tf-keras-stacked-convolutional-autoenconder

Implementation of Stacked Convolutional Autoencoder.
Evaluated the denoising performance using MNIST.
Used Python3.7 + TensorFlow 2.3 + tf.keras.

Settings

I recommend to make virtual environment by venv.
1. python -m venv .venv

Enter to virtual environment .venv/Scripts/activate.ps1 (Windows) or .venv/Scripts/activate (Linux)
pip install -r requirements.txt

Run

python train.py --epochs 20 --batch_size 100 --stacked 1 --snr -40 (defalut options)
- --epochs or -e: The number of epochs
- --batch_size or -b: The number of batch size
- --stacked or -s: 1 or 0 (stacked or not)
- --snr or -n: The value of SNR for denoising autoencoder. The smaller value is noisy.
python eval.py --dir_model path_to_hdf5 --snr -40
- --dir_model or -d: path to model.
  - hdf5 file will generate in ../dist/train_conditions/models
- --snr or n: same to train.py
- SNR value will printed.

Results

Denoising performance improved slightly.
- SNR comparison:
  - stacked: 9.307
  - no_stacked: 9.184

Others

If you have any problems, please let me know.
Please send me the pull request if you find wrong.