About training dataset and model parameter details

[Wp-Zhang]

Hi, I can successfully train the model with my own training code right now but I found that the dataset composition is critical to the model performance.

The original Adobe5K dataset contains 5000*5=25000 image pairs, and according to the paper and the published code, you used hueshift and randomcrop for augmentation.

My questions are:

1. did you perform the augmentations on the original pairs, so the length of your full dataset is 50000 (including the identical pairs)?
2. did you perform other often-used image augmentation methods like flipping and rotation?
3. did the identical pairs only contain pairs from the raw images or also all the images that are retouched by different experts?

Thank you in advance!

[Wp-Zhang]

Here's another question:
You stated in the paper that the optimal bin_num for l-channel and ab-channel is 64, and you set them both as 64 in the experiments. But why is bin_num of the l-channel set as 8 in the pre-trained model? Is it a trade-off for training efficiency?