Our toolbox can support various DCN models, but by default we provide only one - TwitterDCN. This model follows the general autoencoder architecture by Theis et al. but uses custom solutions for quatization, entropy estimation and coding. We illustrate the structure of the model below.
We also provide 3 pre-trained versions of this model with low, medium and high-quality. The rate-distortion trade-off on 3 different datasets is shown below.
To train a different model, we can use the train_dcn script and provide a list of model configurations (csv file with hyper-parameter values) and an RGB training set, e.g.:
> python3 train_dcn.py --dcn TwitterDCN --split 31000:1000:1 --param_list data/config/twitter.csv --epochs 2500 --out data/models/dcn/custom --data data/rgb/compression/We test the DCN using the test_dcn script which can:
- show compression results for a batch of images
- compare against JPEG compression at a matching quality level / bpp
- generate rate-distortion curves for various codecs
For example, to compare our low-quality codec against JPEG at matching bpp:
> python3 test_dcn.py --data ./data/rgb/kodak --dcn 16c jpeg-match-bpp --image 4