Unofficial implementation of "The Surprising Effectiveness of Linear Unsupervised Image-to-Image Translation". Features of this repository are:

• usage of core classes provided by this repository is close to the scikit-learn,
• this repository focus on the efficiency of the inference.

Now this code supports only colorization of the CelebA and FFHQ dataset, however, I will add more features in the future.

• Python 3.6.9
• docopt 0.6.2
• Numpy 1.18.4
• PyTorch 1.5.0
• scikit-image 0.17.2

It is good idea to use docker environment in order to avoide polluting your environment. The code in this repository is executable (and actually developed) under this docker image.

The interface of the linear transformation is almost same as the scikit-learn library.

>>> from linear_image2image_translation import LinearI2I
>>> X_gray  = ...                         # Grayscale images (shape = [n_data, n_features])
>>> X_color = ...                         # Color images     (shape = [n_data, n_features])
>>> lin = LinearI2I()                     # Create linear transformation class instance
>>> lin.fit(X_gray, X_color)              # Training
>>> X_color_pred = lin.transform(X_gray)  # Inference
array([[ -1.59893613e+00,  -2.18870965e-01,  -4.84763930e-02,
...

• Left: Original color image (ground truth)
• Center: Grayscale image (input image)
• Right: Predicted color image (inference result)

At first, please download the zipped images from the official web page.

$ cd dataset/celeba
$ python3 convert_celeba.py

The file celeba_align_128x128.npy will be generated. You can erase other files/directories if not necessary.

Then, run the following command:

$ python3 demo_colorize.py

The test results will be dumped under the output directory. Please see python3 demo_colorize.py --help for more details.

• PCA dimension = 128: PCA color, PCA gray, Linear transform

MIT Licence

[1] E. Richardson and Y. Weiss, "The Surprising Effectiveness of Linear Unsupervised Image-to-Image Translation", arXiv, 2020. PDF

Tetsuya Ishikawa (EMail, Website)