Define the network parameters as shown:

batch_size = 7
window_size = 7
hidden_layer = 256
learning_rate = 0.001

Define placeholders for our input and output:

input = tf.placeholder(tf.float32, [batch_size, window_size, 1])
target = tf.placeholder(tf.float32, [batch_size, 1])

Let's now define all the weights we will use in our LSTM cell.

The weights of the input gate are defined as follows:

U_i = tf.Variable(tf.truncated_normal([1, hidden_layer], stddev=0.05))
W_i = tf.Variable(tf.truncated_normal([hidden_layer, hidden_layer], stddev=0.05))
b_i = tf.Variable(tf.zeros([hidden_layer]))

The weights of the forget gate are defined as follows:

U_f = tf.Variable(tf.truncated_normal([1, hidden_layer], stddev=0.05))
W_f = tf.Variable(tf.truncated_normal([hidden_layer, hidden_layer], stddev=0.05))
b_f = tf.Variable(tf.zeros([hidden_layer]))

The weights of the output gate are defined as given:

U_o = tf.Variable(tf.truncated_normal([1, hidden_layer], stddev=0.05))
W_o = tf.Variable(tf.truncated_normal([hidden_layer, hidden_layer], stddev=0.05))
b_o = tf.Variable(tf.zeros([hidden_layer]))

The weights of the candidate state are defined as follows:

U_g = tf.Variable(tf.truncated_normal([1, hidden_layer], stddev=0.05))
W_g = tf.Variable(tf.truncated_normal([hidden_layer, hidden_layer], stddev=0.05))
b_g = tf.Variable(tf.zeros([hidden_layer]))

The output layer weight is given as follows:

V = tf.Variable(tf.truncated_normal([hidden_layer, 1], stddev=0.05))
b_v = tf.Variable(tf.zeros([1]))