Many to One RNN with Variable Sequence Length:

    In this tutorial we implement


    Fig1. Unfolded representation of the implemented RNN structure


    0. Import the required libraries:

    We will start with importing the required libraries to our Python environment.

    In [1]:
    # imports
    import tensorflow as tf
    import numpy as np
    ImportError                               Traceback (most recent call last)
    <ipython-input-1-2df283633649> in <module>()
          1 # imports
    ----> 2 import tensorflow as tf
          3 import numpy as np
    ImportError: No module named tensorflow 

    1. Generate some data

    For this tutorial ...

    1.1. Data dimension

    Here, we specify the dimensions of the data samples which will be used in the code. Defining these variables makes it easier (compared with using hard-coded number all throughout the code) to modify them later. Ideally these would be inferred from the data that has been read, but here we just write the numbers.

    In [ ]:
    # Data Dimensions
    input_dim = 1           # input dimension
    seq_max_len = 4         # sequence maximum length
    out_dim = 1             # output dimension

    1.2. Generate data and display the sizes

    Now we can use the defined helper function in "train" mode which loads the train and validation images and their corresponding labels. We'll also display their sizes:

    In [ ]:
    def generate_data(count=1000, max_length=4, dim=1):
        x = np.random.randint(0, 10, size=(count, max_length, dim))
        length = np.random.randint(1, max_length+1, count)
        for i in range(count):
            x[i, length[i]:, :] = 0
        y = np.sum(x, axis=1)
        return x, y, length
    In [ ]:
    x_train, y_train, seq_len_train = generate_data(count=1000, max_length=seq_max_len, dim=input_dim)
    x_test, y_test, seq_len_test = generate_data(count=5, max_length=seq_max_len, dim=input_dim)
    print("Size of:")
    print("- Training-set:\t\t{}".format(len(y_train)))
    print("- Test-set:\t{}".format(len(y_test)))

    To get batches of samples:

    In [ ]:
    def next_batch(x, y, seq_len, batch_size):
        N = x.shape[0]
        batch_indeces = np.random.permutation(N)[:batch_size]
        x_batch = x[batch_indeces]
        y_batch = y[batch_indeces]
        seq_len_batch = seq_len[batch_indeces]
        return x_batch, y_batch, seq_len_batch

    2. Hyperparameters

    In [ ]:
    # Parameters
    learning_rate = 0.01    # The optimization initial learning rate
    training_steps = 10000  # Total number of training steps
    batch_size = 10         # batch size
    display_freq = 1000     # Frequency of displaying the training results

    2. Hyperparameters

    In [ ]:
    learning_rate = 0.001 # The optimization initial learning rate
    epochs = 10           # Total number of training epochs
    batch_size = 100      # Training batch size
    display_freq = 100    # Frequency of displaying the training results

    3. Network configuration

    In [ ]:
    num_hidden_units = 10   # number of hidden units

    4. Create network helper functions

    4.1. Helper functions for creating new variables

    In [ ]:
    # weight and bais wrappers
    def weight_variable(shape):
        Create a weight variable with appropriate initialization
        :param name: weight name
        :param shape: weight shape
        :return: initialized weight variable
        initer = tf.truncated_normal_initializer(stddev=0.01)
        return tf.get_variable('W',
    def bias_variable(shape):
        Create a bias variable with appropriate initialization
        :param name: bias variable name
        :param shape: bias variable shape
        :return: initialized bias variable
        initial = tf.constant(0., shape=shape, dtype=tf.float32)
        return tf.get_variable('b',

    4.2. Helper-function for creating a RNN

    In [ ]:
    def RNN(x, weights, biases, n_hidden, seq_max_len, seq_len):
        :param x: inputs of shape [batch_size, max_time, input_dim]
        :param weights: matrix of fully-connected output layer weights
        :param biases: vector of fully-connected output layer biases
        :param n_hidden: number of hidden units
        :param seq_max_len: sequence maximum length
        :param seq_len: length of each sequence of shape [batch_size,]
        cell = tf.nn.rnn_cell.BasicRNNCell(n_hidden)
        outputs, states = tf.nn.dynamic_rnn(cell, x, sequence_length=seq_len, dtype=tf.float32)
        # Hack to build the indexing and retrieve the right output.
        batch_size = tf.shape(outputs)[0]
        # Start indices for each sample
        index = tf.range(0, batch_size) * seq_max_len + (seq_len - 1)
        # Indexing
        outputs = tf.gather(tf.reshape(outputs, [-1, n_hidden]), index)
        out = tf.matmul(outputs, weights) + biases
        return out

    5. Create the network graph

    5.1. Placeholders for the inputs (x), sequence length (seqLen), and corresponding labels (y)

    In [ ]:
    # Placeholders for inputs(x), input sequence lengths (seqLen) and outputs(y)
    x = tf.placeholder(tf.float32, [None, seq_max_len, input_dim])
    seqLen = tf.placeholder(tf.int32, [None])
    y = tf.placeholder(tf.float32, [None, 1])

    5.2. Define the network

    In [ ]:
    # create weight matrix initialized randomly from N~(0, 0.01)
    W = weight_variable(shape=[num_hidden_units, out_dim])
    # create bias vector initialized as zero
    b = bias_variable(shape=[out_dim])
    # Network predictions
    pred_out = RNN(x, W, b, num_hidden_units, seq_max_len, seqLen)

    5.3. Define the loss function and optimizer

    In [ ]:
    # Define the loss function (i.e. mean-squared error loss) and optimizer
    cost = tf.reduce_mean(tf.square(pred_out - y))
    train_op = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

    5.4. Initialize all variables

    In [ ]:
    # Creating the op for initializing all variables
    init = tf.global_variables_initializer()

    6. Train

    In [ ]:
    with tf.Session() as sess:
        for i in range(training_steps):
            x_batch, y_batch, seq_len_batch = next_batch(x_train, y_train, seq_len_train, batch_size)
            _, mse =[train_op, cost], feed_dict={x: x_batch, y: y_batch, seqLen: seq_len_batch})
            if i % display_freq == 0:
                print('Step {0:<6}, MSE={1:.4f}'.format(i, mse))

    7. Test

    7.1. Helper functions for plotting the results

    In [ ]:
        # Test
        y_pred =, feed_dict={x: x_test, seqLen: seq_len_test})
        print('--------Test Results-------')
        for i, x in enumerate(y_test):
            print("When the ground truth output is {}, the model thinks it is {}"
                  .format(y_test[i], y_pred[i]))

    © 2018 Easy-TensorFlow team. All Rights Reserved.