Denoising Likelihood Score Matching for Conditional Score-based Data Generation

This repository includes the official implementation for the motivational experiments in the paper Denoising Likelihood Score Matching for Conditional Score-based Data Generation.

The code implementation for the experiments on the real-world datasets (Cifar-10 and Cifar-100) is in chen-hao-chao/dlsm.

Dependencies

Install the necessary python packages through the following command.

pip install -r requirements.txt

Usage

Train and evaluate the model using the following command.

python main.py --mode {train / eval} --workdir {$directory} --config {$configuration} 

• --mode: train for training eval for evaluation.
• --workdir: the directory created for saving the experimental results such as visualized examples and checkpoints.
• --config: the configuration file that specifies the hyper-parameters.

Examples

• Train the score model

python main.py --workdir score_cond --mode train --config configs/score/inter_twinning_moon_configs.py 
python main.py --workdir score_cond_lower --mode train --config configs/score/inter_twinning_moon_configs_lower.py
python main.py --workdir score_cond_upper --mode train --config configs/score/inter_twinning_moon_configs_upper.py

• Train the classifier

python main.py --workdir classifier_cond_total --mode train --config configs/classifier/inter_twinning_moon_configs_total.py
python main.py --workdir classifier_cond_ce --mode train --config configs/classifier/inter_twinning_moon_configs_ce.py
python main.py --workdir classifier_cond_dlsm --mode train --config configs/classifier/inter_twinning_moon_configs_dlsm.py

Please note that for the setups involving the total loss or the ce loss, please make sure the pretrained score models are placed at training.score_restore_path identified in the configuration file.

Please note that you could add a --plot flag to save the training information for plotting (Fig. 5 in the manuscript). By doing so, run_lib_classifier_plot.py will be executed, and the training loss will be saved in the tensorboard folder.

Please note that noise_conditioned in the configuration files is defaultly turned on for training noise-conditioned score models and classifiers. In particular, weighting_ce and weighting_dlsm in the configuration files refer to the balancing factors for different noise levels.

• Evaluate

python main.py --workdir test_eval --mode eval --config configs/eval/inter_twinning_moon_configs.py

Please note that type in the configuration file controls the evaluation modes. For type = distance, the L2 distances between the estimated scores and the oracle scores will be calculated. For type = sampling, the sampling quality will be evatuated.

Citing DLSM

If you find this code useful, please consider citing our paper.

@inproceedings{
  chao2022denoising,
  title={Denoising Likelihood Score Matching for Conditional Score-based Data Generation},
  author={Chen-Hao Chao and Wei-Fang Sun and Bo-Wun Cheng and Yi-Chen Lo and Chia-Che Chang and Yu-Lun Liu and Yu-Lin Chang and Chia-Ping Chen and Chun-Yi Lee},
  booktitle={International Conference on Learning Representations},
  year={2022},
  url={https://openreview.net/forum?id=LcF-EEt8cCC}
}

License

The code in this repository is licensed under the Apache-2.0 License.