def __init__(self,
input_shape,
output_dim,
hidden_dim,
name=None,
hidden_nonlinearity=tf.nn.relu,
output_w_init=ly.XavierUniformInitializer(),
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_x_init=ly.XavierUniformInitializer(),
recurrent_w_h_init=ly.OrthogonalInitializer(),
lstm_layer_cls=ly.LSTMLayer,
output_nonlinearity=None,
input_var=None,
input_layer=None,
forget_bias=1.0,
use_peepholes=False,
layer_args=None):
with tf.variable_scope(name, 'LSTMNetwork'):
if input_layer is None:
l_in = ly.InputLayer(shape=(None, None) + input_shape,
input_var=input_var,
name='input')
else:
l_in = input_layer
l_step_input = ly.InputLayer(shape=(None, ) + input_shape,
name='step_input')
# contains previous hidden and cell state
l_step_prev_state = ly.InputLayer(shape=(None, hidden_dim * 2),
name='step_prev_state')
if layer_args is None:
layer_args = dict()
l_lstm = lstm_layer_cls(l_in,
num_units=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
gate_nonlinearity=recurrent_nonlinearity,
hidden_init_trainable=False,
name='lstm_layer',
forget_bias=forget_bias,
cell_init_trainable=False,
w_x_init=recurrent_w_x_init,
w_h_init=recurrent_w_h_init,
use_peepholes=use_peepholes,
**layer_args)
l_lstm_flat = ly.ReshapeLayer(l_lstm,
shape=(-1, hidden_dim),
name='lstm_flat')
l_output_flat = ly.DenseLayer(l_lstm_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=output_w_init,
name='output_flat')
l_output = ly.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.stack(
(tf.shape(l_input)[0], tf.shape(l_input)[1], -1))),
shape_op=lambda flat_output_shape, l_input_shape:
(l_input_shape[0], l_input_shape[1], flat_output_shape[-1]),
extras=[l_in],
name='output')
l_step_state = l_lstm.get_step_layer(l_step_input,
l_step_prev_state,
name='step_state')
l_step_hidden = ly.SliceLayer(l_step_state,
indices=slice(hidden_dim),
name='step_hidden')
l_step_cell = ly.SliceLayer(l_step_state,
indices=slice(hidden_dim, None),
name='step_cell')
l_step_output = ly.DenseLayer(l_step_hidden,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=l_output_flat.w,
b=l_output_flat.b,
name='step_output')
self._l_in = l_in
self._hid_init_param = l_lstm.h0
self._cell_init_param = l_lstm.c0
self._l_lstm = l_lstm
self._l_out = l_output
self._l_step_input = l_step_input
self._l_step_prev_state = l_step_prev_state
self._l_step_hidden = l_step_hidden
self._l_step_cell = l_step_cell
self._l_step_state = l_step_state
self._l_step_output = l_step_output
self._hidden_dim = hidden_dim
def __init__(self,
input_shape,
output_dim,
hidden_dim,
name=None,
hidden_nonlinearity=tf.nn.relu,
lstm_layer_cls=ly.LSTMLayer,
output_nonlinearity=None,
input_var=None,
input_layer=None,
forget_bias=1.0,
use_peepholes=False,
layer_args=None):
with tf.variable_scope(name, "LSTMNetwork"):
if input_layer is None:
l_in = ly.InputLayer(
shape=(None, None) + input_shape,
input_var=input_var,
name="input")
else:
l_in = input_layer
l_step_input = ly.InputLayer(
shape=(None, ) + input_shape, name="step_input")
# contains previous hidden and cell state
l_step_prev_state = ly.InputLayer(
shape=(None, hidden_dim * 2), name="step_prev_state")
if layer_args is None:
layer_args = dict()
l_lstm = lstm_layer_cls(
l_in,
num_units=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
hidden_init_trainable=False,
name="lstm_layer",
forget_bias=forget_bias,
cell_init_trainable=False,
use_peepholes=use_peepholes,
**layer_args)
l_lstm_flat = ly.ReshapeLayer(
l_lstm, shape=(-1, hidden_dim), name="lstm_flat")
l_output_flat = ly.DenseLayer(
l_lstm_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output_flat")
l_output = ly.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.stack((tf.shape(l_input)[0], tf.shape(l_input)[1], -1))
),
shape_op=lambda flat_output_shape, l_input_shape: (
l_input_shape[0], l_input_shape[1], flat_output_shape[-1]),
extras=[l_in],
name="output")
l_step_state = l_lstm.get_step_layer(
l_step_input, l_step_prev_state, name="step_state")
l_step_hidden = ly.SliceLayer(
l_step_state, indices=slice(hidden_dim), name="step_hidden")
l_step_cell = ly.SliceLayer(
l_step_state,
indices=slice(hidden_dim, None),
name="step_cell")
l_step_output = ly.DenseLayer(
l_step_hidden,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=l_output_flat.w,
b=l_output_flat.b,
name="step_output")
self._l_in = l_in
self._hid_init_param = l_lstm.h0
self._cell_init_param = l_lstm.c0
self._l_lstm = l_lstm
self._l_out = l_output
self._l_step_input = l_step_input
self._l_step_prev_state = l_step_prev_state
self._l_step_hidden = l_step_hidden
self._l_step_cell = l_step_cell
self._l_step_state = l_step_state
self._l_step_output = l_step_output
self._hidden_dim = hidden_dim
def __init__(self,
input_shape,
output_dim,
hidden_dim,
name=None,
hidden_nonlinearity=tf.nn.relu,
output_w_init=ly.XavierUniformInitializer(),
recurrent_nonlinearity=tf.nn.sigmoid,
recurrent_w_x_init=ly.XavierUniformInitializer(),
recurrent_w_h_init=ly.OrthogonalInitializer(),
gru_layer_cls=ly.GRULayer,
output_nonlinearity=None,
input_var=None,
input_layer=None,
layer_args=None):
with tf.variable_scope(name, 'GRUNetwork'):
if input_layer is None:
l_in = ly.InputLayer(shape=(None, None) + input_shape,
input_var=input_var,
name='input')
else:
l_in = input_layer
l_step_input = ly.InputLayer(shape=(None, ) + input_shape,
name='step_input')
l_step_prev_state = ly.InputLayer(shape=(None, hidden_dim),
name='step_prev_state')
if layer_args is None:
layer_args = dict()
l_gru = gru_layer_cls(l_in,
num_units=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
gate_nonlinearity=recurrent_nonlinearity,
hidden_init_trainable=False,
w_x_init=recurrent_w_x_init,
w_h_init=recurrent_w_h_init,
name='gru',
**layer_args)
l_gru_flat = ly.ReshapeLayer(l_gru,
shape=(-1, hidden_dim),
name='gru_flat')
l_output_flat = ly.DenseLayer(l_gru_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=output_w_init,
name='output_flat')
l_output = ly.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.stack(
(tf.shape(l_input)[0], tf.shape(l_input)[1], -1))),
shape_op=lambda flat_output_shape, l_input_shape:
(l_input_shape[0], l_input_shape[1], flat_output_shape[-1]),
extras=[l_in],
name='output')
l_step_state = l_gru.get_step_layer(l_step_input,
l_step_prev_state,
name='step_state')
l_step_hidden = l_step_state
l_step_output = ly.DenseLayer(l_step_hidden,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=l_output_flat.w,
b=l_output_flat.b,
name='step_output')
self._l_in = l_in
self._hid_init_param = l_gru.h0
self._l_gru = l_gru
self._l_out = l_output
self._l_step_input = l_step_input
self._l_step_prev_state = l_step_prev_state
self._l_step_hidden = l_step_hidden
self._l_step_state = l_step_state
self._l_step_output = l_step_output
self._hidden_dim = hidden_dim
def __init__(self,
input_shape,
output_dim,
hidden_dim,
name=None,
hidden_nonlinearity=tf.nn.relu,
gru_layer_cls=ly.GRULayer,
output_nonlinearity=None,
input_var=None,
input_layer=None,
layer_args=None):
with tf.variable_scope(name, "GRUNetwork"):
if input_layer is None:
l_in = ly.InputLayer(
shape=(None, None) + input_shape,
input_var=input_var,
name="input")
else:
l_in = input_layer
l_step_input = ly.InputLayer(
shape=(None, ) + input_shape, name="step_input")
l_step_prev_state = ly.InputLayer(
shape=(None, hidden_dim), name="step_prev_state")
if layer_args is None:
layer_args = dict()
l_gru = gru_layer_cls(
l_in,
num_units=hidden_dim,
hidden_nonlinearity=hidden_nonlinearity,
hidden_init_trainable=False,
name="gru",
**layer_args)
l_gru_flat = ly.ReshapeLayer(
l_gru, shape=(-1, hidden_dim), name="gru_flat")
l_output_flat = ly.DenseLayer(
l_gru_flat,
num_units=output_dim,
nonlinearity=output_nonlinearity,
name="output_flat")
l_output = ly.OpLayer(
l_output_flat,
op=lambda flat_output, l_input: tf.reshape(
flat_output,
tf.stack((tf.shape(l_input)[0], tf.shape(l_input)[1], -1))
),
shape_op=lambda flat_output_shape, l_input_shape: (
l_input_shape[0], l_input_shape[1], flat_output_shape[-1]),
extras=[l_in],
name="output")
l_step_state = l_gru.get_step_layer(
l_step_input, l_step_prev_state, name="step_state")
l_step_hidden = l_step_state
l_step_output = ly.DenseLayer(
l_step_hidden,
num_units=output_dim,
nonlinearity=output_nonlinearity,
w=l_output_flat.w,
b=l_output_flat.b,
name="step_output")
self._l_in = l_in
self._hid_init_param = l_gru.h0
self._l_gru = l_gru
self._l_out = l_output
self._l_step_input = l_step_input
self._l_step_prev_state = l_step_prev_state
self._l_step_hidden = l_step_hidden
self._l_step_state = l_step_state
self._l_step_output = l_step_output
self._hidden_dim = hidden_dim
