def functions(self):
return collections.OrderedDict([
('conv1_1', [self.conv1_1, relu]),
('conv1_2', [self.conv1_2, relu]),
('pool1', [_max_pooling_2d]),
('conv2_1', [self.conv2_1, relu]),
('conv2_2', [self.conv2_2, relu]),
('pool2', [_max_pooling_2d]),
('conv3_1', [self.conv3_1, relu]),
('conv3_2', [self.conv3_2, relu]),
('conv3_3', [self.conv3_3, relu]),
('conv3_4', [self.conv3_4, relu]),
('pool3', [_max_pooling_2d]),
('conv4_1', [self.conv4_1, relu]),
('conv4_2', [self.conv4_2, relu]),
('conv4_3', [self.conv4_3, relu]),
('conv4_4', [self.conv4_4, relu]),
('pool4', [_max_pooling_2d]),
('conv5_1', [self.conv5_1, relu]),
('conv5_2', [self.conv5_2, relu]),
('conv5_3', [self.conv5_3, relu]),
('conv5_4', [self.conv5_4, relu]),
('pool5', [_max_pooling_2d]),
# ('fc6', [self.fc6, relu, dropout]),
('fc6', [self.fc6, relu, lambda x: dropout(x, ratio=0.0)]),
# ('fc7', [self.fc7, relu, dropout]),
('fc7', [self.fc7, relu, lambda x: dropout(x, ratio=0.0)]),
('fc8', [self.fc8]),
('prob', [softmax]),
])
def __call__(self, x):
h = x
h = relu(self.conv1_1(h))
h = relu(self.conv1_2(h))
h = _max_pooling_2d(h)
h = relu(self.conv2_1(h))
h = relu(self.conv2_2(h))
h = _max_pooling_2d(h)
h = relu(self.conv3_1(h))
h = relu(self.conv3_2(h))
h = relu(self.conv3_3(h))
h = _max_pooling_2d(h)
h = relu(self.conv4_1(h))
h = relu(self.conv4_2(h))
h = relu(self.conv4_3(h))
h = _max_pooling_2d(h)
h = relu(self.conv5_1(h))
h = relu(self.conv5_2(h))
h = relu(self.bn1(self.conv5_3(h)))
h = _max_pooling_2d(h)
h = dropout(relu(self.bn2(self.fc6(h))))
h = dropout(relu(self.bn3(self.fc7(h))))
h = self.fc8(h)
return h
def feed_lstm(self, word, embed_layer, lstm_layer_list, train):
# get embedding for word
embed_id = N.dropout(embed_layer(word),
ratio=DROPOUT_RATIO,
train=train)
# feed into first LSTM layer
hs = N.dropout(self[lstm_layer_list[0]](embed_id),
ratio=DROPOUT_RATIO,
train=train)
# feed into remaining LSTM layers
for lstm_layer in lstm_layer_list[1:]:
hs = N.dropout(self[lstm_layer](hs),
ratio=DROPOUT_RATIO,
train=train)
def __call__(self, x, **kwargs):
"""Applies the lstm layer.
Args:
x (~chainer.Variable): Time-Batch of input vectors.
Returns:
~chainer.Variable: Output of the lstm layer.
"""
dropout_rate = kwargs.get('dropout', 0.)
dropout_rate_hidden_hidden = kwargs.get('dropout_hidden_hidden', 0.)
x = dropout(x, dropout_rate)
lstm_in = sequence_linear_function(x, self.W_x, self.b)
if self.normalized:
lstm_in = sequence_batch_normalization_function(
lstm_in, self.gamma, self.beta)
if self.stateful:
c_prev = self.c_prev
h_prev = self.h_prev
else:
c_prev = None
h_prev = None
lstm_out, self.h_prev, self.c_prev = \
sequence_lstm_function(lstm_in, self.W_h, c_prev, h_prev,
self.reverse, dropout_rate_hidden_hidden)
return lstm_out
def _one_directional_loop(di):
# di=0, forward RNN
# di=1, backward RNN
xs_list = xs_next if di == 0 else reversed(xs_next)
layer_idx = direction * layer + di
h = hx[layer_idx]
h_list = []
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
else:
h_rest = None
if layer > 0:
x = dropout.dropout(x, ratio=dropout_ratio)
rnn_in = (
linear.linear(x, xws[layer_idx], xbs[layer_idx]) +
linear.linear(h, hws[layer_idx], hbs[layer_idx]))
if activation == 'tanh':
h_bar = tanh.tanh(rnn_in)
elif activation == 'relu':
h_bar = relu.relu(rnn_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
else:
h = h_bar
h_list.append(h_bar)
return h, h_list
def _one_directional_loop(di):
# di=0, forward GRU
# di=1, backward GRU
xs_list = xs_next if di == 0 else reversed(xs_next)
layer_idx = direction * layer + di
h = hx[layer_idx]
h_list = []
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
else:
h_rest = None
if layer > 0:
x = dropout.dropout(x, ratio=dropout_ratio)
gru_x = linear.linear(x, xws[layer_idx], xbs[layer_idx])
gru_h = linear.linear(h, hws[layer_idx], hbs[layer_idx])
W_r_x, W_z_x, W_x = split_axis.split_axis(gru_x, 3, axis=1)
U_r_h, U_z_h, U_x = split_axis.split_axis(gru_h, 3, axis=1)
r = sigmoid.sigmoid(W_r_x + U_r_h)
z = sigmoid.sigmoid(W_z_x + U_z_h)
h_bar = tanh.tanh(W_x + r * U_x)
h_bar = (1 - z) * h_bar + z * h
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
else:
h = h_bar
h_list.append(h_bar)
return h, h_list
def __call__(self, x, **kwargs):
"""Applies the lstm layer.
Args:
x (~chainer.Variable): Time-Batch of input vectors.
Returns:
~chainer.Variable: Output of the lstm layer.
"""
dropout_rate = kwargs.get('dropout', 0.)
dropout_rate_hidden_hidden = kwargs.get('dropout_hidden_hidden', 0.)
x = dropout(x, dropout_rate)
lstm_in = sequence_linear_function(x, self.W_x, self.b)
if self.normalized:
lstm_in = sequence_batch_normalization_function(lstm_in, self.gamma,
self.beta)
if self.stateful:
c_prev = self.c_prev
h_prev = self.h_prev
else:
c_prev = None
h_prev = None
lstm_out, self.h_prev, self.c_prev = \
sequence_lstm_function(lstm_in, self.W_h, c_prev, h_prev,
self.reverse, dropout_rate_hidden_hidden)
return lstm_out
def _one_directional_loop(di):
# di=0, forward GRU
# di=1, backward GRU
xs_list = xs_next if di == 0 else reversed(xs_next)
layer_idx = direction * layer + di
h = hx[layer_idx]
h_list = []
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
else:
h_rest = None
if layer > 0:
x = dropout.dropout(x, ratio=dropout_ratio)
gru_x = linear.linear(x, xws[layer_idx], xbs[layer_idx])
gru_h = linear.linear(h, hws[layer_idx], hbs[layer_idx])
W_r_x, W_z_x, W_x = split_axis.split_axis(gru_x, 3, axis=1)
U_r_h, U_z_h, U_x = split_axis.split_axis(gru_h, 3, axis=1)
r = sigmoid.sigmoid(W_r_x + U_r_h)
z = sigmoid.sigmoid(W_z_x + U_z_h)
h_bar = tanh.tanh(W_x + r * U_x)
h_bar = (1 - z) * h_bar + z * h
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
else:
h = h_bar
h_list.append(h_bar)
return h, h_list
def _one_directional_loop(di):
# di=0, forward RNN
# di=1, backward RNN
xs_list = xs_next if di == 0 else reversed(xs_next)
layer_idx = direction * layer + di
h = hx[layer_idx]
h_list = []
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
else:
h_rest = None
if layer > 0:
x = dropout.dropout(x, ratio=dropout_ratio)
rnn_in = (linear.linear(x, xws[layer_idx],
xbs[layer_idx]) +
linear.linear(h, hws[layer_idx], hbs[layer_idx]))
if activation == 'tanh':
h_bar = tanh.tanh(rnn_in)
elif activation == 'relu':
h_bar = relu.relu(rnn_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
else:
h = h_bar
h_list.append(h_bar)
return h, h_list
def __call__(self, frame, prev_word, state, dropout_flag, dropout_ratio):
i1 = self.xi1(dropout(frame, dropout_ratio, dropout_flag))
c1, h1 = lstm(state['c1'], self.ih1(i1) + self.hh1(state['h1']))
i2 = self.xi2(prev_word)
concat = array.concat.concat((i2, h1))
c2, h2 = lstm(state['c2'], self.ih2(concat) + self.hh2(state['h2']))
state = {'c1': c1, 'h1': h1, 'c2': c2, 'h2': h2}
return state
def __call__(self, x):
x = self.embed(x)
xs = split_axis.split_axis(x, x.data.shape[1], 1)
ret = []
for x in xs:
for l in self.rnns:
x = l(x)
x = dropout.dropout(x, 0.25, self.train)
for l in self.linears:
x = l(x)
x = reshape.reshape(x, x.data.shape + (-1, ))
ret.append(x)
ret = concat.concat(ret, axis=2)
return ret
def ASPP(x):
y = [
F.tile(
self.ASPP[0](F.average_pooling_2d(
x, ksize=x.shape[-2:])), x.shape[-2:])
]
y.extend([
self.ASPP[i](x) for i in range(1,
len(self.ASPP) - 1)
])
y = F.concat(y, axis=1)
y = dropout.dropout(y, ratio=0.5)
y = self.ASPP[-1](y)
return y
def __init__(self, pretrained_model='auto'):
super(GoogLeNet, self).__init__(
conv1=Convolution2D(3, 64, 7, stride=2, pad=3),
conv2_reduce=Convolution2D(64, 64, 1),
conv2=Convolution2D(64, 192, 3, stride=1, pad=1),
inc3a=Inception(192, 64, 96, 128, 16, 32, 32),
inc3b=Inception(256, 128, 128, 192, 32, 96, 64),
inc4a=Inception(480, 192, 96, 208, 16, 48, 64),
inc4b=Inception(512, 160, 112, 224, 24, 64, 64),
inc4c=Inception(512, 128, 128, 256, 24, 64, 64),
inc4d=Inception(512, 112, 144, 288, 32, 64, 64),
inc4e=Inception(528, 256, 160, 320, 32, 128, 128),
inc5a=Inception(832, 256, 160, 320, 32, 128, 128),
inc5b=Inception(832, 384, 192, 384, 48, 128, 128),
loss3_fc=Linear(1024, 1000),
loss1_conv=Convolution2D(512, 128, 1),
loss1_fc1=Linear(4 * 4 * 128, 1024),
loss1_fc2=Linear(1024, 1000),
loss2_conv=Convolution2D(528, 128, 1),
loss2_fc1=Linear(4 * 4 * 128, 1024),
loss2_fc2=Linear(1024, 1000),
)
if pretrained_model == 'auto':
_retrieve(
'bvlc_googlenet.npz',
'http://dl.caffe.berkeleyvision.org/bvlc_googlenet.caffemodel',
self)
elif pretrained_model:
npz.load_npz(pretrained_model, self)
self.functions = OrderedDict([
('conv1', [self.conv1, relu]),
('pool1', [
lambda x: max_pooling_2d(x, ksize=3, stride=2),
lambda x: local_response_normalization(x, n=5)
]), ('conv2_reduce', [self.conv2_reduce, relu]),
('conv2', [self.conv2, relu]),
('pool2', [
lambda x: local_response_normalization(x, n=5),
lambda x: max_pooling_2d(x, ksize=3, stride=2)
]), ('inc3a', [self.inc3a]), ('inc3b', [self.inc3b]),
('pool3', [lambda x: max_pooling_2d(x, ksize=3, stride=2)]),
('inc4a', [self.inc4a]), ('inc4b', [self.inc4b]),
('inc4c', [self.inc4c]), ('inc4d', [self.inc4d]),
('inc4e', [self.inc4e]),
('pool4', [lambda x: max_pooling_2d(x, ksize=3, stride=2)]),
('inc5a', [self.inc5a]), ('inc5b', [self.inc5b]),
('pool6', [lambda x: average_pooling_2d(x, ksize=7, stride=1)]),
('prob', [lambda x: dropout(x, ratio=0.4), self.loss3_fc])
])
def __call__(self, x, **kwargs):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Time-Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
dropout_rate = kwargs.get('dropout', 0.)
x = dropout(x, dropout_rate)
x = sequence_linear_function(x, self.W, self.b)
if self.normalized:
x = sequence_batch_normalization_function(x, self.gamma, self.beta)
return x
def __call__(self, x, **kwargs):
"""Applies the linear layer.
Args:
x (~chainer.Variable): Time-Batch of input vectors.
Returns:
~chainer.Variable: Output of the linear layer.
"""
dropout_rate = kwargs.get('dropout', 0.)
x = dropout(x, dropout_rate)
x = sequence_linear_function(x, self.W, self.b)
if self.normalized:
x = sequence_batch_normalization_function(x, self.gamma, self.beta)
return x
def _one_directional_loop(di):
# di=0, forward LSTM
# di=1, backward LSTM
h_list = []
c_list = []
layer_idx = direction * layer + di
h = hx[layer_idx]
c = cx[layer_idx]
if di == 0:
xs_list = xs_next
else:
xs_list = reversed(xs_next)
counter = 0
for x in xs_list:
counter += 1
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
c_rest = None
if layer != 0:
x = dropout.dropout(x, ratio=dropout_ratio)
if counter == 4:
lstm_in = linear.linear(x, xws[layer_idx],
xbs[layer_idx])
else:
lstm_in = linear.linear(
x, xws[layer_idx], xbs[layer_idx]) + linear.linear(
h, hws[layer_idx], hbs[layer_idx])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_list.append(h_bar)
c_list.append(c_bar)
return h, c, h_list, c_list
def _one_directional_loop(di):
# di=0, forward LSTM
# di=1, backward LSTM
h_list = []
c_list = []
layer_idx = direction * layer + di
h = hx[layer_idx]
c = cx[layer_idx]
if di == 0:
xs_list = xs_next
else:
xs_list = reversed(xs_next)
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
c_rest = None
if layer != 0:
x = dropout.dropout(x, ratio=dropout_ratio,
train=train)
lstm_in = linear.linear(x, xws[layer_idx],
xbs[layer_idx]) + \
linear.linear(h, hws[layer_idx], hbs[layer_idx])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_list.append(h_bar)
c_list.append(c_bar)
return h, c, h_list, c_list
def classifier(x, train):
x = A.average_pooling_2d(x, 8)
x = dropout.dropout(x, train=train)
x = self.linear(x)
return x
def _dropout_sequence(xs, dropout_ratio):
return [dropout.dropout(x, ratio=dropout_ratio) for x in xs]
def _dropout(x):
return dropout(x, ratio=0.4)
def _one_directional_loop(di):
# di=0, forward GRU
# di=1, backward GRU
xs_list = xs_next if di == 0 else reversed(xs_next)
layer_idx = direction * layer + di
h = h0[layer_idx]
# h:d_bar_s_1
# h_bar:d_s
'''
print(len(xs_list))
print(len(xs_list[0]))
print(len(xs_list[0][0]))
'''
h_list = []
h_bar_list = []
c_s_list = []
z_s_list = []
for x in xs_list:
batch = x.shape[0]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
else:
h_rest = None
if layer > 0:
x = dropout.dropout(x, ratio=dropout_ratio)
gru_x = linear.linear(x, xws[layer_idx], xbs[layer_idx])
gru_h = linear.linear(h, hws[layer_idx], hbs[layer_idx])
W_r_x, W_z_x, W_x = split_axis.split_axis(gru_x, 3, axis=1)
U_r_h, U_z_h, U_x = split_axis.split_axis(gru_h, 3, axis=1)
r = sigmoid.sigmoid(W_r_x + U_r_h)
z = sigmoid.sigmoid(W_z_x + U_z_h)
h_bar = tanh.tanh(W_x + r * U_x)
h_bar = (1 - z) * h_bar + z * h
phi_d = linear.linear(h_bar, W2, B2)
'''
print(type(phi_ht), len(phi_ht))
print(type(phi_ht[0]), len(phi_ht[0]))
print(type(phi_ht[0][0]), len(phi_ht[0][0]))
print(type(phi_d), len(phi_d))
print(type(phi_d[0]), len(phi_d[0]), phi_d[0].shape)
'''
#phi_ht_len = [t.shape[1] for t in phi_ht]
#phi_ht_section = np.cumsum(phi_ht_len[:-1])
#concat_phi_ht = F.concat(phi_ht, axis=1)
#concat_phi_d = [F.concat([phi_d[i]]*phi_ht_len[i], axis=0) for i in range(batch)]
#concat_phi_d = F.concat(concat_phi_d, axis=0)
#concat_phi_d = F.concat(F.transpose(phi_d), axis=0)
u_st = list(
map(
lambda x, y: reshape.reshape((linear.linear(
x, reshape.reshape(y, (1, len(y))))),
(len(x), )), phi_ht,
phi_d)) #(4)
sum_u = list(map(F.sum, u_st))
alpha_st = list(
map(lambda x, y: x / F.broadcast_to(y, x.shape), u_st,
sum_u)) #(3)
z_s = list(map(F.argmax, alpha_st))
z_s = list(map(lambda x: F.broadcast_to(x, (1, )), z_s))
z_s = F.concat(z_s, axis=0)
'''
print(type(alpha_st),len(alpha_st))
print(type(alpha_st[0]),len(alpha_st[0]))
print(alpha_st[0].shape)
print(ht[0].shape)
'''
c_s = list(
map(
lambda x, y: F.sum(F.broadcast_to(
reshape.reshape(x,
(x.shape[0], 1)), y.shape) * y,
axis=0), alpha_st, ht)) #(2)
c_s_2d = list(
map(lambda x: reshape.reshape(x, (1, len(x))), c_s))
concat_c_s = F.concat(c_s_2d, axis=0)
c_s = list(
map(lambda x: F.broadcast_to(x, (1, len(x))), c_s))
c_s = F.concat(c_s, axis=0)
'''
print(type(c_s), len(c_s))
print(type(c_s[0]), len(c_s[0]), c_s[0].shape)
'''
h = F.relu(
linear.linear(F.concat([concat_c_s, h_bar], axis=1),
W3, B3))
h_list.append(h)
h_bar_list.append(h_bar)
c_s_list.append(c_s)
z_s_list.append(z_s)
#単語数の違いを担保
if h_rest is not None:
h = concat.concat([h, h_rest], axis=0)
h_bar = concat.concat([h_bar, h_rest], axis=0)
return h_list, h_bar_list, c_s_list, z_s_list
def _dropout(x, train):
return dropout(x, ratio=0.4, train=train)
def n_step_lstm(n_layers,
dropout_ratio,
hx,
cx,
ws,
bs,
xs,
train=True,
use_cudnn=True):
"""Stacked Long Short-Term Memory function for sequence inputs.
This function calculates stacked LSTM with sequences. This function gets
an initial hidden state :math:`h_0`, an initial cell state :math:`c_0`,
an input sequence :math:`x`, weight matrices :math:`W`, and bias vectors
:math:`b`.
This function calculates hidden states :math:`h_t` and :math:`c_t` for each
time :math:`t` from input :math:`x_t`.
.. math::
i_t = \sigma(W_0 x_t + W_4 h_{t-1} + b_0 + b_4)
f_t = \sigma(W_1 x_t + W_5 h_{t-1} + b_1 + b_5)
o_t = \sigma(W_2 x_t + W_6 h_{t-1} + b_2 + b_6)
a_t = \tanh(W_3 x_t + W_7 h_{t-1} + b_3 + b_7)
c_t = f_t \dot c_{t-1} + i_t \dot a_t
h_t = o_t \dot \tanh(c_t)
As the function accepts a sequence, it calculates :math:`h_t` for all
:math:`t` with one call. Eight weight matrices and eight bias vectors are
required for each layers. So, when :math:`S` layers exists, you need to
prepare :math:`8S` weigth matrices and :math:`8S` bias vectors.
If the number of layers ``n_layers`` is greather than :math:`1`, input
of ``k``-th layer is hidden state ``h_t`` of ``k-1``-th layer.
Note that all input variables except first layer may have different shape
from the first layer.
Args:
n_layers(int): Number of layers.
dropout_ratio(float): Dropout ratio.
hx (chainer.Variable): Variable holding stacked hidden states.
Its shape is ``(S, B, N)`` where ``S`` is number of layers and is
equal to ``n_layers``, ``B`` is mini-batch size, and ``N`` is
dimention of hidden units.
cx (chainer.Variable): Variable holding stacked cell states.
It has the same shape as ``hx``.
ws (list of list of chainer.Variable): Weight matrices. ``ws[i]``
represents weights for i-th layer.
Each ``ws[i]`` is a list containing eight matrices.
``ws[i][j]`` is corresponding with ``W_j`` in the equation.
Only ``ws[0][j]`` where ``0 <= j < 4`` is ``(I, N)`` shape as they
are multiplied with input variables. All other matrices has
``(N, N)`` shape.
bs (list of list of chainer.Variable): Bias vectors. ``bs[i]``
represnents biases for i-th layer.
Each ``bs[i]`` is a list containing eight vectors.
``bs[i][j]`` is corresponding with ``b_j`` in the equation.
Shape of each matrix is ``(N,)`` where ``N`` is dimention of
hidden units.
xs (list of chainer.Variable): A list of :class:`chainer.Variable`
holding input values. Each element ``xs[t]`` holds input value
for time ``t``. Its shape is ``(B_t, I)``, where ``B_t`` is
mini-batch size for time ``t``, and ``I`` is size of input units.
Note that this functions supports variable length sequences.
When sequneces has different lengths, sort sequences in descending
order by length, and transpose the sorted sequence.
:func:`~chainer.functions.transpose_sequence` transpose a list
of :func:`~chainer.Variable` holding sequence.
So ``xs`` needs to satisfy
``xs[t].shape[0] >= xs[t + 1].shape[0]``.
train (bool): If ``True``, this function executes dropout.
use_cudnn (bool): If ``True``, this function uses cuDNN if available.
Returns:
tuple: This functions returns a tuple concaining three elements,
``hy``, ``cy`` and ``ys``.
- ``hy`` is an updated hidden states whose shape is same as ``hx``.
- ``cy`` is an updated cell states whose shape is same as ``cx``.
- ``ys`` is a list of :class:~chainer.Variable. Each element
``ys[t]`` holds hidden states of the last layer corresponding
to an input ``xs[t]``. Its shape is ``(B_t, N)`` where ``B_t`` is
mini-batch size for time ``t``, and ``N`` is size of hidden
units. Note that ``B_t`` is the same value as ``xs[t]``.
.. seealso::
:func:`chainer.functions.lstm`
"""
xp = cuda.get_array_module(hx, hx.data)
if use_cudnn and xp is not numpy and cuda.cudnn_enabled and \
_cudnn_version >= 5000:
states = get_random_state().create_dropout_states(dropout_ratio)
# flatten all input variables
inputs = tuple(
itertools.chain((hx, cx), itertools.chain.from_iterable(ws),
itertools.chain.from_iterable(bs), xs))
rnn = NStepLSTM(n_layers, states, train=train)
ret = rnn(*inputs)
hy, cy = ret[:2]
ys = ret[2:]
return hy, cy, ys
else:
hx = split_axis.split_axis(hx, n_layers, axis=0, force_tuple=True)
hx = [reshape.reshape(h, h.shape[1:]) for h in hx]
cx = split_axis.split_axis(cx, n_layers, axis=0, force_tuple=True)
cx = [reshape.reshape(c, c.shape[1:]) for c in cx]
xws = [_stack_weight([w[2], w[0], w[1], w[3]]) for w in ws]
hws = [_stack_weight([w[6], w[4], w[5], w[7]]) for w in ws]
xbs = [_stack_weight([b[2], b[0], b[1], b[3]]) for b in bs]
hbs = [_stack_weight([b[6], b[4], b[5], b[7]]) for b in bs]
ys = []
for x in xs:
batch = x.shape[0]
h_next = []
c_next = []
for layer in six.moves.range(n_layers):
h = hx[layer]
c = cx[layer]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
x = dropout.dropout(x, ratio=dropout_ratio, train=train)
h = dropout.dropout(h, ratio=dropout_ratio, train=train)
lstm_in = linear.linear(x, xws[layer], xbs[layer]) + \
linear.linear(h, hws[layer], hbs[layer])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_next.append(h)
c_next.append(c)
x = h_bar
hx = h_next
cx = c_next
ys.append(x)
hy = stack.stack(hx)
cy = stack.stack(cx)
return hy, cy, tuple(ys)
def n_step_lstm(
n_layers, dropout_ratio, hx, cx, ws, bs, xs, train=True,
use_cudnn=True):
"""Stacked Long Short-Term Memory function for sequence inputs.
This function calculates stacked LSTM with sequences. This function gets
an initial hidden state :math:`h_0`, an initial cell state :math:`c_0`,
an input sequence :math:`x`, weight matrices :math:`W`, and bias vectors
:math:`b`.
This function calculates hidden states :math:`h_t` and :math:`c_t` for each
time :math:`t` from input :math:`x_t`.
.. math::
i_t &= \\sigma(W_0 x_t + W_4 h_{t-1} + b_0 + b_4) \\\\
f_t &= \\sigma(W_1 x_t + W_5 h_{t-1} + b_1 + b_5) \\\\
o_t &= \\sigma(W_2 x_t + W_6 h_{t-1} + b_2 + b_6) \\\\
a_t &= \\tanh(W_3 x_t + W_7 h_{t-1} + b_3 + b_7) \\\\
c_t &= f_t \\dot c_{t-1} + i_t \\dot a_t \\\\
h_t &= o_t \\dot \\tanh(c_t)
As the function accepts a sequence, it calculates :math:`h_t` for all
:math:`t` with one call. Eight weight matrices and eight bias vectors are
required for each layers. So, when :math:`S` layers exists, you need to
prepare :math:`8S` weigth matrices and :math:`8S` bias vectors.
If the number of layers ``n_layers`` is greather than :math:`1`, input
of ``k``-th layer is hidden state ``h_t`` of ``k-1``-th layer.
Note that all input variables except first layer may have different shape
from the first layer.
Args:
n_layers(int): Number of layers.
dropout_ratio(float): Dropout ratio.
hx (chainer.Variable): Variable holding stacked hidden states.
Its shape is ``(S, B, N)`` where ``S`` is number of layers and is
equal to ``n_layers``, ``B`` is mini-batch size, and ``N`` is
dimention of hidden units.
cx (chainer.Variable): Variable holding stacked cell states.
It has the same shape as ``hx``.
ws (list of list of chainer.Variable): Weight matrices. ``ws[i]``
represents weights for i-th layer.
Each ``ws[i]`` is a list containing eight matrices.
``ws[i][j]`` is corresponding with ``W_j`` in the equation.
Only ``ws[0][j]`` where ``0 <= j < 4`` is ``(I, N)`` shape as they
are multiplied with input variables. All other matrices has
``(N, N)`` shape.
bs (list of list of chainer.Variable): Bias vectors. ``bs[i]``
represnents biases for i-th layer.
Each ``bs[i]`` is a list containing eight vectors.
``bs[i][j]`` is corresponding with ``b_j`` in the equation.
Shape of each matrix is ``(N,)`` where ``N`` is dimention of
hidden units.
xs (list of chainer.Variable): A list of :class:`~chainer.Variable`
holding input values. Each element ``xs[t]`` holds input value
for time ``t``. Its shape is ``(B_t, I)``, where ``B_t`` is
mini-batch size for time ``t``, and ``I`` is size of input units.
Note that this functions supports variable length sequences.
When sequneces has different lengths, sort sequences in descending
order by length, and transpose the sorted sequence.
:func:`~chainer.functions.transpose_sequence` transpose a list
of :func:`~chainer.Variable` holding sequence.
So ``xs`` needs to satisfy
``xs[t].shape[0] >= xs[t + 1].shape[0]``.
train (bool): If ``True``, this function executes dropout.
use_cudnn (bool): If ``True``, this function uses cuDNN if available.
Returns:
tuple: This functions returns a tuple concaining three elements,
``hy``, ``cy`` and ``ys``.
- ``hy`` is an updated hidden states whose shape is same as ``hx``.
- ``cy`` is an updated cell states whose shape is same as ``cx``.
- ``ys`` is a list of :class:`~chainer.Variable` . Each element
``ys[t]`` holds hidden states of the last layer corresponding
to an input ``xs[t]``. Its shape is ``(B_t, N)`` where ``B_t`` is
mini-batch size for time ``t``, and ``N`` is size of hidden
units. Note that ``B_t`` is the same value as ``xs[t]``.
.. seealso::
:func:`chainer.functions.lstm`
"""
xp = cuda.get_array_module(hx, hx.data)
if use_cudnn and xp is not numpy and cuda.cudnn_enabled and \
_cudnn_version >= 5000:
states = get_random_state().create_dropout_states(dropout_ratio)
# flatten all input variables
inputs = tuple(itertools.chain(
(hx, cx),
itertools.chain.from_iterable(ws),
itertools.chain.from_iterable(bs),
xs))
rnn = NStepLSTM(n_layers, states, train=train)
ret = rnn(*inputs)
hy, cy = ret[:2]
ys = ret[2:]
return hy, cy, ys
else:
hx = split_axis.split_axis(hx, n_layers, axis=0, force_tuple=True)
hx = [reshape.reshape(h, h.shape[1:]) for h in hx]
cx = split_axis.split_axis(cx, n_layers, axis=0, force_tuple=True)
cx = [reshape.reshape(c, c.shape[1:]) for c in cx]
xws = [_stack_weight([w[2], w[0], w[1], w[3]]) for w in ws]
hws = [_stack_weight([w[6], w[4], w[5], w[7]]) for w in ws]
xbs = [_stack_weight([b[2], b[0], b[1], b[3]]) for b in bs]
hbs = [_stack_weight([b[6], b[4], b[5], b[7]]) for b in bs]
ys = []
for x in xs:
batch = x.shape[0]
h_next = []
c_next = []
for layer in six.moves.range(n_layers):
h = hx[layer]
c = cx[layer]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
x = dropout.dropout(x, ratio=dropout_ratio, train=train)
h = dropout.dropout(h, ratio=dropout_ratio, train=train)
lstm_in = linear.linear(x, xws[layer], xbs[layer]) + \
linear.linear(h, hws[layer], hbs[layer])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_next.append(h)
c_next.append(c)
x = h_bar
hx = h_next
cx = c_next
ys.append(x)
hy = stack.stack(hx)
cy = stack.stack(cx)
return hy, cy, tuple(ys)
def _dropout(x):
return dropout(x, ratio=0.4)
def fixed_length_n_step_lstm(
n_layers,
dropout_ratio,
hx,
cx,
ws,
bs,
xs,
train=True,
):
xp = cuda.get_array_module(hx, hx.data)
if xp is not numpy and cuda.cudnn_enabled and _cudnn_version >= 5000:
states = get_random_state().create_dropout_states(dropout_ratio)
# flatten all input variables
inputs = tuple(
itertools.chain((hx, cx), itertools.chain.from_iterable(ws),
itertools.chain.from_iterable(bs), (xs, )))
rnn = FixedLengthNStepLSTMFunction(n_layers, states, train=train)
ret = rnn(*inputs)
hy, cy, ys = ret
_, batch_size, dim = hy.shape
ys_reshape = F.reshape(ys,
(-1, batch_size, dim)) # (length, batch, dim)
return hy, cy, ys_reshape
else:
hx = split_axis.split_axis(hx, n_layers, axis=0, force_tuple=True)
hx = [reshape.reshape(h, h.shape[1:]) for h in hx]
cx = split_axis.split_axis(cx, n_layers, axis=0, force_tuple=True)
cx = [reshape.reshape(c, c.shape[1:]) for c in cx]
xws = [_stack_weight([w[2], w[0], w[1], w[3]]) for w in ws]
hws = [_stack_weight([w[6], w[4], w[5], w[7]]) for w in ws]
xbs = [_stack_weight([b[2], b[0], b[1], b[3]]) for b in bs]
hbs = [_stack_weight([b[6], b[4], b[5], b[7]]) for b in bs]
ys = []
for x in xs:
batch = x.shape[0]
h_next = []
c_next = []
for layer in six.moves.range(n_layers):
h = hx[layer]
c = cx[layer]
if h.shape[0] > batch:
h, h_rest = split_axis.split_axis(h, [batch], axis=0)
c, c_rest = split_axis.split_axis(c, [batch], axis=0)
else:
h_rest = None
x = dropout.dropout(x, ratio=dropout_ratio)
h = dropout.dropout(h, ratio=dropout_ratio)
lstm_in = linear.linear(x, xws[layer], xbs[layer]) + \
linear.linear(h, hws[layer], hbs[layer])
c_bar, h_bar = lstm.lstm(c, lstm_in)
if h_rest is not None:
h = concat.concat([h_bar, h_rest], axis=0)
c = concat.concat([c_bar, c_rest], axis=0)
else:
h = h_bar
c = c_bar
h_next.append(h)
c_next.append(c)
x = h_bar
hx = h_next
cx = c_next
ys.append(x)
hy = stack.stack(hx)
cy = stack.stack(cx)
#return hy, cy, tuple(ys)
ys_concat = F.concat(ys, axis=0)
ys_reshape = F.reshape(
ys_concat,
(-1, ys[0].shape[0], ys[0].shape[1])) # (length, batch, dim)
return hy, cy, ys_reshape
def _dropout(x, train):
return dropout(x, ratio=0.4, train=train)
def _dropout_sequence(xs, dropout_ratio):
return [dropout.dropout(x, ratio=dropout_ratio) for x in xs]