I have code image to image colorizer.

Input shape is grayscale:

(batch_size, width, height, 1)

Output shape is colored:

(batch_size, width, height, 3)

I’m using model regressor (U-net) therefore using regression loss.

Every epoch trained, it will predict and display an unseen image. (an X_test).

As the performance of predicted output is subjective and no available label (no y_test), I want tell the model if this epoch is good model or bad model based on my manual judgement. So it must be label 0, or label 1 (better model for current epoch).

Here is the unseen image where the label of image (y_test) had been never exists since the picture taken 1950.
55E

So, the model no idea what is the better result to evaluate for that data, instead it will judged by my self.

How do I do that in Keras? I expect every epoch, it prompts me whether the current predicted image is better or not (labeling 1 or 0).

Basically this is same concept as GAN, except the discriminator is human.

Here is the code: https://colab.research.google.com/drive/141m91yB6rx4zlLutuRNe4e44h7L552SB?usp=sharing