def check_value_check(self):
if self.valid:
# Check if it throws nothing
functions.vstack(self.xs)
else:
with pytest.raises(type_check.InvalidType):
functions.vstack(self.xs)
def __call__(self, xs, ilens):
"""BRNN forward propagation.
Args:
xs (chainer.Variable): Batch of padded charactor ids. (B, Tmax)
ilens (chainer.Variable): Batch of length of each input batch. (B,)
Returns:
tuple(chainer.Variable): Tuple of `chainer.Variable` objects.
chainer.Variable: `ilens` .
"""
logging.info(self.__class__.__name__ + " input lengths: " + str(ilens))
# need to move ilens to cpu
ilens = cuda.to_cpu(ilens)
if "lstm" in self.typ:
_, _, ys = self.nbrnn(None, None, xs)
else:
_, ys = self.nbrnn(None, xs)
ys = self.l_last(F.vstack(ys)) # (sum _utt frame_utt) x dim
xs = F.split_axis(ys, np.cumsum(ilens[:-1]), axis=0)
# final tanh operation
xs = F.split_axis(F.tanh(F.vstack(xs)), np.cumsum(ilens[:-1]), axis=0)
# 1 utterance case, it becomes an array, so need to make a utt tuple
if not isinstance(xs, tuple):
xs = [xs]
return xs, ilens # x: utt list of frame x dim
def __call__(self, xs, ilens):
'''BLSTMP forward
:param xs:
:param ilens:
:return:
'''
logging.info(self.__class__.__name__ + ' input lengths: ' + str(ilens))
for layer in six.moves.range(self.elayers):
hy, cy, ys = self['bilstm' + str(layer)](None, None, xs)
# ys: utt list of frame x cdim x 2 (2: means bidirectional)
# TODO(watanabe) replace subsample and FC layer with CNN
ys, ilens = _subsamplex(ys, self.subsample[layer + 1])
# (sum _utt frame_utt) x dim
ys = self['bt' + str(layer)](F.vstack(ys))
xs = F.split_axis(ys, _ilens_to_index(ilens), axis=0)
del hy, cy
# final tanh operation
xs = F.split_axis(F.tanh(F.vstack(xs)), _ilens_to_index(ilens), axis=0)
# 1 utterance case, it becomes an array, so need to make a utt tuple
if not isinstance(xs, tuple):
xs = [xs]
return xs, ilens # x: utt list of frame x dim
def __call__(self, xs, ilens):
"""RNNP forward.
Args:
xs (chainer.Variable): Batch of padded charactor ids. (B, Tmax)
ilens (chainer.Variable): Batch of length of each input batch. (B,)
Returns:
xs (chainer.Variable):subsampled vector of xs.
chainer.Variable: Subsampled vector of ilens.
"""
logging.info(self.__class__.__name__ + " input lengths: " + str(ilens))
for layer in six.moves.range(self.elayers):
if "lstm" in self.typ:
_, _, ys = self[self.rnn_label + str(layer)](None, None, xs)
else:
_, ys = self[self.rnn_label + str(layer)](None, xs)
# ys: utt list of frame x cdim x 2 (2: means bidirectional)
# TODO(watanabe) replace subsample and FC layer with CNN
ys, ilens = _subsamplex(ys, self.subsample[layer + 1])
# (sum _utt frame_utt) x dim
ys = self["bt" + str(layer)](F.vstack(ys))
xs = F.split_axis(ys, np.cumsum(ilens[:-1]), axis=0)
# final tanh operation
xs = F.split_axis(F.tanh(F.vstack(xs)), np.cumsum(ilens[:-1]), axis=0)
# 1 utterance case, it becomes an array, so need to make a utt tuple
if not isinstance(xs, tuple):
xs = [xs]
return xs, ilens # x: utt list of frame x dim
def __call__(self, xs, ilens):
"""BRNN forward
:param xs:
:param ilens:
:return:
"""
logging.info(self.__class__.__name__ + ' input lengths: ' + str(ilens))
# need to move ilens to cpu
ilens = cuda.to_cpu(ilens)
if "lstm" in self.typ:
_, _, ys = self.nbrnn(None, None, xs)
else:
_, ys = self.nbrnn(None, xs)
ys = self.l_last(F.vstack(ys)) # (sum _utt frame_utt) x dim
xs = F.split_axis(ys, np.cumsum(ilens[:-1]), axis=0)
# final tanh operation
xs = F.split_axis(F.tanh(F.vstack(xs)), np.cumsum(ilens[:-1]), axis=0)
# 1 utterance case, it becomes an array, so need to make a utt tuple
if not isinstance(xs, tuple):
xs = [xs]
return xs, ilens # x: utt list of frame x dim
def hierarchical_encode(self, ids, xs, xs_embed, position_info, x_spans,
shell_spans, x_position_info):
_, _, ys_l = self.Bilstm(None, None, xs_embed)
if self.lstm_ac:
ac_reps = self.get_span_reps(x_spans, ys_l, split_axis=True)
_, _, ys_l_ac = self.AcBilstm(None, None, ac_reps)
ac_reps = chaFunc.vstack(ys_l_ac)
else:
ac_reps = self.get_span_reps(x_spans, ys_l)
if self.lstm_shell:
shell_reps = self.get_span_reps(shell_spans, ys_l, split_axis=True)
_, _, ys_l_shell = self.ShellBilstm(None, None, shell_reps)
shell_reps = chaFunc.vstack(ys_l_shell)
else:
shell_reps = self.get_span_reps(shell_spans, ys_l)
ac_shell_reps = chaFunc.concat([ac_reps, shell_reps], axis=-1)
span_reps_bow = self.get_bow_reps(ids, xs, xs_embed, position_info,
x_spans, shell_spans,
x_position_info)
ac_shell_reps = chaFunc.concat(
[ac_shell_reps, position_info, span_reps_bow], -1)
assert ac_shell_reps.shape[-1] == self.ac_shell_rep_size_in
return ac_shell_reps
def entropy_filter(self, x, b, ent_T):
xp = cuda.get_array_module(b)
#print "Entropy input pre_data : %s " % b.data
#a = F.softmax(b)
#print "a[0] : %s" % (a[0].data)
#print 'Entropy input : %s' % (a.data)
eb = entropy(F.softmax(b)) / np.log(b.shape[1])
print "Entropy result : %s " % (eb.data)
eb.to_cpu()
if hasattr(eb.data, 'get'):
with cuda.get_device(eb.data):
exited = eb.data < ent_T
exited = exited.get()
else:
exited = eb.data < ent_T
print "exited : %s" % exited
y_exit = []
y_cont = []
for i, idx in enumerate(exited):
if idx:
y_exit.append(b[i:i + 1])
else:
y_cont.append(x[i:i + 1])
if len(y_exit) > 0:
y_exit = F.vstack(y_exit)
if len(y_cont) > 0:
y_cont = F.vstack(y_cont)
#print "y_exit = %s \n y_cont = %s" % (y_exit.data,y_cont.data.shape)
return y_exit, y_cont, exited
def __call__(self, xs, ilens):
"""RNNP forward
:param xs:
:param ilens:
:return:
"""
logging.info(self.__class__.__name__ + ' input lengths: ' + str(ilens))
for layer in six.moves.range(self.elayers):
if "lstm" in self.typ:
_, _, ys = self[self.rnn_label + str(layer)](None, None, xs)
else:
_, ys = self[self.rnn_label + str(layer)](None, xs)
# ys: utt list of frame x cdim x 2 (2: means bidirectional)
# TODO(watanabe) replace subsample and FC layer with CNN
ys, ilens = _subsamplex(ys, self.subsample[layer + 1])
# (sum _utt frame_utt) x dim
ys = self['bt' + str(layer)](F.vstack(ys))
xs = F.split_axis(ys, np.cumsum(ilens[:-1]), axis=0)
# final tanh operation
xs = F.split_axis(F.tanh(F.vstack(xs)), np.cumsum(ilens[:-1]), axis=0)
# 1 utterance case, it becomes an array, so need to make a utt tuple
if not isinstance(xs, tuple):
xs = [xs]
return xs, ilens # x: utt list of frame x dim
def _asign_gt_to_anchor(self, anchors, locs, confs, gt_bboxes, gt_labels):
_anchors, _locs, _confs = [], [], []
_gt_labels, _gt_bboxes = [], []
for anchor, loc, conf, gt_bbox, gt_label in zip(
anchors, locs, confs, gt_bboxes, gt_labels):
if gt_label.shape[0] > 0:
iou = bbox_iou(anchor, gt_bbox)
max_iou = self.xp.max(iou, axis=-1)
max_iou_indices = self.xp.argmax(iou, axis=-1)
else: # guard no annotation
max_iou = self.xp.zeros(conf.shape[0], self.xp.float32)
max_iou_indices = self.xp.empty(conf.shape[0], self.xp.float32)
fg_mask = max_iou > self._fg_thresh
bg_mask = max_iou < self._bg_thresh
n_bg = self.xp.where(bg_mask)[0].shape[0]
max_iou_indices_fg = max_iou_indices[fg_mask]
_gt_label_fg = self.xp.array(
[gt_label[i] + 1 for i in max_iou_indices_fg], self.xp.int32)
_gt_bbox_fg = self.xp.array(
[gt_bbox[i] for i in max_iou_indices_fg], self.xp.float32)
if _gt_bbox_fg.shape[0] == 0: # guard not fg anchor
_gt_bbox_fg = self.xp.empty((0, 4), self.xp.float32)
_anchors.append(F.vstack((anchor[fg_mask], anchor[bg_mask])))
_locs.append(F.vstack((loc[fg_mask], loc[bg_mask])))
_confs.append(F.vstack((conf[fg_mask], conf[bg_mask])))
_gt_bboxes.append(
self.xp.vstack((_gt_bbox_fg, self.xp.zeros((n_bg, 4)))))
_gt_labels.append(
self.xp.hstack((_gt_label_fg, self.xp.zeros(n_bg))))
return _anchors, _locs, _confs, _gt_bboxes, _gt_labels
def check_value_check(self):
if self.valid:
# Check if it throws nothing
functions.vstack(self.xs)
else:
with self.assertRaises(type_check.InvalidType):
functions.vstack(self.xs)
def _forward_spans(hs_flatten,
pairs,
ckeys,
lengths,
use_block=True,
block_size=128):
xp = chainer.cuda.get_array_module(hs_flatten)
begins, ends = pairs.T
if use_block:
def _uniq(start, end):
idxs = defaultdict(lambda: len(idxs))
offset = np.array([idxs[(s, e)] for s, e in zip(start, end)])
start, end = np.array(
[k for k, v in sorted(idxs.items(), key=lambda x: x[1])]).T
return start, end, offset
def _sum(start, end):
size = len(start)
lb, ub = min(start), max(end)
hs = hs_flatten[lb:ub]
mask = xp.zeros((size, ub - lb, 1), dtype=xp.float32)
for i, (s, e) in enumerate(zip(start, end)):
mask[i, s - lb:e - lb] = 1.0
return F.sum(hs * mask, axis=1)
def _extract(start, end):
spans = []
start, end, offset = _uniq(start, end)
ofs, lb, ub = 0, 0, 0
for k in range(len(start)):
lb, ub = min(lb, start[k]), max(ub, end[k])
if ub - lb > block_size and k > 0:
spans.append(_sum(start[ofs:k], end[ofs:k]))
ofs, lb, ub = k, start[k], end[k]
spans.append(_sum(start[ofs:], end[ofs:]))
spans = F.vstack(spans) / xp.asarray(end - start)[:, None]
return F.embed_id(xp.asarray(offset), spans)
left_spans = _extract(begins, ckeys)
right_spans = _extract(ckeys + 1, ends + 1)
else:
@functools.lru_cache(maxsize=None)
def _get_span_v(i, j):
return F.average(hs_flatten[i:j + 1], axis=0)
left_spans = F.vstack([
_get_span_v(begin, ckey_pre)
for begin, ckey_pre in zip(begins, ckeys - 1)
])
right_spans = F.vstack([
_get_span_v(ckey_post, end)
for ckey_post, end in zip(ckeys + 1, ends)
])
return left_spans, right_spans
def calcLoss(G, X, S):
GXr = G*xp.real(X).astype(np.float32)
GXi = G*xp.imag(X).astype(np.float32)
Sr = xp.real(S).astype(np.float32)
Si = xp.imag(S).astype(np.float32)
gxL = [F.expand_dims(iDGTcf.iDGT(GXr[ii],GXi[ii],windowDG,shiftLenG,fftLenG), axis=0) for ii in range(len(G))]
sL = [F.expand_dims(iDGTcf.iDGT(Sr[ii],Si[ii],windowDG,shiftLenG,fftLenG), axis=0) for ii in range(len(G))]
gx = F.vstack(gxL)
s = F.vstack(sL)
loss = F.mean_absolute_error(gx,s)
return loss
def forward(self, *xs, **kwxs):
if kwxs:
h = F.vstack(list(kwxs.values()))
elif len(xs) > 1:
h = F.vstack(xs)
else:
h = xs[0]
h2 = self.l1(h)
h3 = F.relu(h2)
h4 = self.l2(h3)
return F.relu(h4)
def merge_representation(self, index, section, xs, ys):
"""
Merge and average the context representation to prepare the input
for next layer. If the prediction is 'O', its corresponding row of
context representation of xs will be used as the input for next
layer, otherwise its corresponding row of ys will be seleted as the
input for next layer.
+ index: merge index for predicts
+ xs: context representation, input of bi_word_tag BiLSTM layer
+ ys: context representation, output of bi_word_tag BiLSTM layer
e.g. predicts: B-Gene, I-Gene, O,B-protein,B-DNA
index array:
[ 1, -1, -1, -1
1, -1, -1, -1
-1, 0, -1, -1
-1, -1, 1, -1
-1, -1, -1, 1 ]
ys_index clip array:
[ 1, 0, 0, 0
1, 0, 0, 0
0, 1, 0, 0
0, 0, 1, 0
0, 0, 0, 1 ]
xs index(1-|index|) array:
[ 0, 0, 0, 0
0, 0, 0, 0
0, 1, 0, 0
0, 0, 0, 0
0, 0, 0, 0 ]
"""
ys_index = index.copy()
ys_index = F.clip(ys_index.astype('f'), 0., 1.0)
ys = F.matmul(ys_index, F.vstack(ys), transa=True)
xs_index = index.copy()
xs_index = 1 - Fmat.absolute(xs_index.astype('f'))
xs = F.matmul(xs_index, F.vstack(xs), transa=True)
# Sum word vectors
ys = Fmat.add(xs, ys)
# Average word vectors for entity representation
sum_index = F.sum(ys_index, axis=0)
sum_index = F.clip(sum_index, 1.0, 1000000.0)
sum_index = F.tile(sum_index, (ys.shape[1], 1))
sum_index = F.transpose(sum_index)
ys = Fmat.div(ys, sum_index)
ys = F.split_axis(ys, section, axis=0)
return ys
def __call__(self, xs, hx, cx, xxs, train=True):
forward_char_embeds = [[self.char_embed(item) for item in items]
for items in xxs]
backward_char_embeds = [[item[::-1] for item in items]
for items in forward_char_embeds]
# Encode character sequences
forward_encodings = []
backward_encodings = []
for forward, backward in zip(forward_char_embeds,
backward_char_embeds):
hhx = chainer.Variable(
self.xp.zeros((1, len(forward), 25), dtype=self.xp.float32))
ccx = chainer.Variable(
self.xp.zeros((1, len(forward), 25), dtype=self.xp.float32))
_, __, forward_char_encs = self.forward_char(hhx, ccx, forward)
_, __, backward_char_encs = self.backward_char(hhx, ccx, backward)
forward_encodings.append([x[-1] for x in forward_char_encs])
backward_encodings.append([x[-1] for x in backward_char_encs])
forward_encodings = [F.vstack(x) for x in forward_encodings]
backward_encodings = [F.vstack(x) for x in backward_encodings]
# Encode word embeddings
xs = [self.embed(item) for item in xs]
xs_forward = [
F.concat([x, y, z], axis=1)
for x, y, z in zip(xs, forward_encodings, backward_encodings)
]
xs_backward = [x[::-1] for x in xs_forward]
if self.dropout and train:
xs_forward = [F.dropout(item) for item in xs_forward]
xs_backward = [F.dropout(item) for item in xs_backward]
forward_hy, forward_cy, forward_ys = self.forward_l1(hx,
cx,
xs_forward,
train=train)
backward_hy, backward_cy, backward_ys = self.backward_l1(hx,
cx,
xs_backward,
train=train)
ys = [
F.concat([forward, backward[::-1]], axis=1)
for forward, backward in zip(forward_ys, backward_ys)
]
y = [self.l2(item) for item in ys]
return y
def translate(self, hxs, max_length):
"""Generate target sentences given hidden states of source sentences.
Args:
hxs: Hidden states for source sequences.
Returns:
ys: Generated sequences.
"""
batch_size, _, _ = hxs.shape
compute_context = self.attention(hxs)
c = Variable(self.xp.zeros((batch_size, self.n_units), 'f'))
h = F.broadcast_to(self.bos_state, ((batch_size, self.n_units)))
# first character's embedding
previous_embedding = self.embed_y(
Variable(self.xp.full((batch_size, ), EOS, 'i')))
results = []
for _ in range(max_length):
context = compute_context(h)
concatenated = F.concat((previous_embedding, context))
c, h = self.lstm(c, h, concatenated)
concatenated = F.concat((concatenated, h))
logit = self.w(self.maxout(concatenated))
y = F.reshape(F.argmax(logit, axis=1), (batch_size, ))
results.append(y)
previous_embedding = self.embed_y(y)
results = F.separate(F.transpose(F.vstack(results)), axis=0)
ys = [get_subsequence_before_eos(result.data) for result in results]
return ys
def forward(self, xs, ilens):
'''BLSTM forward
:param xs:
:param ilens:
:return:
'''
logging.info(self.__class__.__name__ + ' input lengths: ' + str(ilens))
# need to move ilens to cpu
ilens = cuda.to_cpu(ilens)
hy, cy, ys = self.nblstm(None, None, xs)
ys = self.l_last(F.vstack(ys)) # (sum _utt frame_utt) x dim
# (satos) xs = F.split_axis(ys, np.cumsum(ilens[:-1]), axis=0)
# (satos) del hy, cy
# final tanh operation
# (satos) xs = F.split_axis(F.tanh(F.vstack(xs)), np.cumsum(ilens[:-1]), axis=0)
# ニュアンス的にtanhがほぼidの挙動しかしないぽいな
xs = F.split_axis(F.tanh(ys), np.cumsum(ilens[:-1]), axis=0) # (satos)
# 1 utterance case, it becomes an array, so need to make a utt tuple
# print(isinstance(xs,tuple),xs)
# (satos) たってき無視する(実際tupleなので大丈夫そう)
# if not isinstance(xs, tuple):
# xs = [xs]
# print(xs[2])
return xs, ilens # x: utt list of frame x dim
def merge_representation(self, index, section, ys):
"""
Merge and average the context representation to prepare the input
for next layer. If the prediction is 'O', its corresponding row of
context representation of xs will be used as the input for next
layer, otherwise its corresponding row of ys will be seleted as the
input for next layer.
+ index: merge index for predicts
+ ys: context representation, output of bi_word_tag BiLSTM layer
e.g. predicts: B-Gene, I-Gene, O,B-protein,B-DNA
index array:
[ 1, 0, 0, 0
1, 0, 0, 0
0, 1, 0, 0
0, 0, 1, 0
0, 0, 0, 1 ]
"""
ys = F.matmul(index.astype('f'), F.vstack(ys), transa=True)
# Average word vectors for entity representation
sum_index = F.sum(index.astype('f'), axis=0)
sum_index = F.tile(sum_index, (ys.shape[1], 1))
sum_index = F.transpose(sum_index)
ys = Fmat.div(ys, sum_index)
ys = F.split_axis(ys, section, axis=0)
return ys
def fit(model, data_x, data_y, init_len):
'''
data_x: input with the shape of (samples, channel)
data_y: teacher signal with the shape of (samples, channel)
'''
train_len = len(data_x)
X = xp.zeros(
(1 + model.n_inputs + model.n_reservoir, train_len - init_len),
dtype=xp.float32)
Yt = data_y[init_len:train_len + 1]
# forward computation
for i in range(train_len):
u = data_x[i]
u = F.reshape(u, (-1, 1))
r = model.update(u)
if i >= init_len:
# collect data after initialization
'''
s = F.vstack(
(xp.array([[0]], dtype=xp.float32), u, r
))[:, 0].data
'''
X[:, i - init_len] = F.vstack(
(xp.array([[0]], dtype=xp.float32), u, r))[:, 0].data
# linear regression: X and Yt
model.Wo = xp.dot(Yt[:, 0], scipy.linalg.pinv(X))
def __call__(self, xs, ts):
hy, cy, ys = self.lstm(None, None, xs)
ys = F.relu(self.ll1(F.vstack(ys)))
del hy, cy
ys = self.ll2(ys)
loss = F.mean_squared_error(ys, ts)
return loss
def crop_and_resize(bottom_data, bottom_rois, outh, outw, spatial_scale):
if isinstance(bottom_rois, chainer.Variable):
bottom_rois = bottom_rois.data
bottom_rois = cuda.to_cpu(bottom_rois)
B, C, H, W = bottom_data.shape
N = bottom_rois.shape[0]
ys = collections.defaultdict(list)
for i_roi in six.moves.range(N):
i_batch, x1, y1, x2, y2 = bottom_rois[i_roi].tolist()
i_batch = int(i_batch)
x1 = int(spatial_scale * x1)
x2 = max(int(spatial_scale * x2), x1 + 1)
y1 = int(spatial_scale * y1)
y2 = max(int(spatial_scale * y2), y1 + 1)
x_roi = bottom_data[i_batch][:, y1:y2, x1:x2]
x_roi = x_roi[None, :, :, :] # N, C, H, W
y = functions.resize_images(x_roi, (outh, outw))
ys[i_batch].append(y)
yss = []
for i_batch in six.moves.range(B):
ys = functions.vstack(ys[i_batch])
yss.append(ys)
yss = functions.concat(yss, axis=0)
return yss
def compute_shifts(cell, pbc, cutoff):
xp = cell.xp
reciprocal_cell = F.batch_inv(cell)
inv_distances = F.max(F.sqrt(F.sum(reciprocal_cell**2, axis=1)), axis=0)
num_repeats = F.ceil(cutoff * inv_distances)
num_repeats = F.where(pbc, num_repeats, xp.zeros_like(num_repeats.data))
num_repeats = F.max(num_repeats, axis=0)
r1 = xp.arange(1, num_repeats.data[0] + 1)
r2 = xp.arange(1, num_repeats.data[1] + 1)
r3 = xp.arange(1, num_repeats.data[2] + 1)
o = xp.zeros(1, dtype=r1.dtype)
return F.vstack([
xp.array([[0.0, 0.0, 0.0]]),
cartesian_prod(r1, r2, r3),
cartesian_prod(r1, r2, o),
cartesian_prod(r1, r2, -r3),
cartesian_prod(r1, o, r3),
cartesian_prod(r1, o, o),
cartesian_prod(r1, o, -r3),
cartesian_prod(r1, -r2, r3),
cartesian_prod(r1, -r2, o),
cartesian_prod(r1, -r2, -r3),
cartesian_prod(o, r2, r3),
cartesian_prod(o, r2, o),
cartesian_prod(o, r2, -r3),
cartesian_prod(o, o, r3),
]).data
def translate(self, hxs, max_length=100):
batch_size, _, _ = hxs.shape
compute_context = self.attention(hxs)
c = Variable(self.xp.zeros((batch_size, self.n_units), 'f'))
h = Variable(self.xp.zeros((batch_size, self.n_units), 'f'))
ys = self.xp.full(batch_size, tokens['<SOS>'], np.int32)
results = []
for _ in range(max_length):
eys = self.embed_y(ys)
context = compute_context(h)
concatenated = F.concat([eys, context])
c, h = self.lstm(c, h, concatenated)
concatenated = F.concat([concatenated, h])
logit = self.w(self.maxout(concatenated))
y = F.reshape(F.argmax(logit, axis=1), (batch_size, ))
results.append(y)
results = F.separate(F.transpose(F.vstack(results)), axis=0)
outs = []
for y in results:
inds = np.argwhere(y == tokens['<EOS>'])
if len(inds) > 0:
y = y[:inds[0, 0]]
outs.append(y)
return outs
def translate(
self,
encoded: Variable,
max_length: int = 100
) -> List[ndarray]:
sentence_count = encoded.shape[0]
self.setup(encoded)
cell, state, previous_words = self.get_initial_states(sentence_count)
result = []
for _ in range(max_length):
cell, state, context, concatenated = \
self.advance_one_step(cell, state, previous_words)
logit, state = self.compute_logit(concatenated, state, context)
output_id = F.reshape(F.argmax(logit, axis=1), (sentence_count,))
result.append(output_id)
previous_words = output_id
# Remove words after <EOS>
outputs = F.separate(F.transpose(F.vstack(result)), axis=0)
assert len(outputs) == sentence_count
output_sentences = []
for output in outputs:
assert output.shape == (max_length,)
indexes = np.argwhere(output.data == EOS)
if len(indexes) > 0:
output = output[:indexes[0, 0] + 1]
output_sentences.append(output.data)
return output_sentences
def __call__(self, chars):
"""
Note: we use zero-padding instead of spacial PADDING token.
"""
if isinstance(chars, (tuple, list)):
return F.vstack([self.forward_one(_chars) for _chars in chars])
return self.forward_one(chars)
def parse(self, hs):
self.hs = self.parser_blstm(hs)
self.pads = F.tanh(self.pad_linear(self.pad_embed(self.xp.arange(4))))
states = [(sent_idx, transition.State(h.shape[0]))
for sent_idx, h in enumerate(hs)]
_states = states[:]
while len(_states) > 0:
features = []
fs = []
for sent_idx, state in _states:
feature = self.xp.array([self.extract_feature(state)])
features.append(feature)
fs.append(self._populate_features(feature, sent_idx))
fs = F.vstack(fs)
fs = F.stack(F.split_axis(fs, fs.shape[0] // 4, axis=0), axis=0)
fs = F.reshape(fs, (fs.shape[0], -1))
action_scores = self.parser_mlp(fs)
action_scores.to_cpu()
action_scores = action_scores.data
for i in range(len(_states) - 1, -1, -1):
_, state = _states[i]
best_action, best_score = -1, -np.inf
for action, score in enumerate(action_scores[i]):
if score > best_score and \
self.transition_system.is_allowed(action, state):
best_action, best_score = action, score
self.transition_system.apply(best_action, state)
if self.transition_system.is_terminal(state):
del _states[i]
heads, labels, _states = zip(*[(state.heads, state.labels, state)
for i, state in states])
return heads, labels, _states
def forward(self, ws, ss, ps, dep_ts=None):
batchsize = len(ws)
xp = chainer.cuda.get_array_module(ws[0])
split = scanl(lambda x, y: x + y, 0, [w.shape[0] for w in ws])[1:-1]
wss = self.emb_word(F.hstack(ws))
sss = F.reshape(self.emb_suf(F.vstack(ss)), (-1, 4 * self.afix_dim))
pss = F.reshape(self.emb_prf(F.vstack(ps)), (-1, 4 * self.afix_dim))
ins = F.dropout(F.concat([wss, sss, pss]),
self.dropout_ratio,
train=self.train)
xs_f = list(F.split_axis(ins, split, 0))
xs_b = [x[::-1] for x in xs_f]
cx_f, hx_f, cx_b, hx_b = self._init_state(xp, batchsize)
_, _, hs_f = self.lstm_f(hx_f, cx_f, xs_f, train=self.train)
_, _, hs_b = self.lstm_b(hx_b, cx_b, xs_b, train=self.train)
hs_b = [x[::-1] for x in hs_b]
# ys: [(sentence length, number of category)]
hs = [F.concat([h_f, h_b]) for h_f, h_b in zip(hs_f, hs_b)]
dep_ys = [
self.biaffine_arc(
F.elu(F.dropout(self.arc_dep(h), 0.32, train=self.train)),
F.elu(F.dropout(self.arc_head(h), 0.32, train=self.train)))
for h in hs
]
# if dep_ts is not None and random.random >= 0.5:
if dep_ts is not None:
heads = dep_ts
else:
heads = [F.argmax(y, axis=1) for y in dep_ys]
heads = F.elu(F.dropout(
self.rel_head(
F.vstack([F.embed_id(t, h, ignore_label=IGNORE) \
for h, t in zip(hs, heads)])),
0.32, train=self.train))
childs = F.elu(
F.dropout(self.rel_dep(F.vstack(hs)), 0.32, train=self.train))
cat_ys = self.biaffine_tag(childs, heads)
cat_ys = list(F.split_axis(cat_ys, split, 0))
return cat_ys, dep_ys
def predict_with_mask(self, x, ent_Ts=None, test=True):
num = x.shape[0]
bs = []
exit_i = 0
for i, link in enumerate(self.links):
if isinstance(link, Sequential):
b = link(x, test=test)
b = b[0]
y_exit, y_cont, exited = self.entropy_filter(
x, b, ent_Ts[min(exit_i,
len(ent_Ts) - 1)])
exit_i = exit_i + 1
b = y_exit
bs.append((b, exited))
x = y_cont
if len(x) == 0:
break
elif isinstance(link, function.dropout):
x = link(x, train=not test)
elif isinstance(link, chainer.links.BatchNormalization):
x = link(x, test=test)
elif hasattr(link, '__call__') and 'train' in inspect.getargspec(
link.__call__)[0]:
x = link(x, train=not test)
elif hasattr(link, '__call__') and 'test' in inspect.getargspec(
link.__call__)[0]:
x = link(x, test=test)
else:
x = link(x)
#if len(x) > 0:
# bs.append((x,[True]*x.shape[0]))
ys = [None] * num
exited = [False] * num
# branch exit
for b, ex in bs:
i = 0
j = 0
for exit in ex:
while ys[i] is not None:
i = i + 1
if exit:
ys[i] = b[j]
exited[i] = True #only count the branch exited
j = j + 1
i = i + 1
# main exit
if len(x) > 0:
b, ex = (x, [True] * x.shape[0])
i = 0
j = 0
for exit in ex:
while ys[i] is not None:
i = i + 1
if exit:
ys[i] = b[j]
j = j + 1
i = i + 1
return F.vstack(ys), exited
def forward(self, x):
hy, cy, x = self.h(None, None, x)
x = np.array([_x.data for _x in x])
x = F.transpose(x, axes=(1, 0, 2))
x = F.vstack([np.array([_x.data[-1] for _x in x])])
#x = chainer.Variable(np.array([_h[-1].data for _h in h]))
x = self.out(x)
return x
def __call__(self, h): #h:batch_size*time*channel
batch_size = h.shape[0]
sequence_length = h.shape[1]
h1 = self.lstm(h[:, 0, :]).reshape(1, batch_size, -1)
for i in range(1, sequence_length):
h1 = F.vstack((h1, self.lstm(h[:,
i, :]).reshape(1, batch_size, -1)))
return h1
def __call__(self, h): #h:batch_size*time*channel
batch_size = h.shape[0]
sequence_length = h.shape[1]
h1 = self.lstm_forward(h[:, 0, :]).reshape(1, batch_size, -1)
for i in range(1, sequence_length):
h1 = F.vstack(
(h1, self.lstm_forward(h[:, i, :]).reshape(1, batch_size, -1)))
h2 = self.lstm_backward(h[:, sequence_length - 1, :]).reshape(
1, batch_size, -1)
for i in range(sequence_length - 2, -1, -1):
h2 = F.vstack(
(self.lstm_backward(h[:, i, :]).reshape(1, batch_size,
-1), h2))
return h1 + h2
def forward(self, inputs, device):
x = list(inputs)
y = functions.vstack(x)
return y,
def check_forward(self, xs_data):
xs = [chainer.Variable(x) for x in xs_data]
y = functions.vstack(xs)
expect = numpy.vstack(self.xs)
testing.assert_allclose(y.data, expect)
def func(*xs):
return functions.vstack(xs)
def func(*xs):
y = functions.vstack(xs)
return y * y
