def __init__(self,
inputs,
sequence_length,
embedding,
sampling_probability,
time_major=False,
seed=None,
scheduling_seed=None,
name=None):
"""Initializer.
Args:
inputs: A (structure of) input tensors.
sequence_length: An int32 vector tensor.
embedding: A callable or the `params` argument for `embedding_lookup`.
If a callable, it can take a vector tensor of token `ids`,
or take two arguments (`ids`, `times`), where `ids` is a vector
tensor of token ids, and `times` is a vector tensor of current
time steps (i.e., position ids). The latter case can be used when
attr:`embedding` is a combination of word embedding and position
embedding.
sampling_probability: A 0D `float32` tensor: the probability of sampling
categorically from the output ids instead of reading directly from the
inputs.
time_major: Python bool. Whether the tensors in `inputs` are time major.
If `False` (default), they are assumed to be batch major.
seed: The sampling seed.
scheduling_seed: The schedule decision rule sampling seed.
name: Name scope for any created operations.
Raises:
ValueError: if `sampling_probability` is not a scalar or vector.
"""
with ops.name_scope(name, "ScheduledEmbeddingSamplingWrapper",
[embedding, sampling_probability]):
if callable(embedding):
self._embedding_fn = embedding
else:
self._embedding_fn = (
lambda ids: embedding_ops.embedding_lookup(embedding, ids))
self._embedding_args_cnt = len(get_args(self._embedding_fn))
if self._embedding_args_cnt != 1 and self._embedding_args_cnt != 2:
raise ValueError('`embedding` should expect 1 or 2 arguments.')
self._sampling_probability = ops.convert_to_tensor(
sampling_probability, name="sampling_probability")
if self._sampling_probability.get_shape().ndims not in (0, 1):
raise ValueError(
"sampling_probability must be either a scalar or a vector. "
"saw shape: %s" % (self._sampling_probability.get_shape()))
self._seed = seed
self._scheduling_seed = scheduling_seed
super(ScheduledEmbeddingTrainingHelper,
self).__init__(inputs=inputs,
sequence_length=sequence_length,
time_major=time_major,
name=name)
def get_gradient_clip_fn(hparams=None):
"""Creates a gradient clipping function based on the hyperparameters.
See the :attr:`gradient_clip` field in
:meth:`~texar.tf.core.default_optimization_hparams` for all
hyperparameters and default values.
The gradient clipping function takes a list of `(gradients, variables)`
tuples and returns a list of `(clipped_gradients, variables)` tuples.
Typical examples include
:tf_main:`tf.clip_by_global_norm <clip_by_global_norm>`,
:tf_main:`tf.clip_by_value <clip_by_value>`,
:tf_main:`tf.clip_by_norm <clip_by_norm>`,
:tf_main:`tf.clip_by_average_norm <clip_by_average_norm>`, etc.
Args:
hparams (dict or HParams, optional): hyperparameters. Missing
hyperparameters are set to default values automatically.
Returns:
function or `None`: If hparams["type"] is specified, returns
the respective function. If hparams["type"] is empty,
returns `None`.
"""
if hparams is None or isinstance(hparams, dict):
hparams = HParams(
hparams, default_optimization_hparams()["gradient_clip"])
fn_type = hparams["type"]
if fn_type is None or fn_type == "":
return None
fn_modules = ["tensorflow", "texar.tf.custom"]
clip_fn = utils.get_function(fn_type, fn_modules)
clip_fn_args = utils.get_args(clip_fn)
fn_kwargs = hparams["kwargs"]
if isinstance(fn_kwargs, HParams):
fn_kwargs = fn_kwargs.todict()
def grad_clip_fn(grads_and_vars):
"""Gradient clipping function.
Args:
grads_and_vars (list): A list of `(gradients, variables)` tuples.
Returns:
list: A list of `(clipped_gradients, variables)` tuples.
"""
grads, vars_ = zip(*grads_and_vars)
if clip_fn == tf.clip_by_global_norm:
clipped_grads, _ = clip_fn(t_list=grads, **fn_kwargs)
elif 't_list' in clip_fn_args:
clipped_grads = clip_fn(t_list=grads, **fn_kwargs)
elif 't' in clip_fn_args: # e.g., tf.clip_by_value
clipped_grads = [clip_fn(t=grad, **fn_kwargs) for grad in grads]
return list(zip(clipped_grads, vars_))
return grad_clip_fn
def __init__(self,
embedding,
start_tokens,
end_token,
tau,
embedding_size=None,
stop_gradient=False,
use_finish=True):
if callable(embedding):
self._embedding_fn = embedding
if embedding_size is None:
raise ValueError('`embedding_size` must be provided if '
'`embedding` is a callable.')
self._embedding_size = tf.convert_to_tensor(embedding_size,
dtype=tf.int32,
name="embedding_size")
else:
self._embedding_fn = (
lambda soft_ids: soft_embedding_lookup(embedding, soft_ids))
self._embedding_size = tf.shape(embedding)[0]
self._start_tokens = tf.convert_to_tensor(start_tokens,
dtype=tf.int32,
name="start_tokens")
self._end_token = tf.convert_to_tensor(end_token,
dtype=tf.int32,
name="end_token")
if self._start_tokens.get_shape().ndims != 1:
raise ValueError("start_tokens must be a vector")
self._batch_size = array_ops.size(start_tokens)
if self._end_token.get_shape().ndims != 0:
raise ValueError("end_token must be a scalar")
soft_start_tokens = tf.one_hot(self._start_tokens,
self._embedding_size,
dtype=tf.float32)
self._embedding_args_cnt = len(utils.get_args(self._embedding_fn))
if self._embedding_args_cnt == 1:
self._start_inputs = self._embedding_fn(soft_ids=soft_start_tokens)
elif self._embedding_args_cnt == 2:
# Position index is 0 in the beginning
times = tf.zeros([self._batch_size], dtype=tf.int32)
self._start_inputs = self._embedding_fn(soft_ids=soft_start_tokens,
times=times)
else:
raise ValueError('`embedding` should expect 1 or 2 arguments.')
self._batch_size = tf.size(self._start_tokens)
self._tau = tau
self._stop_gradient = stop_gradient
self._use_finish = use_finish
def __init__(self, embedding, start_tokens, end_token):
"""Initializer.
Args:
embedding: A callable or the `params` argument for `embedding_lookup`.
If a callable, it can take a vector tensor of `ids` (argmax ids),
or take two arguments (`ids`, `times`), where `ids` is a vector
tensor of argmax ids, and `times` is a vector tensor of current
time steps (i.e., position ids). The latter case can be used when
attr:`embedding` is a combination of word embedding and position
embedding.
The returned tensor will be returned by :meth:`next_inputs`.
start_tokens: `int32` vector shaped `[batch_size]`, the start tokens.
end_token: `int32` scalar, the token that marks end of decoding.
Raises:
ValueError: if `start_tokens` is not a 1D tensor or `end_token` is not a
scalar.
"""
if callable(embedding):
self._embedding_fn = embedding
else:
self._embedding_fn = (
lambda ids: embedding_ops.embedding_lookup(embedding, ids))
self._start_tokens = ops.convert_to_tensor(start_tokens,
dtype=dtypes.int32,
name="start_tokens")
self._end_token = ops.convert_to_tensor(end_token,
dtype=dtypes.int32,
name="end_token")
if self._start_tokens.get_shape().ndims != 1:
raise ValueError("start_tokens must be a vector")
self._batch_size = shape_list(start_tokens)[0]
if self._end_token.get_shape().ndims != 0:
raise ValueError("end_token must be a scalar")
self._embedding_args_cnt = len(get_args(self._embedding_fn))
if self._embedding_args_cnt == 1:
self._start_inputs = self._embedding_fn(self._start_tokens)
elif self._embedding_args_cnt == 2:
# Position index is 0 in the beginning
times = tf.zeros([self._batch_size], dtype=tf.int32)
self._start_inputs = self._embedding_fn(self._start_tokens, times)
else:
raise ValueError('`embedding` should expect 1 or 2 arguments.')
def _get_opt(learning_rate=None):
opt_kwargs = hparams["kwargs"].todict()
fn_args = set(utils.get_args(opt_class.__init__))
if 'learning_rate' in fn_args and learning_rate is not None:
opt_kwargs["learning_rate"] = learning_rate
return opt_class(**opt_kwargs)
