def _preprocess_data(self, X, Y=None, idxs=None, train=False):
"""
Preprocess the data:
1. Convert sparse feature matrix to dense matrix for pytorch operation.
2. Make sentence with mention into sequence data for LSTM.
3. Select subset of the input if idxs exists.
:param X: The input data of the model
:param Y: The labels of input data (optional)
:param idxs: The selected indexs of input data
:param train: An indicator for word dictionary to extend new words
"""
C, F = X
# Covert sparse feature matrix to dense matrix
# TODO: the pytorch implementation is taking dense vector as input,
# should optimize later
if issparse(F):
F = F.todense()
# Create word dictionary for LSTM
if not hasattr(self, "word_dict"):
self.word_dict = SymbolTable()
arity = len(C[0])
# Add paddings into word dictionary
for i in range(arity):
# TODO: optimize this
list(
map(self.word_dict.get,
["~~[[" + str(i), str(i) + "]]~~"]))
# Make sequence input for LSTM from candidates
seq_data = []
for candidate in C:
cand_idx = []
for i in range(len(candidate)):
# Add mark for each mention in the original sentence
args = [(
candidate[i].span.get_word_start(),
candidate[i].span.get_word_end(),
i,
)]
s = mark_sentence(mention_to_tokens(candidate[i]), args)
f = self.word_dict.get if train else self.word_dict.lookup
cand_idx.append(list(map(f, s)))
seq_data.append(cand_idx)
# Generate proprcessed the input
if idxs is None:
if Y is not None:
return [(seq_data[i], F[i]) for i in range(len(seq_data))], Y
else:
return [(seq_data[i], F[i]) for i in range(len(seq_data))]
if Y is not None:
return [(seq_data[i], F[i]) for i in idxs], Y[idxs]
else:
return [(seq_data[i], F[i]) for i in idxs]
def _preprocess_data(self, X, Y=None, idxs=None, train=False):
"""
Preprocess the data:
1. Make sentence with mention into sequence data for LSTM.
2. Select subset of the input if idxs exists.
:param X: The input data of the model.
:type X: pair with candidates and corresponding features
:param Y: The labels of input data (optional).
:type Y: list of floats if num_classes = 2
otherwise num_classes-length numpy array
:param idxs: The selected indexs of input data.
:type idxs: list or numpy.array
:param train: An indicator for word dictionary to extend new words.
:type train: bool
:return: Preprocessed data.
:rtype: list of (candidate, features) pairs
"""
C, F = X
# Create word dictionary for LSTM
if not hasattr(self, "word_dict"):
self.word_dict = SymbolTable()
arity = len(C[0])
# Add paddings into word dictionary
for i in range(arity):
# TODO: optimize this
list(
map(self.word_dict.get,
["~~[[" + str(i), str(i) + "]]~~"]))
# Make sequence input for LSTM from candidates
seq_data = []
for candidate in C:
cand_idx = []
for i in range(len(candidate)):
# Add mark for each mention in the original sentence
args = [(
candidate[i].span.get_word_start_index(),
candidate[i].span.get_word_end_index(),
i,
)]
s = mark_sentence(mention_to_tokens(candidate[i]), args)
f = self.word_dict.get if train else self.word_dict.lookup
cand_idx.append(list(map(f, s)))
seq_data.append(cand_idx)
# Generate proprcessed the input
if idxs is None:
if Y is not None:
return (
[(
seq_data[i],
F.indices[F.indptr[i]:F.indptr[i + 1]],
F.data[F.indptr[i]:F.indptr[i + 1]],
) for i in range(len(C))],
Y,
)
else:
return [(
seq_data[i],
F.indices[F.indptr[i]:F.indptr[i + 1]],
F.data[F.indptr[i]:F.indptr[i + 1]],
) for i in range(len(C))]
if Y is not None:
return (
[(
seq_data[i],
F.indices[F.indptr[i]:F.indptr[i + 1]],
F.data[F.indptr[i]:F.indptr[i + 1]],
) for i in idxs],
Y[idxs],
)
else:
return [(
seq_data[i],
F.indices[F.indptr[i]:F.indptr[i + 1]],
F.data[F.indptr[i]:F.indptr[i + 1]],
) for i in idxs]
def _preprocess_data(self, X, Y=None, idxs=None, train=False):
"""
Preprocess the data:
1. Convert sparse feature matrix to dense matrix for pytorch operation.
2. Make sentence with mention into sequence data for LSTM.
3. Select subset of the input if idxs exists.
:param X: The input data of the model.
:type X: pair with candidates and corresponding features
:param Y: The labels of input data (optional).
:type Y: list or numpy.array
:param idxs: The selected indexs of input data.
:type idxs: list or numpy.array
:param train: An indicator for word dictionary to extend new words.
:type train: bool
:return: Preprocessed data.
:rtype: list of (candidate, features) pairs
"""
C, F = X
# Covert sparse feature matrix to dense matrix
if issparse(F):
F = np.array(F.todense(), dtype=np.float32)
if Y is not None:
Y = np.array(Y).astype(np.float32)
# Create word dictionary for LSTM
if not hasattr(self, "word_dict"):
self.word_dict = SymbolTable()
arity = len(C[0])
# Add paddings into word dictionary
for i in range(arity):
list(map(self.word_dict.get, ["~~[[" + str(i), str(i) + "]]~~"]))
# Make sequence input for LSTM from candidates
seq_data = []
for candidate in C:
cand_idx = []
for i in range(len(candidate)):
# Add mark for each mention in the original sentence
args = [
(
candidate[i].context.get_word_start_index(),
candidate[i].context.get_word_end_index(),
i,
)
]
s = mark_sentence(mention_to_tokens(candidate[i]), args)
f = self.word_dict.get if train else self.word_dict.lookup
cand_idx.append(list(map(f, s)))
seq_data.append(cand_idx)
# Generate proprcessed the input
if idxs is None:
if Y is not None:
return [[seq_data[i], F[i]] for i in range(len(seq_data))], Y
else:
return [[seq_data[i], F[i]] for i in range(len(seq_data))]
if Y is not None:
return [[seq_data[i], F[i]] for i in idxs], Y[idxs]
else:
return [[seq_data[i], F[i]] for i in idxs]