Incremental Sampling Without Replacement for Sequence Models

Authors: Kensen Shi, David Bieber, Charles Sutton

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally demonstrate that using Unique Randomizer leads to higher-quality samples in program synthesis and combinatorial optimization.
Researcher Affiliation Industry 1Google. Correspondence to: Kensen Shi <kshi@google.com>.
Pseudocode Yes Algorithm 1: Using Unique Randomizer to sample outputs of P without replacement. Algorithm 2: Random choice operation and trie construction for Unique Randomizer.
Open Source Code Yes Our Python implementations of Unique Randomizer, its batched version, Stochastic Beam Search, and the Hindsight Gumbel Estimator can be found at https://github. com/google-research/unique-randomizer.
Open Datasets Yes We use the Search-based Pseudocode to Code (SPo C) dataset (Kulal et al., 2019)...
Dataset Splits No The paper mentions 'test splits' but does not specify training, validation, and test splits with explicit percentages or sample counts, nor does it refer to predefined validation splits.
Hardware Specification No The paper mentions running experiments 'with a GPU' but does not specify the model, type, or any other detailed hardware specifications.
Software Dependencies No The paper states 'Our Python implementations' but does not provide specific version numbers for Python or any associated libraries/dependencies.
Experiment Setup Yes We use a temperature τ = 0.8 when sampling with replacement and τ = 0.4 when sampling without replacement, which we found to be the best among {0.2, 0.4, 0.6, 0.8, 1.0}. For TSP...We set τ = 0.3, 0.2, and 0.15 for n = 20, 50, and 100 nodes, respectively.