Potpov/grimoire-mnist

Grimoire — MNIST (raw PNGs + ART tokens)

MNIST for the Grimoire SVG pipeline. The stage-1 (VSQ) input is shipped raw (PNGs) — the heavy pre-tiled tensors (2.7 TB) regenerate easily from them, so they are not shipped. The stage-2 (ART) tokens are shipped ready to use.

Contents

• mnist_png.tar.gz (15 MB) — training/{0..9}/*.png + testing/{0..9}/*.png (standard MNIST).
• mnist_tokenized.tar (3.7 GB) — all precomputed VSQ token sets for stage-2 (ART) (the paper variants: VSQ_MNIST_BW_*, color, augmented, …).

Grimoire dataset family — recap

dataset	what's pushed	plug-and-play for VSQ (stage 1)?	size
FIGR-8	processed SVGs + VSQ csv + ART tokens	✅ yes	svgs 345 MB · csv 35 MB · tokens 530 MB
MNIST (this repo)	raw PNGs + ART tokens	⚠️ needs 1 easy step (PNG→tensors)	png 15 MB · tokens 3.7 GB
Emoji	processed SAM masks	✅ yes	5.7 GB
Cartoons	— not pushed (no experiments)	—	—

Post-process the raw PNGs into the VSQ training format (verified)

MNIST is the one dataset that needs a post-processing step for VSQ — but it's light (just PIL tiling, no SAM/deepsvg). The stage-1 loader consumes pre-tiled tensors [N, 3, patch, patch]:

tar xzf mnist_png.tar.gz          # -> mnist_png/{training,testing}/{0..9}/*.png
# in the Grimoire code repo, tile the PNGs into train.pt / test.pt:
python -c "
import torchvision.transforms as T, sys; sys.path.insert(0,'scripts')
from create_precomputed_mnist import Mnister
ps,tiles,pad,th = 128,6,0,0.2; nd=(ps-pad)*tiles
tf=T.Compose([T.Resize(nd,antialias=True), T.RandomInvert(1.0), T.Grayscale(3), T.ToTensor()])
Mnister(source_dir='mnist_png', output_dir='mnist_pretiled/P128_T6_P0_TH0.2',
        patch_size=ps, transform=tf, num_tiles_per_row=tiles, total_padding=pad, filter_th=th)
"

This writes mnist_pretiled/P128_T6_P0_TH0.2/{train.pt,test.pt} (tensors [N,3,128,128]). Point configs/MNIST/MNIST_VSQ_BW.yaml:data_params.data_path there and run python run.py -c configs/MNIST/MNIST_VSQ_BW.yaml. (Verified: PNGs → [N,3,32,32] tensors.)

For stage-2 (ART) you can skip all of the above: tar xf mnist_tokenized.tar and point configs/MNIST/MNIST_ART_BW.yaml at the matching mnist_tokenized/VSQ_MNIST_BW_*/{split.csv,vsq_tokenized.npy} (stage-1 checkpoint at Potpov/grimoire-checkpoints mnist_bw/vsq).