Official repo for Recursion's two recently accepted papers:
See the repo for ingredients required for defining our MAEs. Users seeking to re-implement training will need to stitch together the Encoder and Decoder modules according to their usecase.
Furthermore the baseline Vision Transformer architecture backbone used in this work can be built with the following code snippet from Timm:
import timm.models.vision_transformer as vit
def vit_base_patch16_256(**kwargs):
default_kwargs = dict(
img_size=256,
in_chans=6,
num_classes=0,
fc_norm=None,
class_token=True,
drop_path_rate=0.1,
init_values=0.0001,
block_fn=vit.ParallelScalingBlock,
qkv_bias=False,
qk_norm=True,
)
for k, v in kwargs.items():
default_kwargs[k] = v
return vit.vit_base_patch16_224(**default_kwargs)
A publicly available model for research can be found via Nvidia's BioNemo platform, which handles inference and auto-scaling: https://www.rxrx.ai/phenom
We have partnered with Nvidia to host a publicly-available smaller and more flexible version of the MAE phenomics foundation model, called Phenom-Beta. Interested parties can access it directly through the Nvidia BioNemo API: