def GetTensorRange(self, t_name, ts):
# Always straddle a real zero point.
if self.do_eval:
# At eval/inference time, use the memorized range.
# Important: Don't capture these variables in training mode so as to
# avoid extra/unnecessary captures.
min_var = tf.stop_gradient(self._GetQStateVar(t_name, 'min'))
max_var = tf.stop_gradient(self._GetQStateVar(t_name, 'max'))
return (min_var, max_var)
# Calculate min/max for all tensors.
batch_min = tf.minimum(tf.reduce_min(ts), 0.0)
batch_max = tf.maximum(tf.reduce_max(ts), 0.0)
return (tf.stop_gradient(batch_min), tf.stop_gradient(batch_max))
def QuantizeTensors(self, t_name, ts, eval_only=False):
p = self.params
# Always straddle a real zero point.
if self.do_eval:
# At eval/inference time, use the memorized range.
# Important: Don't capture these variables in training mode so as to
# avoid extra/unnecessary captures.
min_var = self._GetQStateVar(t_name, 'min')
max_var = self._GetQStateVar(t_name, 'max')
return [
self._MaybeFakeQuant(t, min_var, max_var, num_bits=p.bits)
for t in ts
]
else:
# At training time, use the batch calculated min/max.
accumulator_name = self._GetAccumulatorNameForTensor(t_name)
# Calculate min/max for all tensors.
batch_min = 0.0
batch_max = 0.0
for t in ts:
batch_min = tf.minimum(tf.reduce_min(t), batch_min)
batch_max = tf.maximum(tf.reduce_max(t), batch_max)
# New state.
state1 = tf.stack([1.0, batch_min, batch_max])
self.accumulators[accumulator_name].Update(state1)
# Results.
ts_out = []
for i, t in enumerate(ts):
if eval_only:
# If only quantizing at eval time, still record ranges as above
# but don't quantize.
quant_t = t
else:
# If quantizing during training, skip quantization if it produces
# NANs. Sometimes early in the training process, things are unstable
# and ranges can produce numerical instability that makes it
# impossible to perform a fake_quant.
quant_t = self._MaybeFakeQuant(t,
batch_min,
batch_max,
num_bits=p.bits)
# TODO(laurenzo): Plumb quant_t_has_nans through state and report.
quant_t_has_nans = tf.math.is_nan(quant_t)
quant_t = tf.where(quant_t_has_nans, t, quant_t)
ts_out.append(quant_t)
summary_utils.histogram(
'%s/%s_%d' % (self._qvars_scope.name, t_name, i), t)
return ts_out
def FProp(self, theta, inputs, paddings):
"""Apply global spatial pooling to inputs.
Args:
theta: A `.NestedMap` object containing weights' values of this layer and
its children layers.
inputs: The inputs tensor. It is expected to be of shape [batch, time,
frequency, channel]. The time dimension corresponds to the height
dimension as in images and the frequency dimension corresponds to the
width dimension as in images.
paddings: The paddings tensor. It is expected to be of shape [batch,
time]. Defaults to None, which means there no paddings.
Returns:
outputs, out_paddings pair.
- outputs: has shape [batch, 1, 1, channel].
- out_paddings: None or has shape [batch, 1].
"""
p = self.params
assert p.pooling_type in ['MAX', 'AVG'], p.pooling_type
b, t, f = py_utils.GetShape(inputs, ndims=3)
if paddings is not None:
paddings = py_utils.HasShape(paddings, [b, t])
if paddings is not None:
mask = 1.0 - paddings[..., tf.newaxis, tf.newaxis]
else:
mask = tf.ones([b, t, 1, 1], p.dtype)
if p.pooling_type == 'AVG':
global_sum = tf.reduce_sum(inputs * mask,
axis=[1, 2],
keepdims=True)
f = tf.cast(tf.convert_to_tensor(f), p.dtype)
count = f * tf.reduce_sum(mask, axis=[1, 2], keepdims=True)
out_feature = global_sum / tf.maximum(1.0, count)
elif p.pooling_type == 'MAX':
large_negative = (tf.ones_like(inputs) * p.dtype.max *
tf.constant(-0.7, dtype=p.dtype))
padded_inputs = tf.where_v2(mask > 0.0, inputs, large_negative)
out_feature = tf.reduce_max(padded_inputs,
axis=[1, 2],
keepdims=True)
if paddings is None:
out_paddings = None
else:
out_paddings = tf.reduce_min(paddings, axis=1, keepdims=True)
out_feature *= 1.0 - out_paddings[..., tf.newaxis, tf.newaxis]
return out_feature, out_paddings
def QuantizeWeight(self, w):
p = self.params
w_min = tf.reduce_min(w)
w_max = tf.reduce_max(w)
# NOTE: We force a small, non-zero range because otherwise, zero weights
# can cause downstream inference engines to blow up.
w_min = tf.minimum(w_min, -p.quantize_weight_epsilon)
w_max = tf.maximum(w_max, p.quantize_weight_epsilon)
quant_w = self._MaybeFakeQuant(w, w_min, w_max, num_bits=p.bits)
if self.do_eval:
return quant_w
else:
# If quantizing during training, skip quantization if it produces
# NANs. Sometimes early in the training process, things are unstable
# and ranges can produce numerical instability that makes it
# impossible to perform a fake_quant.
quant_w_has_nans = tf.math.is_nan(quant_w)
return tf.where(quant_w_has_nans, w, quant_w)
def FProp(self, theta, input_batch):
"""Embeds source ids and transforms with TransformerStack.
Args:
theta: A `.NestedMap` object containing weights' values of this
layer and its children layers.
input_batch: A `.NestedMap` with fields:
- ids: The inputs tensor. It is expected to be of shape [batch, time].
- paddings: The paddings tensor. Expected shape [batch, time].
- task_ids: If p.task_emb is provided, must contain per-token task
ids of shape [batch, time].
Returns:
A NestedMap containing
- encoded: The encoded features, either a tensor of shape
[time, batch, depth], or a list of tensors if is_transparent is set in
transformer_stack.
- padding: of shape [time, batch]
- segment_id: [time, batch] if packed inputs are supported by the model
(and all layers), or None otherwise.
- embedded_inputs: [time, batch, depth] embedded inputs tokens without
positional encodings.
"""
p = self.params
with tf.name_scope(p.name):
src_segment_id = None
src_segment_pos = None
input_ids = py_utils.with_dependencies([
py_utils.assert_shape_match(tf.shape(input_batch.ids),
tf.shape(input_batch.paddings)),
py_utils.assert_equal(tf.rank(input_batch.ids), 2)
], input_batch.ids)
if (not py_utils.use_tpu()
and tf.flags.FLAGS.transformer_encoder_truncates_inputs):
max_seq_length = tf.cast(
tf.reduce_max(tf.reduce_sum(1.0 - input_batch.paddings,
1)), tf.int32)
paddings = py_utils.with_dependencies([
py_utils.assert_equal(
tf.constant(True, tf.bool),
tf.reduce_all(
input_batch.paddings[:, max_seq_length:] > 0.5))
], input_batch.paddings)
input_ids = input_ids[:, :max_seq_length]
paddings = paddings[:, :max_seq_length]
if p.packed_input:
src_segment_id = input_batch.segment_ids[:, :
max_seq_length]
src_segment_pos = input_batch.segment_pos[:, :
max_seq_length]
else:
paddings = input_batch.paddings
if p.packed_input:
src_segment_id = input_batch.segment_ids
src_segment_pos = input_batch.segment_pos
max_time = tf.shape(input_ids)[1]
# Input token embeddings + positional embeddings
if not p.shared_emb:
input_embs = self.token_emb.EmbLookup(
theta.token_emb, tf.reshape(input_ids, [-1]))
else:
input_embs = self.softmax.EmbLookup(
theta.softmax, tf.reshape(input_ids, [-1]))
input_embs = tf.reshape(input_embs,
[-1, max_time, p.token_emb.embedding_dim])
# [time, batch, dim]
orig_input_embs = tf.transpose(input_embs, [1, 0, 2])
if p.packed_input:
position_embs = self.position_emb.FPropWithPosition(
theta.position_emb, src_segment_pos)
else:
position_embs = self.position_emb.FProp(
theta.position_emb, max_time)
position_embs = tf.reshape(
position_embs, [1, max_time, p.token_emb.embedding_dim])
input_embs += position_embs
if p.task_emb:
input_embs += self.task_emb.EmbLookup(theta.task_emb,
input_batch.task_ids)
if p.model_dim != p.token_emb.embedding_dim:
input_embs = self.emb_proj.FProp(theta.emb_proj, input_embs)
paddings = tf.cast(tf.transpose(paddings), py_utils.FPropDtype(p))
if p.packed_input:
src_segment_id = tf.transpose(src_segment_id)
input_embs = self.input_dropout.FProp(theta.input_dropout,
input_embs)
# [time, batch, dim]
transformer_input = tf.transpose(input_embs, [1, 0, 2])
if not self.do_eval and p.apply_source_mask:
# Augment padding for masked source word positions.
dtype = paddings.dtype
source_mask = tf.where(tf.equal(input_ids, p.source_mask_id),
tf.ones_like(input_ids, dtype=dtype),
tf.zeros_like(input_ids, dtype=dtype))
# Make sure padding is between 0 and 1.
paddings = tf.clip_by_value(paddings + tf.transpose(source_mask),
0.0, 1.0)
encoded, padding, segment_id = self.transformer_stack.FProp(
theta.transformer_stack, transformer_input, paddings,
src_segment_id)
return py_utils.NestedMap(encoded=encoded,
padding=padding,
segment_id=segment_id,
embedded_inputs=orig_input_embs)
def _StringsToIdsImpl(self, strs, max_length, append_eos, languages):
"""Takes a tensor of strings and returns id/padding tensors.
This generates `token_ids`, `target_ids`, and `paddings` in the format that
is expected for tokenizers. This performs padding to a fixed length and
appends the end-of-sentence token as appropriate.
Args:
strs: a string Tensor.
max_length: a python integer. The second dimension of the returned arrays.
All sequences are padded or truncated to that length.
append_eos: a python bool. See `BaseTokenizer` for explanation.
languages: A vector of strings with the same length as `strs`.
Returns:
A tuple of 3 tensors:
- token_ids: a tensor of sequences of WPM ids starting with SOS. Sequences
always end with EOS unless the sequence exceeds the maximum length.
Always padded with EOS.
- target_ids: a tensor of sequences of WPM ids not starting with SOS
but ending with EOS. Always padded with EOS.
- paddings: a tensor of floats indicating, at each position, whether
the corresponding position is padded.
"""
p = self.params
if append_eos is None:
append_eos = p.append_eos
batch_size = py_utils.GetShape(strs)[0]
token_ids_ta = tf.TensorArray(tf.int32, batch_size)
target_ids_ta = tf.TensorArray(tf.int32, batch_size)
paddings_ta = tf.TensorArray(tf.float32, batch_size)
def _TokenizeOneSentence(i, strs, token_ids_ta, target_ids_ta,
paddings_ta):
"""Tokenizes a single sentence."""
ids, _ = self._wpm_encoder.Encode(strs[i])
if append_eos:
ids = tf.concat([ids, [self.eos_id]], axis=0)
# This truncates after the eos is added, so some sentences might
# not have </s> at the end.
token_ids_ta = token_ids_ta.write(
i,
py_utils.PadOrTrimTo(tf.concat([[self.sos_id], ids], axis=0),
[max_length], self.eos_id))
target_ids_ta = target_ids_ta.write(
i, py_utils.PadOrTrimTo(ids, [max_length], self.eos_id))
paddings_ta = paddings_ta.write(
i,
py_utils.PadOrTrimTo(tf.zeros_like(ids, dtype=tf.float32),
[max_length], 1.))
return i + 1, strs, token_ids_ta, target_ids_ta, paddings_ta
_, _, token_ids_ta, target_ids_ta, paddings_ta = tf.while_loop(
lambda i, *_: i < batch_size,
_TokenizeOneSentence,
loop_vars=(tf.constant(0, tf.int32), strs, token_ids_ta,
target_ids_ta, paddings_ta),
parallel_iterations=30,
back_prop=False)
token_ids = token_ids_ta.stack()
target_ids = target_ids_ta.stack()
paddings = paddings_ta.stack()
if not p.pad_to_max_length:
maxlen = tf.cast(
tf.round(tf.reduce_max(tf.reduce_sum(1.0 - paddings, axis=1))),
tf.int32)
token_ids = token_ids[:, :maxlen]
target_ids = target_ids[:, :maxlen]
paddings = paddings[:, :maxlen]
return token_ids, target_ids, paddings
def FProp(self,
theta,
x,
x_paddings=None,
eos_id=1,
force_sample_last_token=True):
"""Applies SymbolInsertionLayer.
We take in a `x`, which represents the groundtruth sequence (i.e., English
sequence). We return a sampled rollin (observed) canvas (i.e., random subset
of the English sequence), as well as the target (indices) for an
insertion-based model (i.e., the targets given the random observed subset).
Args:
theta: Ignored, this can be None.
x: The symbol ids of shape `[batch_size, time_dim]`.
x_paddings: The paddings (1 or 0) of shape `[batch_size, time_dim]` where
0 is valid and 1 is invalid.
eos_id: The <eos> token id to represent end-of-slot.
force_sample_last_token: Set True to force sample the last token of `x`.
Returns:
A `NestedMap`.
- canvas: The canvas (based off of the `rollin_policy`) of shape
[batch_size, c_dim]. Note that, `c_dim` <= `time_dim` but need not be
equal.
- canvas_indices: The canvas indices (into `x`).
- canvas_paddings: The paddings of `canvas_indices`.
- target_indices: The target indices of shape [num_targets, 3].
`num_targets` is the number of total targets in the entire batch.
[:, 0] captures the batch, [:, 1] captures the slot, and [:, 2]
captures the token. Each row [batch, slot, vocab] represents the
indices of the target -- i.e., the batch, slot and vocab combination
of the target. Typical usage of these indices is to tf.gather_nd
the log-probs (from the softmax layer).
- target_weights: The target weights.
Raises:
ValueError: If invalid params.
"""
p = self.params
batch_size = py_utils.GetShape(x)[0]
time_dim = py_utils.GetShape(x)[1]
if x_paddings is None:
x_paddings = tf.zeros([batch_size, time_dim], tf.float32)
oracle_policy = p.oracle_policy
rollin_policy = (oracle_policy
if p.rollin_policy == 'oracle' else p.rollin_policy)
if rollin_policy != 'uniform':
raise ValueError('Unknown or unsupported rollin policy: %s' %
rollin_policy)
if oracle_policy != 'uniform':
raise ValueError('Unknown or unsupported oracle policy: %s' %
oracle_policy)
x_len = tf.cast(tf.round(tf.reduce_sum(1 - x_paddings, 1)), tf.int32)
# Compute the desired length per example in the batch.
ratio = tf.random.uniform([batch_size], 0.0, 1.0, seed=p.random_seed)
if force_sample_last_token:
c_len = tf.minimum(
tf.cast(ratio * tf.cast(x_len, tf.float32), tf.int32),
x_len - 1) + 1
else:
c_len = tf.minimum(
tf.cast(ratio * tf.cast(x_len + 1, tf.float32), tf.int32),
x_len)
# Compute the maximum length across the batch.
c_len_max = tf.reduce_max(c_len)
# Grab subset of random valid indices per example.
z_logits = tf.cast(
tf.expand_dims(tf.range(time_dim), 0) >= tf.expand_dims(x_len, 1),
tf.float32) * -1e9
if force_sample_last_token:
# Force sample the last token -- i.e., as indexed by `x_len - 1`. We can
# accomplish this by add +LARGE_NUMBER to the logits.
z_logits += tf.cast(
tf.equal(tf.expand_dims(tf.range(time_dim), 0),
tf.expand_dims(x_len - 1, 1)), tf.float32) * 1e9
# Gumbel-max trick to sample (we only sample valid positions per sample in
# the batch).
z = -tf.math.log(-tf.math.log(
tf.random.uniform([batch_size, time_dim], seed=p.random_seed)))
unused_c_values, c_indices = tf.nn.top_k(z_logits + z, time_dim)
# Trim everything > c_len_max.
c_indices = c_indices[:, :c_len_max]
# Invalidate any indices >= c_len, we use the last index as the default
# invalid index.
c_indices = tf.where(
tf.expand_dims(tf.range(c_len_max), 0) < tf.expand_dims(c_len, 1),
c_indices, tf.fill(py_utils.GetShape(c_indices), time_dim - 1))
# Materialize the canvas.
c_indices = tf.sort(c_indices)
c = tf.gather_nd(
x,
tf.stack([
tf.reshape(
tf.tile(tf.expand_dims(tf.range(batch_size), 1),
[1, c_len_max]), [-1]),
tf.reshape(c_indices, [-1])
], 1))
c = tf.reshape(c, [batch_size, c_len_max])
# Compute the paddings.
c_paddings = 1 - tf.sequence_mask(
c_len, c_len_max, dtype=x_paddings.dtype)
c *= tf.cast(1 - c_paddings, tf.int32)
indices = tf.concat([
tf.reshape(
tf.tile(tf.expand_dims(tf.range(batch_size), 1),
[1, c_len_max]), [batch_size * c_len_max, 1]),
tf.reshape(c_indices, [batch_size * c_len_max, 1])
], 1)
x_token_is_observed = tf.scatter_nd(
indices, tf.ones([batch_size * c_len_max], tf.int32),
py_utils.GetShape(x))
# `x_segments` captures which slot each `x` belongs to (both observed and
# tokens that need to be observed).
x_segments = tf.cumsum(x_token_is_observed, 1, exclusive=True)
x_token_is_observed = tf.cast(x_token_is_observed, tf.bool)
prev_x_token_is_observed = tf.pad(x_token_is_observed[:, :-1],
[[0, 0], [1, 0]],
constant_values=True)
x_token_is_observed = tf.reshape(x_token_is_observed, [-1])
prev_x_token_is_observed = tf.reshape(prev_x_token_is_observed, [-1])
x_is_valid = tf.cast(1 - x_paddings, tf.bool)
x_is_valid = tf.reshape(x_is_valid, [-1])
# Remap all the observed to <eos>, note some of these need a zero weight
# (or else there would be <eos> and valid token in the same slot).
target_indices = tf.cast(tf.reshape(x, [-1, 1]), tf.int32)
target_indices = tf.where(
x_token_is_observed,
tf.fill(py_utils.GetShape(target_indices), eos_id), target_indices)
# TODO(williamchan): We give uniform 1.0 weight, however, math suggests
# we may want to weigh this term by the original sequence length.
target_weights = tf.ones_like(target_indices, tf.float32)
# We need to set all the weights for <eos> which actually have valid tokens
# in the slot to zero.
target_weights = tf.where(
x_token_is_observed & ~prev_x_token_is_observed,
tf.zeros_like(target_weights), target_weights)
# TODO(williamchan): Consider dropping the entries w/ weight zero.
# Add the batch and slot indices.
target_indices = tf.concat([
tf.reshape(
tf.tile(tf.expand_dims(tf.range(batch_size), 1),
[1, time_dim]), [batch_size * time_dim, 1]),
tf.reshape(x_segments, [-1, 1]), target_indices
], 1)
# Select only the valid indices. The selected valid ones include slots w/
# <eos>.
target_indices = target_indices[x_is_valid]
target_weights = target_weights[x_is_valid]
return py_utils.NestedMap(canvas=c,
canvas_indices=c_indices,
canvas_paddings=c_paddings,
target_indices=target_indices,
target_weights=target_weights)
def beam_search_step(in_scores,
in_atten_probs,
in_best_scores,
in_cumulative_scores,
in_histories,
cur_step,
eos_id,
num_beams,
beam_size,
num_hyps_per_beam,
valid_eos_max_logit_delta=5.0,
local_eos_threshold=-100.0,
merge_paths=False,
is_last_chunk=None,
eoc_id=0):
"""A single step of beam search.
Let "b" be the number of beams, "k" be the number hyps in each beam. This
function supports values with dtypes tf.float32 or tf.bfloat16.
The following data structures are allocated before the first decoding step and
are passed along from cur step to the next step:
Args:
in_scores: A tensor of shape [b * k, vocab_size], where [i, ...] is the
token score of the j-th hyps of the n-th beam. j = (i / k), and n = i % k
in_atten_probs: A tensor of shape [b*k, s_len], where in_atten_probs[i, ...]
is the attention probabilities over the source words of the j-th hyps of
n-th beam (where j, and n are derived as above).
in_best_scores: A vector of size [b], best scores of terminated hyps so far
in each of the beams.
in_cumulative_scores: A vector of size [b * k]. The cumulative score of each
active hyp before the current step.
in_histories: An int32 vector of size [b * k] containing hashes of the
histories of each active hyp. If 'merge_paths' is enabled, the histories
are used to identify hypotheses that are identical modulo epsilons (e.g.
"a <eps> b" and "a b <eps>") and merge them. See 'update_histories'
docstring for details.
cur_step: Current step id.
eos_id: Token id of the special end of sequence token.
num_beams: Number of beams.
beam_size: Search terminates if the delta between the scores of the active
hyps.
num_hyps_per_beam: Number of hyps in a beam.
valid_eos_max_logit_delta: We allow </s> to terminate a hyp only if its
logit is no more than 'valid_eos_max_logit_delta' away from the logit of
the best candidate.
local_eos_threshold: We allow </s> to terminate a hyp if the local score for
</s> is greater than local_eos_threshold.
merge_paths: If true, hyps which are identical when epsilons are removed
will be combined into a single hyp. The probability for that combined hyp
will be the sum of the probabilities of the component hyps. This can only
be applied for epsilon-emitting models (RNN-T and NT).
is_last_chunk: A tensor of shape [b * k, 1]. Used by neural transducer,
determines whether the current hypothesis reaches the last chunk and
should treat the next end-of-chunk symbol as end-of-sentence.
eoc_id: int, the id of the end of chunk (a.k.a epsilon) token used by neural
transducer models. Only relevant if 'merge_paths' is True or
'is_last_chunk' is provided.
Returns:
out_best_scores: A tensor of shape [b] of updated best scores for each of
the beams.
out_cumulative_scores: A tensor of shape [b * k]. The cumulative score of
the new hyps after the current decoding step.
out_scores: A tensor of shape [b * k] with scores of the token selected.
out_eos_scores: A tensor of shape [b * k] with token scores for the EOS, in
case the hyp was terminated, otherwise 0.0.
out_hyps: A tensor of shape [b * k] with ids of the token selected.
out_prev_hyps: A tensor of shape [b * k] with index to the previous hyps
which was selected.
out_done_hyps: A boolean tensor of shape [b * k] where value indicates
if hyps was terminated.
out_atten_probs: A tensor of shape [b * k, seq_len] which contain the
attention probabilities over the source words against word in the previous
hyps.
out_eos_atten_probs: A tensor of shape [b * k, seq_len] which contains the
attention probabilities over the source against word in the current hyp
which was terminated.
out_all_done: A scalar, whether decoding should terminate for all beams.
out_histories: A tensor of shape [b * k] containing new history hashes for
the active hypotheses. See 'update_histories' docstring for details.
Raises:
ValueError: if inputs are invalid.
"""
num_hyps_per_beam = int(num_hyps_per_beam)
if num_hyps_per_beam <= 0:
raise ValueError("num_hyps_per_beam = {} and must be > 0.".format(
num_hyps_per_beam))
in_scores = tf.convert_to_tensor(in_scores)
in_scores.shape.assert_has_rank(2)
num_classes = in_scores.get_shape()[1]
in_atten_probs = tf.convert_to_tensor(in_atten_probs)
in_atten_probs.shape.assert_has_rank(2)
in_best_scores = tf.convert_to_tensor(in_best_scores)
in_best_scores.shape.assert_has_rank(1)
in_cumulative_scores = tf.convert_to_tensor(in_cumulative_scores)
in_cumulative_scores.shape.assert_has_rank(1)
in_histories = tf.convert_to_tensor(in_histories)
in_histories.shape.assert_has_rank(1)
with tf.name_scope("beam_search_step"):
# For k = num_hyps_per_beam
# First step of beam search is to find the top tokens based on its score.
# Normally we select k+1, where the extra +1 is to make sure we have k
# non-eos tokens to select if EOS token is in the top-k. If path merging is
# on, we actually need to select k+2; this ensures there are k+1 tokens left
# after the merge, at least k of which are not EOS.
# TODO(b/118644069): Avoid casts when there is a XLA op available that takes
# in bfloat16.
num_candidates_per_input_hyp = (num_hyps_per_beam + 2 if merge_paths
else num_hyps_per_beam + 1)
# [b * k, num_candidates_per_input_hyp]
local_score_values, local_indices = xla_ops.top_k_with_unique(
tf.cast(in_scores, tf.float32), k=num_candidates_per_input_hyp)
local_score_values = tf.cast(local_score_values, in_scores.dtype)
# Compute the global score which is sum of the local score, and the
# cumulative scores for each of the hyps.
# [b * k, num_candidates_per_input_hyp]
global_score_values = local_score_values + tf.expand_dims(
in_cumulative_scores, 1)
values_dtype = local_score_values.dtype
is_first_step = tf.cast(tf.equal(cur_step, 0), values_dtype)
# Preprocessing to reorder the tensor from `mod` sharding to `div` so that
# we can use matrix/vector operations to complete the beam search.
# [b * k, num_candidates_per_input_hyp]
global_score_values = reorder_tensor("mod_to_div", global_score_values,
num_beams, num_hyps_per_beam)
local_score_values = reorder_tensor("mod_to_div", local_score_values,
num_beams, num_hyps_per_beam)
local_indices = reorder_tensor("mod_to_div",
local_indices,
num_beams,
num_hyps_per_beam,
max_value=num_classes - 1)
# [b * k, 1]
histories = reorder_tensor("mod_to_div",
tf.expand_dims(in_histories, 1), num_beams,
num_hyps_per_beam)
if is_last_chunk is None:
is_last_chunk = tf.zeros([num_beams * num_hyps_per_beam, 1],
tf.bool)
else:
is_last_chunk = tf.cast(
reorder_tensor(
"mod_to_div",
tf.reshape(is_last_chunk,
[num_beams * num_hyps_per_beam, 1]), num_beams,
num_hyps_per_beam), tf.bool)
# For the first step mask everything but the first row.
# [num_hyps_per_beam]
per_example_mask = tf.concat([
tf.constant([1.0], dtype=values_dtype),
tf.zeros([num_hyps_per_beam - 1], dtype=values_dtype)
], 0)
# [num_hyps_per_beam, num_beams] => [b*k, 1]
mask = tf.reshape(
tf.tile(per_example_mask, tf.expand_dims(num_beams, 0)),
[-1, 1]) * is_first_step + (1.0 - is_first_step)
local_score_values *= mask
global_score_values *= mask
# We add a large negative value for the unmasked values.
per_example_additive_mask = tf.concat([
tf.constant([0.0], dtype=values_dtype),
tf.constant(BEST_SCORES_INIT,
shape=[num_hyps_per_beam - 1],
dtype=values_dtype)
], 0)
additive_mask = tf.reshape(
tf.tile(per_example_additive_mask, tf.expand_dims(num_beams, 0)),
[-1, 1]) * is_first_step
local_score_values += additive_mask
global_score_values += additive_mask
if merge_paths:
with tf.name_scope("merge_paths"):
# Compute new history hashes for each hypothesis + new token.
# [b * k, num_candidates_per_input_hyp]
histories = update_histories(histories,
local_indices,
mask,
epsilon_id=eoc_id)
global_score_values, histories = merge_hyps(
global_score_values, histories, mask, num_beams,
num_hyps_per_beam)
# As we keep num_candidates_per_input_hyp, we have a total of
# num_candidates_per_input_hyp * k hyps active per example.
num_candidate_hyps = num_candidates_per_input_hyp * num_hyps_per_beam
batch_shape = [-1, num_candidate_hyps]
# Reshape score values so that each row corresponds to a particular example.
# [num_beams, num_candidate_hyps]
global_score_values_batch = tf.reshape(global_score_values,
batch_shape)
# First for each beam: Find the top 2 * num_hyps_per_beam candidates.
# The factor of 2 is to be able to process non EOS token ids in the case
# where top scoring token for each hyps is EOS token.
# [k * b, 2 * num_hyps_per_beam]
_, candidates_indices_in_top_k = xla_ops.top_k_with_unique(
tf.cast(global_score_values_batch, tf.float32),
k=2 * num_hyps_per_beam)
# Find the previous hyps of the candidate. We divide here by (k+1) to
# identify which hyps this token came from.
hyps_id = candidates_indices_in_top_k // num_candidates_per_input_hyp
# Add in offset so that we can get the candidate index in the [b * k] space.
offset = tf.expand_dims(tf.range(num_beams) * num_candidate_hyps, 1)
flat_candidates_indices_in_top_k = tf.reshape(
candidates_indices_in_top_k + offset, [-1])
flat_local_indices = tf.reshape(local_indices, [1, -1])
flat_token_scores = tf.reshape(local_score_values, [-1, 1])
flat_global_scores = tf.reshape(global_score_values, [-1, 1])
# Gather the token scores for each of 2*k candidates. We use tf.one_hot()
# followed by a tf.matmul() to speedup gather on TPUs.
total_num_candidates = num_beams * num_candidate_hyps
token_scores_for_beam = tf.reshape(
fast_gather(flat_token_scores, flat_candidates_indices_in_top_k,
total_num_candidates),
[num_beams, 2 * num_hyps_per_beam])
token_scores_for_beam_shape = tf.shape(token_scores_for_beam)
global_scores_for_beam = tf.reshape(
fast_gather(flat_global_scores, flat_candidates_indices_in_top_k,
total_num_candidates), token_scores_for_beam_shape)
# Local indices value's are between [0, vocab_size-1], hence we use the
# slower version of gather.
token_ids_for_beam = tf.reshape(
fast_gather(flat_local_indices,
flat_candidates_indices_in_top_k,
total_num_candidates,
max_value=num_classes - 1,
axis=1), token_scores_for_beam_shape)
# We have access to 2*num_hyps_per_beam hyps per beam.
# We shrink back to num_hyps_per_beam that does not include EOS, and move
# EOS that occurs in top-num_hyps_per_beam to the EOS done matrix.
# To determine the threshold at which eos is allowed to terminate a hyp,
# we need to know the maximum global score for that hyp with any additional
# token. If path merging is *not* enabled, the global_score_values are
# by construction in sorted order, so we can just look at its 0th column. If
# path merging is enabled, the global scores of deleted (merged) hyps break
# the sorted order, which means we have to do a full reduce_max.
if merge_paths:
max_global_score_per_input_hyp = tf.reduce_max(global_score_values,
axis=1,
keepdims=True)
else:
max_global_score_per_input_hyp = global_score_values[:, 0:1]
# [num_beams * num_hyps_per_beam, 1]
global_eos_threshold = (max_global_score_per_input_hyp -
valid_eos_max_logit_delta)
local_eos_threshold_tensor = local_eos_threshold * tf.ones_like(
global_eos_threshold)
# Find EOS in top num_hyps_per_beam token ids. We also treat EOC as EOS if
# the model has indicated this is the last chunk.
local_index_is_eos = tf.equal(local_indices, eos_id)
local_index_is_last_chunk_eoc = tf.math.logical_and(
tf.equal(local_indices, eoc_id), is_last_chunk)
eos_mask = tf.math.logical_and(
tf.math.logical_and(
tf.math.logical_and(
tf.greater(
local_score_values,
tf.tile(local_eos_threshold_tensor,
[1, num_candidates_per_input_hyp])),
tf.greater(
global_score_values,
tf.tile(global_eos_threshold,
[1, num_candidates_per_input_hyp]))),
tf.math.logical_or(local_index_is_eos,
local_index_is_last_chunk_eoc)),
tf.cast(mask, tf.bool))
end_hyps_bool_mask = tf.reshape(tf.reduce_any(eos_mask, 1), [-1, 1])
end_hyps_bool_mask = reorder_tensor("div_to_mod", end_hyps_bool_mask,
num_beams, num_hyps_per_beam)
eos_atten_probs = in_atten_probs * tf.cast(end_hyps_bool_mask,
in_atten_probs.dtype)
eos_atten_probs = tf.reshape(eos_atten_probs,
[num_beams * num_hyps_per_beam, -1])
# A boolean tensor of shape [b * k] where value indicates if hyps was
# terminated.
out_done_hyps = tf.reshape(end_hyps_bool_mask, [-1])
# Scores for EOS token.
eos_float_mask = tf.cast(eos_mask, values_dtype)
eos_local_scores = eos_float_mask * local_score_values
eos_additive_float_mask = (1.0 - eos_float_mask) * BEST_SCORES_INIT
eos_local_scores += eos_additive_float_mask
out_eos_scores = tf.reshape(tf.reduce_max(eos_local_scores, 1),
[-1, 1])
out_eos_scores = tf.reshape(
reorder_tensor("div_to_mod", out_eos_scores, num_beams,
num_hyps_per_beam), [-1])
# A tensor of shape [b] of updated best scores for each of the beams.
eos_global_scores = eos_float_mask * global_score_values
eos_global_scores += eos_additive_float_mask
best_scores = tf.reduce_max(
tf.reshape(eos_global_scores, [num_beams, -1]), 1)
# Following operations are to finds the top num_hyps_per_beam that are
# active.
# Active ones are the ones that do not correspond to EOS termination.
# We keep num_hyps_per_beam * 2 in case every hyps is terminated by EOS id.
# Top K with eos removed.
non_eos_mask = tf.not_equal(token_ids_for_beam, eos_id)
num_candidate_hyps = num_hyps_per_beam * 2 * num_beams
index = tf.where(
non_eos_mask,
tf.reshape(tf.range(num_candidate_hyps, dtype=tf.int32),
token_scores_for_beam_shape),
num_candidate_hyps *
tf.ones(dtype=tf.int32, shape=token_scores_for_beam_shape))
# Unrolled TopK.
sorted_indices = []
# Finds the first num_hyps_per_beam unmasked indexes and stores them in
# concated_index (shape: [num_beams, num_candidate_hyps])
# This is done by iteratively record the min index in each row, and reset
# it to the max, so that next iteration reduce_min returns the 2nd minimum
# index.
for _ in range(num_hyps_per_beam):
min_index = tf.reshape(tf.reduce_min(index, [1]), [num_beams, 1])
sorted_indices.append(min_index)
# Replace position with num_candidate_hyps value.
index = tf.where(
tf.equal(index, min_index),
num_candidate_hyps *
tf.ones(dtype=tf.int32, shape=token_scores_for_beam_shape),
index)
# Post processing ops to output expected tensors.
concated_sorted_indices = tf.concat(sorted_indices, 1)
flat_sorted_indices = tf.reshape(concated_sorted_indices, [-1])
# A tensor of shape [b * k] with scores of the token selected.
out_scores = tf.reshape(
fast_gather(tf.reshape(token_scores_for_beam, [-1, 1]),
flat_sorted_indices, num_candidate_hyps), [-1, 1])
out_scores = tf.reshape(
reorder_tensor("div_to_mod", out_scores, num_beams,
num_hyps_per_beam), [-1])
# Gather the updated histories of selected hypotheses if path merging is
# enabled. Otherwise, the histories are unused, so just output in_histories.
if merge_paths:
flat_histories = tf.reshape(histories, [-1, 1])
# [num_beams, 2 * num_hyps_per_beam]
histories_for_beam = tf.reshape(
fast_gather(flat_histories, flat_candidates_indices_in_top_k,
total_num_candidates), token_scores_for_beam_shape)
out_histories = tf.reshape(
fast_gather(tf.reshape(histories_for_beam, [-1, 1]),
flat_sorted_indices, num_candidate_hyps), [-1, 1])
out_histories = tf.reshape(
reorder_tensor("div_to_mod", out_histories, num_beams,
num_hyps_per_beam), [-1])
else:
out_histories = in_histories
prev_hyps_ids = tf.reshape(
tf.reshape(
fast_gather(tf.reshape(hyps_id, [1, -1]),
flat_sorted_indices,
num_candidate_hyps,
max_value=num_hyps_per_beam,
axis=1), [num_beams, -1]) * num_beams +
tf.expand_dims(tf.range(num_beams), 1), [-1, 1])
prev_hyps_ids = reorder_tensor("div_to_mod",
prev_hyps_ids,
num_beams,
num_hyps_per_beam,
max_value=num_hyps_per_beam)
# A tensor of shape [b * k] with index to the previous hyps which was
# selected.
out_prev_hyps = tf.reshape(prev_hyps_ids, [-1])
# A tensor of shape [b * k, seq_len] which contain the attention
# probabilities over the source words against word in the previous hyps.
out_atten_probs = tf.reshape(
fast_gather(in_atten_probs, out_prev_hyps,
num_beams * num_hyps_per_beam),
[num_beams * num_hyps_per_beam, -1])
sorted_top_k_ids = fast_gather(tf.reshape(token_ids_for_beam, [1, -1]),
flat_sorted_indices,
num_candidate_hyps,
max_value=num_classes - 1,
axis=1)
sorted_top_k_ids = reorder_tensor("div_to_mod",
sorted_top_k_ids,
num_beams,
num_hyps_per_beam,
max_value=num_classes - 1,
axis=1)
# A tensor of shape [b * k] with ids of the token selected.
out_hyps = tf.reshape(sorted_top_k_ids, [-1])
# A tensor of shape [b * k]. The cumulative score of the selected hyps after
# the current decoding step.
out_cumulative_scores = tf.reshape(
fast_gather(tf.reshape(global_scores_for_beam, [-1, 1]),
flat_sorted_indices, num_candidate_hyps), [-1, 1])
out_cumulative_scores = tf.reshape(
reorder_tensor("div_to_mod", out_cumulative_scores, num_beams,
num_hyps_per_beam), [-1])
out_best_scores = tf.maximum(best_scores, in_best_scores)
# A scalar, whether decoding should terminate for all beams.
out_all_done = tf.reshape(
tf.math.logical_not(
tf.reduce_any(
tf.greater(
out_cumulative_scores,
tf.reshape(
tf.tile(
tf.reshape(out_best_scores - beam_size,
[-1, 1]), [1, num_hyps_per_beam]),
[-1])))), [])
return (out_best_scores, out_cumulative_scores, out_scores,
out_eos_scores, out_hyps, out_prev_hyps, out_done_hyps,
out_atten_probs, eos_atten_probs, out_all_done, out_histories)