DinoV2 with registers backbone, transformer decoder, classifier free guidance film layers training script

☘️ Shoot an email to sebastian@mbodi.ai if you'd like to tackle this issue and I'll help as often as I can. Can provide A100 access once script is ready.

Starter Code
Example Doing Identical task but with MaxViT

Resources

Highly-Recommended Guide to Follow
Transformer Head Code
DinoV2 Source Code
Text Guidance with Film
RT1: Robotics Transformers paper

Tokenize Actions (x, y, z, roll, pitch, yaw, grasp)

Transform pattern: (b frames action) -> (b f a bins), bins=255

This is just simple classification not sequence to sequence modeling

1. Apply MinMax Scaler

2. Apply kbins

Apply film layers from classifier-free-guidance

Inference pattern: (b f c h w ), str --> (b f a bins)

Example Doing Identical task but with MaxViT

Details

• Use pytorch lightning, transformers, or fastai (transformers preferred but fastai likely easiest)
• Use pretrained ViT-g/14 small or large with registers
• Start with basic encoder-decoder pattern (see the starter code script)

Use the following losses:

• Asymmetric loss: https://timm.fast.ai/asymmetric_loss
• CutUpMix: https://timm.fast.ai/random_resized_crop
• Standard Image Augmentations: https://timm.fast.ai/random_resized_crop
• Repeat with Film Layers added to ViT blocks and no caption caption input using: classifier-free-guidance (again refer to the MaxViT example above).

Follow-On Work

• Ablations with early, middle, late fusion
• Ablations with DinoV2 frozen, dinov2 without registers, smaller or larger dinov2
• Whiten image inputs with PCA
• AutoAugment with timm