Skip to content

Instantly share code, notes, and snippets.

@nikitakit
Last active January 6, 2024 08:48
Show Gist options
  • Save nikitakit/6ab61a73b86c50ad88d409bac3c3d09f to your computer and use it in GitHub Desktop.
Save nikitakit/6ab61a73b86c50ad88d409bac3c3d09f to your computer and use it in GitHub Desktop.
Tensorflow Beam Search
"""
Beam decoder for tensorflow
Sample usage:
```
from tf_beam_decoder import beam_decoder
decoded_sparse, decoded_logprobs = beam_decoder(
cell=cell,
beam_size=7,
stop_token=2,
initial_state=initial_state,
initial_input=initial_input,
tokens_to_inputs_fn=lambda tokens: tf.nn.embedding_lookup(my_embedding, tokens),
)
```
See the `beam_decoder` function for complete documentation. (Only the
`beam_decoder` function is part of the public API here.)
"""
import tensorflow as tf
import numpy as np
from tensorflow.python.util import nest
# %%
def nest_map(func, nested):
if not nest.is_sequence(nested):
return func(nested)
flat = nest.flatten(nested)
return nest.pack_sequence_as(nested, list(map(func, flat)))
# %%
def sparse_boolean_mask(tensor, mask):
"""
Creates a sparse tensor from masked elements of `tensor`
Inputs:
tensor: a 2-D tensor, [batch_size, T]
mask: a 2-D mask, [batch_size, T]
Output: a 2-D sparse tensor
"""
mask_lens = tf.reduce_sum(tf.cast(mask, tf.int32), -1, keep_dims=True)
mask_shape = tf.shape(mask)
left_shifted_mask = tf.tile(
tf.expand_dims(tf.range(mask_shape[1]), 0),
[mask_shape[0], 1]
) < mask_lens
return tf.SparseTensor(
indices=tf.where(left_shifted_mask),
values=tf.boolean_mask(tensor, mask),
shape=tf.cast(tf.pack([mask_shape[0], tf.reduce_max(mask_lens)]), tf.int64) # For 2D only
)
# %%
def flat_batch_gather(flat_params, indices, validate_indices=None,
batch_size=None,
options_size=None):
"""
Gather slices from `flat_params` according to `indices`, separately for each
example in a batch.
output[(b * indices_size + i), :, ..., :] = flat_params[(b * options_size + indices[b, i]), :, ..., :]
The arguments `batch_size` and `options_size`, if provided, are used instead
of looking up the shape from the inputs. This may help avoid redundant
computation (TODO: figure out if tensorflow's optimizer can do this automatically)
Args:
flat_params: A `Tensor`, [batch_size * options_size, ...]
indices: A `Tensor`, [batch_size, indices_size]
validate_indices: An optional `bool`. Defaults to `True`
batch_size: (optional) an integer or scalar tensor representing the batch size
options_size: (optional) an integer or scalar Tensor representing the number of options to choose from
"""
if batch_size is None:
batch_size = indices.get_shape()[0].value
if batch_size is None:
batch_size = tf.shape(indices)[0]
if options_size is None:
options_size = flat_params.get_shape()[0].value
if options_size is None:
options_size = tf.shape(flat_params)[0] // batch_size
else:
options_size = options_size // batch_size
indices_offsets = tf.reshape(tf.range(batch_size) * options_size, [-1] + [1] * (len(indices.get_shape())-1))
indices_into_flat = indices + tf.cast(indices_offsets, indices.dtype)
flat_indices_into_flat = tf.reshape(indices_into_flat, [-1])
return tf.gather(flat_params, flat_indices_into_flat, validate_indices=validate_indices)
def batch_gather(params, indices, validate_indices=None,
batch_size=None,
options_size=None):
"""
Gather slices from `params` according to `indices`, separately for each
example in a batch.
output[b, i, ..., j, :, ..., :] = params[b, indices[b, i, ..., j], :, ..., :]
The arguments `batch_size` and `options_size`, if provided, are used instead
of looking up the shape from the inputs. This may help avoid redundant
computation (TODO: figure out if tensorflow's optimizer can do this automatically)
Args:
params: A `Tensor`, [batch_size, options_size, ...]
indices: A `Tensor`, [batch_size, ...]
validate_indices: An optional `bool`. Defaults to `True`
batch_size: (optional) an integer or scalar tensor representing the batch size
options_size: (optional) an integer or scalar Tensor representing the number of options to choose from
"""
if batch_size is None:
batch_size = params.get_shape()[0].merge_with(indices.get_shape()[0]).value
if batch_size is None:
batch_size = tf.shape(indices)[0]
if options_size is None:
options_size = params.get_shape()[1].value
if options_size is None:
options_size = tf.shape(params)[1]
batch_size_times_options_size = batch_size * options_size
# TODO(nikita): consider using gather_nd. However as of 1/9/2017 gather_nd
# has no gradients implemented.
flat_params = tf.reshape(params, tf.concat(0,[[batch_size_times_options_size], tf.shape(params)[2:]]))
indices_offsets = tf.reshape(tf.range(batch_size) * options_size, [-1] + [1] * (len(indices.get_shape())-1))
indices_into_flat = indices + tf.cast(indices_offsets, indices.dtype)
return tf.gather(flat_params, indices_into_flat, validate_indices=validate_indices)
# %%
class BeamFlattenWrapper(tf.nn.rnn_cell.RNNCell):
def __init__(self, cell, beam_size):
self.cell = cell
self.beam_size = beam_size
def merge_batch_beam(self, tensor):
remaining_shape = tf.shape(tensor)[2:]
res = tf.reshape(tensor, tf.concat(0, [[-1], remaining_shape]))
res.set_shape(tf.TensorShape((None,)).concatenate(tensor.get_shape()[2:]))
return res
def unmerge_batch_beam(self, tensor):
remaining_shape = tf.shape(tensor)[1:]
res = tf.reshape(tensor, tf.concat(0, [[-1, self.beam_size], remaining_shape]))
res.set_shape(tf.TensorShape((None,self.beam_size)).concatenate(tensor.get_shape()[1:]))
return res
def prepend_beam_size(self, element):
return tf.TensorShape(self.beam_size).concatenate(element)
def tile_along_beam(self, state):
if nest.is_sequence(state):
return nest_map(
lambda val: self.tile_along_beam(val),
state
)
if not isinstance(state, tf.Tensor):
raise ValueError("State should be a sequence or tensor")
tensor = state
tensor_shape = tensor.get_shape().with_rank_at_least(1)
new_tensor_shape = tensor_shape[:1].concatenate(self.beam_size).concatenate(tensor_shape[1:])
dynamic_tensor_shape = tf.unpack(tf.shape(tensor))
res = tf.expand_dims(tensor, 1)
res = tf.tile(res, [1, self.beam_size] + [1] * (tensor_shape.ndims-1))
res = tf.reshape(res, [-1, self.beam_size] + list(dynamic_tensor_shape[1:]))
res.set_shape(new_tensor_shape)
return res
def __call__(self, inputs, state, scope=None):
flat_inputs = nest_map(self.merge_batch_beam, inputs)
flat_state = nest_map(self.merge_batch_beam, state)
flat_output, flat_next_state = self.cell(flat_inputs, flat_state, scope=scope)
output = nest_map(self.unmerge_batch_beam, flat_output)
next_state = nest_map(self.unmerge_batch_beam, flat_next_state)
return output, next_state
@property
def state_size(self):
return nest_map(self.prepend_beam_size, self.cell.state_size)
@property
def output_size(self):
return nest_map(self.prepend_beam_size, self.cell.output_size)
# %%
class BeamReplicateWrapper(tf.nn.rnn_cell.RNNCell):
def __init__(self, cell, beam_size):
self.cell = cell
self.beam_size = beam_size
def prepend_beam_size(self, element):
return tf.TensorShape(self.beam_size).concatenate(element)
def tile_along_beam(self, state):
if nest.is_sequence(state):
return nest_map(
lambda val: self.tile_along_beam(val),
state
)
if not isinstance(state, tf.Tensor):
raise ValueError("State should be a sequence or tensor")
tensor = state
tensor_shape = tensor.get_shape().with_rank_at_least(1)
new_tensor_shape = tensor_shape[:1].concatenate(self.beam_size).concatenate(tensor_shape[1:])
dynamic_tensor_shape = tf.unpack(tf.shape(tensor))
res = tf.expand_dims(tensor, 1)
res = tf.tile(res, [1, self.beam_size] + [1] * (tensor_shape.ndims-1))
res = tf.reshape(res, [-1, self.beam_size] + list(dynamic_tensor_shape[1:]))
res.set_shape(new_tensor_shape)
return res
def __call__(self, inputs, state, scope=None):
varscope = scope or tf.get_variable_scope()
flat_inputs = nest.flatten(inputs)
flat_state = nest.flatten(state)
flat_inputs_unstacked = list(zip(*[tf.unstack(tensor, num=self.beam_size, axis=1) for tensor in flat_inputs]))
flat_state_unstacked = list(zip(*[tf.unstack(tensor, num=self.beam_size, axis=1) for tensor in flat_state]))
flat_output_unstacked = []
flat_next_state_unstacked = []
output_sample = None
next_state_sample = None
for i, (inputs_k, state_k) in enumerate(zip(flat_inputs_unstacked, flat_state_unstacked)):
inputs_k = nest.pack_sequence_as(inputs, inputs_k)
state_k = nest.pack_sequence_as(state, state_k)
# TODO(nikita): is this scope stuff correct?
if i == 0:
output_k, next_state_k = self.cell(inputs_k, state_k, scope=scope)
else:
with tf.variable_scope(varscope, reuse=True):
output_k, next_state_k = self.cell(inputs_k, state_k, scope=varscope if scope is not None else None)
flat_output_unstacked.append(nest.flatten(output_k))
flat_next_state_unstacked.append(nest.flatten(next_state_k))
output_sample = output_k
next_state_sample = next_state_k
flat_output = [tf.stack(tensors, axis=1) for tensors in zip(*flat_output_unstacked)]
flat_next_state = [tf.stack(tensors, axis=1) for tensors in zip(*flat_next_state_unstacked)]
output = nest.pack_sequence_as(output_sample, flat_output)
next_state = nest.pack_sequence_as(next_state_sample, flat_next_state)
return output, next_state
@property
def state_size(self):
return nest_map(self.prepend_beam_size, self.cell.state_size)
@property
def output_size(self):
return nest_map(self.prepend_beam_size, self.cell.output_size)
# %%
class BeamSearchHelper(object):
# Our beam scores are stored in a fixed-size tensor, but sometimes the
# tensor size is greater than the number of elements actually on the beam.
# The invalid elements are assigned a highly negative score.
# However, top_k errors if any of the inputs have a score of -inf, so we use
# a large negative constant instead
INVALID_SCORE = -1e18
def __init__(self, cell, beam_size, stop_token, initial_state, initial_input,
score_upper_bound=None,
max_len=100,
outputs_to_score_fn=None,
tokens_to_inputs_fn=None,
cell_transform='default',
scope=None
):
self.beam_size = beam_size
self.stop_token = stop_token
self.max_len = max_len
self.scope = scope
if score_upper_bound is None and outputs_to_score_fn is None:
self.score_upper_bound = 0.0
elif score_upper_bound is None or score_upper_bound > 3e38:
# Note: 3e38 is just a little smaller than the largest float32
# Second condition allows for Infinity as a synonym for None
self.score_upper_bound = None
else:
self.score_upper_bound = float(score_upper_bound)
if self.max_len is None and self.score_upper_bound is None:
raise ValueError("Beam search needs a stopping criterion. Please provide max_len or score_upper_bound.")
if cell_transform == 'default':
if type(cell) in [tf.nn.rnn_cell.LSTMCell,
tf.nn.rnn_cell.GRUCell,
tf.nn.rnn_cell.BasicLSTMCell,
tf.nn.rnn_cell.BasicRNNCell]:
cell_transform = 'flatten'
else:
cell_transform = 'replicate'
if cell_transform == 'none':
self.cell = cell
self.initial_state = initial_state
self.initial_input = initial_input
elif cell_transform == 'flatten':
self.cell = BeamFlattenWrapper(cell, self.beam_size)
self.initial_state = self.cell.tile_along_beam(initial_state)
self.initial_input = self.cell.tile_along_beam(initial_input)
elif cell_transform == 'replicate':
self.cell = BeamReplicateWrapper(cell, self.beam_size)
self.initial_state = self.cell.tile_along_beam(initial_state)
self.initial_input = self.cell.tile_along_beam(initial_input)
else:
raise ValueError("cell_transform must be one of: 'default', 'flatten', 'replicate', 'none'")
self._cell_transform_used = cell_transform
if outputs_to_score_fn is not None:
self.outputs_to_score_fn = outputs_to_score_fn
if tokens_to_inputs_fn is not None:
self.tokens_to_inputs_fn = tokens_to_inputs_fn
batch_size = tf.Dimension(None)
if not nest.is_sequence(self.initial_state):
batch_size = batch_size.merge_with(self.initial_state.get_shape()[0])
else:
for tensor in nest.flatten(self.initial_state):
batch_size = batch_size.merge_with(tensor.get_shape()[0])
if not nest.is_sequence(self.initial_input):
batch_size = batch_size.merge_with(self.initial_input.get_shape()[0])
else:
for tensor in nest.flatten(self.initial_input):
batch_size = batch_size.merge_with(tensor.get_shape()[0])
self.inferred_batch_size = batch_size.value
if self.inferred_batch_size is not None:
self.batch_size = self.inferred_batch_size
else:
if not nest.is_sequence(self.initial_state):
self.batch_size = tf.shape(self.initial_state)[0]
else:
self.batch_size = tf.shape(list(nest.flatten(self.initial_state))[0])[0]
self.inferred_batch_size_times_beam_size = None
if self.inferred_batch_size is not None:
self.inferred_batch_size_times_beam_size = self.inferred_batch_size * self.beam_size
self.batch_size_times_beam_size = self.batch_size * self.beam_size
def outputs_to_score_fn(self, cell_output):
return tf.nn.log_softmax(cell_output)
def tokens_to_inputs_fn(self, symbols):
return tf.expand_dims(symbols, -1)
def beam_setup(self, time):
emit_output = None
next_cell_state = self.initial_state
next_input = self.initial_input
# Set up the beam search tracking state
cand_symbols = tf.fill([self.batch_size, 0], tf.constant(self.stop_token, dtype=tf.int32))
cand_logprobs = tf.ones((self.batch_size,), dtype=tf.float32) * -float('inf')
first_in_beam_mask = tf.equal(tf.range(self.batch_size_times_beam_size) % self.beam_size, 0)
beam_symbols = tf.fill([self.batch_size_times_beam_size, 0], tf.constant(self.stop_token, dtype=tf.int32))
beam_logprobs = tf.select(
first_in_beam_mask,
tf.fill([self.batch_size_times_beam_size], 0.0),
tf.fill([self.batch_size_times_beam_size], self.INVALID_SCORE)
)
# Set up correct dimensions for maintaining loop invariants.
# Note that the last dimension (initialized to zero) is not a loop invariant,
# so we need to clear it. TODO(nikita): is there a public API for clearing shape
# inference so that _shape is not necessary?
cand_symbols._shape = tf.TensorShape((self.inferred_batch_size, None))
cand_logprobs._shape = tf.TensorShape((self.inferred_batch_size,))
beam_symbols._shape = tf.TensorShape((self.inferred_batch_size_times_beam_size, None))
beam_logprobs._shape = tf.TensorShape((self.inferred_batch_size_times_beam_size,))
next_loop_state = (
cand_symbols,
cand_logprobs,
beam_symbols,
beam_logprobs,
)
emit_output = tf.zeros(self.cell.output_size)
elements_finished = tf.zeros([self.batch_size], dtype=tf.bool)
return (elements_finished, next_input, next_cell_state,
emit_output, next_loop_state)
def beam_loop(self, time, cell_output, cell_state, loop_state):
(
past_cand_symbols, # [batch_size, time-1]
past_cand_logprobs,# [batch_size]
past_beam_symbols, # [batch_size*beam_size, time-1], right-aligned
past_beam_logprobs,# [batch_size*beam_size]
) = loop_state
# We don't actually use this, but emit_output is required to match the
# cell output size specfication. Otherwise we would leave this as None.
emit_output = cell_output
# 1. Get scores for all candidate sequences
logprobs = self.outputs_to_score_fn(cell_output)
try:
num_classes = int(logprobs.get_shape()[-1])
except:
# Shape inference failed
num_classes = tf.shape(logprobs)[-1]
logprobs_batched = tf.reshape(logprobs + tf.expand_dims(tf.reshape(past_beam_logprobs, [self.batch_size, self.beam_size]), 2),
[self.batch_size, self.beam_size * num_classes])
# 2. Determine which states to pass to next iteration
# TODO(nikita): consider using slice+fill+concat instead of adding a mask
nondone_mask = tf.reshape(
tf.cast(tf.equal(tf.range(num_classes), self.stop_token), tf.float32) * self.INVALID_SCORE,
[1, 1, num_classes])
nondone_mask = tf.reshape(tf.tile(nondone_mask, [1, self.beam_size, 1]),
[-1, self.beam_size*num_classes])
beam_logprobs, indices = tf.nn.top_k(logprobs_batched + nondone_mask, self.beam_size)
beam_logprobs = tf.reshape(beam_logprobs, [-1])
# For continuing to the next symbols
symbols = indices % num_classes # [batch_size, self.beam_size]
parent_refs = indices // num_classes # [batch_size, self.beam_size]
symbols_history = flat_batch_gather(past_beam_symbols, parent_refs, batch_size=self.batch_size, options_size=self.beam_size)
beam_symbols = tf.concat(1, [symbols_history, tf.reshape(symbols, [-1, 1])])
# Handle the output and the cell state shuffling
next_cell_state = nest_map(
lambda element: batch_gather(element, parent_refs, batch_size=self.batch_size, options_size=self.beam_size),
cell_state
)
next_input = self.tokens_to_inputs_fn(tf.reshape(symbols, [-1, self.beam_size]))
# 3. Update the candidate pool to include entries that just ended with a stop token
logprobs_done = tf.reshape(logprobs_batched, [-1, self.beam_size, num_classes])[:,:,self.stop_token]
done_parent_refs = tf.argmax(logprobs_done, 1)
done_symbols = flat_batch_gather(past_beam_symbols, done_parent_refs, batch_size=self.batch_size, options_size=self.beam_size)
logprobs_done_max = tf.reduce_max(logprobs_done, 1)
cand_symbols_unpadded = tf.select(logprobs_done_max > past_cand_logprobs,
done_symbols,
past_cand_symbols)
cand_logprobs = tf.maximum(logprobs_done_max, past_cand_logprobs)
cand_symbols = tf.concat(1, [cand_symbols_unpadded, tf.fill([self.batch_size, 1], self.stop_token)])
# 4. Check the stopping criteria
if self.max_len is not None:
elements_finished_clip = (time >= self.max_len)
if self.score_upper_bound is not None:
elements_finished_bound = tf.reduce_max(tf.reshape(beam_logprobs, [-1, self.beam_size]), 1) < (cand_logprobs - self.score_upper_bound)
if self.max_len is not None and self.score_upper_bound is not None:
elements_finished = elements_finished_clip | elements_finished_bound
elif self.score_upper_bound is not None:
elements_finished = elements_finished_bound
elif self.max_len is not None:
# this broadcasts elements_finished_clip to the correct shape
elements_finished = tf.zeros([self.batch_size], dtype=tf.bool) | elements_finished_clip
else:
assert False, "Lack of stopping criterion should have been caught in constructor"
# 5. Prepare return values
# While loops require strict shape invariants, so we manually set shapes
# in case the automatic shape inference can't calculate these. Even when
# this is redundant is has the benefit of helping catch shape bugs.
for tensor in list(nest.flatten(next_input)) + list(nest.flatten(next_cell_state)):
tensor.set_shape(tf.TensorShape((self.inferred_batch_size, self.beam_size)).concatenate(tensor.get_shape()[2:]))
for tensor in [cand_symbols, cand_logprobs, elements_finished]:
tensor.set_shape(tf.TensorShape((self.inferred_batch_size,)).concatenate(tensor.get_shape()[1:]))
for tensor in [beam_symbols, beam_logprobs]:
tensor.set_shape(tf.TensorShape((self.inferred_batch_size_times_beam_size,)).concatenate(tensor.get_shape()[1:]))
next_loop_state = (
cand_symbols,
cand_logprobs,
beam_symbols,
beam_logprobs,
)
return (elements_finished, next_input, next_cell_state,
emit_output, next_loop_state)
def loop_fn(self, time, cell_output, cell_state, loop_state):
if cell_output is None:
return self.beam_setup(time)
else:
return self.beam_loop(time, cell_output, cell_state, loop_state)
def decode_dense(self):
emit_ta, final_state, final_loop_state = tf.nn.raw_rnn(self.cell, self.loop_fn, scope=self.scope)
cand_symbols, cand_logprobs, beam_symbols, beam_logprobs = final_loop_state
return cand_symbols, cand_logprobs
def decode_sparse(self, include_stop_tokens=True):
dense_symbols, logprobs = self.decode_dense()
mask = tf.not_equal(dense_symbols, self.stop_token)
if include_stop_tokens:
mask = tf.concat(1, [tf.ones_like(mask[:,:1]), mask[:,:-1]])
return sparse_boolean_mask(dense_symbols, mask), logprobs
# %%
def beam_decoder(
cell,
beam_size,
stop_token,
initial_state,
initial_input,
tokens_to_inputs_fn,
outputs_to_score_fn=None,
score_upper_bound=None,
max_len=None,
cell_transform='default',
output_dense=False,
scope=None
):
"""Beam search decoder
Args:
cell: tf.nn.rnn_cell.RNNCell defining the cell to use
beam_size: the beam size for this search
stop_token: the index of the symbol used to indicate the end of the
output
initial_state: initial cell state for the decoder
initial_input: initial input into the decoder (typically the embedding
of a START token)
tokens_to_inputs_fn: function to go from token numbers to cell inputs.
A typical implementation would look up the tokens in an embedding
matrix.
(signature: [batch_size, beam_size, num_classes] int32 -> [batch_size, beam_size, ...])
outputs_to_score_fn: function to go from RNN cell outputs to scores for
different tokens. If left unset, log-softmax is used (i.e. the cell
outputs are treated as unnormalized logits).
Inputs to the function are cell outputs, i.e. a possibly nested
structure of items with shape [batch_size, beam_size, ...].
Must return a single Tensor with shape [batch_size, beam_size, num_classes]
score_upper_bound: (float or None). An upper bound on sequence scores.
Used to determine a stopping criterion for beam search: the search
stops if the highest-scoring complete sequence found so far beats
anything on the beam by at least score_upper_bound. For typical
sequence decoder models, outputs_to_score_fn returns normalized
logits and this upper bound should be set to 0. Defaults to 0 if
outputs_to_score_fn is not provided, otherwise defaults to None.
max_len: (default None) maximum length after which to abort beam search.
This provides an alternative stopping criterion.
cell_transform: 'flatten', 'replicate', 'none', or 'default'. Most RNN
primitives require inputs/outputs/states to have a shape that starts
with [batch_size]. Beam search instead relies on shapes that start
with [batch_size, beam_size]. This parameter controls how the arguments
cell/initial_state/initial_input are transformed to comply with this.
* 'flatten' creates a virtual batch of size batch_size*beam_size, and
uses the cell with such a batch size. This transformation is only
valid for cells that do not rely on the batch ordering in any way.
(This is true of most RNNCells, but notably excludes cells that
use attention.)
The values of initial_state and initial_input are expanded and
tiled along the beam_size dimension.
* 'replicate' creates beam_size virtual replicas of the cell, each
one of which is applied to batch_size elements. This should yield
correct results (even for models with attention), but may not have
ideal performance.
The values of initial_state and initial_input are expanded and
tiled along the beam_size dimension.
* 'none' passes along cell/initial_state/initial_input as-is.
Note that this requires initial_state and initial_input to already
have a shape [batch_size, beam_size, ...] and a custom cell type
that can handle this
* 'default' selects 'flatten' for LSTMCell, GRUCell, BasicLSTMCell,
and BasicRNNCell. For all other cell types, it selects 'replicate'
output_dense: (default False) toggles returning the decoded sequence as
dense tensor.
scope: VariableScope for the created subgraph; defaults to "RNN".
Returns:
A tuple of the form (decoded, log_probabilities) where:
decoded: A SparseTensor (or dense Tensor if output_dense=True), of
underlying shape [batch_size, ?] that contains the decoded sequence
for each batch element
log_probability: a [batch_size] tensor containing sequence
log-probabilities
"""
with tf.variable_scope(scope or "RNN") as varscope:
helper = BeamSearchHelper(
cell=cell,
beam_size=beam_size,
stop_token=stop_token,
initial_state=initial_state,
initial_input=initial_input,
tokens_to_inputs_fn=tokens_to_inputs_fn,
outputs_to_score_fn=outputs_to_score_fn,
score_upper_bound=score_upper_bound,
max_len=max_len,
cell_transform=cell_transform,
scope=varscope
)
if output_dense:
return helper.decode_dense()
else:
return helper.decode_sparse()
import tensorflow as tf
from tensorflow.python.platform import test
import numpy as np
from tf_beam_decoder import beam_decoder, BeamSearchHelper
# %%
sess = tf.InteractiveSession()
# %%
class MarkovChainCell(tf.nn.rnn_cell.RNNCell):
"""
This cell type is only used for testing the beam decoder.
It represents a Markov chain characterized by a probability table p(x_t|x_{t-1},x_{t-2}).
"""
def __init__(self, table):
"""
table[a,b,c] = p(x_t=c|x_{t-1}=b,x_{t-2}=a)
"""
assert len(table.shape) == 3 and table.shape[0] == table.shape[1] == table.shape[2]
with np.errstate(divide='ignore'): # ignore warning for log(0)
self.log_table = np.log(np.asarray(table, dtype=np.float32))
self.log_table_var = None
self._output_size = table.shape[0]
def __call__(self, inputs, state, scope=None):
"""
inputs: [batch_size, 1] int tensor
state: [batch_size, 1] int tensor
"""
# Simulate variable creation, to ensure scoping works correctly
log_table = tf.get_variable('log_table',
shape=(3,3,3),
dtype=tf.float32,
initializer=tf.constant_initializer(self.log_table))
if self.log_table_var is None:
self.log_table_var = log_table
else:
assert self.log_table_var == log_table
logits = tf.reshape(log_table, [-1, self.output_size])
indices = state[0] * self.output_size + inputs
return tf.gather(logits, tf.reshape(indices, [-1])), (inputs,)
@property
def state_size(self):
return (1,)
@property
def output_size(self):
return self._output_size
class BeamSearchTest(test.TestCase):
def test1(self):
"""
test correct decode in sequence
"""
with self.test_session() as sess:
table = np.array([[[0.0, 0.6, 0.4],
[0.0, 0.4, 0.6],
[0.0, 0.0, 1.0]]] * 3)
for cell_transform in ['default', 'flatten', 'replicate']:
cell = MarkovChainCell(table)
initial_state = cell.zero_state(1, tf.int32)
initial_input = initial_state[0]
with tf.variable_scope('test1_{}'.format(cell_transform)):
best_sparse, best_logprobs = beam_decoder(
cell=cell,
beam_size=7,
stop_token=2,
initial_state=initial_state,
initial_input=initial_input,
tokens_to_inputs_fn=lambda x:tf.expand_dims(x, -1),
max_len=5,
cell_transform=cell_transform,
output_dense=False,
)
tf.variables_initializer([cell.log_table_var]).run()
assert all(best_sparse.eval().values == [2])
assert np.isclose(np.exp(best_logprobs.eval())[0], 0.4)
def test2(self):
"""
test correct intermediate beam states
"""
with self.test_session() as sess:
table = np.array([[[0.9, 0.1, 0],
[0, 0.9, 0.1],
[0, 0, 1.0]]] * 3)
for cell_transform in ['default', 'flatten', 'replicate']:
cell = MarkovChainCell(table)
initial_state = cell.zero_state(1, tf.int32)
initial_input = initial_state[0]
with tf.variable_scope('test2_{}'.format(cell_transform)):
helper = BeamSearchHelper(
cell=cell,
beam_size=10,
stop_token=2,
initial_state=initial_state,
initial_input=initial_input,
tokens_to_inputs_fn=lambda x:tf.expand_dims(x, -1),
max_len=3,
cell_transform=cell_transform
)
_, _, final_loop_state = tf.nn.raw_rnn(helper.cell, helper.loop_fn)
_, _, beam_symbols, beam_logprobs = final_loop_state
tf.variables_initializer([cell.log_table_var]).run()
candidates, candidate_logprobs = sess.run((beam_symbols, beam_logprobs))
assert all(candidates[0,:] == [0,0,0])
assert np.isclose(np.exp(candidate_logprobs[0]), 0.9 * 0.9 * 0.9)
# Note that these three candidates all have the same score, and the sort order
# may change in the future
assert all(candidates[1,:] == [0,0,1])
assert np.isclose(np.exp(candidate_logprobs[1]), 0.9 * 0.9 * 0.1)
assert all(candidates[2,:] == [0,1,1])
assert np.isclose(np.exp(candidate_logprobs[2]), 0.9 * 0.1 * 0.9)
assert all(candidates[3,:] == [1,1,1])
assert np.isclose(np.exp(candidate_logprobs[3]), 0.1 * 0.9 * 0.9)
assert all(np.isclose(np.exp(candidate_logprobs[4:]), 0.0))
def test3(self):
"""
test that variable reuse works as expected
"""
with self.test_session() as sess:
table = np.array([[[0.0, 0.6, 0.4],
[0.0, 0.4, 0.6],
[0.0, 0.0, 1.0]]] * 3)
for cell_transform in ['default', 'flatten', 'replicate']:
cell = MarkovChainCell(table)
initial_state = cell.zero_state(1, tf.int32)
initial_input = initial_state[0]
with tf.variable_scope('test3_{}'.format(cell_transform)) as scope:
best_sparse, best_logprobs = beam_decoder(
cell=cell,
beam_size=7,
stop_token=2,
initial_state=initial_state,
initial_input=initial_input,
tokens_to_inputs_fn=lambda x:tf.expand_dims(x, -1),
max_len=5,
cell_transform=cell_transform,
output_dense=False,
scope=scope
)
tf.variables_initializer([cell.log_table_var]).run()
with tf.variable_scope(scope, reuse=True) as varscope:
best_sparse_2, best_logprobs_2 = beam_decoder(
cell=cell,
beam_size=7,
stop_token=2,
initial_state=initial_state,
initial_input=initial_input,
tokens_to_inputs_fn=lambda x:tf.expand_dims(x, -1),
max_len=5,
cell_transform=cell_transform,
output_dense=False,
scope=varscope
)
assert all(sess.run(tf.equal(best_sparse.values, best_sparse_2.values)))
assert np.isclose(*sess.run((best_logprobs, best_logprobs_2)))
def test4(self):
"""
test batching, with statically unknown batch size
"""
with self.test_session() as sess:
table = np.array([[[0.9, 0.1, 0],
[0, 0.9, 0.1],
[0, 0, 1.0]]] * 3)
for cell_transform in ['default', 'flatten', 'replicate']:
cell = MarkovChainCell(table)
initial_state = (tf.constant([[2],[0]]),)
initial_input = initial_state[0]
initial_input._shape = tf.TensorShape([None, 1])
with tf.variable_scope('test4_{}'.format(cell_transform)):
helper = BeamSearchHelper(
cell=cell,
beam_size=10,
stop_token=2,
initial_state=initial_state,
initial_input=initial_input,
tokens_to_inputs_fn=lambda x:tf.expand_dims(x, -1),
max_len=3,
cell_transform=cell_transform
)
_, _, final_loop_state = tf.nn.raw_rnn(helper.cell, helper.loop_fn)
_, _, beam_symbols, beam_logprobs = final_loop_state
tf.variables_initializer([cell.log_table_var]).run()
candidates, candidate_logprobs = sess.run((beam_symbols, beam_logprobs))
assert all(candidates[10,:] == [0,0,0])
assert np.isclose(np.exp(candidate_logprobs[10]), 0.9 * 0.9 * 0.9)
# Note that these three candidates all have the same score, and the sort order
# may change in the future
assert all(candidates[11,:] == [0,0,1])
assert np.isclose(np.exp(candidate_logprobs[11]), 0.9 * 0.9 * 0.1)
assert all(candidates[12,:] == [0,1,1])
assert np.isclose(np.exp(candidate_logprobs[12]), 0.9 * 0.1 * 0.9)
assert all(candidates[13,:] == [1,1,1])
assert np.isclose(np.exp(candidate_logprobs[13]), 0.1 * 0.9 * 0.9)
assert all(np.isclose(np.exp(candidate_logprobs[14:]), 0.0))
if __name__ == '__main__':
test.main()
@therne
Copy link

therne commented Aug 12, 2016

Thanks for you to sharing the nice source. However, this seems not working correctly. beam path tracking seems to be working well but cand_symbols is fixed at full of zero at first step. Can you look at it, please?

@nikitakit
Copy link
Author

nikitakit commented Aug 29, 2016

@therne Can you explain what the problem is in more detail? cand_symbols refers to already-completed beams (i.e. a STOP token was already encountered), so it's not surprising that it would be filled with filler values at the first timestep. The filler value used is the STOP token, because multiple such tokens will never occur in actually-valid entries.

I also just fixed a bug in the code, so you can try running again and see your issue has gone away.

@suryabhupa
Copy link

@nikitakit Is there a clean way to incorporate this code into the translate/seq2seq_model.py code in the official TensorFlow repo?

@avostryakov
Copy link

avostryakov commented Jan 13, 2017

@nikitakit It looks like there is a bug in a case when I define my own outputs_to_score_fn. In this case num_classes is not the same like in cell_output. So, I propose to change lines:

num_classes = int(cell_output.get_shape()[-1])

logprobs = self.outputs_to_score_fn(cell_output)

to:

logprobs = self.outputs_to_score_fn(cell_output)
num_classes = int(logprobs.get_shape()[-1])

What do you think?

@nikitakit
Copy link
Author

@avostryakov Thanks for pointing that out, I'll make that change in the next version.

There's also another bug with outputs_to_score_fn, which is that the stopping criterion is wrong if scores are allowed to be positive. I have code that fixes this, but it still needs to be cleaned up before I post it here.

@avostryakov
Copy link

avostryakov commented Jan 16, 2017

@nikitakit It looks like there is one more bug:

In line 515, you "return cand_symbols, cand_logprobs"

but in your test file (line 107, 193) you "return beam_symbols, beam_logprobs"

If I return cand_symbols, cand_logprobs every time I see the same sequences with single stop symbol within it. But if I return beam_symbols, beam_logprobs I see some sequences with length more than 1.

UPD: I understood that cand_symbols returns the best candidate for every sequence and beam_symbols returns all candidates for every sequence. But anyway cand_symbols returns only single stop token for every example. I don't understand what's wrong.

@avostryakov
Copy link

There's also another bug with outputs_to_score_fn, which is that the stopping criterion is wrong if scores are allowed to be positive.

Does it mean that I can use it but sequences can be longer? (it will be cutted later than nesessary)

@avostryakov
Copy link

avostryakov commented Jan 16, 2017

@nikitakit I'm sorry. It looks like it was my fault. I defined my outputs_to_score_fn like this:

def score_fn(self, cell_output):
        cell_output_2 = tf.reshape(cell_output, (-1, self.utterance_dec_hidden))
        output = tf.matmul(cell_output_2, tf.transpose(self.w_out)) + self.b_out
        output = tf.reshape(output, (-1, self.beam_size, self.vocab_size))
        return log_softmax(tf.nn.relu(output))

When I changed log_softmax to softmax all started to work. Anyway, can you add an example of outputs_to_score_fn?

@nikitakit
Copy link
Author

I just pushed a new version which fixes the num_classes bug. It also allows for globally-normalized models via the addition of a score_upper_bound parameter.

@nikitakit
Copy link
Author

@avostryakov

The score_fn above looks fine, except for the following two points:

  1. You want log_softmax in your outputs_to_score_fn. Scores are added together inside beam search, and adding together raw probabilities is extremely unlikely to be what you want. (Do not use softmax, I can guarantee it will give you the wrong results)
  2. I've never seen anyone put a softmax layer right after a ReLU layer. This combination makes little sense -- it's a no-op at best and degenerates to the uniform distribution at worst. I would very much appreciate it if you could point me to a paper that does this, but I suspect that this is actually a bug in your model and that you would want to get rid of the ReLU layer.

Does it mean that I can use it but sequences can be longer? (it will be cutted later than nesessary)

The bug I was describing is only applicable to certain classes of models, and your model is not one of them. In any case, this bug is now fixed.

If I return cand_symbols, cand_logprobs every time I see the same sequences with single stop symbol within it.

This is probably an issue with your model, not with the beam search decoder. You should try calculating the score of the single-stop-symbol sequence under your model, as well as the score of the other sequence you're comparing to -- I think you'll find that the single-stop-symbol sequence scores higher. (This is typical for a randomly-initialized model, but should go away once you've trained the model for long enough.)

@avostryakov
Copy link

avostryakov commented Jan 18, 2017

@nikitakit Thank you for answers and improvements! I'll try again.

I would very much appreciate it if you could point me to a paper that does this, but I suspect that this is actually a bug in your model and that you would want to get rid of the ReLU layer.

I can't point a paper. I didn't think that it could hurt performance. Thank you that you noticed this.

@gidim
Copy link

gidim commented Feb 7, 2017

@nikitakit
Can you provide an example on how to use this with the seq2seq interface provided with tensorflow?

seq2seq.py

@nikitakit
Copy link
Author

@gidim There are usage instructions at the top of tf_beam_decoder.py about how to use beam search. Plugging it into your model should be extremely straightforward. If you are having issues, could you post a code snippet here? (I can try to help you debug)

@fodrh1201
Copy link

there is a bug in calculating logprobs_done.
it has to be divided by time length.

it should be logprobs_done = tf.reshape(logprobs_batched, [-1, self.beam_size, num_classes])[:, :, self.stop_token] / tf.float(time)

@nikitakit
Copy link
Author

@fodrh1201 Why do you say this is a bug? I've certainly seen people normalize scores by sequence length before, but as far as I know that's a non-standard modification of beam search. Vanilla beam search just finds the most-probable sequence without considering length at all.

@wanpifeng4ever
Copy link

@nikitakit
hi, I use your beam_decoder and find that when setting beam_size to 1, the predicted result is still different from that predicted by a model which uses greedy search. Is it a normal situation?
Here is the code for tokens_to_inputs_fn and outputs_to_score_fn:

  1. tokens_to_inputs_fn
def tokens_to_intputs_fn(model_predictions): 
    _input = tf.nn.embedding_lookup(embeddings, model_predictions)
    return _input
  1. outputs_to_score_fn
def outputs_to_score_fn(model_output): 
    output_logits = tf.reshape(model_output, (-1, self.num_lstm_units)) 
    # fully connected layer to creat the probability ditribution over the vocabulary
    output_logits = tf.add(tf.matmul(output_logits, W), b) 
    output_logits = tf.reshape(output_logits, (-1, self.beam_size, self.vocab_size)) 
    return tf.nn.log_softmax(output_logits)

@nikitakit
Copy link
Author

@wanpifeng4ever Maybe the difference has to do with handling of stop tokens?

Imagine a model that always outputs a 90% probability of CONTINUE token and 10% probability of STOP token. Under greedy decode the output will be an infinite sequence of CONTINUE tokens (though in practice will probably be a hard-enforced length limit), but under beam search the output will just be a single STOP token.

I recommend you find a case where the two behaviors aren't the same for your model and examine in more detail the outputs from greedy and beam decode, their scores, and the scores along the way to the outputs.

@prashantserai
Copy link

@nikitakit hey! does this work with TF 1.0? thanks!

@yangshao
Copy link

@nikitakit I tried this with version 1.1.0. After modifying some api changes, I'm stuck with the error:
alueError: Attempt to reuse RNNCell <tensorflow.contrib.rnn.python.ops.core_rnn_cell_impl.BasicLSTMCell object at 0x7f16c0ca3588> with a different variable scope than its first use. First use of cell was with scope 'seq2seq_att/social_bot/Decoder/decoder/decoder/basic_lstm_cell', this attempt is with scope 'seq2seq_att/social_bot/Decoder/decoder/rnn/basic_lstm_cell'

at this line: output_k, next_state_k = self.cell(inputs_k, state_k,scope=scope)

@yangshao
Copy link

@nikitakit I have some questions about outputs_to_score_fn and the return value.

  1. what's the expect input shape and output shape of this function?
  2. the output should be the log_softmax results instead of the logits before the softmax, right?
  3. if I want to get n best list, what should I do ?
    Thanks!

@yangshao
Copy link

@avostryakov did you solve your problem? I think I have exactly same problem with you. the cand_symbols is just sequence of stop tokens...

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment