Download this tutorial

Sequence Generation with Beam Search Sampler and Sequence Sampler

This tutorial demonstrates how to sample sequences using a pre-trained language model in the following two ways: with a beam search sampler and with a sequence sampler.

Let’s use V to denote the vocabulary size, and T to denote the sequence length. Given a language model, we can sample sequences according to the probability that they would occur according to our model. At each time step, a language model predicts the likelihood of each word occurring, given the context from prior time steps. The outputs at any time step can be any word from the vocabulary whose size is V and thus the number of all possible outcomes for a sequence of length T is thus



While sometimes we might want to sample sentences according to their probability of occurring, at other times we want to find the sentences that are most likely to occur. This is especially true in the case of language translation where we don’t just want to see a translation. We want the best translation. While finding the optimal outcome quickly becomes intractable as time increases, there are still many ways to sample reasonably good sequences. GluonNLP provides two samplers for generating from a language model: BeamSearchSampler and SequenceSampler.

Loading a pre-trained language model (LM)

Firstly, let’s load a pre-trained language model, from which we will sample sequences. GluonNLP makes this a painless process.

import mxnet as mx
import gluonnlp as nlp

ctx = mx.cpu()
lm_model, vocab = nlp.model.get_model(name='awd_lstm_lm_1150',
Vocab file is not found. Downloading.
Downloading /root/.mxnet/models/3823008510234756239/ from
Downloading /root/.mxnet/models/awd_lstm_lm_1150_wikitext-2-f9562ed0.zip7a8a1657-337b-40ab-83ac-676c2f2d3af7 from

Sampling a Sequence with BeamSearchSampler

To overcome the exponential complexity in sequence decoding, beam search decodes greedily, keeping those sequences that are most likely based on the probability up to the current time step. The size of this subset is called the beam size. Suppose the beam size is K and the output vocabulary size is V. When selecting the beams to keep, the beam search algorithm first predicts all possible successor words from the previous K beams, each of which has V possible outputs. This becomes a total of K*V paths. Out of these K*V paths, beam search ranks them by their score keeping only the top K paths.

Let’s take a look how to construct a BeamSearchSampler. The nlp.model.BeamSearchSampler class takes the following arguments for customization and extension:

  • beam_size : the beam size

  • decoder : callable function of the one-step-ahead decoder

  • eos_id : the id of the EOS token

  • scorer: the score function used in beam search

  • max_length: the maximum search length

For beam search to work, we need a scorer function.

The scorer function

In this tutorial, we will use the BeamSearchScorer as the scorer function, which implements the scoring function with length penalty in the Google NMT paper:

scorer = nlp.model.BeamSearchScorer(alpha=0, K=5, from_logits=False)

Defining the scorer is as simple as this one line.

The decoder function

Next, we define the decoder based on the pre-trained language model.

class LMDecoder(object):
    def __init__(self, model):
        self._model = model
    def __call__(self, inputs, states):
        outputs, states = self._model(mx.nd.expand_dims(inputs, axis=0), states)
        return outputs[0], states
    def state_info(self, *arg, **kwargs):
        return self._model.state_info(*arg, **kwargs)
decoder = LMDecoder(lm_model)

Beam Search Sampler

Given a scorer and a decoder, we are ready to create a sampler. We use the symbol . to indicate the end of sentence (EOS). We can use vocab to get the index of the EOS to then feed the index to the sampler. The following code shows how to construct a beam search sampler. We will create a sampler with 4 beams and a maximum sample length of 20.

eos_id = vocab['.']
beam_sampler = nlp.model.BeamSearchSampler(beam_size=5,

It’s really that simple!

Sampling a Sequence with SequenceSampler

The previous generation results may look a bit boring. Instead, let’s now use the sequence sampler to get relatively more interesting results.

A SequenceSampler samples from the contextual multinomial distribution produced by the language model at each time step. Since we may want to control how “sharp” the distribution is to tradeoff diversity with correctness, we can use the temperature option in SequenceSampler, which controls the temperature of the softmax activation function.

For each input, sequence sampler can sample multiple independent sequences at once. The number of independent sequences to sample can be specified through the argument beam_size.

Defining the SequenceSampler is as simple as this:

seq_sampler = nlp.model.SequenceSampler(beam_size=5,

Generate Sequences with Sequence Sampler

Now, instead of using the beam sampler for our generate_sequences function, we can use the SequenceSampler instead to sample sequences based on the same inputs used previously.

generate_sequences(seq_sampler, inputs, begin_states, 5)
Generation Result:
['I love it and enjoy one of their series , Schafer or Clive .', -56.322582]
['I love it in a television vein and rid him of his adventures .', -46.985336]
['I love it for news .', -12.868879]
['I love it ; it is not until the end of the year that a relative <unk> to a EEC owner is now raised .', -68.12103]
['I love it in an AADT .', -20.544458]

Et voila! We’ve generated the most likely sentences based on our given input.

Exercises for the keen reader

  • Tweak alpha and K in BeamSearchScorer, how are the results changed? Does it do relatively better or worse than the sequence SequenceSampler?

  • Try different samples to decode and figure out which results the BeamSearchSampler does better than the SequenceSampler