stylegan3_training_and_inference_2022_02_11.ipynb

stylegan3_fine_tuning_with_colab_2022_01_01.ipynb

Thanks for this. I'm having some issues with the training, I think there might be a string issue with rclone trying to output to google drive, was wondering if you could take a peek. I haven't modified it, it just goes to training run ended immediately regardless of whether I run it from scratch or if I resume from a pickle

training from scratch!
running until kimg: 2000
/content
*** training run with gamma=6.0 and ema_factor=1.0***
quoteargs(): ignoring deleted or empty or negative-integer long-option 'resume' with value -1!
quoteargs(): ignoring deleted or empty or negative-integer long-option 'resume_kimg' with value -1!
quoteargs(): ignoring deleted or empty or negative-integer long-option 'mbstd_group' with value -1!
quoteargs(): ignoring deleted or empty or negative-integer long-option 'freezed' with value None!
/content/stylegan3/train.py '--outdir=/content/training-runs --data=/content/dataset --cfg=stylegan3-t --cond=True --gpus=1 --workers=1 --kimg=2000 --gamma=6.0 --ema-factor=1.0 --batch=32 --batch-gpu=16 --aug=ada --augpipe=blit --mirror=False --tick=5 --snap=5 --img-snap=1 --snap-res=4k --cbase=16384 --seed=1983 --metrics=none --img-cmd=\'rclone --config=/content/rclone.config copy "$1" driveapi:/scratch/training-run-"$DESC"/; echo "$1"\' --snap-cmd=\'rclone --config=/content/rclone.config copy "$1" driveapi:/scratch/training-run-"$DESC"/; echo "$1"\''

running train.py script...
Usage: train.py [OPTIONS]
*** training run ended ***

@un1tz3r0 let me know if I can provide anything else.

did you generate a dataset? the first time you run it, you must check the "generate_missing_dataset" toggle in the "Prepare Dataset" cell in order to create a dataset of cropped images in my pixelscapes-dataset repo. it will download realesrgan and some dependencies and grab the source images from the repo, upscale them, and then produce the crops and the class conditional labels which i found crucial to successfully training on diverse datasets like the pixelscapes. once it has been generated, it will upload a zip of the dataset to your google drive as dataset.zip and future invocations should download that and use it. please include the output of the Prepare Dataset cell and check that /content/dataset exists and is not empty if you are still having issues after checking that toggle. Hopefully this helps!

I already had a dataset prepared with the dataset tool that I use to train on another colab without problem which I uploaded, and ran the cell without gen missing dataset checkbox. The cell said it found and processed the dataset over.