def my_gather(tensor):
return mtf.gather(tensor,
top_beam_index,
beam_dim,
output_shape=mtf.Shape([
double_beam if d == beam_dim else d
for d in tensor.shape.dims
]))
def gather(tensor, name):
with tf.name_scope(prefix + name):
output_shape = mtf.Shape([
beam_dim if d == old_beam_dim else d for d in tensor.shape.dims
])
return mtf.gather(tensor,
topk_indices,
old_beam_dim,
output_shape=output_shape)
def _is_finished(i, unused_alive_seq, alive_log_probs, unused_finished_seq,
finished_scores, finished_in_finished, *unused_states):
"""Checking termination condition.
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
by the max length penalty
Args:
i: loop index
alive_log_probs: probabilities of the beams. [batch_size, beam_size]
finished_scores: scores for each of these sequences.
[batch_size, beam_size]
finished_in_finished: finished bools for each of these sequences.
[batch_size, beam_size]
Returns:
Bool.
"""
# TODO(noam): support a different decode length...
# decode_length = mtf.constant(mesh, length_dim.size, dtype=tf.int32)
# del alive_log_probs, finished_scores, finished_in_finished
# return mtf.less(i, length_dim.size)
if not stop_early:
return mtf.less(i, decode_length)
max_length_penalty = mtf.pow(
((5. + mtf.cast(decode_length, finished_scores.dtype)) / 6.),
alpha)
# The best possible score of the most likely alive sequence.
lower_bound_alive_scores = mtf.gather(
alive_log_probs, mtf.constant(mesh, 0, dtype=tf.int32),
beam_dim) / max_length_penalty
# 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
# scores are all -ve, taking the min will give us the score of the lowest
# finished item.
lowest_score_of_finished_in_finished = mtf.reduce_min(
finished_scores *
mtf.cast(finished_in_finished, finished_scores.dtype),
reduced_dim=beam_dim)
# 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
# will be much higher than -ve INF and the termination condition will not
# be met.
lowest_score_of_finished_in_finished += ((1. - mtf.cast(
mtf.reduce_any(finished_in_finished, reduced_dim=beam_dim),
finished_scores.dtype)) * -INF)
bound_is_met = mtf.reduce_all(
mtf.greater(lowest_score_of_finished_in_finished,
lower_bound_alive_scores))
return mtf.logical_and(mtf.less(i, decode_length),
mtf.logical_not(bound_is_met))
def body_fn(step_num, ids, *states):
"""Body function for greedy decoding.
Args:
step_num: a mtf.Tensor
ids: a mtf.Tensor
*states: additional mtf.Tensors
Returns:
new_step_num, new_ids, *new_states
"""
logits, new_states = logits_fn(step_num, ids, states)
vocab_dim = logits.shape.dims[-1]
new_ids = mtf.sample_with_temperature(logits, vocab_dim, temperature)
if forced_ids is not None:
# force the new ids to equal the partial targets where specified
# (positions where partial_targets contain nonzero values)
forced = mtf.gather(forced_ids, step_num, length_dim)
new_ids = forced + new_ids * mtf.to_int32(mtf.equal(forced, 0))
ids += new_ids * mtf.one_hot(step_num, length_dim, dtype=tf.int32)
new_step_num = step_num + 1
return [new_step_num, ids] + new_states
def grow_topk(i, alive_seq, alive_log_probs, states=None):
r"""Inner beam search loop.
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
number of sequences that might hit <EOS> and there will be no alive
sequences to continue. With 2*beam_size, this will not happen. This relies
on the assumption the vocab size is > beam size. If this is true, we'll
have at least beam_size non <EOS> extensions if we extract the next top
2*beam words.
Length penalty is given by = (5+len(decode)/6) ^ -\alpha. Pls refer to
https://arxiv.org/abs/1609.08144.
Args:
i: loop index
alive_seq: Topk sequences decoded so far [batch, beam, length]
alive_log_probs: probabilities of these sequences. [batch, beam]
states: optional list of mtf.Tensor
Returns:
Tuple of
(Topk sequences extended by the next word,
The log probs of these sequences,
The scores with length penalty of these sequences,
Flags indicating which of these sequences have finished decoding,
list of transformed decoding states)
"""
logits, new_states = logits_fn(i, alive_seq, states)
batch_dim, beam_dim, vocab_dim = logits.shape.dims
# Convert logits to normalized log probs
candidate_log_probs = mtf.log_softmax(logits, vocab_dim)
# Multiply the probabilities by the current probabilities of the beam.
# (batch_size, beam_size, vocab_size) + (batch_size, beam_size, 1)
log_probs = candidate_log_probs + alive_log_probs
length_penalty = mtf.pow(((5. + mtf.cast(i + 1, logits.dtype)) / 6.),
alpha)
# scores have shape [batch, beam, vocab]
curr_scores = log_probs / length_penalty
# We find the top 2k sequences to make sure we get k alive sequences.
#
# TODO(noam): This is inefficient. We should separately compute the k
# finished sequences (previously alive sequences + EOS), and the top k new
# alive sequences.
double_beam = mtf.Dimension("double_beam", beam_dim.size * 2)
if use_tpu and layout is not None and mesh_shape is not None:
# Do some partial top-k-ing first locally to avoid communication.
# We reshape the logits from:
# [batch, beam, vocab] to
# [batch, beam, major_vocab, minor_vocab]
# We first reduce (locally) across the minor_vocab dimension. This makes
# the thing we need to broadcast smaller.
# This also enables our shortcut of only picking the top num_prefilter
# sequences per beam per major_vocab in the first pass.
major_vocab_size = mtf.tensor_dim_to_mesh_dim_size(
layout, mesh_shape, vocab_dim)
major_vocab = mtf.Dimension(vocab_dim.name, major_vocab_size)
minor_vocab = mtf.Dimension("minor_vocab",
vocab_dim.size // major_vocab_size)
curr_scores = mtf.reshape(
curr_scores, [batch_dim, beam_dim, major_vocab, minor_vocab])
prefilter = mtf.Dimension("prefilter", num_prefilter
or double_beam.size)
# shape = [batch_dim, beam_dim, major_vocab, prefilter]
top_scores, top_minor_vocab_ids = mtf.top_k(
curr_scores, reduced_dim=minor_vocab, k_dim=prefilter)
combined = mtf.Dimension(
"combined", beam_dim.size * major_vocab.size * prefilter.size)
top_scores = mtf.reshape(top_scores, [batch_dim, combined])
top_minor_vocab_ids = mtf.reshape(top_minor_vocab_ids,
[batch_dim, combined])
# shpae = [batch_dim, double_beam]
# ids are indices representing (beam, major_vocab, prefilter)
top_scores, top_combined_ids = mtf.top_k(top_scores,
reduced_dim=combined,
k_dim=double_beam)
top_minor_vocab_ids = mtf.gather(
top_minor_vocab_ids,
top_combined_ids,
combined,
output_shape=[batch_dim, double_beam])
top_beam_index = top_combined_ids // (major_vocab.size *
prefilter.size)
top_combined_ids -= top_beam_index * (major_vocab.size *
prefilter.size)
top_major_vocab_ids = top_combined_ids // prefilter.size
top_combined_ids -= top_major_vocab_ids * prefilter.size
top_ids = top_major_vocab_ids * minor_vocab.size + top_minor_vocab_ids
else:
beam_and_vocab_dim = mtf.Dimension("beam_and_vocab",
beam_dim.size * vocab_dim.size)
flat_shape = mtf.Shape([batch_dim, beam_and_vocab_dim])
# Flatten out (beam_size, vocab_size) probs into a list of possibilities
flat_curr_scores = mtf.reshape(curr_scores,
flat_shape,
name="flatten_scores")
top_scores, top_ids = mtf.top_k(flat_curr_scores,
reduced_dim=beam_and_vocab_dim,
k_dim=double_beam)
# Work out what beam the top probs are in.
top_beam_index = top_ids // vocab_dim.size
top_ids %= vocab_dim.size # Unflatten the ids
# Recovering the log probs because we will need to send them back
top_log_probs = top_scores * length_penalty
selector = mtf.one_hot(top_beam_index, beam_dim, dtype=tf.float32)
def my_gather(tensor):
return mtf.gather(tensor,
top_beam_index,
beam_dim,
output_shape=mtf.Shape([
double_beam if d == beam_dim else d
for d in tensor.shape.dims
]))
# Gather up the most probable 2*beams both for the ids and finished_in_alive
# bools
top_seq = my_gather(alive_seq)
# Append the most probable alive
top_seq += top_ids * mtf.one_hot(i, length_dim, dtype=tf.int32)
top_finished = mtf.equal(top_ids, eos_id)
return (top_seq, top_log_probs, top_scores, top_finished, new_states,
selector)
def embedding(indices, vocab_dim, output_dim, variable_dtype, name="embedding"):
"""Embedding layer."""
weights = embedding_weights(
indices.mesh, vocab_dim, output_dim, variable_dtype, name)
return mtf.gather(weights, indices, vocab_dim)
