Skip to content

Autoencoder coloring gray scale pictures. An artificial neural network constructed and trained with Tensorflow 2.

Notifications You must be signed in to change notification settings

DominikKossinski/ColoringAutoencoder

Repository files navigation

Coloring Autoencoder

An artificial neural network implemented with Tensorflow 2. Contains only conventional layers. Network trained on 6000 images (75 x 75 px) from Oxford Flowers 102.

Training

To start model training type command:

python train.py 
    -h, --help                                  show this help message and exit
    -e EPOCHS, --epochs EPOCHS                  Number of training epochs
    --path PATH                                 Models directory path
    --name NAME                                 Model name
    -bs BATCH_SIZE, --batch-size BATCH_SIZE     Batch size
    -f {RGB,HSV,LAB}, --format {RGB,HSV,LAB}    Output image format
    -da, --data-ag                              Data agumentation
                

Test

To test pretrained models type command:

python test.py
    -h, --help                                  show this help message and exit
    --path PATH                                 Models directory path
    --name NAME                                 Model name
    -bs BATCH_SIZE, --batch-size BATCH_SIZE     Batch size

During the training open next command line and type, to start Tensorboard:

tensorboard --logdir models/<Your model name><Image format>/logs

Then open url localhost:6006/ to see generated summaries.

Jupyter-notebook

The jupyter-notebook examples.ipynb contains code, that allows to test pretrained model on images given by user.

To start jupyter server type:

jupyter-notebook examples.ipynb

Images examples

Input

Original image Input Image
Input image Black and white image

Output

RGB HSV LAB
RGB RGB RGB

About

Autoencoder coloring gray scale pictures. An artificial neural network constructed and trained with Tensorflow 2.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published