Non-Autoregressive Electron Redistribution Modeling for Reaction Prediction

Authors: Hangrui Bi, Hengyi Wang, Chence Shi, Connor Coley, Jian Tang, Hongyu Guo

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on the USPTO-MIT dataset show that, our approach has established a new state-of-the-art top-1 accuracy and achieves at least 27 times inference speedup over the state-of-the-art methods.
Researcher Affiliation Academia 1Peking University 2Mila Quebec AI Institute 3University of Montr eal 4MIT 5CIFAR AI Research Chair 6HEC Montr eal 7National Research Council Canada.
Pseudocode No The paper describes the framework and components of the model but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statements about open-source code availability or links to code repositories.
Open Datasets Yes We evaluate our model using the popular benchmarking USPTO-MIT dataset. This dataset was created by Jin et al. (2017) by removing duplicate and erroneous reactions from Lowe (2012) s original data and filtering to those with contiguous reaction centers.
Dataset Splits No The paper mentions using the USPTO-MIT dataset and excluding certain reactants from training and testing, implying a split, but does not provide explicit details on the training, validation, or test set sizes or percentages.
Hardware Specification Yes We train the models for 100 epochs with a batch size of 128 using two Nvidia V100 GPUs (took about 3 days). All models are evaluated on a single Nvidia V100 GPU and a Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz with 28 cores.
Software Dependencies No The paper states 'We implement NERF using Pytorch (Paszke et al., 2019)' but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes The node embedding dimension is 256 and the dimension of the latent representation is 64. The model is optimized with Adam W (Kingma & Ba, 2014) optimizer at learning rate 10 4 with linear warmup and linear learning rate decay. We train the models for 100 epochs with a batch size of 128 using two Nvidia V100 GPUs (took about 3 days).