def createnet():
return createmodel(
rnn_layer_layers=RNN_HIDDEN_LAYERS,
isbrnn=RNN_IS_BIDIRECTIONAL,
batch_size=BATCH_SIZE,
n_features=N_FEATURES,
n_classes=N_CLASSES,
layer_type_rnn=RNN_LAYER_TYPE,
padded_seq_len=MAX_SEQ_LENGTH,
input_layers=INPUT_LAYERS,
input_layer_dropout=INPUT_DROPOUT_RATE,
output_layers=OUTPUT_LAYERS,
final_output_layer=OUTPUT,
learn_init=LEARN_INIT,
dropout_rnn=RNN_DROPOUT_RATE,
output_layer_dropout=OUTPUT_DROPOUT_RATE,
unittype_rnn=UNITTYPE_RNN,
relucap=RELU_CAP,
reluleakyness=RELU_LEAKYNESS)
l_in, num_units=102, peepholes=peepholes, learn_init=True)
recout = lasagne.layers.BidirectionalLSTMLayer(
recout, num_units=156, peepholes=peepholes, learn_init=True)
l_reshape = lasagne.layers.ReshapeLayer(
recout, (BATCH_SIZE*MAX_SEQ_LENGTH, recout.get_output_shape()[-1]))
l_rec_out = lasagne.layers.DenseLayer(
l_reshape, num_units=N_CLASSES, nonlinearity=lasagne.nonlinearities.softmax)
l_out = lasagne.layers.ReshapeLayer(
l_rec_out, (BATCH_SIZE, MAX_SEQ_LENGTH, N_CLASSES))
else:
l_out, _, l_in = createmodel(rnn_layer_layers=[3,4,5],
isbrnn=True,
batch_size=BATCH_SIZE,
n_features=N_FEATURES,
n_classes=N_CLASSES,
layer_type_rnn="LSTMFAST",
padded_seq_len=MAX_SEQ_LENGTH,
output_layers=[7],
input_layers=None,
learn_init=True,
final_output_layer="softmax")
# createnet in LSTMTrainingFunctions can setup networks with a number of
# different architectures.
# Cross entropy cost function.
# Note that we use the mask to ignore masked sequences during cost calculation
def costfun(p_y_given_x, y, mask,db='COST:'):
shape = (BATCH_SIZE*MAX_SEQ_LENGTH, N_CLASSES)
y_reshape = y.flatten()
