def decode(self, targets, encoder_outputs, attention_bias):
"""Generate logits for each value in the target sequence.
Args:
targets: target values for the output sequence.
int tensor with shape [batch_size, target_length]
encoder_outputs: continuous representation of input sequence.
float tensor with shape [batch_size, input_length, hidden_size]
attention_bias: float tensor with shape [batch_size, 1, 1, input_length]
Returns:
float32 tensor with shape [batch_size, target_length, vocab_size]
"""
with tf.name_scope("decode"):
# Prepare inputs to decoder layers by shifting targets, adding positional
# encoding and applying dropout.
decoder_inputs = self.embedding_softmax_layer(targets)
with tf.name_scope("shift_targets"):
# Shift targets to the right, and remove the last element
decoder_inputs = tf.pad(decoder_inputs, [[0, 0], [1, 0], [0, 0]])[:, :-1, :]
with tf.name_scope("add_pos_encoding"):
length = tf.shape(decoder_inputs)[1]
decoder_inputs += model_utils.get_position_encoding(length, self.params["hidden_size"])
if self.train:
decoder_inputs = tf.nn.dropout(
decoder_inputs, 1 - self.params["layer_postprocess_dropout"])
# Run values
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(length)
outputs = self.decoder_stack(decoder_inputs, encoder_outputs, decoder_self_attention_bias, attention_bias)
logits = self.embedding_softmax_layer.linear(outputs)
return logits
def _get_symbols_to_logits_fn(self, max_decode_length, training):
"""Returns a decoding function that calculates logits of the next tokens."""
timing_signal = model_utils.get_position_encoding(
max_decode_length + 1, self.params["hidden_size"])
timing_signal = tf.cast(timing_signal, self.params["dtype"])
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
max_decode_length, dtype=self.params["dtype"])
# TODO(b/139770046): Refactor code with better naming of i.
def symbols_to_logits_fn(ids, i, cache):
"""Generate logits for next potential IDs.
Args:
ids: Current decoded sequences. int tensor with shape [batch_size *
beam_size, i + 1].
i: Loop index.
cache: dictionary of values storing the encoder output, encoder-decoder
attention bias, and previous decoder attention values.
Returns:
Tuple of
(logits with shape [batch_size * beam_size, vocab_size],
updated cache values)
"""
# Set decoder input to the last generated IDs
decoder_input = ids[:, -1:]
# Preprocess decoder input by getting embeddings and adding timing signal.
decoder_input = self.embedding_softmax_layer(decoder_input)
if self.params["padded_decode"]:
timing_signal_shape = timing_signal.shape.as_list()
decoder_input += tf.slice(timing_signal, [i, 0],
[1, timing_signal_shape[1]])
bias_shape = decoder_self_attention_bias.shape.as_list()
self_attention_bias = tf.slice(
decoder_self_attention_bias, [0, 0, i, 0],
[bias_shape[0], bias_shape[1], 1, bias_shape[3]])
else:
decoder_input += timing_signal[i:i + 1]
self_attention_bias = decoder_self_attention_bias[:, :, i:i +
1, :i + 1]
decoder_outputs = self.decoder_stack(
decoder_input,
cache.get("encoder_outputs"),
self_attention_bias,
cache.get("encoder_decoder_attention_bias"),
training=training,
cache=cache,
decode_loop_step=i if self.params["padded_decode"] else None)
logits = self.embedding_softmax_layer(decoder_outputs,
mode="linear")
logits = tf.squeeze(logits, axis=[1])
return logits, cache
return symbols_to_logits_fn
def decode(self, _, inputs, encoder_outputs, attention_bias):
"""Generate logits for each value in the target sequence.
Args:
inputs:
int tensor (old dst sentence) with shape [batch_size, input_length].
encoder_outputs: continuous representation of diff sequence.
float tensor with shape [batch_size, input_length, hidden_size]
attention_bias: float tensor with shape [batch_size, 1, 1, input_length]
Returns:
float32 tensor with shape [batch_size, target_length, vocab_size]
"""
with tf.name_scope("decode"):
# Prepare inputs to decoder layers by adding positional
# encoding and applying dropout.
decoder_inputs = self.embedding_softmax_layer(inputs)
with tf.name_scope("add_pos_encoding"):
length = tf.shape(decoder_inputs)[1]
decoder_inputs += model_utils.get_position_encoding(
length, self.params["hidden_size"])
if self.train:
decoder_inputs = tf.nn.dropout(
decoder_inputs, 1 - self.params["layer_postprocess_dropout"])
# Run values
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
length)
outputs = self.decoder_stack(
decoder_inputs, encoder_outputs, decoder_self_attention_bias,
attention_bias)
logits = self.embedding_softmax_layer.linear(outputs)
return logits
def decode(self, start_tokens, targets, encoder_outputs, attention_bias):
with tf.name_scope("decode"):
with tf.name_scope("shift_targets"):
decoder_inputs = tf.concat(
[tf.expand_dims(start_tokens, axis=1), targets[:, :-1]],
axis=1)
decoder_inputs = self.decoder_embedding_layer(decoder_inputs)
with tf.name_scope("add_pos_encoding"):
length = tf.shape(decoder_inputs)[1]
decoder_inputs += model_utils.get_position_encoding(
length, self.params["hidden_size"])
if self.train:
decoder_inputs = tf.nn.dropout(
decoder_inputs,
1 - self.params["layer_postprocess_dropout"])
# Run values
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
length)
decoder_outputs = self.decoder_stack(decoder_inputs,
encoder_outputs,
decoder_self_attention_bias,
attention_bias)
outputs = self.output_embedding_layer(decoder_outputs)
return outputs
def test_get_decoder_self_attention_bias(self):
length = 5
bias = model_utils.get_decoder_self_attention_bias(length)
with self.test_session() as sess:
bias = sess.run(bias)
self.assertAllEqual(
[[[[0, NEG_INF, NEG_INF, NEG_INF, NEG_INF],
[0, 0, NEG_INF, NEG_INF, NEG_INF], [0, 0, 0, NEG_INF, NEG_INF],
[0, 0, 0, 0, NEG_INF], [0, 0, 0, 0, 0]]]], bias)
def _get_symbols_to_logits_fn(self, max_decode_length):
"""Returns a decoding function that calculates logits of the next tokens."""
# 返回一个能够计算下一个token的decode函数
timing_signal = model_utils.get_position_encoding( # 时序信息，形状是[length, hidden_size]
max_decode_length + 1, self.params["hidden_size"])
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
max_decode_length) # self attention 的偏差, 形状是[1, 1, length, length]
def symbols_to_logits_fn(ids, i, cache):
"""Generate logits for next potential IDs.
这个函数可以做到，给出已经预测的tokens的id，使用decoder和encode的信息，预测下一个token
ids表示已经预测出来的tokens
i表示当前是第i个位置，要被预测
cache应该是因为：训练时的decode只需要做一次，但是inference时的decode需要做多次，因为要逐个单词预测，多次decode用的encode信息是一样的，因此需要提前存储好。
Args:
ids: Current decoded sequences.
int tensor with shape [batch_size * beam_size, i + 1]，忽略batch_size，可以看出，这个ids不是整个句子的ids，而是从开始到某一位置的候选tokens的id
i: Loop index
cache: dictionary of values storing the encoder output, encoder-decoder
attention bias, and previous decoder attention values.
Returns:
Tuple of
(logits with shape [batch_size * beam_size, vocab_size],
updated cache values)
"""
# Set decoder input to the last generated IDs
decoder_input = ids[:,
-1:] # 貌似是想要获得句子中当前位置的候选tokens的ids，也就是获得形状 [batch_size * beam_size, 1]
# Preprocess decoder input by getting embeddings and adding timing signal.
# 做embedding，也就是[batch_size * beam_size, 1, hidden_size]
# 从这一步可以看出，在inference的decode的输入，就是用已经预测的tokens的最后一个token，构成句子长度为1的句子，做embedding，输入到decode进行解码
decoder_input = self.embedding_softmax_layer(decoder_input)
decoder_input += timing_signal[i:i + 1] # 加上第i个token的时序信息
self_attention_bias = decoder_self_attention_bias[:, :,
i:i + 1, :i +
1] # self attention，形状是[1, 1, 1, i+1]
decoder_outputs = self.decoder_stack( # 进行decode，输出tensor的形状和输入decoder_input一样，也是[batch_size * beam_size, 1, hidden_size]
decoder_input,
cache.get("encoder_outputs"), self_attention_bias,
cache.get("encoder_decoder_attention_bias"), cache)
# softmax，从[batch_size * beam_size, 1, hidden_size]映射到[batch_size * beam_size, 1, vocab_size]
logits = self.embedding_softmax_layer.linear(decoder_outputs)
logits = tf.squeeze(
logits, axis=[1]
) # 去掉中间那个长度为1的维度，即由[batch_size * beam_size, 1, vocab_size]变为[batch_size * beam_size, vocab_size]
return logits, cache
return symbols_to_logits_fn
def _get_symbols_to_logits_fn(self, max_decode_length):
"""Returns a decoding function that calculates logits of the next tokens."""
timing_signal = model_utils.get_position_encoding(
max_decode_length + 1, self.params["hidden_size"])
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
max_decode_length)
def symbols_to_logits_fn(ids, i, cache):
"""Generate logits for next potential IDs.
Args:
ids: Current decoded sequences.
int tensor with shape [batch_size * beam_size, i + 1]
i: Loop index
cache: dictionary of values storing the encoder output, encoder-decoder
attention bias, and previous decoder attention values.
Returns:
Tuple of
(logits with shape [batch_size * beam_size, vocab_size],
updated cache values)
"""
# Set decoder input to the last generated IDs
decoder_input = ids[:, -1:]
### domyounglee 2020.2.12
cls_dec_bias = model_utils.get_cls_dec_attention_bias(
tf.cast(tf.equal(decoder_input, 2), tf.int64))
#self.cls_attention_bias=None
# Preprocess decoder input by getting embeddings and adding timing signal.
decoder_input = self.embedding_softmax_layer(decoder_input)
decoder_input += timing_signal[i:i + 1]
self_attention_bias = decoder_self_attention_bias[:, :,
i:i + 1, :i + 1]
decoder_outputs = self.decoder_stack(
decoder_input,
cache.get("encoder_outputs"),
self_attention_bias,
cache.get("encoder_decoder_attention_bias"),
cls_attention_bias=None,
cls_dec_attention_bias=None,
identity_mask=None,
cache=cache)
logits = self.embedding_softmax_layer.linear(decoder_outputs)
logits = tf.squeeze(logits, axis=[1])
return logits, cache
return symbols_to_logits_fn
def decode(self, targets, encoder_outputs, attention_bias, training):
"""Generate logits for each value in the target sequence.
Args:
targets: target values for the output sequence. int tensor with shape
[batch_size, target_length]
encoder_outputs: continuous representation of input sequence. float tensor
with shape [batch_size, input_length, hidden_size]
attention_bias: float tensor with shape [batch_size, 1, 1, input_length]
training: boolean, whether in training mode or not.
Returns:
float32 tensor with shape [batch_size, target_length, vocab_size]
"""
with tf.name_scope("decode"):
# Prepare inputs to decoder layers by shifting targets, adding positional
# encoding and applying dropout.
decoder_inputs = self.embedding_softmax_layer(targets)
decoder_inputs = tf.cast(decoder_inputs, self.params["dtype"])
attention_bias = tf.cast(attention_bias, self.params["dtype"])
with tf.name_scope("shift_targets"):
# Shift targets to the right, and remove the last element
# 第二维在维首加了一行padding，去掉最后的EOS，看来输入没有BOS
decoder_inputs = tf.pad(decoder_inputs,
[[0, 0], [1, 0], [0, 0]])[:, :-1, :]
with tf.name_scope("add_pos_encoding"):
length = tf.shape(decoder_inputs)[1]
pos_encoding = model_utils.get_position_encoding(
length, self.params["hidden_size"])
pos_encoding = tf.cast(pos_encoding, self.params["dtype"])
decoder_inputs += pos_encoding
if training:
decoder_inputs = tf.nn.dropout(
decoder_inputs,
rate=self.params["layer_postprocess_dropout"])
# Run values
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
length, dtype=self.params["dtype"])
outputs = self.decoder_stack(decoder_inputs,
encoder_outputs,
decoder_self_attention_bias,
attention_bias,
training=training)
logits = self.embedding_softmax_layer(outputs, mode="linear")
logits = tf.cast(logits, tf.float32)
return logits
def predict(self, start_tokens, encoder_outputs,
encoder_decoder_attention_bias):
"""Return predicted sequence."""
with tf.name_scope('decode'):
batch_size = tf.shape(encoder_outputs)[0]
max_decode_length = self.params['sequence_length']
timing_signal = model_utils.get_position_encoding(
max_decode_length, self.params['hidden_size'])
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
max_decode_length)
# Create cache storing decoder attention values for each layer.
cache = {
'layer_%d' % layer: {
'k': tf.zeros([batch_size, 0, self.params['hidden_size']]),
'v': tf.zeros([batch_size, 0, self.params['hidden_size']])
}
for layer in range(self.params['num_hidden_layers'])
}
# Add encoder output and attention bias to the cache.
cache['encoder_outputs'] = encoder_outputs
cache[
'encoder_decoder_attention_bias'] = encoder_decoder_attention_bias
# Forward decoder_inputs to decoder_stack max_decode_length times instead of applying beam search.
decoder_outputs = tf.zeros(
[batch_size, 0, self.params['output_size']])
decoder_inputs = tf.expand_dims(start_tokens, axis=1)
for i in range(max_decode_length):
decoder_inputs = self.decoder_embedding_layer(decoder_inputs)
decoder_inputs += timing_signal[i:i + 1]
self_attention_bias = decoder_self_attention_bias[:, :, i:i +
1, :i + 1]
decoder_inputs = self.decoder_stack(
decoder_inputs, cache.get('encoder_outputs'),
self_attention_bias,
cache.get('encoder_decoder_attention_bias'), cache)
decoder_inputs = self.output_embedding_layer(decoder_inputs)
decoder_outputs = tf.concat([decoder_outputs, decoder_inputs],
axis=1)
return decoder_outputs
def _get_symbols_to_logits_fn(self, max_decode_length):
"""Returns a decoding function that calculates logits of the next tokens."""
timing_signal = model_utils.get_position_encoding(
max_decode_length + 1, self.params.hidden_size)
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
max_decode_length)
def symbols_to_logits_fn(ids, i, cache):
"""Generate logits for next potential IDs.
Args:
ids: Current decoded sequences.
int tensor with shape [batch_size * beam_size, i + 1]
i: Loop index
cache: dictionary of values storing the encoder output, encoder-decoder
attention bias, and previous decoder attention values.
Returns:
Tuple of
(logits with shape [batch_size * beam_size, vocab_size],
updated cache values)
"""
# Set decoder input to the last generated IDs
decoder_input = ids[:, -1:]
# Preprocess decoder input by getting embeddings and adding timing signal.
decoder_input = self.embedding_softmax_layer(decoder_input)
decoder_input += timing_signal[i:i + 1]
self_attention_bias = decoder_self_attention_bias[:, :, i:i + 1, :i + 1]
decoder_outputs = self.decoder_stack(
decoder_input, cache.get("encoder_outputs"), self_attention_bias,
cache.get("encoder_decoder_attention_bias"), cache)
logits = self.embedding_softmax_layer.linear(decoder_outputs)
logits = tf.squeeze(logits, axis=[1])
return logits, cache
return symbols_to_logits_fn
def decode(self, targets, encoder_outputs, attention_bias):
"""Generate logits for each value in the target sequence.
Args:
targets: target values for the output sequence.
int tensor with shape [batch_size, target_length]
encoder_outputs: continuous representation of input sequence.
float tensor with shape [batch_size, input_length, hidden_size]
attention_bias: float tensor with shape [batch_size, 1, 1, input_length]
Returns:
float32 tensor with shape [batch_size, target_length, vocab_size]
"""
with tf.name_scope("decode"):
# Prepare inputs to decoder layers by shifting targets, adding positional
# encoding and applying dropout.
decoder_inputs = self.embedding_softmax_layer(targets)
with tf.name_scope("shift_targets"):
# Shift targets to the right, and remove the last element
decoder_inputs = tf.pad(
decoder_inputs, [[0, 0], [1, 0], [0, 0]])[:, :-1, :]
with tf.name_scope("add_pos_encoding"):
length = tf.shape(decoder_inputs)[1]
decoder_inputs += model_utils.get_position_encoding(
length, self.params.hidden_size)
if self.train:
decoder_inputs = tf.nn.dropout(
decoder_inputs, 1 - self.params.layer_postprocess_dropout)
# Run values
decoder_self_attention_bias = model_utils.get_decoder_self_attention_bias(
length)
outputs = self.decoder_stack(
decoder_inputs, encoder_outputs, decoder_self_attention_bias,
attention_bias)
logits = self.embedding_softmax_layer.linear(outputs)
return logits
