ColorNet

Introduction:

ColorNet is a sequential image coloring model created for the AMD Pervasive AI Developer Contest.

Dependencies:

Install required packages using:

pip install -r requirements.txt

Training

The final model was trained on the blackclover\colored subdirectory of the japanese manga dataset which can be obtained from kaggle. Link to Dataset

The model can be trained by running:

python train.py --data DATASET_PATH

Other parameters:

--batch_size           default: 2
--size                 default: 256
--tempo_length         default: 5
--pretrain_epochs      default: 5
--train_epochs         default: 10

Inference

If you just wish to try out the model, the trained weights can be downloaded here: Download Weights. Copy the downloaded file to the weights subdirectory.

After obtaining the weights(through training or from the link provided above), the inference app can be run using:

python app.py

A script to convert colored images to greyscale has also been included

Usage:

python convert.py (Path to directory contining colored images) (Path to store colored images)

Example:
python convert.py ./Dataset/ds/onepiece/colored ./Dataset/ds/onepiece/L