2022-2 캡스톤디자인1

KHUVCM

- Color Space Convertion Neural Compression

Inatall compressai -> pip install compressai

Directory

./save -> checkpoint (.pth) file saved
./log  -> tensorboard log files saved
./data -> you can change

Dataset Preparation

Vimeo90k (82GB) for training (http://data.csail.mit.edu/tofu/dataset/vimeo_septuplet.zip)
Kodak24
For downloading Vimeo90K dataset, recommend using 'Axel' for parallel precessing

train_mean_scale_hyperprior.py -> trainig

test_mean_scale_hyperprior.py -> test

custom_model.py -> You can customizing Mean-Scale Hyperprior from compressai project

load_model.py -> customed load_model function from compressai project (for custom model)

1. train

   1. CUDA_VISIBLE_DEVICES={your_gpu_number} python train.py --quality {1-8}

   2. for each quality level, lambda is determined

2. test

   1. CUDA_VISIBLE_DEVICES={your_gpu_number} python test_mean_scale_hyperprior.py --quality {1-8} --checkpoint {your model dir}

   2. with no checkpoint, pre-trained model loaded

   3. validation is performed with first 1000 categories in Vimeo90K datasets

3. Arguments read som arguments in code for customizing

4. scheduling ReduceLROnPlateau is used and if learning rate is lower than 1e-6,training process stop early and save the last step.

5. best model is also saved but recommand to use last saved model weights

6. Model customizing

   1. forward() in CustomMeanScaleHyperprior class is used in training

   2. compress() and decompress() in CustomMeanScaleHyperprior class is used in inference (real entropy coding)

7. Configuration "-q", "--quality", type=int, default=0, help="quality of the model" "-save_dir", "--save_dir", type=str, default='save/', help="save_dir" "-log_dir", "--log_dir", type=str, default='log/', help="log_dir" "-total_step", default=5000000, type=int, help="total_step (default: %(default)s)" "-test_step", "--test_step", default=5000, "-save_step", default=100000, "-lr", "--learning-rate", default=1e-4, "-n", "--num-workers", type=int, default=4, "--patch-size", default=(256, 256), "--batch-size", type=int, default=16, help="Batch size (default: %(default)s)" "--test-batch-size", default=1, help="Test batch size (default: %(default)s)", "--aux-learning-rate", default=1e-3, help="Auxiliary loss learning rate (default: %(default)s)", "--checkpoint", type=str, help="Path to a checkpoint"

8. In log directory, enter this command: tensorboard dev upload --logdir ./

9. There can be some errors. if any problem occurs, please contact me

제안된 모델

원본과 압축 이미지 비교(Kodak24)

PSNR 성능 비교

- (6PSNR_Y + 1PSNR_Cb + 1PSNR_Cr) / 6

RD 커브