All files contain such big images that github can't load them. please download and see them on Google colab.
dataset : flower102
you can download at https://www.kaggle.com/c/oxford-102-flower-pytorch/data
If you are interested, you can try another dataset. However, you should change models(ex. change encoder structure for 102 categories to other).
we have 3 sequential steps to train each modules instead of traning the whole model at one step.
- encoder
- Link the dataset to use for training, validation, test.
- It uses resnet50 as the baseline.
- Encoder is defined as classification model for 102 categories. After training, encoder is assumed to make good feature vectors for images.
- generator
- Use the trained encoder from 1.
- The model contains same structure for the encoder with 1. The decoder part is added for generating ab(NOT RGB) colors.
- Generator structure
- GAN
- Use the trained generator from 2.
- L1 loss is added to GAN loss. This is for training stability, also the right color.
- GAN structure
with L1 loss integrated to GAN, final model shows good performances : right color, good impression(색감) .
[1] Colorful Image Colorization Richard Zhang, Phillip Isola, Alexei A. Efros (Submitted on 28 Mar 2016 (v1), last revised 5 Oct 2016 (this version, v5)) https://arxiv.org/abs/1603.08511
[2] Deep Colorization, Zezhou Cheng, Qingxiong Yang, Bin Sheng (Submitted on 30 Apr 2016) https://arxiv.org/abs/1605.00075
[3] Auto Colorization of Black and White Images using Machine Learning “Auto-encoders” technique https://medium.com/@mahmoudeljiddawi/auto-colorization-of-black-and-white-images-using-machine-learning-auto-encoders-technique-a213b47f7339