Top-N: Equivariant Set and Graph Generation without Exchangeability

Authors: Clement Vignac, Pascal Frossard

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, our method outperforms i.i.d. generation by 15% at Set MNIST reconstruction, by 33% at object detection on CLEVR, generates sets that are 74% closer to the true distribution on a synthetic molecule-like dataset, and generates more valid molecules on QM9.
Researcher Affiliation Academia Cl ement Vignac, Pascal Frossard LTS4, EPFL Lausanne, Switzerland
Pseudocode No The paper provides mathematical equations for the Top-n creation module but does not include a formally labeled 'Algorithm' or 'Pseudocode' block.
Open Source Code Yes Source code is available at github.com/cvignac/Top-N
Open Datasets Yes We first perform experiments on the Set MNIST benchmark, introduced in Zhang et al. (2019)... We further benchmark Top-n on object detection with the CLEVR dataset, made of 70k training images and 15k validation images... Finally, we evaluate Top-n on a graph generation task. We train a graph VAE (detailed in Appendix D.3) on QM9 molecules...
Dataset Splits Yes We further benchmark Top-n on object detection with the CLEVR dataset, made of 70k training images and 15k validation images...
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU/CPU models, memory specifications).
Software Dependencies No The paper mentions general software components like 'Transformer layers', 'MLP', 'PNA global pooling layer', and 'Adam' optimizer, but does not provide specific version numbers for any software or libraries.
Experiment Setup Yes TSPN was therefore trained for 100 epochs with a learning rate of 5e-4, and DSPN with a learning rate of 1e-4 for 200 epochs... We use a learning rate of 2e 4 and a scheduler that halves it when reconstruction performance does not improve significantly after 750 epochs... The reference set contains 35 points... The model is trained over 600 epochs with a batch size of 512 and a learning rate of 2e 3. It is halved after 100 epochs when the loss does not improve anymore. The reference set has 12 points.