-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
190 lines (162 loc) · 9.12 KB
/
model.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
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
from tensorflow.keras.layers import Dropout
import tensorflow as tf
from tensorflow.keras.preprocessing.sequence import pad_sequences
import numpy as np
import json
"""initialize transformer"""
def get_angles(pos, i, d_model):
angle_rates = 1 / np.power(10000, (2 * (i//2)) / np.float32(d_model))
return pos * angle_rates
def pos_encoding(position, d_model):
angle_rads = get_angles(np.arange(position)[:, np.newaxis],
np.arange(d_model)[np.newaxis, :],
d_model)
angle_rads[:, 0::2] = np.sin(angle_rads[:, 0::2])
angle_rads[:, 1::2] = np.cos(angle_rads[:, 1::2])
pos_encoding = angle_rads[np.newaxis, ...]
return tf.cast(pos_encoding, dtype=tf.float32)
class MultiheadSelfAttention(tf.keras.Model):
def __init__(self, d_model, num_heads, mask=False):
super(MultiheadSelfAttention, self).__init__()
self.depth = d_model // num_heads
self.num_heads = num_heads
self.query_dense = tf.keras.layers.Dense(d_model)
self.key_dense = tf.keras.layers.Dense(d_model)
self.value_dense = tf.keras.layers.Dense(d_model)
self.dense = tf.keras.layers.Dense(d_model)
self.mask = mask
def create_mask(self, length, width):
mask = np.zeros((length, width), dtype='float32')
for i in range(length):
for j in range(width):
if j > i:
mask[i, j] = -1e9
return tf.cast(mask, dtype=tf.float32)
def call(self, query, key, value):
query = self.query_dense(query)
key = self.key_dense(key)
value = self.value_dense(value)
batch_size = query.shape[0]
query = tf.reshape(query, [batch_size, -1, self.num_heads, self.depth])
query = tf.transpose(query, [0, 2, 1, 3])
key = tf.reshape(key, [batch_size, -1, self.num_heads, self.depth])
key = tf.transpose(key, [0, 2, 1, 3])
value = tf.reshape(value, [batch_size, -1, self.num_heads, self.depth])
value = tf.transpose(value, [0, 2, 1, 3])
score = tf.matmul(query, key, transpose_b=True)
score /= tf.math.sqrt(tf.dtypes.cast(self.depth, dtype=tf.float32))
if self.mask:
score += self.create_mask(score.shape[-2], score.shape[-1])
alignment = tf.nn.softmax(score, axis=-1)
context = tf.matmul(alignment, value)
context = tf.transpose(context, [0, 2, 1, 3])
context = tf.reshape(context, [batch_size, -1, self.depth * self.num_heads])
output = self.dense(context)
# heads has shape (batch, decoder_len, model_size)
return output
class Encoder(tf.keras.Model):
def __init__(self, vocab_size, d_model, num_heads, num_layers=1, dropout=0.1):
super(Encoder, self).__init__()
self.d_model = d_model
self.num_layers = num_layers
self.dropouts1 = [Dropout(dropout) for i in range(num_layers)]
self.dropouts2 = [Dropout(dropout) for i in range(num_layers)]
self.embedding = tf.keras.layers.Embedding(vocab_size + 1, d_model)
self.mha = [MultiheadSelfAttention(d_model, num_heads) for _ in range(num_layers)]
self.layer_normalization_1 = [tf.keras.layers.BatchNormalization() for _ in range(num_layers)]
self.layer_normalization_2 = [tf.keras.layers.BatchNormalization() for _ in range(num_layers)]
self.dense_1 = [tf.keras.layers.Dense(d_model * 4, activation='relu') for _ in range(num_layers)]
self.dense_2 = [tf.keras.layers.Dense(d_model) for _ in range(num_layers)]
def call(self, encoder_inputs):
encoder_outputs = self.embedding(encoder_inputs)
encoder_outputs *= tf.math.sqrt(tf.cast(self.d_model, tf.float32))
encoder_outputs += pos_encoding(encoder_outputs.shape[1], encoder_outputs.shape[2])
for i in range(self.num_layers):
attention = self.mha[i](query=encoder_outputs, key=encoder_outputs, value=encoder_outputs)
attention = self.dropouts1[i](attention)
attention = self.layer_normalization_1[i](encoder_outputs + attention)
encoder_outputs = self.dense_1[i](attention)
encoder_outputs = self.dense_2[i](encoder_outputs)
encoder_outputs = self.dropouts2[i](encoder_outputs)
encoder_outputs = self.layer_normalization_2[i](attention + encoder_outputs)
return encoder_outputs
class Decoder(tf.keras.Model):
def __init__(self, vocab_size, d_model, num_heads, num_layers=1, dropout=0.1):
super(Decoder, self).__init__()
self.d_model = d_model
self.num_layers = num_layers
self.embedding = tf.keras.layers.Embedding(vocab_size+1, d_model)
self.masked_mha = [MultiheadSelfAttention(d_model, num_heads, mask=True) for _ in range(num_layers)]
self.layer_normalization_1 = [tf.keras.layers.BatchNormalization() for _ in range(num_layers)]
self.layer_normalization_2 = [tf.keras.layers.BatchNormalization() for _ in range(num_layers)]
self.layer_normalization_3 = [tf.keras.layers.BatchNormalization() for _ in range(num_layers)]
self.dropouts1 = [tf.keras.layers.Dropout(dropout) for _ in range(num_layers)]
self.dropouts2 = [tf.keras.layers.Dropout(dropout) for _ in range(num_layers)]
self.dropouts3 = [tf.keras.layers.Dropout(dropout) for _ in range(num_layers)]
self.mha = [MultiheadSelfAttention(d_model, num_heads) for _ in range(num_layers)]
self.dense_1 = [tf.keras.layers.Dense(d_model * 4, activation='relu') for _ in range(num_layers)]
self.dense_2 = [tf.keras.layers.Dense(d_model) for _ in range(num_layers)]
self.dense = tf.keras.layers.Dense(vocab_size+1, activation = 'softmax')
def call(self, encoder_outputs, decoder_inputs):
decoder_outputs = self.embedding(decoder_inputs)
decoder_outputs *= tf.math.sqrt(tf.cast(self.d_model, tf.float32))
decoder_outputs += pos_encoding(decoder_outputs.shape[1], decoder_outputs.shape[2])
for i in range(self.num_layers):
masked_attention = self.masked_mha[i](query=decoder_outputs, key=decoder_outputs, value=decoder_outputs)
masked_attention = self.dropouts1[i](masked_attention)
masked_attention = self.layer_normalization_1[i](masked_attention + decoder_outputs)
attention = self.mha[i](query=masked_attention, key=encoder_outputs, value=encoder_outputs)
attention = self.dropouts2[i](attention)
attention = self.layer_normalization_2[i](attention+masked_attention)
decoder_outputs = self.dense_1[i](attention)
decoder_outputs = self.dense_2[i](decoder_outputs)
decoder_outputs = self.dropouts3[i](decoder_outputs)
decoder_outputs = self.layer_normalization_3[i](decoder_outputs + attention)
decoder_outputs = self.dense(decoder_outputs)
return decoder_outputs
class Transformer(tf.keras.Model):
def __init__(self, input_vocab_size, target_vocab_size, d_model, num_heads, num_layers_encoder=1,
num_layers_decoder=1):
super().__init__()
self.encoder = Encoder(vocab_size=input_vocab_size, d_model=d_model, num_heads=num_heads,
num_layers=num_layers_encoder)
self.decoder = Decoder(vocab_size=target_vocab_size, d_model=d_model, num_heads=num_heads,
num_layers=num_layers_decoder)
def call(self, inputs):
encoder_outputs = self.encoder(inputs[0])
output = self.decoder(encoder_outputs, inputs[1])
return output
with open('C:\\Users\\anton\\PycharmProjects\\transformer\\venv\\transformer_model\\input_dict.json', 'r') as fp:
input_dict = json.load(fp)
with open('C:\\Users\\anton\\PycharmProjects\\transformer\\venv\\transformer_model\\target_dict.json', 'r') as fp:
target_dict = json.load(fp)
with open('C:\\Users\\anton\\PycharmProjects\\transformer\\venv\\transformer_model\\target_dict_r.json', 'r') as fp:
target_dict_r = json.load(fp)
def encode(sentence):
sentence = sentence.split()
for i, el in enumerate(sentence):
sentence[i] = input_dict[el]
sentence = pad_sequences([sentence], maxlen=20, padding='post')
return sentence
def predict(sentence):
enc_input = encode(sentence)
print(sentence)
out_words = []
de_input = tf.constant([[1]], dtype=tf.int64)
while True:
de_output = transformer([enc_input, de_input])
new_word = tf.expand_dims(tf.argmax(de_output, -1)[:, -1], axis=1)
de_input = tf.concat((de_input, new_word), axis=-1)
out_words.append(target_dict_r[str(new_word.numpy()[0][0])])
if out_words[-1] == 'end_token' or len(out_words) >= 20:
break
return ' '.join(out_words[:-1])
num_heads = 8
d_model = 256
input_vocab_size = len(input_dict)
target_vocab_size = len(target_dict)
batch_size=100
transformer = Transformer(input_vocab_size=input_vocab_size, target_vocab_size=target_vocab_size,
d_model=d_model, num_heads=num_heads, num_layers_encoder=1, num_layers_decoder=1)
"""load weights imported from colab file"""
transformer.load_weights('C:\\Users\\anton\\PycharmProjects\\transformer\\venv\\transformer_model\\my_model')