Commit aaa7a715 by libei

revise some transformer decoding configuration, support relative position…

revise some transformer decoding configuration, support relative position reprasentation training, add transformer_rpr_base
parent ad6cc8e3
# coding=utf-8 # Copyright 2017 The Tensor2Tensor Authors.
# Copyright 2018 The Tensor2Tensor Authors.
# #
# Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
...@@ -12,81 +11,28 @@ ...@@ -12,81 +11,28 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
"""Implementation of beam search with penalties."""
"""Implemetation of beam seach with penalties."""
from __future__ import absolute_import from __future__ import absolute_import
from __future__ import division from __future__ import division
from __future__ import print_function from __future__ import print_function
from tensor2tensor.models import common_layers
# Dependency imports
import tensorflow as tf import tensorflow as tf
from tensorflow.python.util import nest
# Assuming EOS_ID is 1 # Assuming EOS_ID is 1
EOS_ID = 1 EOS_ID = 1
# Default value for INF # Default value for INF
INF = 1. * 1e7 INF = 1. * 1e7
def _merge_beam_dim(tensor): def log_prob_from_logits(logits):
"""Reshapes first two dimensions in to single dimension. return logits - tf.reduce_logsumexp(logits, axis=2, keep_dims=True)
Args:
tensor: Tensor to reshape of shape [A, B, ...]
Returns:
Reshaped tensor of shape [A*B, ...]
"""
shape = common_layers.shape_list(tensor)
shape[0] *= shape[1] # batch -> batch * beam_size
shape.pop(1) # Remove beam dim
return tf.reshape(tensor, shape)
def _unmerge_beam_dim(tensor, batch_size, beam_size):
"""Reshapes first dimension back to [batch_size, beam_size].
Args:
tensor: Tensor to reshape of shape [batch_size*beam_size, ...]
batch_size: Tensor, original batch size.
beam_size: int, original beam size.
Returns:
Reshaped tensor of shape [batch_size, beam_size, ...]
"""
shape = common_layers.shape_list(tensor)
new_shape = [batch_size] + [beam_size] + shape[1:]
return tf.reshape(tensor, new_shape)
def _expand_to_beam_size(tensor, beam_size):
"""Tiles a given tensor by beam_size.
Args:
tensor: tensor to tile [batch_size, ...]
beam_size: How much to tile the tensor by.
Returns:
Tiled tensor [batch_size, beam_size, ...]
"""
tensor = tf.expand_dims(tensor, axis=1)
tile_dims = [1] * tensor.shape.ndims
tile_dims[1] = beam_size
return tf.tile(tensor, tile_dims)
def get_state_shape_invariants(tensor):
"""Returns the shape of the tensor but sets middle dims to None."""
shape = tensor.shape.as_list()
for i in range(1, len(shape) - 1):
shape[i] = None
return tf.TensorShape(shape)
def compute_batch_indices(batch_size, beam_size): def compute_batch_indices(batch_size, beam_size):
"""Computes the i'th coordinate that contains the batch index for gathers. """Computes the i'th coodinate that contains the batch index for gathers.
Batch pos is a tensor like [[0,0,0,0,],[1,1,1,1],..]. It says which Batch pos is a tensor like [[0,0,0,0,],[1,1,1,1],..]. It says which
batch the beam item is in. This will create the i of the i,j coordinate batch the beam item is in. This will create the i of the i,j coordinate
...@@ -104,20 +50,13 @@ def compute_batch_indices(batch_size, beam_size): ...@@ -104,20 +50,13 @@ def compute_batch_indices(batch_size, beam_size):
def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags, def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags,
beam_size, batch_size, prefix="default", beam_size, batch_size):
states_to_gather=None):
"""Given sequences and scores, will gather the top k=beam size sequences. """Given sequences and scores, will gather the top k=beam size sequences.
This function is used to grow alive, and finished. It takes sequences, This function is used to grow alive, and finished. It takes sequences,
scores, and flags, and returns the top k from sequences, scores_to_gather, scores, and flags, and returns the top k from sequences, scores_to_gather,
and flags based on the values in scores. and flags based on the values in scores.
This method permits easy introspection using tfdbg. It adds three named ops
that are prefixed by `prefix`:
- _topk_seq: the tensor for topk_seq returned by this method.
- _topk_flags: the tensor for topk_finished_flags returned by this method.
- _topk_scores: the tensor for tokp_gathered_scores returned by this method.
Args: Args:
sequences: Tensor of sequences that we need to gather from. sequences: Tensor of sequences that we need to gather from.
[batch_size, beam_size, seq_length] [batch_size, beam_size, seq_length]
...@@ -127,13 +66,11 @@ def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags, ...@@ -127,13 +66,11 @@ def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags,
[batch_size, beam_size]. We will return the gathered scores from here. [batch_size, beam_size]. We will return the gathered scores from here.
Scores to gather is different from scores because for grow_alive, we will Scores to gather is different from scores because for grow_alive, we will
need to return log_probs, while for grow_finished, we will need to return need to return log_probs, while for grow_finished, we will need to return
the length penalized scores. the length penalized scors.
flags: Tensor of bools for sequences that say whether a sequence has reached flags: Tensor of bools for sequences that say whether a sequence has reached
EOS or not EOS or not
beam_size: int beam_size: int
batch_size: int batch_size: int
prefix: string that will prefix unique names for the ops run.
states_to_gather: dict (possibly nested) of decoding states.
Returns: Returns:
Tuple of Tuple of
(topk_seq [batch_size, beam_size, decode_length], (topk_seq [batch_size, beam_size, decode_length],
...@@ -153,20 +90,11 @@ def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags, ...@@ -153,20 +90,11 @@ def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags,
# last dimension contains the i,j gathering coordinates. # last dimension contains the i,j gathering coordinates.
top_coordinates = tf.stack([batch_pos, topk_indexes], axis=2) top_coordinates = tf.stack([batch_pos, topk_indexes], axis=2)
# Gather up the highest scoring sequences. For each operation added, give it # Gather up the highest scoring sequences
# a concrete name to simplify observing these operations with tfdbg. Clients topk_seq = tf.gather_nd(sequences, top_coordinates)
# can capture these tensors by watching these node names. topk_flags = tf.gather_nd(flags, top_coordinates)
def gather(tensor, name): topk_gathered_scores = tf.gather_nd(scores_to_gather, top_coordinates)
return tf.gather_nd(tensor, top_coordinates, name=(prefix + name)) return topk_seq, topk_gathered_scores, topk_flags
topk_seq = gather(sequences, "_topk_seq")
topk_flags = gather(flags, "_topk_flags")
topk_gathered_scores = gather(scores_to_gather, "_topk_scores")
if states_to_gather:
topk_gathered_states = nest.map_structure(
lambda state: gather(state, "_topk_states"), states_to_gather)
else:
topk_gathered_states = states_to_gather
return topk_seq, topk_gathered_scores, topk_flags, topk_gathered_states
def beam_search(symbols_to_logits_fn, def beam_search(symbols_to_logits_fn,
...@@ -175,35 +103,14 @@ def beam_search(symbols_to_logits_fn, ...@@ -175,35 +103,14 @@ def beam_search(symbols_to_logits_fn,
decode_length, decode_length,
vocab_size, vocab_size,
alpha, alpha,
states=None, eos_id=EOS_ID):
eos_id=EOS_ID,
stop_early=True):
"""Beam search with length penalties. """Beam search with length penalties.
Requires a function that can take the currently decoded symbols and return Uses an interface specific to the sequence cnn models;
Requires a function that can take the currently decoded sybmols and return
the logits for the next symbol. The implementation is inspired by the logits for the next symbol. The implementation is inspired by
https://arxiv.org/abs/1609.08144. https://arxiv.org/abs/1609.08144.
When running, the beam search steps can be visualized by using tfdbg to watch
the operations generating the output ids for each beam step. These operations
have the pattern:
(alive|finished)_topk_(seq,scores)
Operations marked `alive` represent the new beam sequences that will be
processed in the next step. Operations marked `finished` represent the
completed beam sequences, which may be padded with 0s if no beams finished.
Operations marked `seq` store the full beam sequence for the time step.
Operations marked `scores` store the sequence's final log scores.
The beam search steps will be processed sequentially in order, so when
capturing observed from these operations, tensors, clients can make
assumptions about which step is being recorded.
WARNING: Assumes 2nd dimension of tensors in `states` and not invariant, this
means that the shape of the 2nd dimension of these tensors will not be
available (i.e. set to None) inside symbols_to_logits_fn.
Args: Args:
symbols_to_logits_fn: Interface to the model, to provide logits. symbols_to_logits_fn: Interface to the model, to provide logits.
Shoud take [batch_size, decoded_ids] and return [batch_size, vocab_size] Shoud take [batch_size, decoded_ids] and return [batch_size, vocab_size]
...@@ -215,34 +122,27 @@ def beam_search(symbols_to_logits_fn, ...@@ -215,34 +122,27 @@ def beam_search(symbols_to_logits_fn,
vocab_size: Size of the vocab, must equal the size of the logits returned by vocab_size: Size of the vocab, must equal the size of the logits returned by
symbols_to_logits_fn symbols_to_logits_fn
alpha: alpha for length penalty. alpha: alpha for length penalty.
states: dict (possibly nested) of decoding states.
eos_id: ID for end of sentence. eos_id: ID for end of sentence.
stop_early: a boolean - stop once best sequence is provably determined.
Returns: Returns:
Tuple of Tuple of
(decoded beams [batch_size, beam_size, decode_length] (decoded beams [batch_size, beam_size, decode_length]
decoding probabilities [batch_size, beam_size]) decoding probablities [batch_size, beam_size])
""" """
batch_size = common_layers.shape_list(initial_ids)[0] batch_size = tf.shape(initial_ids)[0]
# Assume initial_ids are prob 1.0 # Assume initial_ids are prob 1.0
initial_log_probs = tf.constant([[0.] + [-float("inf")] * (beam_size - 1)]) initial_log_probs = tf.constant([[0.] + [-float("inf")] * (beam_size - 1)])
# Expand to beam_size (batch_size, beam_size) # Expand to beam_size (batch_size, beam_size)
alive_log_probs = tf.tile(initial_log_probs, [batch_size, 1]) alive_log_probs = tf.tile(initial_log_probs, [batch_size, 1])
# Expand each batch and state to beam_size # Expand each batch to beam_size
alive_seq = _expand_to_beam_size(initial_ids, beam_size) alive_seq = tf.tile(tf.expand_dims(initial_ids, 1), [1, beam_size])
alive_seq = tf.expand_dims(alive_seq, axis=2) # (batch_size, beam_size, 1) alive_seq = tf.expand_dims(alive_seq, 2) # (batch_size, beam_size, 1)
if states:
states = nest.map_structure(
lambda state: _expand_to_beam_size(state, beam_size), states)
else:
states = {}
# Finished will keep track of all the sequences that have finished so far # Finished will keep track of all the sequences that have finished so far
# Finished log probs will be negative infinity in the beginning # Finished log probs will be negative infinity in the beginning
# finished_flags will keep track of booleans # finished_flags will keep track of booleans
finished_seq = tf.zeros(common_layers.shape_list(alive_seq), tf.int32) finished_seq = tf.zeros(tf.shape(alive_seq), tf.int32)
# Setting the scores of the initial to negative infinity. # Setting the scores of the initial to negative infinity.
finished_scores = tf.ones([batch_size, beam_size]) * -INF finished_scores = tf.ones([batch_size, beam_size]) * -INF
finished_flags = tf.zeros([batch_size, beam_size], tf.bool) finished_flags = tf.zeros([batch_size, beam_size], tf.bool)
...@@ -284,9 +184,9 @@ def beam_search(symbols_to_logits_fn, ...@@ -284,9 +184,9 @@ def beam_search(symbols_to_logits_fn,
curr_finished_flags = tf.concat([finished_flags, curr_finished], axis=1) curr_finished_flags = tf.concat([finished_flags, curr_finished], axis=1)
return compute_topk_scores_and_seq( return compute_topk_scores_and_seq(
curr_finished_seq, curr_finished_scores, curr_finished_scores, curr_finished_seq, curr_finished_scores, curr_finished_scores,
curr_finished_flags, beam_size, batch_size, "grow_finished") curr_finished_flags, beam_size, batch_size)
def grow_alive(curr_seq, curr_scores, curr_log_probs, curr_finished, states): def grow_alive(curr_seq, curr_scores, curr_log_probs, curr_finished):
"""Given sequences and scores, will gather the top k=beam size sequences. """Given sequences and scores, will gather the top k=beam size sequences.
Args: Args:
...@@ -297,7 +197,6 @@ def beam_search(symbols_to_logits_fn, ...@@ -297,7 +197,6 @@ def beam_search(symbols_to_logits_fn,
[batch_size, beam_size] [batch_size, beam_size]
curr_finished: Finished flags for each of these sequences. curr_finished: Finished flags for each of these sequences.
[batch_size, beam_size] [batch_size, beam_size]
states: dict (possibly nested) of decoding states.
Returns: Returns:
Tuple of Tuple of
(Topk sequences based on scores, (Topk sequences based on scores,
...@@ -308,11 +207,10 @@ def beam_search(symbols_to_logits_fn, ...@@ -308,11 +207,10 @@ def beam_search(symbols_to_logits_fn,
# values # values
curr_scores += tf.to_float(curr_finished) * -INF curr_scores += tf.to_float(curr_finished) * -INF
return compute_topk_scores_and_seq(curr_seq, curr_scores, curr_log_probs, return compute_topk_scores_and_seq(curr_seq, curr_scores, curr_log_probs,
curr_finished, beam_size, batch_size, curr_finished, beam_size, batch_size)
"grow_alive", states)
def grow_topk(i, alive_seq, alive_log_probs, states): def grow_topk(i, alive_seq, alive_log_probs):
r"""Inner beam search loop. r"""Inner beam seach loop.
This function takes the current alive sequences, and grows them to topk This function takes the current alive sequences, and grows them to topk
sequences where k = 2*beam. We use 2*beam because, we could have beam_size sequences where k = 2*beam. We use 2*beam because, we could have beam_size
...@@ -328,45 +226,36 @@ def beam_search(symbols_to_logits_fn, ...@@ -328,45 +226,36 @@ def beam_search(symbols_to_logits_fn,
i: loop index i: loop index
alive_seq: Topk sequences decoded so far [batch_size, beam_size, i+1] alive_seq: Topk sequences decoded so far [batch_size, beam_size, i+1]
alive_log_probs: probabilities of these sequences. [batch_size, beam_size] alive_log_probs: probabilities of these sequences. [batch_size, beam_size]
states: dict (possibly nested) of decoding states.
Returns: Returns:
Tuple of Tuple of
(Topk sequences extended by the next word, (Topk sequences extended by the next word,
The log probs of these sequences, The log probs of these sequences,
The scores with length penalty of these sequences, The scores with length penalty of these sequences,
Flags indicating which of these sequences have finished decoding, Flags indicating which of these sequences have finished decoding)
dict of transformed decoding states)
""" """
# Get the logits for all the possible next symbols # Get the logits for all the possible next symbols
flat_ids = tf.reshape(alive_seq, [batch_size * beam_size, -1]) flat_ids = tf.reshape(alive_seq, [batch_size * beam_size, -1])
# (batch_size * beam_size, decoded_length) # (batch_size * beam_size, decoded_length)
if states: flat_logits = symbols_to_logits_fn(flat_ids)
flat_states = nest.map_structure(_merge_beam_dim, states) logits = tf.reshape(flat_logits, (batch_size, beam_size, -1))
flat_logits, flat_states = symbols_to_logits_fn(flat_ids, i, flat_states)
states = nest.map_structure(
lambda t: _unmerge_beam_dim(t, batch_size, beam_size), flat_states)
else:
flat_logits = symbols_to_logits_fn(flat_ids)
logits = tf.reshape(flat_logits, [batch_size, beam_size, -1])
# Convert logits to normalized log probs # Convert logits to normalized log probs
candidate_log_probs = common_layers.log_prob_from_logits(logits) candidate_log_probs = log_prob_from_logits(logits)
# Multiply the probabilities by the current probabilities of the beam. # Multiply the probabilites by the current probabilites of the beam.
# (batch_size, beam_size, vocab_size) + (batch_size, beam_size, 1) # (batch_size, beam_size, vocab_size) + (batch_size, beam_size, 1)
log_probs = candidate_log_probs + tf.expand_dims(alive_log_probs, axis=2) log_probs = candidate_log_probs + tf.expand_dims(alive_log_probs, axis=2)
length_penalty = tf.pow(((5. + tf.to_float(i + 1)) / 6.), alpha) length_penalty = tf.pow(((5. + tf.to_float(i + 1)) / 6.), alpha)
curr_scores = log_probs / length_penalty curr_scores = log_probs / length_penalty
# Flatten out (beam_size, vocab_size) probs in to a list of possibilities # Flatten out (beam_size, vocab_size) probs in to a list of possibilites
flat_curr_scores = tf.reshape(curr_scores, [-1, beam_size * vocab_size]) flat_curr_scores = tf.reshape(curr_scores, [-1, beam_size * vocab_size])
topk_scores, topk_ids = tf.nn.top_k(flat_curr_scores, k=beam_size * 2) topk_scores, topk_ids = tf.nn.top_k(flat_curr_scores, k=beam_size * 2)
# Recovering the log probs because we will need to send them back # Recovering the log probs becuase we will need to send them back
topk_log_probs = topk_scores * length_penalty topk_log_probs = topk_scores * length_penalty
# Work out what beam the top probs are in. # Work out what beam the top probs are in.
...@@ -374,7 +263,7 @@ def beam_search(symbols_to_logits_fn, ...@@ -374,7 +263,7 @@ def beam_search(symbols_to_logits_fn,
topk_ids %= vocab_size # Unflatten the ids topk_ids %= vocab_size # Unflatten the ids
# The next three steps are to create coordinates for tf.gather_nd to pull # The next three steps are to create coordinates for tf.gather_nd to pull
# out the correct sequences from id's that we need to grow. # out the correct seqences from id's that we need to grow.
# We will also use the coordinates to gather the booleans of the beam items # We will also use the coordinates to gather the booleans of the beam items
# that survived. # that survived.
batch_pos = compute_batch_indices(batch_size, beam_size * 2) batch_pos = compute_batch_indices(batch_size, beam_size * 2)
...@@ -387,20 +276,17 @@ def beam_search(symbols_to_logits_fn, ...@@ -387,20 +276,17 @@ def beam_search(symbols_to_logits_fn,
# Gather up the most probable 2*beams both for the ids and finished_in_alive # Gather up the most probable 2*beams both for the ids and finished_in_alive
# bools # bools
topk_seq = tf.gather_nd(alive_seq, topk_coordinates) topk_seq = tf.gather_nd(alive_seq, topk_coordinates)
if states:
states = nest.map_structure(
lambda state: tf.gather_nd(state, topk_coordinates), states)
# Append the most probable alive # Append the most probable alive
topk_seq = tf.concat([topk_seq, tf.expand_dims(topk_ids, axis=2)], axis=2) topk_seq = tf.concat([topk_seq, tf.expand_dims(topk_ids, axis=2)], axis=2)
topk_finished = tf.equal(topk_ids, eos_id) topk_finished = tf.equal(topk_ids, eos_id)
return topk_seq, topk_log_probs, topk_scores, topk_finished, states return topk_seq, topk_log_probs, topk_scores, topk_finished
def inner_loop(i, alive_seq, alive_log_probs, finished_seq, finished_scores, def inner_loop(i, alive_seq, alive_log_probs, finished_seq, finished_scores,
finished_flags, states): finished_flags):
"""Inner beam search loop. """Inner beam seach loop.
There are three groups of tensors, alive, finished, and topk. There are three groups of tensors, alive, finished, and topk.
The alive group contains information about the current alive sequences The alive group contains information about the current alive sequences
...@@ -431,7 +317,6 @@ def beam_search(symbols_to_logits_fn, ...@@ -431,7 +317,6 @@ def beam_search(symbols_to_logits_fn,
[batch_size, beam_size] [batch_size, beam_size]
finished_flags: finished bools for each of these sequences. finished_flags: finished bools for each of these sequences.
[batch_size, beam_size] [batch_size, beam_size]
states: dict (possibly nested) of decoding states.
Returns: Returns:
Tuple of Tuple of
...@@ -440,31 +325,30 @@ def beam_search(symbols_to_logits_fn, ...@@ -440,31 +325,30 @@ def beam_search(symbols_to_logits_fn,
Log probs of the alive sequences, Log probs of the alive sequences,
New finished sequences, New finished sequences,
Scores of the new finished sequences, Scores of the new finished sequences,
Flags indicating which sequence in finished as reached EOS, Flags inidicating which sequence in finished as reached EOS)
dict of final decoding states)
""" """
# Each inner loop, we carry out three steps: # Each inner loop, we carry out three steps:
# 1. Get the current topk items. # 1. Get the current topk items.
# 2. Extract the ones that have finished and haven't finished # 2. Extract the ones that have finished and haven't finished
# 3. Recompute the contents of finished based on scores. # 3. Recompute the contents of finished based on scores.
topk_seq, topk_log_probs, topk_scores, topk_finished, states = grow_topk( topk_seq, topk_log_probs, topk_scores, topk_finished = grow_topk(
i, alive_seq, alive_log_probs, states) i, alive_seq, alive_log_probs)
alive_seq, alive_log_probs, _, states = grow_alive( alive_seq, alive_log_probs, _ = grow_alive(topk_seq, topk_scores,
topk_seq, topk_scores, topk_log_probs, topk_finished, states) topk_log_probs, topk_finished)
finished_seq, finished_scores, finished_flags, _ = grow_finished( finished_seq, finished_scores, finished_flags = grow_finished(
finished_seq, finished_scores, finished_flags, topk_seq, topk_scores, finished_seq, finished_scores, finished_flags, topk_seq, topk_scores,
topk_finished) topk_finished)
return (i + 1, alive_seq, alive_log_probs, finished_seq, finished_scores, return (i + 1, alive_seq, alive_log_probs, finished_seq, finished_scores,
finished_flags, states) finished_flags)
def _is_finished(i, unused_alive_seq, alive_log_probs, unused_finished_seq, def _is_finished(i, unused_alive_seq, alive_log_probs, unused_finished_seq,
finished_scores, finished_in_finished, unused_states): finished_scores, finished_in_finished):
"""Checking termination condition. """Checking termination condition.
We terminate when we decoded up to decode_length or the lowest scoring item We terminate when we decoded up to decode_length or the lowest scoring item
in finished has a greater score that the highest prob item in alive divided in finished has a greater score that the higest prob item in alive divided
by the max length penalty by the max length penalty
Args: Args:
...@@ -478,38 +362,41 @@ def beam_search(symbols_to_logits_fn, ...@@ -478,38 +362,41 @@ def beam_search(symbols_to_logits_fn,
Returns: Returns:
Bool. Bool.
""" """
if not stop_early:
return tf.less(i, decode_length)
max_length_penalty = tf.pow(((5. + tf.to_float(decode_length)) / 6.), alpha) max_length_penalty = tf.pow(((5. + tf.to_float(decode_length)) / 6.), alpha)
# The best possible score of the most likely alive sequence. # The best possible score of the most likley alive sequence
lower_bound_alive_scores = alive_log_probs[:, 0] / max_length_penalty lower_bound_alive_scores = alive_log_probs[:, 0] / max_length_penalty
# Now to compute the lowest score of a finished sequence in finished # Now to compute the lowest score of a finished sequence in finished
# If the sequence isn't finished, we multiply it's score by 0. since # If the sequence isn't finished, we multiply it's score by 0. since
# scores are all -ve, taking the min will give us the score of the lowest # scores are all -ve, taking the min will give us the score of the lowest
# finished item. # finished item.
lowest_score_of_finished_in_finished = tf.reduce_min( lowest_score_of_fininshed_in_finished = tf.reduce_min(
finished_scores * tf.to_float(finished_in_finished), axis=1) finished_scores * tf.to_float(finished_in_finished), axis=1)
# If none of the sequences have finished, then the min will be 0 and # If none of the sequences have finished, then the min will be 0 and
# we have to replace it by -ve INF if it is. The score of any seq in alive # we have to replace it by -ve INF if it is. The score of any seq in alive
# will be much higher than -ve INF and the termination condition will not # will be much higher than -ve INF and the termination condition will not
# be met. # be met.
lowest_score_of_finished_in_finished += ( lowest_score_of_fininshed_in_finished += (
(1. - tf.to_float(tf.reduce_any(finished_in_finished, 1))) * -INF) (1. - tf.to_float(tf.reduce_any(finished_in_finished, 1))) * -INF)
bound_is_met = tf.reduce_all( bound_is_met = tf.reduce_all(
tf.greater(lowest_score_of_finished_in_finished, tf.greater(lowest_score_of_fininshed_in_finished,
lower_bound_alive_scores)) lower_bound_alive_scores))
return tf.logical_and( return tf.logical_and(
tf.less(i, decode_length), tf.logical_not(bound_is_met)) tf.less(i, decode_length), tf.logical_not(bound_is_met))
"""
(_, alive_seq, alive_log_probs, finished_seq, finished_scores,
finished_flags) = inner_loop(tf.constant(0), alive_seq, alive_log_probs, finished_seq,
finished_scores, finished_flags)
"""
(_, alive_seq, alive_log_probs, finished_seq, finished_scores, (_, alive_seq, alive_log_probs, finished_seq, finished_scores,
finished_flags, _) = tf.while_loop( finished_flags) = tf.while_loop(
_is_finished, _is_finished,
inner_loop, [ inner_loop, [
tf.constant(0), alive_seq, alive_log_probs, finished_seq, tf.constant(0), alive_seq, alive_log_probs, finished_seq,
finished_scores, finished_flags, states finished_scores, finished_flags
], ],
shape_invariants=[ shape_invariants=[
tf.TensorShape([]), tf.TensorShape([]),
...@@ -517,8 +404,7 @@ def beam_search(symbols_to_logits_fn, ...@@ -517,8 +404,7 @@ def beam_search(symbols_to_logits_fn,
alive_log_probs.get_shape(), alive_log_probs.get_shape(),
tf.TensorShape([None, None, None]), tf.TensorShape([None, None, None]),
finished_scores.get_shape(), finished_scores.get_shape(),
finished_flags.get_shape(), finished_flags.get_shape()
nest.map_structure(get_state_shape_invariants, states),
], ],
parallel_iterations=1, parallel_iterations=1,
back_prop=False) back_prop=False)
......
# Copyright 2017 The Tensor2Tensor Authors. # coding=utf-8
# Copyright 2018 The Tensor2Tensor Authors.
# #
# Licensed under the Apache License, Version 2.0 (the "License"); # Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License. # you may not use this file except in compliance with the License.
...@@ -11,28 +12,81 @@ ...@@ -11,28 +12,81 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
"""Implementation of beam search with penalties."""
"""Implemetation of beam seach with penalties."""
from __future__ import absolute_import from __future__ import absolute_import
from __future__ import division from __future__ import division
from __future__ import print_function from __future__ import print_function
from tensor2tensor.models import common_layers
# Dependency imports
import tensorflow as tf import tensorflow as tf
from tensorflow.python.util import nest
# Assuming EOS_ID is 1 # Assuming EOS_ID is 1
EOS_ID = 1 EOS_ID = 1
# Default value for INF # Default value for INF
INF = 1. * 1e7 INF = 1. * 1e7
def log_prob_from_logits(logits): def _merge_beam_dim(tensor):
return logits - tf.reduce_logsumexp(logits, axis=2, keep_dims=True) """Reshapes first two dimensions in to single dimension.
Args:
tensor: Tensor to reshape of shape [A, B, ...]
Returns:
Reshaped tensor of shape [A*B, ...]
"""
shape = common_layers.shape_list(tensor)
shape[0] *= shape[1] # batch -> batch * beam_size
shape.pop(1) # Remove beam dim
return tf.reshape(tensor, shape)
def _unmerge_beam_dim(tensor, batch_size, beam_size):
"""Reshapes first dimension back to [batch_size, beam_size].
Args:
tensor: Tensor to reshape of shape [batch_size*beam_size, ...]
batch_size: Tensor, original batch size.
beam_size: int, original beam size.
Returns:
Reshaped tensor of shape [batch_size, beam_size, ...]
"""
shape = common_layers.shape_list(tensor)
new_shape = [batch_size] + [beam_size] + shape[1:]
return tf.reshape(tensor, new_shape)
def _expand_to_beam_size(tensor, beam_size):
"""Tiles a given tensor by beam_size.
Args:
tensor: tensor to tile [batch_size, ...]
beam_size: How much to tile the tensor by.
Returns:
Tiled tensor [batch_size, beam_size, ...]
"""
tensor = tf.expand_dims(tensor, axis=1)
tile_dims = [1] * tensor.shape.ndims
tile_dims[1] = beam_size
return tf.tile(tensor, tile_dims)
def get_state_shape_invariants(tensor):
"""Returns the shape of the tensor but sets middle dims to None."""
shape = tensor.shape.as_list()
for i in range(1, len(shape) - 1):
shape[i] = None
return tf.TensorShape(shape)
def compute_batch_indices(batch_size, beam_size): def compute_batch_indices(batch_size, beam_size):
"""Computes the i'th coodinate that contains the batch index for gathers. """Computes the i'th coordinate that contains the batch index for gathers.
Batch pos is a tensor like [[0,0,0,0,],[1,1,1,1],..]. It says which Batch pos is a tensor like [[0,0,0,0,],[1,1,1,1],..]. It says which
batch the beam item is in. This will create the i of the i,j coordinate batch the beam item is in. This will create the i of the i,j coordinate
...@@ -50,13 +104,20 @@ def compute_batch_indices(batch_size, beam_size): ...@@ -50,13 +104,20 @@ def compute_batch_indices(batch_size, beam_size):
def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags, def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags,
beam_size, batch_size): beam_size, batch_size, prefix="default",
states_to_gather=None):
"""Given sequences and scores, will gather the top k=beam size sequences. """Given sequences and scores, will gather the top k=beam size sequences.
This function is used to grow alive, and finished. It takes sequences, This function is used to grow alive, and finished. It takes sequences,
scores, and flags, and returns the top k from sequences, scores_to_gather, scores, and flags, and returns the top k from sequences, scores_to_gather,
and flags based on the values in scores. and flags based on the values in scores.
This method permits easy introspection using tfdbg. It adds three named ops
that are prefixed by `prefix`:
- _topk_seq: the tensor for topk_seq returned by this method.
- _topk_flags: the tensor for topk_finished_flags returned by this method.
- _topk_scores: the tensor for tokp_gathered_scores returned by this method.
Args: Args:
sequences: Tensor of sequences that we need to gather from. sequences: Tensor of sequences that we need to gather from.
[batch_size, beam_size, seq_length] [batch_size, beam_size, seq_length]
...@@ -66,11 +127,13 @@ def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags, ...@@ -66,11 +127,13 @@ def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags,
[batch_size, beam_size]. We will return the gathered scores from here. [batch_size, beam_size]. We will return the gathered scores from here.
Scores to gather is different from scores because for grow_alive, we will Scores to gather is different from scores because for grow_alive, we will
need to return log_probs, while for grow_finished, we will need to return need to return log_probs, while for grow_finished, we will need to return
the length penalized scors. the length penalized scores.
flags: Tensor of bools for sequences that say whether a sequence has reached flags: Tensor of bools for sequences that say whether a sequence has reached
EOS or not EOS or not
beam_size: int beam_size: int
batch_size: int batch_size: int
prefix: string that will prefix unique names for the ops run.
states_to_gather: dict (possibly nested) of decoding states.
Returns: Returns:
Tuple of Tuple of
(topk_seq [batch_size, beam_size, decode_length], (topk_seq [batch_size, beam_size, decode_length],
...@@ -90,11 +153,20 @@ def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags, ...@@ -90,11 +153,20 @@ def compute_topk_scores_and_seq(sequences, scores, scores_to_gather, flags,
# last dimension contains the i,j gathering coordinates. # last dimension contains the i,j gathering coordinates.
top_coordinates = tf.stack([batch_pos, topk_indexes], axis=2) top_coordinates = tf.stack([batch_pos, topk_indexes], axis=2)
# Gather up the highest scoring sequences # Gather up the highest scoring sequences. For each operation added, give it
topk_seq = tf.gather_nd(sequences, top_coordinates) # a concrete name to simplify observing these operations with tfdbg. Clients
topk_flags = tf.gather_nd(flags, top_coordinates) # can capture these tensors by watching these node names.
topk_gathered_scores = tf.gather_nd(scores_to_gather, top_coordinates) def gather(tensor, name):
return topk_seq, topk_gathered_scores, topk_flags return tf.gather_nd(tensor, top_coordinates, name=(prefix + name))
topk_seq = gather(sequences, "_topk_seq")
topk_flags = gather(flags, "_topk_flags")
topk_gathered_scores = gather(scores_to_gather, "_topk_scores")
if states_to_gather:
topk_gathered_states = nest.map_structure(
lambda state: gather(state, "_topk_states"), states_to_gather)
else:
topk_gathered_states = states_to_gather
return topk_seq, topk_gathered_scores, topk_flags, topk_gathered_states
def beam_search(symbols_to_logits_fn, def beam_search(symbols_to_logits_fn,
...@@ -103,14 +175,35 @@ def beam_search(symbols_to_logits_fn, ...@@ -103,14 +175,35 @@ def beam_search(symbols_to_logits_fn,
decode_length, decode_length,
vocab_size, vocab_size,
alpha, alpha,
eos_id=EOS_ID): states=None,
eos_id=EOS_ID,
stop_early=True):
"""Beam search with length penalties. """Beam search with length penalties.
Uses an interface specific to the sequence cnn models; Requires a function that can take the currently decoded symbols and return
Requires a function that can take the currently decoded sybmols and return
the logits for the next symbol. The implementation is inspired by the logits for the next symbol. The implementation is inspired by
https://arxiv.org/abs/1609.08144. https://arxiv.org/abs/1609.08144.
When running, the beam search steps can be visualized by using tfdbg to watch
the operations generating the output ids for each beam step. These operations
have the pattern:
(alive|finished)_topk_(seq,scores)
Operations marked `alive` represent the new beam sequences that will be
processed in the next step. Operations marked `finished` represent the
completed beam sequences, which may be padded with 0s if no beams finished.
Operations marked `seq` store the full beam sequence for the time step.
Operations marked `scores` store the sequence's final log scores.
The beam search steps will be processed sequentially in order, so when
capturing observed from these operations, tensors, clients can make
assumptions about which step is being recorded.
WARNING: Assumes 2nd dimension of tensors in `states` and not invariant, this
means that the shape of the 2nd dimension of these tensors will not be
available (i.e. set to None) inside symbols_to_logits_fn.
Args: Args:
symbols_to_logits_fn: Interface to the model, to provide logits. symbols_to_logits_fn: Interface to the model, to provide logits.
Shoud take [batch_size, decoded_ids] and return [batch_size, vocab_size] Shoud take [batch_size, decoded_ids] and return [batch_size, vocab_size]
...@@ -122,27 +215,34 @@ def beam_search(symbols_to_logits_fn, ...@@ -122,27 +215,34 @@ def beam_search(symbols_to_logits_fn,
vocab_size: Size of the vocab, must equal the size of the logits returned by vocab_size: Size of the vocab, must equal the size of the logits returned by
symbols_to_logits_fn symbols_to_logits_fn
alpha: alpha for length penalty. alpha: alpha for length penalty.
states: dict (possibly nested) of decoding states.
eos_id: ID for end of sentence. eos_id: ID for end of sentence.
stop_early: a boolean - stop once best sequence is provably determined.
Returns: Returns:
Tuple of Tuple of
(decoded beams [batch_size, beam_size, decode_length] (decoded beams [batch_size, beam_size, decode_length]
decoding probablities [batch_size, beam_size]) decoding probabilities [batch_size, beam_size])
""" """
batch_size = tf.shape(initial_ids)[0] batch_size = common_layers.shape_list(initial_ids)[0]
# Assume initial_ids are prob 1.0 # Assume initial_ids are prob 1.0
initial_log_probs = tf.constant([[0.] + [-float("inf")] * (beam_size - 1)]) initial_log_probs = tf.constant([[0.] + [-float("inf")] * (beam_size - 1)])
# Expand to beam_size (batch_size, beam_size) # Expand to beam_size (batch_size, beam_size)
alive_log_probs = tf.tile(initial_log_probs, [batch_size, 1]) alive_log_probs = tf.tile(initial_log_probs, [batch_size, 1])
# Expand each batch to beam_size # Expand each batch and state to beam_size
alive_seq = tf.tile(tf.expand_dims(initial_ids, 1), [1, beam_size]) alive_seq = _expand_to_beam_size(initial_ids, beam_size)
alive_seq = tf.expand_dims(alive_seq, 2) # (batch_size, beam_size, 1) alive_seq = tf.expand_dims(alive_seq, axis=2) # (batch_size, beam_size, 1)
if states:
states = nest.map_structure(
lambda state: _expand_to_beam_size(state, beam_size), states)
else:
states = {}
# Finished will keep track of all the sequences that have finished so far # Finished will keep track of all the sequences that have finished so far
# Finished log probs will be negative infinity in the beginning # Finished log probs will be negative infinity in the beginning
# finished_flags will keep track of booleans # finished_flags will keep track of booleans
finished_seq = tf.zeros(tf.shape(alive_seq), tf.int32) finished_seq = tf.zeros(common_layers.shape_list(alive_seq), tf.int32)
# Setting the scores of the initial to negative infinity. # Setting the scores of the initial to negative infinity.
finished_scores = tf.ones([batch_size, beam_size]) * -INF finished_scores = tf.ones([batch_size, beam_size]) * -INF
finished_flags = tf.zeros([batch_size, beam_size], tf.bool) finished_flags = tf.zeros([batch_size, beam_size], tf.bool)
...@@ -184,9 +284,9 @@ def beam_search(symbols_to_logits_fn, ...@@ -184,9 +284,9 @@ def beam_search(symbols_to_logits_fn,
curr_finished_flags = tf.concat([finished_flags, curr_finished], axis=1) curr_finished_flags = tf.concat([finished_flags, curr_finished], axis=1)
return compute_topk_scores_and_seq( return compute_topk_scores_and_seq(
curr_finished_seq, curr_finished_scores, curr_finished_scores, curr_finished_seq, curr_finished_scores, curr_finished_scores,
curr_finished_flags, beam_size, batch_size) curr_finished_flags, beam_size, batch_size, "grow_finished")
def grow_alive(curr_seq, curr_scores, curr_log_probs, curr_finished): def grow_alive(curr_seq, curr_scores, curr_log_probs, curr_finished, states):
"""Given sequences and scores, will gather the top k=beam size sequences. """Given sequences and scores, will gather the top k=beam size sequences.
Args: Args:
...@@ -197,6 +297,7 @@ def beam_search(symbols_to_logits_fn, ...@@ -197,6 +297,7 @@ def beam_search(symbols_to_logits_fn,
[batch_size, beam_size] [batch_size, beam_size]
curr_finished: Finished flags for each of these sequences. curr_finished: Finished flags for each of these sequences.
[batch_size, beam_size] [batch_size, beam_size]
states: dict (possibly nested) of decoding states.
Returns: Returns:
Tuple of Tuple of
(Topk sequences based on scores, (Topk sequences based on scores,
...@@ -207,10 +308,11 @@ def beam_search(symbols_to_logits_fn, ...@@ -207,10 +308,11 @@ def beam_search(symbols_to_logits_fn,
# values # values
curr_scores += tf.to_float(curr_finished) * -INF curr_scores += tf.to_float(curr_finished) * -INF
return compute_topk_scores_and_seq(curr_seq, curr_scores, curr_log_probs, return compute_topk_scores_and_seq(curr_seq, curr_scores, curr_log_probs,
curr_finished, beam_size, batch_size) curr_finished, beam_size, batch_size,
"grow_alive", states)
def grow_topk(i, alive_seq, alive_log_probs): def grow_topk(i, alive_seq, alive_log_probs, states):
r"""Inner beam seach loop. r"""Inner beam search loop.
This function takes the current alive sequences, and grows them to topk This function takes the current alive sequences, and grows them to topk
sequences where k = 2*beam. We use 2*beam because, we could have beam_size sequences where k = 2*beam. We use 2*beam because, we could have beam_size
...@@ -226,36 +328,45 @@ def beam_search(symbols_to_logits_fn, ...@@ -226,36 +328,45 @@ def beam_search(symbols_to_logits_fn,
i: loop index i: loop index
alive_seq: Topk sequences decoded so far [batch_size, beam_size, i+1] alive_seq: Topk sequences decoded so far [batch_size, beam_size, i+1]
alive_log_probs: probabilities of these sequences. [batch_size, beam_size] alive_log_probs: probabilities of these sequences. [batch_size, beam_size]
states: dict (possibly nested) of decoding states.
Returns: Returns:
Tuple of Tuple of
(Topk sequences extended by the next word, (Topk sequences extended by the next word,
The log probs of these sequences, The log probs of these sequences,
The scores with length penalty of these sequences, The scores with length penalty of these sequences,
Flags indicating which of these sequences have finished decoding) Flags indicating which of these sequences have finished decoding,
dict of transformed decoding states)
""" """
# Get the logits for all the possible next symbols # Get the logits for all the possible next symbols
flat_ids = tf.reshape(alive_seq, [batch_size * beam_size, -1]) flat_ids = tf.reshape(alive_seq, [batch_size * beam_size, -1])
# (batch_size * beam_size, decoded_length) # (batch_size * beam_size, decoded_length)
flat_logits = symbols_to_logits_fn(flat_ids) if states:
logits = tf.reshape(flat_logits, (batch_size, beam_size, -1)) flat_states = nest.map_structure(_merge_beam_dim, states)
flat_logits, flat_states = symbols_to_logits_fn(flat_ids, i, flat_states)
states = nest.map_structure(
lambda t: _unmerge_beam_dim(t, batch_size, beam_size), flat_states)
else:
flat_logits = symbols_to_logits_fn(flat_ids)
logits = tf.reshape(flat_logits, [batch_size, beam_size, -1])
# Convert logits to normalized log probs # Convert logits to normalized log probs
candidate_log_probs = log_prob_from_logits(logits) candidate_log_probs = common_layers.log_prob_from_logits(logits)
# Multiply the probabilites by the current probabilites of the beam. # Multiply the probabilities by the current probabilities of the beam.
# (batch_size, beam_size, vocab_size) + (batch_size, beam_size, 1) # (batch_size, beam_size, vocab_size) + (batch_size, beam_size, 1)
log_probs = candidate_log_probs + tf.expand_dims(alive_log_probs, axis=2) log_probs = candidate_log_probs + tf.expand_dims(alive_log_probs, axis=2)
length_penalty = tf.pow(((5. + tf.to_float(i + 1)) / 6.), alpha) length_penalty = tf.pow(((5. + tf.to_float(i + 1)) / 6.), alpha)
curr_scores = log_probs / length_penalty curr_scores = log_probs / length_penalty
# Flatten out (beam_size, vocab_size) probs in to a list of possibilites # Flatten out (beam_size, vocab_size) probs in to a list of possibilities
flat_curr_scores = tf.reshape(curr_scores, [-1, beam_size * vocab_size]) flat_curr_scores = tf.reshape(curr_scores, [-1, beam_size * vocab_size])
topk_scores, topk_ids = tf.nn.top_k(flat_curr_scores, k=beam_size * 2) topk_scores, topk_ids = tf.nn.top_k(flat_curr_scores, k=beam_size * 2)
# Recovering the log probs becuase we will need to send them back # Recovering the log probs because we will need to send them back
topk_log_probs = topk_scores * length_penalty topk_log_probs = topk_scores * length_penalty
# Work out what beam the top probs are in. # Work out what beam the top probs are in.
...@@ -263,7 +374,7 @@ def beam_search(symbols_to_logits_fn, ...@@ -263,7 +374,7 @@ def beam_search(symbols_to_logits_fn,
topk_ids %= vocab_size # Unflatten the ids topk_ids %= vocab_size # Unflatten the ids
# The next three steps are to create coordinates for tf.gather_nd to pull # The next three steps are to create coordinates for tf.gather_nd to pull
# out the correct seqences from id's that we need to grow. # out the correct sequences from id's that we need to grow.
# We will also use the coordinates to gather the booleans of the beam items # We will also use the coordinates to gather the booleans of the beam items
# that survived. # that survived.
batch_pos = compute_batch_indices(batch_size, beam_size * 2) batch_pos = compute_batch_indices(batch_size, beam_size * 2)
...@@ -276,17 +387,20 @@ def beam_search(symbols_to_logits_fn, ...@@ -276,17 +387,20 @@ def beam_search(symbols_to_logits_fn,
# Gather up the most probable 2*beams both for the ids and finished_in_alive # Gather up the most probable 2*beams both for the ids and finished_in_alive
# bools # bools
topk_seq = tf.gather_nd(alive_seq, topk_coordinates) topk_seq = tf.gather_nd(alive_seq, topk_coordinates)
if states:
states = nest.map_structure(
lambda state: tf.gather_nd(state, topk_coordinates), states)
# Append the most probable alive # Append the most probable alive
topk_seq = tf.concat([topk_seq, tf.expand_dims(topk_ids, axis=2)], axis=2) topk_seq = tf.concat([topk_seq, tf.expand_dims(topk_ids, axis=2)], axis=2)
topk_finished = tf.equal(topk_ids, eos_id) topk_finished = tf.equal(topk_ids, eos_id)
return topk_seq, topk_log_probs, topk_scores, topk_finished return topk_seq, topk_log_probs, topk_scores, topk_finished, states
def inner_loop(i, alive_seq, alive_log_probs, finished_seq, finished_scores, def inner_loop(i, alive_seq, alive_log_probs, finished_seq, finished_scores,
finished_flags): finished_flags, states):
"""Inner beam seach loop. """Inner beam search loop.
There are three groups of tensors, alive, finished, and topk. There are three groups of tensors, alive, finished, and topk.
The alive group contains information about the current alive sequences The alive group contains information about the current alive sequences
...@@ -317,6 +431,7 @@ def beam_search(symbols_to_logits_fn, ...@@ -317,6 +431,7 @@ def beam_search(symbols_to_logits_fn,
[batch_size, beam_size] [batch_size, beam_size]
finished_flags: finished bools for each of these sequences. finished_flags: finished bools for each of these sequences.
[batch_size, beam_size] [batch_size, beam_size]
states: dict (possibly nested) of decoding states.
Returns: Returns:
Tuple of Tuple of
...@@ -325,30 +440,31 @@ def beam_search(symbols_to_logits_fn, ...@@ -325,30 +440,31 @@ def beam_search(symbols_to_logits_fn,
Log probs of the alive sequences, Log probs of the alive sequences,
New finished sequences, New finished sequences,
Scores of the new finished sequences, Scores of the new finished sequences,
Flags inidicating which sequence in finished as reached EOS) Flags indicating which sequence in finished as reached EOS,
dict of final decoding states)
""" """
# Each inner loop, we carry out three steps: # Each inner loop, we carry out three steps:
# 1. Get the current topk items. # 1. Get the current topk items.
# 2. Extract the ones that have finished and haven't finished # 2. Extract the ones that have finished and haven't finished
# 3. Recompute the contents of finished based on scores. # 3. Recompute the contents of finished based on scores.
topk_seq, topk_log_probs, topk_scores, topk_finished = grow_topk( topk_seq, topk_log_probs, topk_scores, topk_finished, states = grow_topk(
i, alive_seq, alive_log_probs) i, alive_seq, alive_log_probs, states)
alive_seq, alive_log_probs, _ = grow_alive(topk_seq, topk_scores, alive_seq, alive_log_probs, _, states = grow_alive(
topk_log_probs, topk_finished) topk_seq, topk_scores, topk_log_probs, topk_finished, states)
finished_seq, finished_scores, finished_flags = grow_finished( finished_seq, finished_scores, finished_flags, _ = grow_finished(
finished_seq, finished_scores, finished_flags, topk_seq, topk_scores, finished_seq, finished_scores, finished_flags, topk_seq, topk_scores,
topk_finished) topk_finished)
return (i + 1, alive_seq, alive_log_probs, finished_seq, finished_scores, return (i + 1, alive_seq, alive_log_probs, finished_seq, finished_scores,
finished_flags) finished_flags, states)
def _is_finished(i, unused_alive_seq, alive_log_probs, unused_finished_seq, def _is_finished(i, unused_alive_seq, alive_log_probs, unused_finished_seq,
finished_scores, finished_in_finished): finished_scores, finished_in_finished, unused_states):
"""Checking termination condition. """Checking termination condition.
We terminate when we decoded up to decode_length or the lowest scoring item We terminate when we decoded up to decode_length or the lowest scoring item
in finished has a greater score that the higest prob item in alive divided in finished has a greater score that the highest prob item in alive divided
by the max length penalty by the max length penalty
Args: Args:
...@@ -362,41 +478,38 @@ def beam_search(symbols_to_logits_fn, ...@@ -362,41 +478,38 @@ def beam_search(symbols_to_logits_fn,
Returns: Returns:
Bool. Bool.
""" """
if not stop_early:
return tf.less(i, decode_length)
max_length_penalty = tf.pow(((5. + tf.to_float(decode_length)) / 6.), alpha) max_length_penalty = tf.pow(((5. + tf.to_float(decode_length)) / 6.), alpha)
# The best possible score of the most likley alive sequence # The best possible score of the most likely alive sequence.
lower_bound_alive_scores = alive_log_probs[:, 0] / max_length_penalty lower_bound_alive_scores = alive_log_probs[:, 0] / max_length_penalty
# Now to compute the lowest score of a finished sequence in finished # Now to compute the lowest score of a finished sequence in finished
# If the sequence isn't finished, we multiply it's score by 0. since # If the sequence isn't finished, we multiply it's score by 0. since
# scores are all -ve, taking the min will give us the score of the lowest # scores are all -ve, taking the min will give us the score of the lowest
# finished item. # finished item.
lowest_score_of_fininshed_in_finished = tf.reduce_min( lowest_score_of_finished_in_finished = tf.reduce_min(
finished_scores * tf.to_float(finished_in_finished), axis=1) finished_scores * tf.to_float(finished_in_finished), axis=1)
# If none of the sequences have finished, then the min will be 0 and # If none of the sequences have finished, then the min will be 0 and
# we have to replace it by -ve INF if it is. The score of any seq in alive # we have to replace it by -ve INF if it is. The score of any seq in alive
# will be much higher than -ve INF and the termination condition will not # will be much higher than -ve INF and the termination condition will not
# be met. # be met.
lowest_score_of_fininshed_in_finished += ( lowest_score_of_finished_in_finished += (
(1. - tf.to_float(tf.reduce_any(finished_in_finished, 1))) * -INF) (1. - tf.to_float(tf.reduce_any(finished_in_finished, 1))) * -INF)
bound_is_met = tf.reduce_all( bound_is_met = tf.reduce_all(
tf.greater(lowest_score_of_fininshed_in_finished, tf.greater(lowest_score_of_finished_in_finished,
lower_bound_alive_scores)) lower_bound_alive_scores))
return tf.logical_and( return tf.logical_and(
tf.less(i, decode_length), tf.logical_not(bound_is_met)) tf.less(i, decode_length), tf.logical_not(bound_is_met))
"""
(_, alive_seq, alive_log_probs, finished_seq, finished_scores,
finished_flags) = inner_loop(tf.constant(0), alive_seq, alive_log_probs, finished_seq,
finished_scores, finished_flags)
"""
(_, alive_seq, alive_log_probs, finished_seq, finished_scores, (_, alive_seq, alive_log_probs, finished_seq, finished_scores,
finished_flags) = tf.while_loop( finished_flags, _) = tf.while_loop(
_is_finished, _is_finished,
inner_loop, [ inner_loop, [
tf.constant(0), alive_seq, alive_log_probs, finished_seq, tf.constant(0), alive_seq, alive_log_probs, finished_seq,
finished_scores, finished_flags finished_scores, finished_flags, states
], ],
shape_invariants=[ shape_invariants=[
tf.TensorShape([]), tf.TensorShape([]),
...@@ -404,7 +517,8 @@ def beam_search(symbols_to_logits_fn, ...@@ -404,7 +517,8 @@ def beam_search(symbols_to_logits_fn,
alive_log_probs.get_shape(), alive_log_probs.get_shape(),
tf.TensorShape([None, None, None]), tf.TensorShape([None, None, None]),
finished_scores.get_shape(), finished_scores.get_shape(),
finished_flags.get_shape() finished_flags.get_shape(),
nest.map_structure(get_state_shape_invariants, states),
], ],
parallel_iterations=1, parallel_iterations=1,
back_prop=False) back_prop=False)
......
Markdown 格式
0%
您添加了 0 到此讨论。请谨慎行事。
请先完成此评论的编辑!
注册 或者 后发表评论