# Copyright 2018 MLBenchmark Group. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Transformer model helper methods."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import math
import tensorflow as tf
_NEG_INF = -1e9
#_NEG_INF_FP16 = -1e4
[docs]def get_position_encoding(
length, hidden_size, min_timescale=1.0, max_timescale=1.0e4):
"""Return positional encoding.
Calculates the position encoding as a mix of sine and cosine functions with
geometrically increasing wavelengths.
Defined and formulized in Attention is All You Need, section 3.5.
Args:
length: Sequence length.
hidden_size: Size of the
min_timescale: Minimum scale that will be applied at each position
max_timescale: Maximum scale that will be applied at each position
Returns:
Tensor with shape [length, hidden_size]
"""
position = tf.cast(tf.range(length),dtype=tf.float32)
num_timescales = hidden_size // 2
log_timescale_increment = (
math.log(float(max_timescale) / float(min_timescale)) /
(tf.cast((num_timescales) - 1, dtype=tf.float32)))
inv_timescales = min_timescale * tf.exp(
tf.cast(tf.range(num_timescales),dtype=tf.float32 ) * -log_timescale_increment)
scaled_time = tf.expand_dims(position, 1) * tf.expand_dims(inv_timescales, 0)
signal = tf.concat([tf.sin(scaled_time), tf.cos(scaled_time)], axis=1)
return signal
[docs]def get_decoder_self_attention_bias(length, dtype=tf.float32):
"""Calculate bias for decoder that maintains model's autoregressive property.
Creates a tensor that masks out locations that correspond to illegal
connections, so prediction at position i cannot draw information from future
positions.
Args:
length: int length of sequences in batch.
Returns:
float tensor of shape [1, 1, length, length]
"""
#print("get_decoder_self_attention_bias", dtype)
with tf.name_scope("decoder_self_attention_bias"):
#valid_locs = tf.matrix_band_part(tf.ones([length, length], dtype=dtype), -1, 0)
valid_locs = tf.matrix_band_part(tf.ones([length, length], dtype=tf.float32), -1, 0)
valid_locs = tf.reshape(valid_locs, [1, 1, length, length])
neg_inf=_NEG_INF #if (dtype==tf.float32) else _NEG_INF_FP16
bias = neg_inf * (1.0 - valid_locs)
#bias=tf.saturate_cast(bias, dtype=dtype)
return bias
[docs]def get_padding(x, padding_value=0, dtype=tf.float32):
"""Return float tensor representing the padding values in x.
Args:
x: int tensor with any shape
padding_value: int value that
dtype: type of the output
Returns:
float tensor with same shape as x containing values 0 or 1.
0 -> non-padding, 1 -> padding
"""
#print("get_padding", dtype)
with tf.name_scope("padding"):
return tf.cast(tf.equal(x, padding_value), dtype=dtype)
[docs]def get_padding_bias(x, res_rank=4, pad_sym=0, dtype=tf.float32):
"""Calculate bias tensor from padding values in tensor.
Bias tensor that is added to the pre-softmax multi-headed attention logits,
which has shape [batch_size, num_heads, length, length]. The tensor is zero at
non-padding locations, and -1e9 (negative infinity) at padding locations.
Args:
x: int tensor with shape [batch_size, length]
res_rank: int indicates the rank of attention_bias.
dtype: type of the output attention_bias
pad_sym: int the symbol used for padding
Returns:
Attention bias tensor of shape
[batch_size, 1, 1, length] if res_rank = 4 - for Transformer
or [batch_size, 1, length] if res_rank = 3 - for ConvS2S
"""
#print("get_padding_bias", dtype)
with tf.name_scope("attention_bias"):
padding = get_padding(x, padding_value=pad_sym, dtype=tf.float32)
# padding = get_padding(x, padding_value=pad_sym, dtype=dtype)
neg_inf=_NEG_INF #if dtype==tf.float32 else _NEG_INF_FP16
attention_bias = padding * neg_inf
if res_rank == 4:
attention_bias = tf.expand_dims(tf.expand_dims(attention_bias, axis=1), axis=1)
elif res_rank == 3:
attention_bias = tf.expand_dims(attention_bias, axis=1)
else:
raise ValueError("res_rank should be 3 or 4 but got {}".format(res_rank))
return attention_bias