Discrete Tree Flows via Tree-Structured Permutations
Authors: Mai Elkady, Jim Lim, David I. Inouye
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically demonstrate the feasibility of our method on multiple datasets. and 4. Experiments To test our approach, we compare different results including the negative log likelihood (NLL), training time (TT), and the number of model parameters (NP), against models for discrete flows that are suitable for categorical data and that allow for discrete latent spaces to be achieved by training by optimizing the exact likelihood. |
| Researcher Affiliation | Academia | 1Department of Computer Science, Purdue University 2Department of Electrical and Computer Engineering, Purdue University. |
| Pseudocode | Yes | Algorithm 1 Construct Tree: Learn node splits and count data at leaves, Algorithm 2 Find Best Split: Find best split of data, Algorithm 3 Learn Local Permutations: Learn local permutations to sort local counts, Algorithm 4 Construct Equivalent Tree: Construct a new tree with equivalent splits and permutations based on local permutations |
| Open Source Code | No | The paper cites and modifies a third-party open-source code (Bricken, 2021) for baseline models but does not explicitly state that the authors' own implementation of Discrete Tree Flows (DTF) or their modifications are open-sourced or provide a link to them. |
| Open Datasets | Yes | Mushroom Dataset This dataset includes different attributes of mushrooms1,and has n = 8, 124, d = 22 and the maximum number of categories in any column is k = 12. 1https://archive.ics.uci.edu/ml/datasets/Mushroom and Vectorized binary MNIST Dataset We then investigate the performance of our algorithm for the the vectorized and binarized MNIST dataset (Deng, 2012) and Genetic Dataset Specifically, we use a dataset of n = 2504 individuals with d = 805 sampled single nucleotide polymorphism (SNPs) that was made available by (Yelmen et al., 2021). |
| Dataset Splits | Yes | We choose to draw 12,800 samples in total, reserving 10,240 samples for training and 2,560 for testing. and We present results for the real-world datasets averaged across 3 folds for MNIST and Genetic data and across 5 folds for the mushroom data in Table 3. |
| Hardware Specification | Yes | All the timing results we report are from running our experiments on an Intel Core TM i9-10920X 3.50GHz for CPU, and Nvidia Ge Force RTXTM 3090 for GPU. |
| Software Dependencies | No | The paper mentions 'PyTorch Discrete Flows' in the context of a baseline implementation (Bricken, 2021) that was modified, but it does not specify version numbers for PyTorch or any other software dependencies used for their own DTF model. |
| Experiment Setup | Yes | AF and BF details: For the AF and BF models, we use a linear layer architecture where 4 flow layers were used for the synthetic data experiments, and 6 flow layers were used for the mushroom, MNIST, and genetic data experiments experiments. We ran for a total of 200 epochs sampling 250 samples in each epoch. and DDF s details: We used the mlp architecture for all experiments of that we tested for the DDF model, and varied the number of coupling layers and hidden layers across the different experiments. We trained all models using 10 epochs for both the prior and the neural network training since increasing this number further didn t lead to a significant improvement in the results. and Table 7 titled The model parameters for the results presented in tables 2 and 3. listing specific n H, n C, n TSP, M, alpha, beta, init values. |