class ResidualBiRNNBuilder:
"""
A residual network with bidirectional first layer
"""
def __init__(self, num_layers, input_dim, hidden_dim, add_to_output=False, dropout=None, exp_global=Ref(Path("exp_global"))):
assert num_layers > 1
assert hidden_dim % 2 == 0
self.forward_layer = UniLSTMSeqTransducer(exp_global=exp_global, input_dim=input_dim, hidden_dim=hidden_dim/2, dropout=dropout)
self.backward_layer = UniLSTMSeqTransducer(exp_global=exp_global, input_dim=input_dim, hidden_dim=hidden_dim/2, dropout=dropout)
self.residual_network = ResidualRNNBuilder(exp_global=exp_global, num_layers=num_layers - 1, input_dim=hidden_dim, hidden_dim=hidden_dim,
add_to_output=add_to_output, dropout=dropout)
def get_final_states(self):
return self._final_states
def add_inputs(self, es):
return PseudoState(self, self.transduce(es))
def transduce(self, es):
forward_e = self.forward_layer(es)
backward_e = self.backward_layer(ReversedExpressionSequence(es))
self._final_states = [FinalTransducerState(dy.concatenate([self.forward_layer.get_final_states()[0].main_expr(),
self.backward_layer.get_final_states()[0].main_expr()]),
dy.concatenate([self.forward_layer.get_final_states()[0].cell_expr(),
self.backward_layer.get_final_states()[0].cell_expr()]))]
output = self.residual_network.transduce(ExpressionSequence(expr_list=[dy.concatenate([f,b]) for f,b in zip(forward_e, ReversedExpressionSequence(backward_e))]))
self._final_states += self.residual_network.get_final_states()
return output
def initial_state(self):
return PseudoState(self)
class ResidualBiRNNBuilder(object):
"""
A residual network with bidirectional first layer.
Note: This is currently not serializable and therefore can't be directly specified in the config file; use :class:`xnmt.residual.ResidualLSTMSeqTransducer` instead.
Args:
num_layers (int): number of layers
input_dim (int): input dimension; if None, use exp_global.default_layer_dim
hidden_dim (int): hidden dimension; if None, use exp_global.default_layer_dim
add_to_output (bool): whether to add a residual connection to the output layer
dropout (float): dropout probability; if None, use exp_global.dropout
exp_global (ExpGlobal): ExpGlobal object to acquire DyNet params and global settings. By default, references the experiment's top level exp_global object.
"""
def __init__(self, num_layers, input_dim, hidden_dim, add_to_output=False, dropout=None, exp_global=Ref(Path("exp_global"))):
assert num_layers > 1
assert hidden_dim % 2 == 0
self.forward_layer = UniLSTMSeqTransducer(exp_global=exp_global, input_dim=input_dim, hidden_dim=hidden_dim//2, dropout=dropout)
self.backward_layer = UniLSTMSeqTransducer(exp_global=exp_global, input_dim=input_dim, hidden_dim=hidden_dim//2, dropout=dropout)
self.residual_network = ResidualRNNBuilder(exp_global=exp_global, num_layers=num_layers - 1, input_dim=hidden_dim, hidden_dim=hidden_dim,
add_to_output=add_to_output, dropout=dropout)
def get_final_states(self):
return self._final_states
def add_inputs(self, es):
return PseudoState(self, self.transduce(es))
def transduce(self, es):
forward_e = self.forward_layer(es)
backward_e = self.backward_layer(ReversedExpressionSequence(es))
self._final_states = [FinalTransducerState(dy.concatenate([self.forward_layer.get_final_states()[0].main_expr(),
self.backward_layer.get_final_states()[0].main_expr()]),
dy.concatenate([self.forward_layer.get_final_states()[0].cell_expr(),
self.backward_layer.get_final_states()[0].cell_expr()]))]
output = self.residual_network.transduce(ExpressionSequence(expr_list=[dy.concatenate([f,b]) for f,b in zip(forward_e, ReversedExpressionSequence(backward_e))]))
self._final_states += self.residual_network.get_final_states()
return output
def initial_state(self):
return PseudoState(self)