def birnn_dynamic(x, config, training, attention=False):
# reshape outputs to [batch_size, max_time_steps, n_features]
#max_time = tf.shape(x)[1]
#rnn_inputs = tf.reshape(x, [-1, max_time, config.embedding_dim])
#rnn_inputs = x
rnn_inputs = tf.reshape(x, [-1, config.max_seq_len, config.n_features])
sequence_lengths = Diagnostics.length(rnn_inputs)
init = tf.contrib.layers.xavier_initializer()
# Choose rnn cell type
if config.rnn_cell == 'lstm':
args = {
'num_units': config.hidden_units,
'forget_bias': 1.0,
'state_is_tuple': True
}
base_cell = tf.nn.rnn_cell.LSTMCell
elif config.rnn_cell == 'gru':
args = {'num_units': config.hidden_units}
base_cell = tf.nn.rnn_cell.GRUCell
elif config.rnn_cell == 'layer_norm':
args = {
'num_units': config.hidden_units,
'forget_bias': 1.0,
'dropout_keep_prob': config.recurrent_keep_prob
}
base_cell = tf.contrib.rnn.LayerNormBasicLSTMCell
cell = base_cell
if config.output_keep_prob < 1:
# rnn_inputs = tf.nn.dropout(rnn_inputs, self.keep_prob)
fwd_cells = [
tf.nn.rnn_cell.DropoutWrapper(
cell(**args),
output_keep_prob=config.output_keep_prob,
variational_recurrent=True,
dtype=tf.float32) for _ in range(config.rnn_layers)
]
bwd_cells = [
tf.nn.rnn_cell.DropoutWrapper(
cell(**args),
output_keep_prob=config.output_keep_prob,
variational_recurrent=True,
dtype=tf.float32) for _ in range(config.rnn_layers)
]
else:
fwd_cells = [cell(**args) for _ in range(config.rnn_layers)]
bwd_cells = [cell(**args) for _ in range(config.rnn_layers)]
fwd_cells = tf.contrib.rnn.MultiRNNCell(fwd_cells)
bwd_cells = tf.contrib.rnn.MultiRNNCell(bwd_cells)
outputs, output_states = tf.nn.bidirectional_dynamic_rnn(
cell_fw=fwd_cells,
cell_bw=bwd_cells,
inputs=rnn_inputs,
sequence_length=sequence_lengths,
dtype=tf.float32,
parallel_iterations=128)
birnn_output = tf.concat(outputs, 2)
if attention: # invoke soft attention mechanism - attend to different particles
summary_vector = attention_A(birnn_output,
config.attention_dim,
my_method=False)
else: # Select last relevant output
summary_vector = Diagnostics.last_relevant(birnn_output,
sequence_lengths)
print(summary_vector.get_shape().as_list())
# Fully connected layer for classification
with tf.variable_scope("fc"):
logits_RNN = tf.layers.dense(summary_vector,
units=config.n_classes,
kernel_initializer=init)
return logits_RNN
