How to train?

[atugar]

Can you please explain what the below params mean? They are not needed for StyleGAN 2.

python run_training.py
--config=config-ada-sv-flat
--data-dir=/path/to/datasets
--dataset-train=path/to/train \ - What is this?
--dataset-eval=path/to/eval \ - What is this?
--total-kimg=30
--metrics=None

[atugar]

Ok. I read the readme again. When you say split dataset into training and eval, are those directories supposed to hold the raw JPG or TFRecord? And how do I specify what network to start from? I assume for 1024 this will be FFHQ config-f, but where to specify this?

[e-271]

Hi, so they should both be paths to TFRecord dirs. You first create the TFRecord dirs from images using the dataset_tool.py script, and then run training with paths to those dirs. --resume-pkl-dir controls which path pickles are found in and --resume-pkl' controls the name, but if this is not given it is inferred based on the name of --dataset-train`. I'll update the readme with these details.

[atugar]

Hello. Thanks for getting back to me. If --dataset-train is the path to the training dir, then what is --data-dir? I thought that was the path to the training dir.

[AndreTInfante]

I am also unclear on what data-dir is, and would appreciate clarification!

[AndreTInfante]

@atugar I actually figured it out. data-dir is the parent directory for the datasets, and data-train and data-eval are subdirectories relative to that parent directory