LSTM inputs for Tensorflow

Question

I'm trying to create an LSTM network in Tensorflow and I'm lost in terminology/basics. I have n time series examples so X=x_n, where x_i=[[x₁¹x₁²,x₁³],...,[x_m¹x_m²,x_m³]] and where x_iⁱ is a float. First of all I want to train a model that given the start of a sequence ([x₁¹x₁²,x₁³]) I can predict the rest of the sequence. Then later I hope to include a classifier to predict which binary class each x_i belongs to.

So my problem is what do I feed in to the start and pull out the end of my model? So far I have something that looks like the below

class ETLSTM(object):
    """docstring for ETLSTM"""
    def __init__(self, isTraining, config):
        super(ETLSTM, self).__init__()

        # This needs to be tidied
        self.batchSize = batchSize = config.batchSize
        self.numSteps = numSteps = config.numSteps
        self.numInputs = numInputs = config.numInputs
        self.numLayers = numLayers = config.numLayers

        lstmSize = config.lstm_size
        DORate = config.keep_prob

        self.input_data = tf.placeholder(tf.float32, [batchSize, numSteps,
                                                      numInputs])
        self.targets = tf.placeholder(tf.float32, [batchSize, numSteps,
                                                   numInputs])
        lstmCell = rnn_cell.BasicLSTMCell(lstmSize, forgetbias=0.0)
        if(isTraining and DORate < 1):
            lstmCell = tf.nn.rnn_cell.DropoutWrapper(lstmCell,
                                                     output_keep_prob=DORate)
        cell = tf.nn.rnn_cell.MultiRNNCell([lstmCell]*numLayers)

        self._initial_state = cell.zero_state(batchSize, tf.float32)

        # This won't work with my data, need to find what goes in...
        with tf.device("/cpu:0"):
            embedding = tf.get_variable("embedding", [vocab_size, size])
            inputs = tf.nn.embedding_lookup(embedding, self._input_data)

        if(isTraining and DORate < 1):
            inputs = tf.nn.dropout(inputs, DORate)

EDIT: Specifically, how to I finish the __init__ function so that it is compatible with my data?

Answer 1

An RNN predicts the value of N+1 given the values from 1 to N so far. (LSTM is just one way to implement an RNN cell.)

The short answer is:

• train your model using back propagation on your complete sequences [[x₁¹x₁²,x₁³],...,[x_m¹x_m²,x_m³]]
• run your trained model forward on the start of your sequence [x₁¹x₁²,x₁³,...] then sample from the model to predict the rest of your sequence [x_m¹x_m²,x_m³,...].

The longer answer is:

Your example just shows the initialization of the model. You also need to implement a training function to run back propagation as well as a sample function that predicts the results.

The following code snippets are mix & match and are for illustration purposes only...

For training just feed in your complete sequences with start + rest in your data iterator.

For example in the sample code tensorflow/models/rnn/ptb_word_lm.py the training loop computes a cost function for batches of input_data against targets (which are the input_data shifted by one timestep)

        # compute a learning rate decay
        session.run(tf.assign(self.learning_rate_variable, learning_rate))

        logger.info("Epoch: %d Learning rate: %.3f" % (i + 1, session.run(self.learning_rate_variable)))


        """Runs the model on the given data."""
        epoch_size = ((len(training_data) // self.batch_size) - 1) // self.num_steps
        costs = 0.0
        iters = 0
        state = self.initial_state.eval()
        for step, (x, y) in enumerate(self.data_iterator(training_data, self.batch_size, self.num_steps)):

            # x and y should have shape [batch_size, num_steps]
            cost, state, _ = session.run([self.cost_function, self.final_state, self.train_op],
                                     {self.input_data: x,
                                      self.targets: y,
                                      self.initial_state: state})
            costs += cost
            iters += self.num_steps

Note the data iterator in tensorflow/models/rnn/reader.py returns the input data as 'x' and the targets as 'y' which are just shifted one step forward from x. (You would need to create a data iterator like this that packages your set of training sequences.)

def ptb_iterator(raw_data, batch_size, num_steps):
  raw_data = np.array(raw_data, dtype=np.int32)

  data_len = len(raw_data)
  batch_len = data_len // batch_size
  data = np.zeros([batch_size, batch_len], dtype=np.int32)
  for i in range(batch_size):
    data[i] = raw_data[batch_len * i:batch_len * (i + 1)]

  epoch_size = (batch_len - 1) // num_steps

  if epoch_size == 0:
    raise ValueError("epoch_size == 0, decrease batch_size or num_steps")

  for i in range(epoch_size):
    x = data[:, i*num_steps:(i+1)*num_steps]
    y = data[:, i*num_steps+1:(i+1)*num_steps+1]
    yield (x, y)

After training, you run the model forward to make predictions for sequences by feeding in the start of your sequence start_x=[X1, X2, X3,...]...this snippets assumes binary values representing classes, you'd have to adjust the sampling function for float values.

def sample(self, sess, num=25, start_x):

    # return state tensor with batch size 1 set to zeros, eval
    state = self.rnn_layers.zero_state(1, tf.float32).eval()

    # run model forward through the start of the sequence
    for char in start_x:

        # create a 1,1 tensor/scalar set to zero
        x = np.zeros((1, 1))

        # set to the vocab index
        x[0, 0] = char


        # fetch: final_state
        # input_data = x, initial_state = state
        [state] = sess.run([self.final_state], {self.input_data: x, self.initial_state:state})

    def weighted_pick(weights):

        # an array of cummulative sum of weights
        t = np.cumsum(weights)

        # scalar sum of tensor
        s = np.sum(weights)

        # randomly selects a value from the probability distribution
        return(int(np.searchsorted(t, np.random.rand(1)*s)))

    # PREDICT REST OF SEQUENCE
    rest_x = []

    # get last character in init
    char = start_x[-1]

    # sample next num chars in the sequence after init
    score = 0.0

    for n in xrange(num):

        # init input to zeros
        x = np.zeros((1, 1))

        # lookup character index
        x[0, 0] = char

        # probs = tf.nn.softmax(self.logits)
        # fetch: probs, final_state
        # input_data = x, initial_state = state
        [probs, state] = sess.run([self.output_layer, self.final_state], {self.input_data: x, self.initial_state:state})

        p = probs[0]
        logger.info("output=%s" % np.shape(p))
        # sample = int(np.random.choice(len(p), p=p))

        # select a random value from the probability distribution
        sample = weighted_pick(p)
        score += p[sample]
        # look up the key with the index
        logger.debug("sample[%d]=%d" % (n, sample))
        pred = self.vocabulary[sample]
        logger.debug("pred=%s" % pred)

        # add the car to the output
        rest_x.append(pred) 

        # set the next input character
        char = pred
    return rest_x, score