def forward_rnn_encode_proj(self, X):
# Reset rnn state
self.reset_rnn_state()
# Get input shape
in_size, batch_size, in_dim = X.shape
enc_states = X
for currL in range(len(self.rnn_enc)):
for i in range(in_size):
temp_f = F.expand_dims(
F.dropout(self[self.rnn_enc[currL]](enc_states[i]),
ratio=self.cfg["dropout"]["rnn"]), 0)
# if bi-directional
if self.bi_rnn:
temp_r = F.expand_dims(
F.dropout(self[self.rnn_rev_enc[currL]](
enc_states[-1]),
ratio=self.cfg["dropout"]["rnn"]), 0)
if i > 0:
h_fwd = F.concat((h_fwd, temp_f), axis=0)
if self.bi_rnn:
h_rev = F.concat((h_rev, temp_r), axis=0)
else:
h_fwd = temp_f
if self.bi_rnn:
h_rev = temp_r
# end current rnn layer
if self.bi_rnn:
h_rev = F.flipud(h_rev)
rnn_states = F.concat((h_fwd, h_rev), axis=2)
else:
rnn_states = h_fwd
"""
Apply linear projection
"""
# print(f"Applying rnn {currL}")
if currL < (len(self.rnn_enc) - 1):
# print(f"Applying linear linear_proj {currL}")
for i in range(0, in_size):
currH = F.relu(self[f"enc_proj{currL}_bn"](
self[f"enc_proj{currL}"](rnn_states[i])))
if i > 0:
enc_states = F.concat(
(enc_states, F.expand_dims(currH, 0)), axis=0)
else:
enc_states = F.expand_dims(currH, 0)
# end for all hidden states
# end all layers
# Make the batch size as the first dimension
self.enc_states = F.swapaxes(enc_states, 0, 1)
def forward_rnn_encode(self, X):
# Reset rnn state
self.reset_rnn_state()
# Get input shape
in_size, batch_size, in_dim = X.shape
# For each time step
for i in range(in_size):
# Store all hidden states
if i > 0:
h_fwd = F.concat(
(h_fwd, F.expand_dims(self.feed_rnn(X[i], self.rnn_enc),
0)),
axis=0)
if self.bi_rnn:
h_rev = F.concat(
(h_rev,
F.expand_dims(self.feed_rnn(X[-i], self.rnn_rev_enc),
0)),
axis=0)
else:
h_fwd = F.expand_dims(self.feed_rnn(X[i], self.rnn_enc), 0)
if self.bi_rnn:
h_rev = F.expand_dims(
self.feed_rnn(X[-i], self.rnn_rev_enc), 0)
"""
Concatenate fwd and rev RNN hidden states
Flip reverse RNN hidden state order
"""
if self.bi_rnn:
h_rev = F.flipud(h_rev)
rnn_states = F.concat((h_fwd, h_rev), axis=2)
else:
rnn_states = h_fwd
"""
Check if linear projection layer required
"""
self.enc_states = rnn_states
# Make the batch size as the first dimension
self.enc_states = F.swapaxes(self.enc_states, 0, 1)
def forward(self, inputs, device):
x, = inputs
return functions.flipud(x),
def forward(self, inputs, device):
x, = inputs
return functions.flipud(x),
def f(x):
y = functions.flipud(x)
return y * y
def check_forward(self, x_data):
x = chainer.Variable(x_data)
y = functions.flipud(x)
testing.assert_allclose(y.data, numpy.flipud(self.x))