-
Notifications
You must be signed in to change notification settings - Fork 192
/
rnn_s2s.py
155 lines (121 loc) · 5.1 KB
/
rnn_s2s.py
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
import tensorflow as tf
from tensorflow.keras.models import *
from tensorflow.keras.layers import *
from tensorflow.keras.callbacks import *
import tensorflow.keras.backend as K
from dataloader import TokenList, pad_to_longest
class Encoder():
def __init__(self, i_token_num, latent_dim, layers=3):
self.emb_layer = Embedding(i_token_num, latent_dim, mask_zero=True)
cells = [GRUCell(latent_dim) for _ in range(layers)]
self.rnn_layer = RNN(cells, return_state=True)
def __call__(self, x):
x = self.emb_layer(x)
xh = self.rnn_layer(x)
x, h = xh[0], xh[1:]
return x, h
class Decoder():
def __init__(self, o_token_num, latent_dim, layers=3):
self.emb_layer = Embedding(o_token_num, latent_dim, mask_zero=True)
cells = [GRUCell(latent_dim) for _ in range(layers)]
self.rnn_layer = RNN(cells, return_sequences=True, return_state=True)
self.out_layer = Dense(o_token_num)
def __call__(self, x, state):
x = self.emb_layer(x)
xh = self.rnn_layer(x, initial_state=state)
x, h = xh[0], xh[1:]
x = TimeDistributed(self.out_layer)(x)
return x, h
def oloss(y_true, y_pred): return K.mean(y_pred)
class RNNSeq2Seq:
def __init__(self, i_tokens, o_tokens, latent_dim, layers=3):
self.i_tokens = i_tokens
self.o_tokens = o_tokens
encoder_inputs = Input(shape=(None,), dtype='int32')
decoder_inputs = Input(shape=(None,), dtype='int32')
encoder = Encoder(i_tokens.num(), latent_dim, layers)
decoder = Decoder(o_tokens.num(), latent_dim, layers)
encoder_outputs, encoder_states = encoder(encoder_inputs)
dinputs = Lambda(lambda x:x[:,:-1])(decoder_inputs)
dtargets = Lambda(lambda x:x[:,1:])(decoder_inputs)
decoder_outputs, decoder_state_h = decoder(dinputs, encoder_states)
def get_loss(args):
y_pred, y_true = args
y_true = tf.cast(y_true, 'int32')
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y_true, logits=y_pred)
mask = tf.cast(tf.not_equal(y_true, 0), 'float32')
loss = tf.reduce_sum(loss * mask, -1) / tf.reduce_sum(mask, -1)
loss = K.mean(loss)
return loss
def get_accu(args):
y_pred, y_true = args
mask = tf.cast(tf.not_equal(y_true, 0), 'float32')
corr = K.cast(K.equal(K.cast(y_true, 'int32'), K.cast(K.argmax(y_pred, axis=-1), 'int32')), 'float32')
corr = K.sum(corr * mask, -1) / K.sum(mask, -1)
return K.mean(corr)
loss = Lambda(get_loss)([decoder_outputs, dtargets])
self.ppl = Lambda(K.exp)(loss)
self.accu = Lambda(get_accu)([decoder_outputs, dtargets])
self.model = Model([encoder_inputs, decoder_inputs], loss)
self.model.add_loss([K.mean(loss)])
encoder_model = Model(encoder_inputs, encoder_states)
decoder_states_inputs = [Input(shape=(latent_dim,)) for _ in range(3)]
decoder_outputs, decoder_states = decoder(decoder_inputs, decoder_states_inputs)
decoder_model = Model([decoder_inputs] + decoder_states_inputs,
[decoder_outputs] + decoder_states)
self.encoder_model = encoder_model
self.decoder_model = decoder_model
def compile(self, optimizer):
self.model.add_metric(self.ppl, name='ppl')
self.model.add_metric(self.accu, name='accu')
self.model.compile(optimizer, None)
def decode_sequence(self, input_seq, delimiter=''):
input_mat = np.zeros((1, len(input_seq)+3))
input_mat[0,0] = self.i_tokens.id('<S>')
for i, z in enumerate(input_seq): input_mat[0,1+i] = self.i_tokens.id(z)
input_mat[0,len(input_seq)+1] = self.i_tokens.id('</S>')
state_value = self.encoder_model.predict_on_batch(input_mat)
target_seq = np.zeros((1, 1))
target_seq[0,0] = self.o_tokens.id('<S>')
decoded_tokens = []
while True:
output_tokens_and_h = self.decoder_model.predict_on_batch([target_seq] + state_value)
output_tokens, h = output_tokens_and_h[0], output_tokens_and_h[1:]
sampled_token_index = np.argmax(output_tokens[0,-1,:])
sampled_token = self.o_tokens.token(sampled_token_index)
decoded_tokens.append(sampled_token)
if sampled_token == '</S>' or len(decoded_tokens) > 50: break
target_seq = np.zeros((1, 1))
target_seq[0,0] = sampled_token_index
state_value = h
return delimiter.join(decoded_tokens[:-1])
import random
import numpy as np
if __name__ == '__main__':
itokens = TokenList(list('0123456789'))
otokens = TokenList(list('0123456789abcdefx'))
def GenSample():
x = random.randint(0, 99999)
y = hex(x); x = str(x)
return x, y
X, Y = [], []
for _ in range(100000):
x, y = GenSample()
X.append(list(x))
Y.append(list(y))
X, Y = pad_to_longest(X, itokens), pad_to_longest(Y, otokens)
print(X.shape, Y.shape)
s2s = RNNSeq2Seq(itokens, otokens, 128)
s2s.compile('rmsprop')
s2s.model.summary()
class TestCallback(Callback):
def on_epoch_end(self, epoch, logs = None):
print('\n')
for test in [123, 12345, 34567]:
ret = s2s.decode_sequence(str(test))
print(test, ret, hex(test))
print('\n')
#s2s.model.load_weights('model.h5')
s2s.model.fit([X, Y], None, batch_size=1024, epochs=10,
validation_split=0.05, callbacks=[TestCallback()])
s2s.model.save_weights('model.h5')