def forward(self, xs, ys, reduce='mean', transpose=False):
"""Computes negative log-likelihood of linear-chain CRF
Args:
xs (list of Variable): Input vector for each label
ys (list of Variable): Expected output labels.
transpose (bool): If ``True``, input/output sequences
will be sorted in descending order of length.
Returns:
~chainer.Variable: A variable holding the average negative
log-likelihood of the input sequences.
.. seealso:: See :func:`~chainer.frunctions.crf1d` for more detail.
"""
if transpose:
indices = argsort_list_descent(xs)
xs = permutate_list(xs, indices, inv=False)
ys = permutate_list(ys, indices, inv=False)
trans_x = transpose_sequence.transpose_sequence(xs)
trans_y = transpose_sequence.transpose_sequence(ys)
loss = crf1d.crf1d(self.cost, trans_x, trans_y, reduce)
else:
loss = crf1d.crf1d(self.cost, xs, ys, reduce)
return loss
def argmax(self, xs, transpose=False):
"""Computes a state that maximizes a joint probability.
Args:
xs (list of Variable): Input vector for each label.
transpose (bool): If ``True``, input/output sequences
will be sorted in descending order of length.
Returns:
tuple: A tuple of :class:`~chainer.Variable` representing each
log-likelihood and a list representing the argmax path.
.. seealso:: See :func:`~chainer.frunctions.crf1d_argmax` for more
detail.
"""
if transpose:
indices = argsort_list_descent(xs)
xs = permutate_list(xs, indices, inv=False)
trans_x = transpose_sequence.transpose_sequence(xs)
score, path = crf1d.argmax_crf1d(self.cost, trans_x)
path = transpose_sequence.transpose_sequence(path)
path = [p.array for p in path]
path = permutate_list(path, indices, inv=True)
else:
score, path = crf1d.argmax_crf1d(self.cost, xs)
return score, path
def argmax(self, xs, transpose=False):
"""Computes a state that maximizes a joint probability.
Args:
xs (list of Variable): Input vector for each label.
transpose (bool): If ``True``, input/output sequences
will be sorted in descending order of length.
Returns:
tuple: A tuple of :class:`~chainer.Variable` representing each
log-likelihood and a list representing the argmax path.
.. seealso:: See :func:`~chainer.frunctions.crf1d_argmax` for more
detail.
"""
if transpose:
indices = argsort_list_descent(xs)
xs = permutate_list(xs, indices, inv=False)
trans_x = transpose_sequence.transpose_sequence(xs)
score, path = crf1d.argmax_crf1d(self.cost, trans_x)
path = transpose_sequence.transpose_sequence(path)
path = [p.array for p in path]
path = permutate_list(path, indices, inv=True)
else:
score, path = crf1d.argmax_crf1d(self.cost, xs)
return score, path
def forward(self, xs, ys, reduce='mean', transpose=False):
"""Computes negative log-likelihood of linear-chain CRF
Args:
xs (list of Variable): Input vector for each label
ys (list of Variable): Expected output labels.
transpose (bool): If ``True``, input/output sequences
will be sorted in descending order of length.
Returns:
~chainer.Variable: A variable holding the average negative
log-likelihood of the input sequences.
.. seealso:: See :func:`~chainer.frunctions.crf1d` for more detail.
"""
if transpose:
indices = argsort_list_descent(xs)
xs = permutate_list(xs, indices, inv=False)
ys = permutate_list(ys, indices, inv=False)
trans_x = transpose_sequence.transpose_sequence(xs)
trans_y = transpose_sequence.transpose_sequence(ys)
loss = crf1d.crf1d(self.cost, trans_x, trans_y, reduce)
else:
loss = crf1d.crf1d(self.cost, xs, ys, reduce)
return loss
def _Transpose(X, indices=None, inv=False):
if not inv:
indices = argsort_list_descent(X)
transpose = list(transpose_sequence.transpose_sequence(permutate_list(X, indices, inv=inv)))
return indices, transpose
else:
transpose = permutate_list(transpose_sequence.transpose_sequence(X), indices, inv=inv)
return transpose
def __call__(self, hx, cx, xs, **kwargs):
"""__call__(self, hx, cx, xs)
Calculate all hidden states and cell states.
.. warning::
``train`` argument is not supported anymore since v2.
Instead, use ``chainer.using_config('train', train)``.
See :func:`chainer.using_config`.
Args:
hx (~chainer.Variable or None): Initial hidden states. If ``None``
is specified zero-vector is used.
cx (~chainer.Variable or None): Initial cell states. If ``None``
is specified zero-vector is used.
xs (list of ~chainer.Variable): List of input sequences.
Each element ``xs[i]`` is a :class:`chainer.Variable` holding
a sequence.
"""
argument.check_unexpected_kwargs(
kwargs,
train='train argument is not supported anymore. '
'Use chainer.using_config')
argument.assert_kwargs_empty(kwargs)
assert isinstance(xs, (list, tuple))
xp = cuda.get_array_module(hx, *xs)
indices = n_step_rnn.argsort_list_descent(xs)
indices_array = xp.array(indices)
xs = n_step_rnn.permutate_list(xs, indices, inv=False)
if hx is None:
hx = self.init_hx(xs)
else:
hx = permutate.permutate(hx, indices_array, axis=1, inv=False)
if cx is None:
cx = self.init_hx(xs)
else:
cx = permutate.permutate(cx, indices_array, axis=1, inv=False)
trans_x = transpose_sequence.transpose_sequence(xs)
ws = [[w.w0, w.w1, w.w2, w.w3, w.w4, w.w5, w.w6, w.w7] for w in self]
bs = [[w.b0, w.b1, w.b2, w.b3, w.b4, w.b5, w.b6, w.b7] for w in self]
hy, cy, trans_y = self.rnn(self.n_layers, self.dropout, hx, cx, ws, bs,
trans_x)
hy = permutate.permutate(hy, indices_array, axis=1, inv=True)
cy = permutate.permutate(cy, indices_array, axis=1, inv=True)
ys = transpose_sequence.transpose_sequence(trans_y)
ys = n_step_rnn.permutate_list(ys, indices, inv=True)
return hy, cy, ys
def __call__(self, hx, cx, xs, **kwargs):
"""__call__(self, hx, cx, xs)
Calculate all hidden states and cell states.
.. warning::
``train`` argument is not supported anymore since v2.
Instead, use ``chainer.using_config('train', train)``.
See :func:`chainer.using_config`.
Args:
hx (~chainer.Variable or None): Initial hidden states. If ``None``
is specified zero-vector is used.
cx (~chainer.Variable or None): Initial cell states. If ``None``
is specified zero-vector is used.
xs (list of ~chianer.Variable): List of input sequences.
Each element ``xs[i]`` is a :class:`chainer.Variable` holding
a sequence.
"""
argument.check_unexpected_kwargs(
kwargs, train='train argument is not supported anymore. '
'Use chainer.using_config')
argument.assert_kwargs_empty(kwargs)
assert isinstance(xs, (list, tuple))
indices = n_step_rnn.argsort_list_descent(xs)
xs = n_step_rnn.permutate_list(xs, indices, inv=False)
if hx is None:
hx = self.init_hx(xs)
else:
hx = permutate.permutate(hx, indices, axis=1, inv=False)
if cx is None:
cx = self.init_hx(xs)
else:
cx = permutate.permutate(cx, indices, axis=1, inv=False)
trans_x = transpose_sequence.transpose_sequence(xs)
ws = [[w.w0, w.w1, w.w2, w.w3, w.w4, w.w5, w.w6, w.w7] for w in self]
bs = [[w.b0, w.b1, w.b2, w.b3, w.b4, w.b5, w.b6, w.b7] for w in self]
hy, cy, trans_y = self.rnn(
self.n_layers, self.dropout, hx, cx, ws, bs, trans_x)
hy = permutate.permutate(hy, indices, axis=1, inv=True)
cy = permutate.permutate(cy, indices, axis=1, inv=True)
ys = transpose_sequence.transpose_sequence(trans_y)
ys = n_step_rnn.permutate_list(ys, indices, inv=True)
return hy, cy, ys
def _accuracy(self, ys, ts):
ys = permutate_list(ys, argsort_list_descent(ys), inv=False)
ts = permutate_list(ts, argsort_list_descent(ts), inv=False)
correct = 0
total = 0
exact = 0
for _y, _t in zip(ys, ts):
y = _y.data
t = _t.data
_correct = (y == cuda.to_gpu(t)).sum()
_total = t.size
if _correct == _total:
exact += 1
correct += _correct
total += _total
accuracy = correct / total
self._eval = {
'accuracy': accuracy,
'correct': correct,
'total': total,
'exact': exact
}
return accuracy
def __call__(self, hx, cx, xs, train=True):
"""Calculate all hidden states and cell states.
Args:
hx (~chainer.Variable or None): Initial hidden states. If ``None``
is specified zero-vector is used.
cx (~chainer.Variable or None): Initial cell states. If ``None``
is specified zero-vector is used.
xs (list of ~chianer.Variable): List of input sequences.
Each element ``xs[i]`` is a :class:`chainer.Variable` holding
a sequence.
"""
assert isinstance(xs, (list, tuple))
indices = argsort_list_descent(xs)
xs = permutate_list(xs, indices, inv=False)
if hx is None:
hx = self.init_hx(xs)
else:
hx = permutate.permutate(hx, indices, axis=1, inv=False)
if cx is None:
cx = self.init_hx(xs)
else:
cx = permutate.permutate(cx, indices, axis=1, inv=False)
trans_x = transpose_sequence.transpose_sequence(xs)
ws = [[w.w0, w.w1, w.w2, w.w3, w.w4, w.w5, w.w6, w.w7] for w in self]
bs = [[w.b0, w.b1, w.b2, w.b3, w.b4, w.b5, w.b6, w.b7] for w in self]
hy, cy, trans_y = self.rnn(self.n_layers,
self.dropout,
hx,
cx,
ws,
bs,
trans_x,
train=train,
use_cudnn=self.use_cudnn)
hy = permutate.permutate(hy, indices, axis=1, inv=True)
cy = permutate.permutate(cy, indices, axis=1, inv=True)
ys = transpose_sequence.transpose_sequence(trans_y)
ys = permutate_list(ys, indices, inv=True)
return hy, cy, ys
def predict_crf(self, hs, ls=None, calculate_loss=True):
indices = argsort_list_descent(hs)
trans_hs = F.transpose_sequence(permutate_list(hs, indices, inv=False))
score, trans_ys = self.crf.argmax(trans_hs)
ys = permutate_list(F.transpose_sequence(trans_ys), indices, inv=True)
ps = [y.data for y in ys]
xp = cuda.get_array_module(hs[0])
if calculate_loss:
tmp = permutate_list(ls, indices, inv=False)
trans_ls = F.transpose_sequence(
permutate_list(ls, indices, inv=False)) # sort by length
loss = self.crf(trans_hs, trans_ls)
else:
loss = chainer.Variable(xp.array(0, dtype='f'))
return loss, ps
def __call__(self, xs):
"""
Args:
xs (list or tuple): batch-length list of Variable, each with shape
(
Returns:
"""
assert isinstance(xs, (list, tuple))
indices = argsort_list_descent(xs)
xs = permutate_list(xs, indices, inv=False)
h = self.rnn.init_h(xs)
c = self.rnn.init_c(xs)
# list of Variable, each with shape (batch, features)
trans_x = transpose_sequence.transpose_sequence(xs)
z_lst = []
pz_lst = []
for x in trans_x:
if len(x) < len(h):
# a sequence ended
h = h[:len(x)]
c = [c_i[:len(x)] for c_i in c]
pz, z, h, c = self._forward_single(x, h, c)
z_lst.append(z)
pz_lst.append(pz)
# h_lst and c_lst basically have same order as x
z_lst = transpose_sequence.transpose_sequence(z_lst)
z_lst = permutate_list(z_lst, indices, inv=True)
z_lst = [F.squeeze(z, 1).data for z in z_lst]
# h_lst and c_lst basically have same order as x
pz_lst = transpose_sequence.transpose_sequence(pz_lst)
pz_lst = permutate_list(pz_lst, indices, inv=True)
pz_lst = [F.squeeze(pz, 1) for pz in pz_lst]
return pz_lst, z_lst
def __call__(self, xs):
"""
Args:
xs (list or tuple): batch-length list of Variable, each with shape
(
Returns:
"""
assert isinstance(xs, (list, tuple))
indices = argsort_list_descent(xs)
xs = permutate_list(xs, indices, inv=False)
h = self.init_h(xs)
c = self.init_c(xs)
# list of Variable, each with shape (batch, features)
trans_x = transpose_sequence.transpose_sequence(xs)
c_lst = []
h_lst = []
for x in trans_x:
if len(x) < len(h):
# a sequence ended
h = h[:len(x)]
c = [c_i[:len(x)] for c_i in c]
h, c = self.forward_single(x, h, c)
c_lst.append(F.hstack(c))
h_lst.append(h)
# h_lst and c_lst basically have same order as x
h_lst = transpose_sequence.transpose_sequence(h_lst)
h_lst = permutate_list(h_lst, indices, inv=True)
c_lst = transpose_sequence.transpose_sequence(c_lst)
c_lst = permutate_list(c_lst, indices, inv=True)
return h_lst, c_lst
def __call__(self, h0, hx, xs, **kwargs):
"""__call__(self, hx, xs)
Calculate all hidden states and cell states.
.. warning::
``train`` argument is not supported anymore since v2.
Instead, use ``chainer.using_config('train', train)``.
See :func:`chainer.using_config`.
Args:
hx (~chainer.Variable or None): Initial hidden states. If ``None``
is specified zero-vector is used.
xs (list of ~chianer.Variable): List of input sequences.
Each element ``xs[i]`` is a :class:`chainer.Variable` holding
a sequence.
"""
argument.check_unexpected_kwargs(
kwargs,
train='train argument is not supported anymore. '
'Use chainer.using_config')
argument.assert_kwargs_empty(kwargs)
assert isinstance(
xs, (list, tuple)), "xs in not a list or tupple: %r" % type(xs)
indices = argsort_list_descent(xs)
xs = permutate_list(xs, indices, inv=False)
if h0 is None:
h0 = self.init_hx(xs)
else:
h0 = permutate.permutate(h0, indices, axis=1, inv=False)
hx = permutate_list(hx, indices, inv=False)
trans_x = transpose_sequence.transpose_sequence(xs)
ws = [[w.w0, w.w1, w.w2, w.w3, w.w4, w.w5] for w in self]
bs = [[w.b0, w.b1, w.b2, w.b3, w.b4, w.b5] for w in self]
for w in self:
w1 = w.w6
w2 = w.w7
w3 = w.w8
b1 = w.b6
b2 = w.b7
b3 = w.b8
W = [w1, w2, w3]
B = [b1, b2, b3]
h_list, h_bar_list, c_s_list, z_s_list = self.rnn(
self.n_layers, self.dropout, h0, hx, ws, bs, trans_x, W, B)
'''
print(type(h_list),len(h_list))
print(type(h_list[0]),len(h_list[0]))
print(type(h_list[0][0]),len(h_list[0][0]))
'''
#batch内で入れ替え
h_list = transpose_sequence.transpose_sequence(h_list)
h_list = permutate_list(h_list, indices, inv=True)
'''
print(type(h_list),len(h_list))
print(type(h_list[0]),len(h_list[0]))
print(type(h_list[0][0]),len(h_list[0][0]))
'''
h_bar_list = transpose_sequence.transpose_sequence(h_bar_list)
h_bar_list = permutate_list(h_bar_list, indices, inv=True)
'''
print(type(c_s_list),len(c_s_list))
print(type(c_s_list[0]),len(c_s_list[0]))
print(type(c_s_list[0][0]),len(c_s_list[0][0]), c_s_list[0][0])
print(type(z_s_list), len(z_s_list))
print(type(z_s_list[0]), len(z_s_list[0]), z_s_list[0])
'''
c_s_list = transpose_sequence.transpose_sequence(c_s_list)
c_s_list = permutate_list(c_s_list, indices, inv=True)
z_s_list = transpose_sequence.transpose_sequence(z_s_list)
z_s_list = permutate_list(z_s_list, indices, inv=True)
return h_list, h_bar_list, c_s_list, z_s_list
