model.py

def model(batch_x, seq_length, dropout):
    def clipped_relu(x):
        return tf.minimum(tf.nn.relu(x), relu_clip)
    
    with slim.arg_scope([slim.variable], device="/cpu:0"):
        with slim.arg_scope([slim.fully_connected], activation_fn=clipped_relu):
            with slim.arg_scope([slim.dropout], keep_prob=(1.0 - dropout)):
                fc_1 = slim.dropout(slim.fully_connected(batch_x, n_hidden_1))
                fc_2 = slim.dropout(slim.fully_connected(fc_1, n_hidden_2))
                fc_3 = slim.dropout(slim.fully_connected(fc_2, n_hidden_3))
                
                lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(n_cell_dim, forget_bias=1.0, state_is_tuple=True)
                
                rnn_4, _ = tf.nn.bidirectional_dynamic_rnn(cell_fw=lstm_cell,
                                                           cell_bw=lstm_cell,
                                                           inputs=fc_3,
                                                           dtype=tf.float32,
                                                           sequence_length=seq_length)
                rnn_4 = tf.concat(2, rnn_4)
                
                fc_5 = slim.dropout(slim.fully_connected(rnn_4, n_hidden_5))
                logits = slim.fully_connected(fc_5, n_hidden_6)
                
                # Reshape final layer to be time major as CTC expects                
                logits = tf.transpose(logits, [1, 0, 2])
                
                return logits