After the installation of FlashAttention2, the model can be loaded using:
from scooby.modeling import Scooby
flooby_neurips = Scooby.from_pretrained(
'johahi/neurips-scooby-flash',
cell_emb_dim=14,
embedding_dim=1920,
n_tracks=3,
return_center_bins_only=True,
disable_cache=False,
use_transform_borzoi_emb=False,
)
Since the transform_borzoi_emb is not used, the models performance is slightly (0.52 vs 0.54 across cell types) worse than neurips-scooby.