Anime Sketch Coloring with Swish-Gated Residual U-Net

Pytorch unofficial port of SGRUnet(the official: here)

Performance

This is the performance of training 13 epochs, config is consistent with this config.py. google drive
Training this model takes a lot of time, so I only trained 13 epochs, which does not represent the best performance.

Feature

LayerNorm requires a lot of memory, so BatchNorm was implemented, which greatly speeds up the training, but may have an impact on performance. You can choose which to use in config.py.
For save your memory, you can choose bilinear or transpose convolution(paper) to upsample.
Two datasets are supported. Anime Sketch Colorization Pair and another one that was used in the paper. Also optional in config.py.
The network that calculates the loss is different(ResNet family vs VGG family). Also optional in config.py.
Support mini-batch training.

Data folder

colorgram

anime_colorization
└── data/
    ├── train/
    |   
    └── val/

safebooru (paper)

anime_colorization
└── data/
    ├── train/
    |    ├── img/
    |    └── label/
    └── val/
         ├── img/
         └── label/

Setup

pytorch >= 1.1.0

Use the requirements.txt file to install the necessary depedencies for this project.

$ pip install -r requirements.txt

Config

Modify config.pyas needed.

Train

python  main.py

Inference

Modify the

model_path
file_name
file_path
output_path

in the inference.pyas needed.

and

python inference.py

gakkiri/SGRUnet-pytorch