)
    H_dis = layers.fully_connected(tf.nn.dropout(H, keep_prob=dropout), num_outputs=opt.H_dis,
                                   biases_initializer=biasInit, activation_fn=tf.nn.relu, scope=prefix + 'dis_1',
                                   reuse=is_reuse)
    H_dis = layers.fully_connected(tf.nn.dropout(H_dis, keep_prob=dropout), num_outputs=opt.H_dis,
                                   biases_initializer=biasInit, activation_fn=tf.nn.relu, scope=prefix + 'dis_2',
                                   reuse=is_reuse)
    logits = layers.linear(tf.nn.dropout(H_dis, keep_prob=dropout), num_outputs=num_outputs,
                           biases_initializer=biasInit, scope=prefix + 'dis_3', reuse=is_reuse)
    return logits

def partial_softmax(logits, weights, dim, name,):
    with tf.name_scope('partial_softmax'):
        exp_logits = tf.exp(logits)
        if len(exp_logits.get_shape()) == len(weights.get_shape()):
            exp_logits_weighted = tf.multiply(exp_logits, weights)
        else:
            exp_logits_weighted = tf.multiply(exp_logits, tf.expand_dims(weights, -1))
        exp_logits_sum = tf.reduce_sum(exp_logits_weighted, axis=dim, keep_dims=True)
        partial_softmax_score = tf.div(exp_logits_weighted, exp_logits_sum, name=name)
        return partial_softmax_score