def __init__(self,
hidden_size=768,
seq_length=512,
num_attention_heads=12,
intermediate_size=3072,
attention_probs_dropout_prob=0.02,
use_one_hot_embeddings=False,
initializer_range=0.02,
hidden_dropout_prob=0.1,
use_relative_positions=False,
hidden_act="gelu",
compute_type=ts.float32):
super(BertEncoderLayer, self).__init__()
self.attention = BertSelfAttention(
hidden_size=hidden_size,
seq_length=seq_length,
num_attention_heads=num_attention_heads,
attention_probs_dropout_prob=attention_probs_dropout_prob,
use_one_hot_embeddings=use_one_hot_embeddings,
initializer_range=initializer_range,
hidden_dropout_prob=hidden_dropout_prob,
use_relative_positions=use_relative_positions,
compute_type=compute_type)
self.intermediate = layers.Dense(in_channels=hidden_size,
out_channels=intermediate_size,
activation=hidden_act,
weight_init=TruncatedNormal(initializer_range)).to_float(compute_type)
self.output = BertOutput(in_channels=intermediate_size,
out_channels=hidden_size,
initializer_range=initializer_range,
dropout_prob=hidden_dropout_prob,
compute_type=compute_type)
def init_net_param(network, initialize_mode='TruncatedNormal'):
"""Init the parameters in net."""
params = network.trainable_params()
for p in params:
if 'beta' not in p.name and 'gamma' not in p.name and 'bias' not in p.name:
if initialize_mode == 'TruncatedNormal':
p.set_data(
initializer(TruncatedNormal(0.02), p.data.shape,
p.data.dtype))
else:
p.set_data(initialize_mode, p.data.shape, p.data.dtype)
def __init__(self, config, is_training, num_labels=2, dropout_prob=0.0, use_one_hot_embeddings=False):
super(BertSquadModel, self).__init__()
if not is_training:
config.hidden_dropout_prob = 0.0
config.hidden_probs_dropout_prob = 0.0
self.bert = Bert(config, is_training, use_one_hot_embeddings)
self.weight_init = TruncatedNormal(config.initializer_range)
self.dense1 = layers.Dense(config.hidden_size, num_labels, weight_init=self.weight_init,
has_bias=True).to_float(config.compute_type)
self.num_labels = num_labels
self.dtype = config.dtype
self.log_softmax = P.LogSoftmax(axis=1)
self.is_training = is_training
def __init__(self,
embedding_size,
embedding_shape,
use_relative_positions=False,
use_token_type=False,
token_type_vocab_size=16,
use_one_hot_embeddings=False,
initializer_range=0.02,
max_position_embeddings=512,
dropout_prob=0.1):
super(EmbeddingPostprocessor, self).__init__()
self.use_token_type = use_token_type
self.token_type_vocab_size = token_type_vocab_size
self.use_one_hot_embeddings = use_one_hot_embeddings
self.max_position_embeddings = max_position_embeddings
self.embedding_table = Parameter(initializer
(TruncatedNormal(initializer_range),
[token_type_vocab_size,
embedding_size]),
name='embedding_table')
self.shape_flat = (-1,)
self.one_hot = layers.OneHot()
self.on_value = Tensor(1.0, ts.float32)
self.off_value = Tensor(0.1, ts.float32)
self.array_mul = P.MatMul()
self.reshape = P.Reshape()
self.shape = tuple(embedding_shape)
self.layernorm = layers.LayerNorm((embedding_size,))
self.dropout = layers.Dropout(1 - dropout_prob)
self.gather = P.Gather()
self.use_relative_positions = use_relative_positions
self.slice = P.StridedSlice()
self.full_position_embeddings = Parameter(initializer
(TruncatedNormal(initializer_range),
[max_position_embeddings,
embedding_size]),
name='full_position_embeddings')
def __init__(self,
in_channels,
out_channels,
initializer_range=0.02,
dropout_prob=0.1,
compute_type=ts.float32):
super(BertOutput, self).__init__()
self.dense = layers.Dense(in_channels, out_channels,
weight_init=TruncatedNormal(initializer_range)).to_float(compute_type)
self.dropout = layers.Dropout(1 - dropout_prob)
self.dropout_prob = dropout_prob
self.add = P.Add()
self.layernorm = layers.LayerNorm((out_channels,)).to_float(compute_type)
self.cast = P.Cast()
def __init__(self, config, is_training, num_labels=2, dropout_prob=0.0, use_one_hot_embeddings=False,
assessment_method=""):
super(BertCLSModel, self).__init__()
if not is_training:
config.hidden_dropout_prob = 0.0
config.hidden_probs_dropout_prob = 0.0
self.bert = Bert(config, is_training, use_one_hot_embeddings)
self.cast = P.Cast()
self.weight_init = TruncatedNormal(config.initializer_range)
self.log_softmax = P.LogSoftmax(axis=-1)
self.dtype = config.dtype
self.num_labels = num_labels
self.dense_1 = layers.Dense(config.hidden_size, self.num_labels, weight_init=self.weight_init,
has_bias=True).to_float(config.compute_type)
self.dropout = layers.Dropout(1 - dropout_prob)
self.assessment_method = assessment_method
def __init__(self, config, is_training, num_labels=11, use_crf=False, dropout_prob=0.0,
use_one_hot_embeddings=False):
super(BertNERModel, self).__init__()
if not is_training:
config.hidden_dropout_prob = 0.0
config.hidden_probs_dropout_prob = 0.0
self.bert = Bert(config, is_training, use_one_hot_embeddings)
self.cast = P.Cast()
self.weight_init = TruncatedNormal(config.initializer_range)
self.log_softmax = P.LogSoftmax(axis=-1)
self.dtype = config.dtype
self.num_labels = num_labels
self.dense_1 = layers.Dense(config.hidden_size, self.num_labels, weight_init=self.weight_init,
has_bias=True).to_float(config.compute_type)
self.dropout = layers.Dropout(1 - dropout_prob)
self.reshape = P.Reshape()
self.shape = (-1, config.hidden_size)
self.use_crf = use_crf
self.origin_shape = (-1, config.seq_length, self.num_labels)
def __init__(self,
vocab_size,
embedding_size,
embedding_shape,
use_one_hot_embeddings=False,
initializer_range=0.02):
super(EmbeddingLookup, self).__init__()
self.vocab_size = vocab_size
self.use_one_hot_embeddings = use_one_hot_embeddings
self.embedding_table = Parameter(initializer
(TruncatedNormal(initializer_range),
[vocab_size, embedding_size]))
self.expand = P.ExpandDims()
self.shape_flat = (-1,)
self.gather = P.Gather()
self.one_hot = P.OneHot()
self.on_value = Tensor(1.0, ts.float32)
self.off_value = Tensor(0.0, ts.float32)
self.array_mul = P.MatMul()
self.reshape = P.Reshape()
self.shape = tuple(embedding_shape)
def __init__(self,
length,
depth,
max_relative_position,
initializer_range,
use_one_hot_embeddings=False):
super(RelaPosEmbeddingsGenerator, self).__init__()
self.depth = depth
self.vocab_size = max_relative_position * 2 + 1
self.use_one_hot_embeddings = use_one_hot_embeddings
self.embeddings_table = Parameter(
initializer(TruncatedNormal(initializer_range),
[self.vocab_size, self.depth]))
self.relative_positions_matrix = RelaPosMatrixGenerator(length=length,
max_relative_position=max_relative_position)
self.reshape = P.Reshape()
self.one_hot = layers.OneHot(depth=self.vocab_size)
self.shape = P.Shape()
self.gather = P.Gather() # index_select
self.matmul = P.BatchMatMul()
def __init__(self,
config,
is_training,
use_one_hot_embeddings=False):
super(Bert, self).__init__()
config = copy.deepcopy(config)
if not is_training:
config.hidden_dropout_prob = 0.0
config.attention_probs_dropout_prob = 0.0
self.seq_length = config.seq_length
self.hidden_size = config.hidden_size
self.num_hidden_layers = config.num_hidden_layers
self.embedding_size = config.hidden_size
self.token_type_ids = None
self.last_idx = self.num_hidden_layers - 1
output_embedding_shape = [-1, self.seq_length, self.embedding_size]
self.bert_embedding_lookup = layers.Embedding(
vocab_size=config.vocab_size,
embedding_size=self.embedding_size,
use_one_hot=use_one_hot_embeddings,
embedding_table=TruncatedNormal(config.initializer_range))
self.bert_embedding_postprocessor = EmbeddingPostprocessor(
embedding_size=self.embedding_size,
embedding_shape=output_embedding_shape,
use_relative_positions=config.use_relative_positions,
use_token_type=True,
token_type_vocab_size=config.type_vocab_size,
use_one_hot_embeddings=use_one_hot_embeddings,
initializer_range=0.02,
max_position_embeddings=config.max_position_embeddings,
dropout_prob=config.hidden_dropout_prob)
self.bert_encoder = BertTransformer(
hidden_size=self.hidden_size,
seq_length=self.seq_length,
num_attention_heads=config.num_attention_heads,
num_hidden_layers=self.num_hidden_layers,
intermediate_size=config.intermediate_size,
attention_probs_dropout_prob=config.attention_probs_dropout_prob,
use_one_hot_embeddings=use_one_hot_embeddings,
initializer_range=config.initializer_range,
hidden_dropout_prob=config.hidden_dropout_prob,
use_relative_positions=config.use_relative_positions,
hidden_act=config.hidden_act,
compute_type=config.compute_type,
return_all_encoders=True)
self.cast = P.Cast()
self.dtype = config.dtype
self.cast_compute_type = SaturateCast(dst_type=config.compute_type)
self.slice = P.StridedSlice()
self.squeeze_1 = P.Squeeze(axis=1)
self.dense = layers.Dense(self.hidden_size, self.hidden_size,
activation="tanh",
weight_init=TruncatedNormal(config.initializer_range)).to_float(config.compute_type)
self._create_attention_mask_from_input_mask = CreateAttentionMaskFromInputMask(config)
def __init__(self,
from_tensor_width,
to_tensor_width,
from_seq_length,
to_seq_length,
num_attention_heads=1,
size_per_head=512,
query_act=None,
key_act=None,
value_act=None,
has_attention_mask=False,
attention_probs_dropout_prob=0.0,
use_one_hot_embeddings=False,
initializer_range=0.02,
do_return_2d_tensor=False,
use_relative_positions=False,
compute_type=ts.float32):
super(BertAttention, self).__init__()
self.from_seq_length = from_seq_length
self.to_seq_length = to_seq_length
self.num_attention_heads = num_attention_heads
self.size_per_head = size_per_head
self.has_attention_mask = has_attention_mask
self.use_relative_positions = use_relative_positions
self.scores_mul = 1.0 / math.sqrt(float(self.size_per_head))
self.reshape = P.Reshape()
self.shape_from_2d = (-1, from_tensor_width)
self.shape_to_2d = (-1, to_tensor_width)
weight = TruncatedNormal(initializer_range)
units = num_attention_heads * size_per_head
self.query_layer = layers.Dense(from_tensor_width,
units,
activation=query_act,
weight_init=weight).to_float(compute_type)
self.key_layer = layers.Dense(to_tensor_width,
units,
activation=key_act,
weight_init=weight).to_float(compute_type)
self.value_layer = layers.Dense(to_tensor_width,
units,
activation=value_act,
weight_init=weight).to_float(compute_type)
self.shape_from = (-1, from_seq_length, num_attention_heads, size_per_head)
self.shape_to = (-1, to_seq_length, num_attention_heads, size_per_head)
self.matmul_trans_b = P.BatchMatMul(transpose_b=True)
self.multiply = P.Mul()
self.transpose = P.Transpose()
self.trans_shape = (0, 2, 1, 3)
self.trans_shape_relative = (2, 0, 1, 3)
self.trans_shape_position = (1, 2, 0, 3)
self.multiply_data = -10000.0
self.matmul = P.BatchMatMul()
self.softmax = layers.Softmax()
self.dropout = layers.Dropout(1 - attention_probs_dropout_prob)
if self.has_attention_mask:
self.expand_dims = P.ExpandDims()
self.sub = P.Sub()
self.add = P.Add()
self.cast = P.Cast()
self.get_dtype = P.DType()
if do_return_2d_tensor:
self.shape_return = (-1, num_attention_heads * size_per_head)
else:
self.shape_return = (-1, from_seq_length, num_attention_heads * size_per_head)
self.cast_compute_type = SaturateCast(dst_type=compute_type)
if self.use_relative_positions:
self._generate_relative_positions_embeddings = \
RelaPosEmbeddingsGenerator(length=to_seq_length,
depth=size_per_head,
max_relative_position=16,
initializer_range=initializer_range,
use_one_hot_embeddings=use_one_hot_embeddings)