Lossy image compression using compressive autoencoders

Train

You can train a new custom model by executing the following command. You can provide any images dataset that you want. For now, dataset must contain 720x1280 images. This will probably be expanded in the future.

python3 src/train.py \
  --output-path trained_models \
  --dataset-path test.bmp \
  --log-freq 5 \
  --save-freq 15

Test

You can test a trained model by executing the following command.

trained_models should be that folder containing the model.

python3 src/test.py \
  --model-path trained_models \
  --image-path test.bmp

Results

Original	Result
2.8 MB	1.2 MB
631 KB	21 KB
787 KB	38 KB

References

• [1] DeepZip: Lossless Compression using Recurrent Networks
• [2] Lossy Image Compression with Compressive Autoencoders, Theis et al.
• [3] Variable Rate Image Compression with Recurrent Neural Networks, Toderici et al.