This is the author's official PyTorch CT2 implementation.
We present Colorization Transformer via Color Tokens (CT2) to colorize grayish images while dealing with incorrect semantic colors and undersaturation without any additional external priors.
- Python 3.6
- PyTorch 1.10
- NVIDIA GPU + CUDA cuDNN
Clone this repo:
git clone https://github.com/shuchenweng/CT2.git
Install PyTorch and dependencies
http://pytorch.org
Install other python requirements
pip install -r requirement.txt
Download the pretrained vit model and move it to segm/resources/vit_large_patch16_384.npz
https://storage.googleapis.com/vit_models/augreg/L_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.1-sd_0.1--imagenet2012-steps_20k-lr_0.01-res_384.npz
We use ImageNet as our dataset, which includes 1.3M training data and covers 1000 categories. We test on the first 5k images of the public validation set, which is consistent with the previous Colorization transformer. All the test images are center cropped and resized into 256 × 256 resolution.
A training script example is below:
python -m torch.distributed.launch --nproc_per_node=8 -m segm.train --log-dir segm/vit-large --batch-size 48 --local_rank 0 --partial_finetune False --backbone vit_large_patch16_384 --color_position True --add_l1_loss True --l1_conv True --l1_weight 10 --amp
To test your training weights, you could excute the script below:
python -m torch.distributed.launch --nproc_per_node=1 -m segm.test --log-dir segm/vit-large --local_rank 0 --only_test True
We also publish the pretrained weights here. Download and decompress it with extraction code v4ay. After renaming the weight as checkpoint.pth, move it to segm/vit-large to enjoy the colorization! Here is the alternative link.
If the colorization results are weired, please carefully check whether the pretrained weight path is correctly set.
Don't worry! We're doing it!
Licensed under a Creative Commons Attribution-NonCommercial 4.0 International.
Except where otherwise noted, this content is published under a CC BY-NC license, which means that you can copy, remix, transform and build upon the content as long as you do not use the material for commercial purposes and give appropriate credit and provide a link to the license.
If you use this code for your research, please cite our papers CT2: Colorization Transformer via Color Tokens
@InProceedings{CT2,
author = {Weng, Shuchen and Sun, Jimeng and Li, Yu and Li, Si and Shi, Boxin},
title = {CT2: Colorization Transformer via Color Tokens},
booktitle = {{ECCV}},
year = {2022}
}