Simple RNN in TensorFlow

The workflow to define and train a simple RNN in TensorFlow is as follows:

1. Define the hyper-parameters for the model:

state_size = 4
n_epochs = 100
n_timesteps = n_x 
learning_rate = 0.1

The new hyper-parameter here is the state_size. The state_size represents the number of weight vectors of an RNN cell.

2. Define the placeholders for X and Y parameters for the model. The shape of X placeholder is (batch_size, number_of_input_timesteps, number_of_inputs) and the shape of Y placeholder is (batch_size, number_of_output_timesteps, number_of_outputs). For batch_size, we use None so that we can input the batch of any size later.

X_p = tf.placeholder(tf.float32, [None, n_timesteps, n_x_vars], 
    name='X_p') 
Y_p = tf.placeholder(tf.float32, [None, n_timesteps, n_y_vars], 
    name='Y_p')

3. Transform the input placeholder X_p into a list of tensors of length equal to the number of time steps, which is n_x or 1 in this example:

# make a list of tensors of length n_timesteps
rnn_inputs...