NAOMI

Code for NeurIPS 2019 paper titled NAOMI: Non-Autoregressive Multiresolution Sequence Imputation

Code is written with PyTorch v0.4.1 (Python 3.6.5). Billiards data can be downloaded here, basketball data is available from STATS.

To train the model:

First open visdom, then adjust hyperparameters in train_model.sh and run the shell file.

Detailed explanations of hyperparameters:

• --model: “NAOMI” or “SingleRes”

• --task: “basketball” or “billiard”

• --y_dim: 10 for basketball and 2 for billiard

• --rnn_dim and --n_layers: gru cell size for all models, including forward and backward rnns

• --dec1_dim to --dec16_dim: For NAOMI, these values correspond to dimensions of different decoders. For SingleRes, only dec1_dim is used for decoder.

• --pre_start_lr: initial learning rate for supervised pretrain

• --pretrain: supervised pretrain epochs

• --highest: largest stepsize for NAOMI decoders, should be 2^n

• --discrim_rnn_dim and --discrim_layers: discriminator rnn size

• --policy_learning_rate: learning rate for generator in adversarial training

• --discrim_learning_rate: learning rate for discriminator in adversarial training

• --pretrain_disc_iter: number of iterations to pretrain discriminator

• --max_iter_num: number of adversarial training iterations

Citation

If you find this repository, e.g., the code and the datasets, useful in your research, please cite the following paper:

@inproceedings{liu2019naomi,
  title={NAOMI: Non-Autoregressive Multiresolution Sequence Imputation},
  author={Liu, Yukai and Yu, Rose and Zheng, Stephan and Zhan, Eric and Yue, Yisong},
  booktitle={Advances in Neural Information Processing Systems(NeurIPS '19)},
  year={2019}
}