YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Experiment 4 v1 — Interpolation Model (DiT)
Overview
Video inpainting DiT for hourly interpolation between 6-hour forecast endpoints. Fills in 5 intermediate hourly frames given known endpoints.
Model
Interpolation DiT (198.3M params)
- Architecture: hidden=768, 12 heads, 5 double + 10 single blocks
- Best T2 MAE at RKSI: 2.48 K (epoch 1/6, early stopped)
- Optimizer: Muon (LR=4e-4) + AdamW
- Uses frozen VAE from forecast model for encode/decode
Training Config
- 6 epochs (early stopped at patience=5), batch=12 x 8 GPU
- Muon LR=4e-4, warmup=2 epochs
- LPIPS lambda=0.1, gradient loss lambda=0.1
- Hardware: 8x NVIDIA B200
- Data: 43,818 7-frame sequences (2015-2019)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support