Syntax-Directed Variational Autoencoder for Structured Data
Authors: Hanjun Dai, Yingtao Tian, Bo Dai, Steven Skiena, Le Song
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed model with applications in programming language and molecules, including reconstruction and program/molecule optimization. The results demonstrate the effectiveness in incorporating syntactic and semantic constraints in discrete generative models, which is significantly better than current state-of-the-art approaches. |
| Researcher Affiliation | Collaboration | 1 College of Computing, Georgia Institute of Technology 2 Department of Computer Science, Stony Brook University 3 Ant Financial |
| Pseudocode | Yes | Algorithm 1 Decoding with Stochastic Syntax-Directed Decoder |
| Open Source Code | Yes | Code is available at https://github.com/Hanjun-Dai/sdvae. |
| Open Datasets | Yes | For molecule experiments, we use the same dataset as in Kusner et al. (2017). It contains 250, 000 SMILES strings, which are extracted from the ZINC database (G omez-Bombarelli et al., 2016). |
| Dataset Splits | Yes | We hold 2000 programs out for testing and the rest for training and validation. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running its experiments. |
| Software Dependencies | No | The paper mentions implementing the model in 'Py Torch framework' but does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | In encoder, we use 3-layer one-dimension convolution neural networks (CNNs) followed by a full connected layer, whose output would be fed into two separate affine layers for producing µ and σ respectively as in reparameterization trick; and in decoder we use 3-layer RNNs followed by a affine layer activated by softmax that gives probability for each production rule. In detail, we use 56 dimensions the latent space and the dimension of layers as the same number as in Kusner et al. (2017). |