superject/cards/neural

various loose ends of the generative card making approach

dream-the-dreamer.sh

setup script for linode GPU instances

• from debian 10 starting point ⟶ various GAN stuff

lightweight & styleGAN

backups are on posthoc

* styleGAN2-ADA

ongoing experiments with styleGAN2

current staging on Linode GPU instances (aka grapheme )

ps ax | grep train.py

generate and mix

python3 style_mixing.py \
    --outdir=/root/results/flickr6/generated/001600 \
    --network=/root/results/flickr6/00003-512x512-auto1-resumecustom/network-snapshot-001600.pkl \
    --rows=6000-6020 \
    --cols=600-620

python3 generate.py \
    --outdir=/root/results/flickr6/generated/001600 \
    --network=/root/results/flickr6/00003-512x512-auto1-resumecustom/network-snapshot-001600.pkl \
    --seeds=3300-3900

projection

image projection to explore latent space

–target=https://live.staticflickr.com/65535/50698595796_8accf83473_k_d.jpg \

python3 projector.py --outdir=/root/results/flickr6/generated/101405 \
    --target=/root/data/_transfer_test/50698595796_8accf83473_k.jpg \
    --network=/root/results/flickr6/00003-512x512-auto1-resumecustom/network-snapshot-001440.pkl \
    --num-steps=5000

rsync -av grapheme:results/ \
      /Users/zzk/code/foam-repo/superject/cards/neural/styleGAN2-ADA/results