DALL-E trainer and generator

Setup

• Install DALLE-pytorch

  python3 -m pip install dalle-pytorch

Training

• Train VQVAE:

python3 dalle.py --source path/to/images/ \
                 --vae vae.pt \
                 --train_vae \
                 --batch_size=16 \
                 --samples_out samples/vae/ \
                 --epochs=2

• Train DALL-E

--source is expected to contain image files and tag files with the same name (ie, foo.png, foo.txt).

Tag files are a single-line file with comma separated tags. Example:

1girl, white_swimsuit, red_hair

python3 dalle.py --source path/to/images/and/tags/ \
                 --vocab curated_512.vocab \
                 --vae vae.pt \
                 --train_dalle \
                 --dalle dalle.pt \
                 --batch_size=16 \
                 --samples_out samples/dalle/ \
                 --epochs=2

Generating Images

Generate a single image from tag file:

python3 dalle.py --generate \
                 --vocab curated_512.vocab \
                 --vae vae.pt \
                 --dalle dalle.pt \
                 --source my_tag.txt \
                 --output results/my_image.png

Generate images from a directory containing tag files:

python3 dalle.py --generate \
                 --vocab curated_512.vocab \
                 --vae vae.pt \
                 --dalle dalle.pt \
                 --source my/tags/ \
                 --output results

Generate an image using specified tags:

python3 dalle.py --generate \
                 --vocab curated_512.vocab \
                 --vae vae.pt \
                 --dalle dalle.pt \
                 --tags="1girl, white_swimsuit, red_hair" \
                 --output samples/my_image.png