UniqueRandomizer

Overview

UniqueRandomizer is a data structure for sampling outputs of a randomized program, such as a neural sequence model, incrementally and without replacement.

• Incremental sampling: Instead of sampling a large batch of outputs all at once, as with beam search, UniqueRandomizer provides samples one at a time. This enables flexibility in stopping criteria, such as stopping the sampling process as soon as a satisfactory output is found.
• Sampling without replacement: In many applications, a neural model is used to produce candidate solutions to some search or optimization problem. In such applications it is usually desirable to consider unique candidate solutions, since duplicates are typically not useful.

For more details, refer to our paper, Incremental Sampling Without Replacement for Sequence Models, published at ICML 2020.

BibTeX entry:

@article{shi2020uniquerandomizer, title = {Incremental Sampling Without Replacement for Sequence Models}, author = {Kensen Shi and David Bieber and Charles Sutton}, booktitle = {Proceedings of the 37th International Conference on Machine Learning}, year = {2020} }

Installation

python3 -m pip install --user unique-randomizer

This package requires Python 3. The above command automatically installs the following dependencies as well:

• absl-py >= 0.6.1
• numpy >= 1.15.4
• scipy >= 1.1.0

Usage

To use UniqueRandomizer, first identify the program or function that you wish to draw unique samples from, such as the draw_sample function in the following example:

def draw_sample(sequence_model, state): """Draws a sample (a sequence of token indices) from the sequence model.""" tokens = [] token = BOS for i in range(MAX_LEN): probs, state = sequence_model(token, state) token = np.random.choice(np.arange(len(probs)), p=probs) if token == EOS: break tokens.append(token) return tokens

Note that draw_sample can take inputs and can use control flow such as loops, conditionals, and recursion. There are only two constraints on the draw_sample function:

1. It must be deterministic given the inputs, except for random choices provided by np.random.choice (or some other method of selecting a random index given a discrete probability distribution).
2. Two different sequences of random choices must lead to draw_sample returning different outputs.

Next, add a UniqueRandomizer object as an input to draw_sample, and use its sample_distribution function to replace np.random.choice:

diff - def draw_sample(sequence_model, state): + def draw_sample(sequence_model, state, randomizer): """Draws a sample (a sequence of token indices) from the sequence model.""" tokens = [] token = BOS for i in range(MAX_LEN): probs, state = sequence_model(token, state) - token = np.random.choice(np.arange(len(probs)), p=probs) + token = randomizer.sample_distribution(probs) if token == EOS: break tokens.append(token) return tokens

Finally, a simple loop around draw_sample can collect unique samples, as follows:

def draw_unique_samples(model, state, num_samples): """Draws multiple unique samples from the sequence model.""" samples = [] randomizer = unique_randomizer.UniqueRandomizer() for _ in range(num_samples): samples.append(draw_sample(model, state, randomizer)) randomizer.mark_sequence_complete() return samples

Code Samples

We include a few code samples that demonstrate how to use UniqueRandomizer:

• examples/weighted_coin_flips.py: This provides a very simple example of using UniqueRandomizer. The function flip_two_weighted_coins simulates flipping a pair of weighted coins. The sample_flips_without_replacement function then uses UniqueRandomizer to efficiently sample outputs of flip_two_weighted_coins without replacement.

• examples/expand_grammar.py: This defines a Probabilistic Context-Free Grammar (PCFG), as well as methods to sample elements of the grammar without replacement by using UniqueRandomizer, rejection sampling, and Stochastic Beam Search (SBS). The script examples/expand_grammar_main.py enables easy comparison between the different sampling methods under different scenarios.

• examples/sequence_example.py: This implements sampling without replacement from a sequence model, using UniqueRandomizer, Batched UniqueRandomizer, rejection sampling, and SBS. The script examples/sequence_example_main.py enables easy comparison between the different sampling methods under different scenarios.

Disclaimer

This is not an officially supported Google product.