File size: 10,514 Bytes
e8e72fb | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 | import tensorflow as tf
from tensorflow.keras import layers, Sequential
import tensorflow_hub as hub
from tensorflow.keras.utils import plot_model
import sys
import os
parent_root = os.path.abspath(os.path.join(os.path.dirname(os.path.abspath(__file__)), os.pardir))
sys.path.append(parent_root)
from config.configs import Params
params = Params()
class TransformerEncoderLayer(tf.keras.layers.Layer):
def __init__(self, d_model, nhead, dim_feedforward, dropout_rate=0.1):
"""
args:
- d_model: Embedding dim
- nhead: Number of heads in MultiHeadAttention
- dim_feedforward: Dense layer's dim
"""
super(TransformerEncoderLayer, self).__init__()
self.attention = layers.MultiHeadAttention(num_heads=nhead, key_dim=d_model // nhead)
self.feed_forward = tf.keras.Sequential([
layers.Dense(dim_feedforward, activation='relu'),
layers.Dense(d_model)
])
self.norm1 = layers.LayerNormalization(epsilon=1e-6)
self.norm2 = layers.LayerNormalization(epsilon=1e-6)
self.dropout = layers.Dropout(dropout_rate)
def call(self, inputs, training=True, mask=None):
attention_output = self.attention(inputs, inputs, inputs, attention_mask=mask)
attention_output = self.dropout(attention_output)
output1 = self.norm1(inputs + attention_output)
feed_forward_output = self.feed_forward(output1)
feed_forward_output = self.dropout(feed_forward_output)
output2 = self.norm2(output1 + feed_forward_output)
return output2
class TransformerEncoder(tf.keras.layers.Layer):
"""
Stack of TransformerEncoderLayer
"""
def __init__(self, num_layers, d_model, nhead, dim_feedforward, dropout_rate=0.1):
super(TransformerEncoder, self).__init__()
self.encoder_layers = [TransformerEncoderLayer(d_model, nhead, dim_feedforward, dropout_rate)
for _ in range(num_layers)]
def call(self, inputs, training=True, mask=None):
x = inputs
for layer in self.encoder_layers:
x = layer(x, training=training, mask=mask)
return x
class TransformerModel(object):
def __init__(self, word_vectorizer = None, char_vectorizer = None, word_embed = None, char_embed= None,
num_layers = None, d_model = None, nhead = None, dim_feedforward = None, pretrained_embedding = None,
glove_embed = None, bert_process = None, bert_layer = None, dropout_rate=0.1, num_classes=5):
super(TransformerModel, self).__init__()
"""
args:
- word_vectorizer: Word-level vectorizer
- char_vectorizer: Char-level vectorizer
- word_embed: Word-level embedding layer
- char_embed: Char-level embedding layer
- pretrained_embedding: "bert", "glove" or None. Default: None
- glove_embed: glove embedding layer
- bert_process: BERT input processing layer
- bert_layer: BERT embedding layer
- num_classes: Number of classes. Default: 5 ("Do not change")
- num_layers: Number of TransformerEncoder in the model
- d_model: Embedding dim
- nhead: Number of heads in MultiHeadAttention
- dim_feedforward: Dense layer's dim
"""
# Params
self.pretrained_embedding = pretrained_embedding
self.num_classes = num_classes
self.word_output_dim = params.WORD_OUTPUT_DIM
self.char_output_dim = params.CHAR_OUTPUT_DIM
self.concate_dim = self.word_output_dim + 2 * self.char_output_dim
self.line_ids_input_dim = params.LINE_IDS_DEPTH
self.length_lines_input_dim = params.LENGTH_LINES_DEPTH
self.total_lines_input_dim = params.TOTAL_LINES_DEPTH
# Vectorizer
self.word_vectorizer = word_vectorizer
self.char_vectorizer = char_vectorizer
# Embedding
self.word_embed = word_embed
self.char_embed = char_embed
self.glove_embed = glove_embed
self.bert_process = bert_process
self.bert_layer = bert_layer
# Define the TransformerEncoder
self.encoder = TransformerEncoder(num_layers, d_model, nhead, dim_feedforward, dropout_rate)
def word_level_branch(self, word_input):
"""
Pretrained BERT don't need vectorization layer
"""
if str(self.pretrained_embedding).lower() == "bert":
# Pretrained Bert embeddings
bert_input = self.bert_process(word_input)
bert_output = self.bert_layer(bert_input, training = False)
word_embeddings = bert_output['sequence_output']
else:
if (self.word_vectorizer):
if (str(self.pretrained_embedding).lower() == "glove"):
# Get glove embedding
word_vectors = self.word_vectorizer(word_input)
word_embeddings = self.glove_embed(word_vectors)
else:
# Original word_embeddings
word_vectors = self.word_vectorizer(word_input)
word_embeddings = self.word_embed(word_vectors)
else:
raise Exception("Please provide word vectorizer.")
x = layers.Dense(128, activation = "relu")(word_embeddings)
word_outputs = layers.BatchNormalization()(x)
return word_outputs
def char_level_branch(self, char_input):
"""
arg:
- char_input: char-level tokens embedding
"""
char_vectors = self.char_vectorizer(char_input)
char_embeddings = self.char_embed(char_vectors)
x = layers.Dense(128, activation ="relu")(char_embeddings)
x = layers.BatchNormalization()(x)
return x
def line_ids_branch(self, line_id_in):
"""
arg:
- line_id_in: line_ids one-hot embedding
"""
x = layers.Dense(64, activation = "relu")(line_id_in)
x = layers.BatchNormalization()(x)
return x
def length_lines_branch(self, length_line_in):
"""
arg:
- length_line_in: length_lines one-hot embedding
"""
x = layers.Dense(64, activation = "relu")(length_line_in)
x = layers.BatchNormalization()(x)
return x
def total_lines_branch(self, total_line_in):
"""
arg:
-total_lines_in: total_lines one-hot embedding
"""
x = layers.Dense(64, activation = "relu")(total_line_in)
x = layers.BatchNormalization()(x)
return x
def _get_model(self):
# 5 inputs
## Token-inputs
word_inputs = layers.Input(shape = [], dtype = tf.string, name = "token_input")
char_inputs = layers.Input(shape = (1, ), dtype = tf.string, name = "char_input")
## Positional inputs
line_ids_inputs = layers.Input(shape = (self.line_ids_input_dim, ), name = "line_ids_input")
length_lines_inputs = layers.Input(shape = (self.length_lines_input_dim, ), name = "length_lines_input")
total_lines_inputs = layers.Input(shape = (self.total_lines_input_dim, ), name = "total_lines_input")
#-----------------------------------------------
# Branch outputs
## Word-level
word_level_output = self.word_level_branch(word_inputs)
## Char-level branch
char_level_output = self.char_level_branch(char_inputs)
##line_ids, length_lines, total_lines branch
line_ids_output = self.line_ids_branch(line_ids_inputs)
length_lines_output = self.length_lines_branch(length_lines_inputs)
total_lines_output = self.total_lines_branch(total_lines_inputs)
#---------------------------------------------------------
#Concate two tokens-embeddings
word_char_concat = tf.concat([word_level_output,char_level_output], axis = 1)
#----------------------------------------------------------
# Concanate last three input
position_embed = tf.concat([length_lines_output, line_ids_output, total_lines_output], axis = 1)
# Reshape axis = 2 dimension
position_embed = layers.Dense(128, activation = "relu")(position_embed)
# Expand-dim
position_embed = tf.expand_dims(position_embed, axis = 1)
#------------------------------------------------------------
# concatenate 5 inputs
total_embed = tf.concat([word_char_concat, position_embed], axis = 1)
# TransformerEncoder layer
x = self.encoder(total_embed, training=True, mask= None)
# Bi-LSTM decoder layer
x = layers.Bidirectional(layers.LSTM(64))(x)
# FCN
x = layers.Dense(64, activation="relu")(x)
x = layers.BatchNormalization()(x)
x = layers.Dropout(0.5)(x)
output= layers.Dense(self.num_classes, activation = "softmax")(x)
model= tf.keras.Model(inputs=[word_inputs, char_inputs, line_ids_inputs, length_lines_inputs, total_lines_inputs],
outputs= output,
name="transformer_encoder_based_model")
model.compile(optimizer = params.OPTIMIZER, loss = params.LOSS, metrics = params.METRICS)
return model
def _define_checkpoint(self):
"""
Define checkpoint for model
"""
if str(self.pretrained_embedding).lower() == "glove":
if not os.path.exists(params.TF_BASED_GLOVE_MODEL_DIR):
os.makedirs(params.TF_BASED_GLOVE_MODEL_DIR)
checkpoint_dir = params.TF_BASED_GLOVE_MODEL_DIR
elif str(self.pretrained_embedding).lower() == "bert":
if not os.path.exists(params.TF_BASED_BERT_MODEL_DIR):
os.makedirs(params.TF_BASED_BERT_MODEL_DIR)
checkpoint_dir = params.TF_BASED_BERT_MODEL_DIR
else:
if not os.path.exists(params.TF_BASED_NOR_MODEL_DIR):
os.makedirs(params.TF_BASED_NOR_MODEL_DIR)
checkpoint_dir = params.TF_BASED_NOR_MODEL_DIR
checkpoint= tf.keras.callbacks.ModelCheckpoint(
filepath = checkpoint_dir + '/best_model.ckpt',
monitor = "val_categorical_accuracy",
save_best_only = True,
save_weights_only = True,
verbose = 1
)
return checkpoint
def _plot_model(self, model):
plot_model(model)
return
|