def build(self, input_shape):
self.input_spec = tf.keras.layers.InputSpec(shape=input_shape)
attention_head_size = int(self.hidden_size / self.num_attention_heads)
self.attention_layer = MultiAttentionLayer(
num_attention_heads=self.num_attention_heads,
size_per_head=attention_head_size,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
initializer_range=self.initializer_range,
do_return_2d_tensor=True,
batch_size=self.batch_size,
from_seq_length=self.seq_length,
to_seq_length=self.seq_length,
name="self")
self.attention_output_layer = tf.keras.layers.Dense(
self.hidden_size,
kernel_initializer=create_initializer(self.initializer_range),
name="dense")
self.inter_output = tf.keras.layers.Dense(
self.intermediate_size,
activation=self.intermediate_act_fn,
kernel_initializer=create_initializer(self.initializer_range),
name="dense")
self.layer_out = tf.keras.layers.Dense(
self.hidden_size,
kernel_initializer=create_initializer(self.initializer_range),
name="dense")
self.dropout = tf.keras.layers.Dropout(self.hidden_dropout_prob)
self.layer_norm = tf.keras.layers.LayerNormalization(axis=-1,
name="LayerNorm")
self.out_layer_norm = tf.keras.layers.LayerNormalization(
axis=-1, name="LayerNorm")
self.built = True
def build(self, input_shape):
hidden_size = input_shape[-1]
self.c_fc = tf.keras.layers.Dense(
hidden_size * 4,
name="c_fc",
kernel_initializer=create_initializer(self.initializer_range))
self.c_proj = tf.keras.layers.Dense(
hidden_size,
name="c_proj",
kernel_initializer=create_initializer(self.initializer_range))
self.act = get_activation('gelu')
self.dropout = tf.keras.layers.Dropout(self.resid_pdrop_rate)
def build(self, input_shape):
self.token_embedding = WTEmbedding(
vocab_size=self.vocab_size,
embedding_size=self.hidden_size,
initializer_range=self.initializer_range,
word_embedding_name="embedding",
name="wte")
# position embedding
self.posembedding = tf.keras.layers.Embedding(
self.max_position_length,
self.hidden_size,
embeddings_initializer=create_initializer(self.initializer_range),
name="wpe",
)
self.embedding_drop = tf.keras.layers.Dropout(self.embedding_drop_rate)
self.encoder_layers = []
for layer_idx in range(self.num_hidden_layers):
self.encoder_layer = GPT2Transformer(
num_attention_heads=self.num_attention_heads,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
initializer_range=self.initializer_range,
epsilon=self.layer_norm_epsilon,
resid_out_rate=self.resid_out_rate,
name="h{}".format(layer_idx))
self.encoder_layers.append(self.encoder_layer)
self.ln_f = tf.keras.layers.LayerNormalization(
epsilon=self.layer_norm_epsilon, name='ln_f')
# self.ln_f = normalization.GPTNorm(epsilon=self.layer_norm_epsilon, name='ln_f')
self.built = True
def build(self, input_shape):
self.size_per_head = int(input_shape[-1] / self.num_attention_heads)
self.c_att = tf.keras.layers.Dense(
# 12*64
self.num_attention_heads * self.size_per_head * 3,
name="c_attn",
kernel_initializer=create_initializer(self.initializer_range))
self.c_proj = tf.keras.layers.Dense(
self.num_attention_heads * self.size_per_head,
name="c_proj",
kernel_initializer=create_initializer(self.initializer_range))
self.resid_out = tf.keras.layers.Dropout(self.resid_out_rate)
self.drop_out = tf.keras.layers.Dropout(
self.attention_probs_dropout_prob)
self.built = True
def build(self, input_shape):
self.input_spec = tf.keras.layers.InputSpec(shape=input_shape)
size_per_head = int(input_shape[2] / self.num_attention_heads)
self.q = dense.DenseLayer3d(self.num_attention_heads, size_per_head,
create_initializer(self.initializer_range),
self.query_act, self.use_einsum, "query")
self.k = dense.DenseLayer3d(self.num_attention_heads, size_per_head,
create_initializer(self.initializer_range),
self.key_act, self.use_einsum, "key")
self.v = dense.DenseLayer3d(self.num_attention_heads, size_per_head,
create_initializer(self.initializer_range),
self.value_act, self.use_einsum, "value")
self.drop_out = tf.keras.layers.Dropout(
self.attention_probs_dropout_prob)
self.built = True
def build(self, input_shape):
self.nx = input_shape[-1]
self.weight = self.add_weight("w",
shape=[1, self.nx, self.nf],
initializer=create_initializer(
self.initializer_range))
self.bias = self.add_weight("b",
shape=[self.nf],
initializer=tf.zeros_initializer())
self.built = True
def build(self, input_shape):
self.input_spec = tf.keras.layers.InputSpec(shape=input_shape)
self.embedding_table = self.add_weight(
name=self.word_embedding_name,
dtype=tf.keras.backend.floatx(),
shape=[self.vocab_size, self.embedding_size],
initializer=create_initializer(self.initializer_range),
trainable=True,
)
self.built = True
def build(self, input_shape):
self.input_spec = tf.keras.layers.InputSpec(shape=input_shape)
self.attention = ALBERTAttention(
num_attention_heads=self.num_attention_heads,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
initializer_range=self.initializer_range,
use_einsum=True,
name='self',
)
self.dense_layer_3d_proj = dense.DenseLayer3dProj(
self.hidden_size,
self.attention_head_size,
create_initializer(self.initializer_range),
None,
use_einsum=self.use_einsum,
name="dense")
self.dense_layer_2d = dense.DenseLayer2d(
self.intermediate_size,
create_initializer(self.initializer_range),
self.intermediate_act_fn,
use_einsum=self.use_einsum,
num_attention_heads=self.num_attention_heads,
name="dense")
self.out_dense_layer_2d = dense.DenseLayer2d(
self.hidden_size,
create_initializer(self.initializer_range),
None,
use_einsum=self.use_einsum,
num_attention_heads=self.num_attention_heads,
name="dense")
self.attdropout = tf.keras.layers.Dropout(self.hidden_dropout_prob)
self.ffdropout = tf.keras.layers.Dropout(self.hidden_dropout_prob)
self.attlayer_norm = tf.keras.layers.LayerNormalization(
axis=-1, name="LayerNorm")
self.ffnlayer_norm = tf.keras.layers.LayerNormalization(
axis=-1, name="LayerNorm")
self.built = True
def build(self, input_shape):
self.input_spec = tf.keras.layers.InputSpec(shape=input_shape)
# `query_layer` =[B*F, N*H]
self._query_layer = tf.keras.layers.Dense(
self.num_attention_heads * self.size_per_head,
activation=self.query_act,
name="query",
kernel_initializer=create_initializer(self.initializer_range))
# `value_layer` = [B*T, N*H]
self._key_layer = tf.keras.layers.Dense(
self.num_attention_heads * self.size_per_head,
activation=self.key_act,
name="key",
kernel_initializer=create_initializer(self.initializer_range))
# `query_layer` =[B*T, N*H]
self._value_layer = tf.keras.layers.Dense(
self.num_attention_heads * self.size_per_head,
activation=self.value_act,
name="value",
kernel_initializer=create_initializer(self.initializer_range))
self.drop_out = tf.keras.layers.Dropout(
self.attention_probs_dropout_prob)
self.built = True
def build(self, input_shape):
input_ids_shape = input_shape
self.input_spec = tf.keras.layers.InputSpec(shape=input_ids_shape)
self.token_type_table = self.add_weight(
name=self.token_type_embedding_name,
shape=[self.token_type_vocab_size, input_shape[2]],
dtype=tf.keras.backend.floatx(),
initializer=create_initializer(self.initializer_range),
trainable=True)
self.full_position_embeddings = self.add_weight(
name=self.position_embedding_name,
shape=[self.max_position_embeddings, input_shape[2]],
dtype=tf.keras.backend.floatx(),
initializer=create_initializer(self.initializer_range),
trainable=True)
self.drop_out = tf.keras.layers.Dropout(self.hidden_dropout_prob)
self.layer_norm = tf.keras.layers.LayerNormalization(axis=-1,
name="LayerNorm")
self.built = True
def build(self, input_shape):
self.token_embedding = WDEmbedding(vocab_size=self.vocab_size,
embedding_size=self.embedding_size,
initializer_range=self.initializer_range,
word_embedding_name="word_embeddings",
use_one_hot_embedding=self.use_one_hot_embedding,
name="embeddings")
# segment and position embedding
self.segposembedding = SegPosEmbedding(use_token_type=True,
hidden_dropout_prob=self.hidden_dropout_prob,
token_type_vocab_size=self.type_vocab_size,
token_type_embedding_name="token_type_embeddings",
use_position_embeddings=True,
position_embedding_name="position_embeddings",
initializer_range=self.initializer_range,
max_position_embeddings=self.max_position_embeddings,
use_one_hot_embedding=self.use_one_hot_embedding,
name="embeddings"
)
self.shape_change = dense.DenseLayer2d(
self.hidden_size,
create_initializer(self.initializer_range),
None,
use_einsum=self.use_einsum,
name="embedding_hidden_mapping_in",
)
self.encoder_layer = AlbertTransformer(
hidden_size=self.hidden_size,
num_attention_heads=self.num_attention_heads,
attention_head_size=self.attention_head_size,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
intermediate_size=self.intermediate_size,
intermediate_act_fn=get_activation(self.hidden_act),
initializer_range=self.initializer_range,
hidden_dropout_prob=self.hidden_dropout_prob,
use_einsum=True,
name="inner_group_{}".format(0)
)
self.pool_out = tf.keras.layers.Dense(
self.hidden_size,
activation=tf.tanh,
# kernel_constraint=create_initializer(self.initializer_range),
name="dense")
self.built = True
