def __call__(
self, dataset: Dataset,
generated_series: Dict[str, Iterable[Any]]) -> Iterable[List[str]]:
source_series = generated_series.get(
self._source_id, dataset.get_series(self._source_id))
edits_series = generated_series.get(self._edits_id,
dataset.get_series(self._edits_id))
for src_seq, edit_seq in zip(source_series, edits_series):
yield reconstruct(src_seq, edit_seq)
def _do_postprocess(
self, dataset: Dataset,
generated_series: Dict[str, Iterable[Any]]) -> Iterable[List[str]]:
source_series = generated_series.get(self._source_id)
if source_series is None:
source_series = dataset.get_series(self._source_id)
edits_series = generated_series.get(self._edits_id)
if edits_series is None:
edits_series = dataset.get_series(self._edits_id)
for src_seq, edit_seq in zip(source_series, edits_series):
reconstructed = reconstruct(src_seq, edit_seq)
yield reconstructed
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
"""Populate the feed dictionary with the encoder inputs.
Arguments:
dataset: The dataset to use
train: Boolean flag telling whether it is training time
"""
# pylint: disable=invalid-name
fd = {} # type: FeedDict
fd[self.train_mode] = train
series = list(dataset.get_series(self.data_id))
lengths = []
inputs = []
max_len = max(x.shape[0] for x in series)
if self.max_input_len is not None:
max_len = min(self.max_input_len, max_len)
for x in series:
length = min(max_len, x.shape[0])
x_padded = np.zeros(shape=(max_len, ) + x.shape[1:], dtype=x.dtype)
x_padded[:length] = x[:length]
lengths.append(length)
inputs.append(x_padded)
fd[self.inputs] = inputs
fd[self._input_lengths] = lengths
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
"""Feed the placholders with the data.
Arguments:
dataset: The dataset.
train: A flag whether the train mode is enabled.
Returns:
The constructed feed dictionary that contains the factor data and
the mask.
"""
fd = {} # type: FeedDict
# for checking the lengths of individual factors
arr_strings = []
last_paddings = None
for factor_plc, name, vocabulary in zip(
self.input_factors, self.data_ids, self.vocabularies):
factors = dataset.get_series(name)
vectors, paddings = vocabulary.sentences_to_tensor(
list(factors), self.max_length, pad_to_max_len=False,
train_mode=train)
fd[factor_plc] = list(zip(*vectors))
arr_strings.append(paddings.tostring())
last_paddings = paddings
if len(set(arr_strings)) > 1:
raise ValueError("The lenghts of factors do not match")
fd[self.mask] = list(zip(*last_paddings))
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
"""Populate the feed dictionary with the encoder inputs.
Encoder input placeholders:
``encoder_input``: Stores indices to the vocabulary,
shape (batch, time)
``encoder_padding``: Stores the padding (ones and zeros,
indicating valid words and positions after the end
of sentence, shape (batch, time)
``train_mode``: Boolean scalar specifying the mode (train
vs runtime)
Arguments:
dataset: The dataset to use
train: Boolean flag telling whether it is training time
"""
# pylint: disable=invalid-name
fd = {} # type: FeedDict
fd[self.train_mode] = train
sentences = dataset.get_series(self.data_id)
vectors, paddings = self.vocabulary.sentences_to_tensor(
list(sentences),
self.max_input_len,
pad_to_max_len=False,
train_mode=train)
# as sentences_to_tensor returns lists of shape (time, batch),
# we need to transpose
fd[self.inputs] = list(zip(*vectors))
fd[self.input_mask] = list(zip(*paddings))
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = {} # type: FeedDict
sentences = cast(Iterable[List[str]],
dataset.get_series(self.data_id, allow_none=True))
fd[self.train_mode] = train
if sentences is not None:
vectors, paddings = self.vocabulary.sentences_to_tensor(
list(sentences), train_mode=train)
# sentences_to_tensor returns time-major tensors, targets need to
# be batch-major
vectors = vectors.T
paddings = paddings.T
# Need to convert the data to a sparse representation
bool_mask = (paddings > 0.5)
indices = np.stack(np.where(bool_mask), axis=1)
values = vectors[bool_mask]
fd[self.train_targets] = tf.SparseTensorValue(
indices=indices, values=values, dense_shape=vectors.shape)
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = {} # type: FeedDict
fd[self.train_mode] = train
# for checking the lengths of individual factors
arr_strings = []
last_paddings = None
for name, vocabulary in zip(self.data_ids, self.vocabularies):
factors = dataset.get_series(name)
vectors, paddings = vocabulary.sentences_to_tensor(
list(factors), self.max_input_len, pad_to_max_len=False,
train_mode=train)
# pylint: disable=unsubscriptable-object
fd[self.input_factors[name]] = list(zip(*vectors))
# pylint: enable=unsubscriptable-object
arr_strings.append(paddings.tostring())
last_paddings = paddings
if len(set(arr_strings)) > 1:
raise ValueError("The lenghts of factors do not match")
fd[self.input_mask] = list(zip(*last_paddings))
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
"""Populate the feed dictionary for the decoder object.
Arguments:
dataset: The dataset to use for the decoder.
train: Boolean flag, telling whether this is a training run.
"""
sentences = cast(Iterable[List[str]],
dataset.get_series(self.data_id, allow_none=True))
if sentences is None and train:
raise ValueError("When training, you must feed "
"reference sentences")
sentences_list = list(sentences) if sentences is not None else None
fd = {} # type: FeedDict
fd[self.train_mode] = train
go_symbol_idx = self.vocabulary.get_word_index(START_TOKEN)
fd[self.go_symbols] = np.full([len(dataset)], go_symbol_idx,
dtype=np.int32)
if sentences is not None:
# train_mode=False, since we don't want to <unk>ize target words!
inputs, weights = self.vocabulary.sentences_to_tensor(
sentences_list, self.max_output_len, train_mode=False,
add_start_symbol=False, add_end_symbol=True,
pad_to_max_len=False)
fd[self.train_inputs] = inputs
fd[self.train_mask] = weights
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = ModelPart.feed_dict(self, dataset, train)
# if it is from the pickled file, it is a list, not a numpy tensor,
# so convert it as as a prevention
images = np.array(list(dataset.get_series(self.data_id)))
fd[self.image_input] = images / 255.0
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = ModelPart.feed_dict(self, dataset, train)
images = np.array(dataset.get_series(self.data_id))
assert images.shape[1:] == (self.height, self.width, 3)
fd[self.input_image] = images
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = ModelPart.feed_dict(self, dataset, train)
# if it is from the pickled file, it is a list, not a numpy tensor,
# so convert it as as a prevention
images = np.array(list(dataset.get_series(self.data_id)))
fd[self.image_input] = images / 255.0
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
# if it is from the pickled file, it is list, not numpy tensor,
# so convert it as as a prevention
images = np.array(dataset.get_series(self.data_id))
f_dict = {}
f_dict[self.image_input] = images / 225.0
f_dict[self.train_mode] = train
return f_dict
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
res = {} # type: FeedDict
res[self.image_features] = dataset.get_series(self.data_id)
if train:
res[self.dropout_placeholder] = self.dropout_keep_prob
else:
res[self.dropout_placeholder] = 1.0
return res
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
"""Populate the feed dictionary with the encoder inputs.
Arguments:
dataset: The dataset to use
train: Boolean flag telling whether it is training time
"""
fd = ModelPart.feed_dict(self, dataset, train)
sentences = dataset.get_series(self.data_id)
fd[self.input_tokens] = pad_batch(list(sentences), self.max_input_len)
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
sentences = cast(Iterable[List[str]],
dataset.get_series(self.data_id, allow_none=True))
sentences_list = list(sentences) if sentences is not None else None
fd = {} # type: FeedDict
if sentences_list is not None:
fd[self.train_inputs] = list(zip(*sentences_list))[0]
fd[self.train_mode] = train
return fd
def _print_examples(dataset: Dataset,
outputs: Dict[str, List[Any]],
num_examples=15) -> None:
"""Print examples of the model output."""
log_print(colored("Examples:", attrs=['bold']))
# for further indexing we need to make sure, all relevant
# dataset series are lists
target_series = {
series_id: list(dataset.get_series(series_id))
for series_id in outputs.keys() if dataset.has_series(series_id)
}
source_series = {
series_id: list(dataset.get_series(series_id))
for series_id in dataset.series_ids if series_id not in outputs
}
for i in range(min(len(dataset), num_examples)):
log_print(
colored(" [{}]".format(i + 1), color='magenta', attrs=['bold']))
def print_line(prefix, color, content):
colored_prefix = colored(prefix, color=color)
formated = _data_item_to_str(content)
log_print(" {}: {}".format(colored_prefix, formated))
for series_id, data in sorted(source_series.items(),
key=lambda x: x[0]):
print_line(series_id, 'yellow', data[i])
for series_id, data in sorted(outputs.items(), key=lambda x: x[0]):
model_output = data[i]
print_line(series_id, 'magenta', model_output)
if series_id in target_series:
desired_output = target_series[series_id][i]
print_line(series_id + " (ref)", "red", desired_output)
log_print("")
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = {}
alignment = dataset.get_series(self.data_id, allow_none=True)
if alignment is None:
if train:
log("Warning: training alignment not present", color="red")
alignment = np.zeros((len(dataset), self.decoder.max_output_len,
self.encoder.max_input_len), np.float32)
fd[self.ref_alignment] = alignment
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = ModelPart.feed_dict(self, dataset, train)
# if it is from the pickled file, it is a list, not a numpy tensor,
# so convert it as as a prevention
images = np.array(list(dataset.get_series(self.data_id)))
fd[self.image_input] = images / 255.0
# the image mask is one everywhere where the image is non-zero, i.e.
# zero pixels are masked out
fd[self.image_mask] = np.sign(np.sum(images, axis=3, keepdims=True))
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = {}
alignment = dataset.get_series(self.data_id, allow_none=True)
if alignment is None:
if train:
warn("Training alignment not present!")
alignment = np.zeros((len(dataset), self.decoder.max_output_len,
self.encoder.input_sequence.max_length),
np.float32)
fd[self.ref_alignment] = alignment
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = {} # type: FeedDict
sentences = cast(Iterable[List[str]],
dataset.get_series(self.data_id, allow_none=True))
fd[self.train_mode] = train
if sentences is not None:
vectors, paddings = self.vocabulary.sentences_to_tensor(
list(sentences), pad_to_max_len=False, train_mode=train)
fd[self.train_targets] = vectors.T
fd[self.train_weights] = paddings.T
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
# if it is from the pickled file, it is list, not numpy tensor,
# so convert it as as a prevention
images = np.array(dataset.get_series(self.data_id))
f_dict = {}
f_dict[self.input_op] = images / 225.0
# it is one everywhere where non-zero, i.e. zero columns are masked out
f_dict[self.padding_masks] = \
np.sum(np.sign(images), axis=3, keepdims=True)
if train:
f_dict[self.dropout_placeholder] = self.dropout_keep_prob
else:
f_dict[self.dropout_placeholder] = 1.0
f_dict[self.is_training] = train
return f_dict
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
sentences = cast(Iterable[List[str]],
dataset.get_series(self.data_id, allow_none=True))
sentences_list = list(sentences) if sentences is not None else None
fd = {} # type: FeedDict
label_tensors, _ = self.vocabulary.sentences_to_tensor(
sentences_list, self.max_output_len)
fd[self.gt_inputs[0]] = label_tensors[0]
if train:
fd[self.dropout_placeholder] = self.dropout_keep_prob
else:
fd[self.dropout_placeholder] = 1.0
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
sentences = cast(Iterable[List[str]],
dataset.get_series(self.data_id, allow_none=True))
sentences_list = list(sentences) if sentences is not None else None
fd = {} # type: FeedDict
if sentences is not None:
label_tensors, _ = self.vocabulary.sentences_to_tensor(
sentences_list, self.max_output_len)
# pylint: disable=unsubscriptable-object
fd[self.gt_inputs[0]] = label_tensors[0]
# pylint: enable=unsubscriptable-object
fd[self.train_mode] = train
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
"""Populate the feed dictionary with the encoder inputs.
Arguments:
dataset: The dataset to use
train: Boolean flag telling whether it is training time
"""
fd = ModelPart.feed_dict(self, dataset, train)
sentences = dataset.get_series(self.data_id)
vectors, _ = self.vocabulary.sentences_to_tensor(
list(sentences), self.max_input_len, pad_to_max_len=False,
train_mode=train)
# as sentences_to_tensor returns lists of shape (time, batch),
# we need to transpose
fd[self.inputs] = list(zip(*vectors))
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
"""Feed the placholders with the data.
Arguments:
dataset: The dataset.
train: A flag whether the train mode is enabled.
Returns:
The constructed feed dictionary that contains the factor data and
the mask.
"""
fd = ModelPart.feed_dict(self, dataset, train)
# for checking the lengths of individual factors
for factor_plc, name in zip(self.input_factors, self.data_ids):
sentences = dataset.get_series(name)
fd[factor_plc] = pad_batch(
list(sentences), self.max_length, self.add_start_symbol,
self.add_end_symbol)
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
"""Feed the placholders with the data.
Arguments:
dataset: The dataset.
train: A flag whether the train mode is enabled.
Returns:
The constructed feed dictionary that contains the factor data and
the mask.
"""
fd = ModelPart.feed_dict(self, dataset, train)
# for checking the lengths of individual factors
for factor_plc, name in zip(self.input_factors, self.data_ids):
sentences = dataset.get_series(name)
fd[factor_plc] = pad_batch(list(sentences), self.max_length,
self.add_start_symbol,
self.add_end_symbol)
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = ModelPart.feed_dict(self, dataset, train)
series = list(dataset.get_series(self.data_id))
lengths = []
inputs = []
max_len = max(x.shape[0] for x in series)
if self.max_input_len is not None:
max_len = min(self.max_input_len, max_len)
for x in series:
length = min(max_len, x.shape[0])
x_padded = np.zeros(shape=(max_len, ) + x.shape[1:], dtype=x.dtype)
x_padded[:length] = x[:length]
lengths.append(length)
inputs.append(x_padded)
fd[self.temporal_states] = inputs
fd[self._input_lengths] = lengths
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = {} # type: FeedDict
sentences = cast(Iterable[List[str]],
dataset.get_series(self.data_id, allow_none=True))
if sentences is not None:
sentences_list = list(sentences) if sentences is not None else None
inputs, weights = self.vocabulary.sentences_to_tensor(
sentences_list, self.max_output_len)
assert len(weights) == len(self.train_weights)
assert len(inputs) == len(self.train_targets)
for placeholder, weight in zip(self.train_weights, weights):
fd[placeholder] = weight
for placeholder, tensor in zip(self.train_targets, inputs):
fd[placeholder] = tensor
if not train:
fd[self.dropout_placeholder] = 1.0
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = ModelPart.feed_dict(self, dataset, train)
series = list(dataset.get_series(self.data_id))
lengths = []
inputs = []
max_len = max(x.shape[0] for x in series)
if self.max_input_len is not None:
max_len = min(self.max_input_len, max_len)
for x in series:
length = min(max_len, x.shape[0])
x_padded = np.zeros(shape=(max_len,) + x.shape[1:],
dtype=x.dtype)
x_padded[:length] = x[:length]
lengths.append(length)
inputs.append(x_padded)
fd[self.temporal_states] = inputs
fd[self._input_lengths] = lengths
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = ModelPart.feed_dict(self, dataset, train)
fd[self.vector] = dataset.get_series(self.data_id)
return fd
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
res = {} # type: FeedDict
res[self.image_features] = dataset.get_series(self.data_id)
return res
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
return {self.vector: dataset.get_series(self.data_id)}
def __call__(self, dataset: Dataset) -> Iterable[List[str]]:
source_series = dataset.get_series(self._source_id)
target_series = dataset.get_series(self._target_id)
for src_seq, tgt_seq in zip(source_series, target_series):
yield convert_to_edits(src_seq, tgt_seq)
def _print_examples(dataset: Dataset,
outputs: Dict[str, List[Any]],
val_preview_input_series: Optional[List[str]] = None,
val_preview_output_series: Optional[List[str]] = None,
num_examples=15) -> None:
"""Print examples of the model output.
Arguments:
dataset: The dataset from which to take examples
outputs: A mapping from the output series ID to the list of its
contents
val_preview_input_series: An optional list of input series to include
in the preview. An input series is a data series that is present in
the dataset. It can be either a target series (one that is also
present in the outputs, i.e. reference), or a source series (one
that is not among the outputs). In the validation preview, source
input series and preprocessed target series are yellow and target
(reference) series are red. If None, all series are written.
val_preview_output_series: An optional list of output series to include
in the preview. An output series is a data series that is present
among the outputs. In the preview, magenta is used as the font
color for output series
"""
log_print(colored("Examples:", attrs=["bold"]))
source_series_names = [s for s in dataset.series_ids if s not in outputs]
target_series_names = [s for s in dataset.series_ids if s in outputs]
output_series_names = list(outputs.keys())
assert outputs
if val_preview_input_series is not None:
target_series_names = [
s for s in target_series_names if s in val_preview_input_series
]
source_series_names = [
s for s in source_series_names if s in val_preview_input_series
]
if val_preview_output_series is not None:
output_series_names = [
s for s in output_series_names if s in val_preview_output_series
]
# for further indexing we need to make sure, all relevant
# dataset series are lists
target_series = {
series_id: list(dataset.get_series(series_id))
for series_id in target_series_names
}
source_series = {
series_id: list(dataset.get_series(series_id))
for series_id in source_series_names
}
if not isinstance(dataset, LazyDataset):
num_examples = min(len(dataset), num_examples)
for i in range(num_examples):
log_print(
colored(" [{}]".format(i + 1), color="magenta", attrs=["bold"]))
def print_line(prefix, color, content):
colored_prefix = colored(prefix, color=color)
formatted = _data_item_to_str(content)
log_print(" {}: {}".format(colored_prefix, formatted))
# Input source series = yellow
for series_id, data in sorted(source_series.items(),
key=lambda x: x[0]):
print_line(series_id, "yellow", data[i])
# Output series = magenta
for series_id in sorted(output_series_names):
data = list(outputs[series_id])
model_output = data[i]
print_line(series_id, "magenta", model_output)
# Input target series (a.k.a. references) = red
for series_id in sorted(target_series_names):
data = outputs[series_id]
desired_output = target_series[series_id][i]
print_line(series_id + " (ref)", "red", desired_output)
log_print("")
def feed_dict(self, dataset: Dataset, train: bool = False) -> FeedDict:
fd = ModelPart.feed_dict(self, dataset, train)
fd[self.spatial_input] = list(dataset.get_series(self.data_id))
return fd