def build_deep_lstm_tbptt(input_var,
mask_var,
input_dim,
num_layers,
num_units,
num_proj_units,
output_dim,
batch_size,
context,
grad_clipping,
is_bidir,
use_layer_norm,
ivector_dim,
ivector_var=None,
backward_on_top=False):
input_layer, mask_layer = build_input_layer(input_dim, input_var, mask_var)
if ivector_var:
ivector_layer = build_ivector_layer(ivector_dim, ivector_var)
input_layer = concatenate_layers(input_layer, ivector_layer)
prev_input_layer = input_layer
tbptt_layers = []
for layer_idx in range(1, num_layers + 1):
fwd_layer = TBPTTLSTMLayer(incoming=prev_input_layer,
mask_input=mask_layer,
num_units=num_units,
batch_size=batch_size,
grad_clipping=grad_clipping,
backwards=False,
num_proj_units=num_proj_units,
use_layer_norm=use_layer_norm,
context=context)
tbptt_layers.append(fwd_layer)
if backward_on_top and layer_idx < num_layers:
prev_input_layer = fwd_layer
continue
if is_bidir:
bwd_layer = LSTMLayer(incoming=prev_input_layer,
mask_input=mask_layer,
num_units=num_units,
grad_clipping=grad_clipping,
backwards=True,
num_proj_units=num_proj_units,
use_layer_norm=use_layer_norm)
prev_input_layer = ConcatLayer(incomings=[fwd_layer, bwd_layer],
axis=-1)
else:
prev_input_layer = fwd_layer
return build_sequence_dense_layer(input_var, prev_input_layer,
output_dim), tbptt_layers
def build_deep_uni_lhuclstm_seqsum_tbptt(layer_name,
input_var,
mask_var,
input_dim,
num_layers,
num_units,
num_proj_units,
output_dim,
num_pred_layers,
num_pred_units,
num_seqsum_units,
num_seqsum_layers,
seqsum_output_dim,
batch_size,
grad_clipping,
use_layer_norm,
use_ivector_input,
ivector_dim,
ivector_var=None):
input_layer, mask_layer = build_input_layer(input_dim, input_var, mask_var)
speaker_layer = None
if ivector_var:
speaker_layer = build_ivector_layer(ivector_dim, ivector_var)
if use_ivector_input:
input_layer = concatenate_layers(input_layer, speaker_layer)
speaker_layer = build_sequence_summarizing_layer(
input_var,
input_layer,
num_units=num_seqsum_units,
num_layers=num_seqsum_layers,
output_dim=seqsum_output_dim)
prev_input_layer = input_layer
tbptt_layers = []
for layer_idx in range(1, num_layers + 1):
prev_fwd_layer = SeqSumLHUCLSTMLayer(prev_input_layer,
speaker_layer,
num_units,
num_pred_units,
num_pred_layers,
batch_size,
backwards=False,
grad_clipping=grad_clipping,
mask_input=mask_layer,
num_proj_units=num_proj_units,
use_layer_norm=use_layer_norm)
prev_input_layer = prev_fwd_layer
tbptt_layers.append(prev_fwd_layer)
return build_sequence_dense_layer(input_var, prev_input_layer,
output_dim), speaker_layer, tbptt_layers
def build_deep_lhuclstm_ivector(layer_name, input_var, mask_var, input_dim,
num_layers, num_units, num_proj_units,
output_dim, num_pred_layers, num_pred_units,
num_seqsum_units, num_seqsum_layers,
seqsum_output_dim, batch_size, grad_clipping,
is_bidir, use_layer_norm, use_ivector_input,
ivector_dim, ivector_var):
input_layer, mask_layer = build_input_layer(input_dim, input_var, mask_var)
speaker_layer = None
if ivector_var:
speaker_layer = build_ivector_layer(ivector_dim, ivector_var)
if use_ivector_input:
input_layer = concatenate_layers(input_layer, speaker_layer)
prev_input_layer = input_layer
for layer_idx in range(1, num_layers + 1):
prev_fwd_layer = IVectorLHUCLSTMLayer(prev_input_layer,
speaker_layer,
num_units,
num_pred_units,
num_pred_layers,
batch_size,
backwards=False,
grad_clipping=grad_clipping,
mask_input=mask_layer,
num_proj_units=num_proj_units,
use_layer_norm=use_layer_norm)
if is_bidir:
prev_bwd_layer = IVectorLHUCLSTMLayer(
prev_input_layer,
speaker_layer,
num_units,
num_pred_units,
num_pred_layers,
batch_size,
backwards=True,
grad_clipping=grad_clipping,
mask_input=mask_layer,
num_proj_units=num_proj_units,
use_layer_norm=use_layer_norm)
prev_input_layer = ConcatLayer(
incomings=[prev_fwd_layer, prev_bwd_layer], axis=-1)
else:
prev_input_layer = prev_fwd_layer
return build_sequence_dense_layer(input_var, prev_input_layer,
output_dim), speaker_layer
def build_deep_lstm(input_var,
mask_var,
input_dim,
num_layers,
num_units,
num_proj_units,
output_dim,
grad_clipping,
is_bidir,
use_layer_norm,
ivector_dim,
ivector_var=None):
input_layer, mask_layer = build_input_layer(input_dim, input_var, mask_var)
if ivector_var:
ivector_layer = build_ivector_layer(ivector_dim, ivector_var)
input_layer = concatenate_layers(input_layer, ivector_layer)
prev_input_layer = input_layer
for layer_idx in range(1, num_layers + 1):
fwd_layer = LSTMLayer(incoming=prev_input_layer,
mask_input=mask_layer,
num_units=num_units,
grad_clipping=grad_clipping,
backwards=False,
num_proj_units=num_proj_units,
use_layer_norm=use_layer_norm)
if is_bidir:
bwd_layer = LSTMLayer(incoming=prev_input_layer,
mask_input=mask_layer,
num_units=num_units,
grad_clipping=grad_clipping,
backwards=True,
num_proj_units=num_proj_units,
use_layer_norm=use_layer_norm)
prev_input_layer = ConcatLayer(incomings=[fwd_layer, bwd_layer],
axis=-1)
else:
prev_input_layer = fwd_layer
return build_sequence_dense_layer(input_var, prev_input_layer, output_dim)