Denoising Likelihood Score Matching for Conditional Score-based Data Generation

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

The code implementation for the motivational example (the toy experiment) is in chen-hao-chao/dlsm-toy.

Pretrained Weights

Model	Dataset	Checkpoint
Classifier	Cifar-10	Cifar10/ours/checkpoint.pth
Classifier	Cifar-10	Cifar10/base/checkpoint.pth
Classifier	Cifar-100	Cifar100/ours/checkpoint.pth
Classifier	Cifar-100	Cifar100/base/checkpoint.pth
Score Model	Cifar-10	Cifar10/score/checkpoint.pth
Score Model	Cifar-100	Cifar100/score/checkpoint.pth

Dependencies

Install the necessary python packages through the following command.

pip3 install -r requirements.txt

Training and Evaluation Procedure

Please follow the below procedure to reproduce this work:

Training: (1) Train a score model. (2) Load the pretrained score model and train a classifier.

Evaluation: (3) Sample data points using run_lib_sampling.py. (4) Evaluate the performance using run_lib_evaluation.py and run_lib_CAS.py.

The details about these training and evaluation procedures are elaborated below.

Usage

Train score models

Use the following command to train a score model with default setting on the Cifar-10 dataset:

python3 main.py --config configs/score/cifar10_score_configs.py --model score --workdir score_cifar10

Use the following command to train a score model with default setting on the Cifar-100 dataset:

python3 main.py --config configs/score/cifar100_score_configs.py --model score --workdir score_cifar100

Train classifiers

Use the following command to train a classifier with default setting on the Cifar-10 dataset:

python3 main.py --config configs/classifier/cifar10_classifier_resnet18_cond_configs.py --model classifier --workdir classifier_cifar10_ours_resnet18_cond

Use the following command to train a classifier with default setting on the Cifar-100 dataset:

python3 main.py --config configs/classifier/cifar100_classifier_resnet34_cond_configs.py --model classifier --workdir classifier_cifar100_ours_resnet34_cond 

Please make sure that the pretrained score model is placed at sampling.score_restore_path specified in the configuration file for the above two settings. It is also possible to add an additional argument --restore {PATH_TO_SCORE_CHECKPOINT} to identify sampling.score_restore_path. For example, the following command describes the training procedure of DLSM with the score model placed at results/score_cifar10/checkpoints/checkpoint.pth:

python3 main.py --config configs/classifier/cifar10_classifier_resnet18_cond_configs.py --model classifier --workdir classifier_cifar10_ours_resnet18_cond --restore results/score_cifar10/checkpoints/checkpoint.pth

Sampling (run_lib_sampling.py)

Sampling with default setting with models pretrained on the Cifar-10 dataset:

python3 run_lib_sampling.py --config configs/eval/cifar10/eval_cifar10_configs.py --mode full --workdir classifier_cifar10_ours_resnet18_cond

Sampling with scaling factor 10:

python3 run_lib_sampling.py --config configs/eval/cifar10/eval_cifar10_configs.py --mode full --workdir classifier_cifar10_ours_resnet18_cond --scale 10

Sampling class-wisely with class index from 5 (inclusive) to 7 (exclusive):

python3 run_lib_sampling.py --config configs/eval/cifar10/eval_cifar10_configs.py --mode class --classid 5 --classidend 7 --workdir classifier_cifar10_ours_resnet18_cond

Sampling with default setting using models pretrained on the Cifar-100 dataset:

python3 run_lib_sampling.py --config configs/eval/cifar100/eval_cifar100_configs.py --mode full --workdir classifier_cifar100_ours_resnet34_cond

Please make sure that the pretrained score model and classifier are placed at model.score_restore_path and model.classifier_restore_path specified in the configuration file and the arguments in main.py for the above settings. It is also possible to add additional arguments --restore {PATH_TO_SCORE_CHECKPOINT} and --restore_classifier {PATH_TO_CLASSIFIER_CHECKPOINT} to identify model.score_restore_path and model.classifier_restore_path. For example, the following command describes the sampling procedure with a pretrained score model and a pretrained classifier.

python3 run_lib_sampling.py --config configs/eval/cifar10/eval_cifar10_configs.py --mode full --workdir classifier_cifar10_ours_resnet18_cond --restore results/score_cifar10/checkpoints/checkpoint.pth --restore results/classifier_cifar100_ours_resnet34_cond/checkpoints/checkpoint.pth

Evaluatation (run_lib_evaluation.py)

Examine the sampling quality (FID/IS/PRDC) with default setting:

python3 run_lib_evaluation.py --config configs/eval/cifar10/eval_cifar10_configs.py --workdir classifier_cifar10_ours_resnet18_cond --stat --latent --fidis --prdc --mode full 

• The flag stat indicates whether to create stats files. If it is set to True, the images from a dataset will be stacked class-wisely and the resulting vector will be saved in {WORKDIR}/stat/stat.pnz.
• The flag latent indicates whether to create latent files. If it is set to True, the data vector stored in {WORKDIR}/stat/stat.pnz will be encoded as latents and saved in {WORKDIR}/stat/stat.pnz. Meanwhile, the generated samples will also be encoded as latents and saved.
• The flag fidis indicates whether to calculate the FID / IS metrics. If it is set to True, the FID / IS metrics will be calculated and reported.
• The flag prdc indicates whether to calculate the P / R / D / C metrics. If it is set to True, the P / R / D / C metrics will be calculated and reported.
• The flag mode indicates the evaluation mode, which can be full or class. If it is set to full, the metrics will be calculated unconditionally. If it is set to class, the metrics will be calculated conditionally, i.e., the CW metrics in our paper.

Evaluatation (run_lib_evaluation_cas.py)

Examine the sampling quality (CAS) with default setting:

python3 run_lib_evaluation_cas.py --config configs/eval/cifar10/eval_cifar10_configs_cas.py --workdir classifier_cifar10_ours_resnet18_cond --setup ours

Please make sure the generated samples are placed at {WORKDIR}/results/classifier_cifar10_{SETUP}_resnet18_cond/sample for Cifar-10, and {WORKDIR}/results/classifier_cifar100_{SETUP}_resnet34_cond/sample for Cifar-100. It is also possible to change the paths specified in samples/samples, i.e., SAMPLES_DIR_BASE, SAMPLES_DIR_OURS, SAMPLES_DIR_SCALE.

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

To maintain reproducibility, we freezed the following repository and list its license below:

• yang-song/score_sde_pytorch (at commit 1618dde) is licensed under the Apache-2.0 License

Further changes based on the repository above are licensed under the Apache-2.0 License.