YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
fastwam-egodex-uncond-1cam320-30k (v2, constant lr)
Same corpus and setup as fastwam-egodex-uncond-1cam320-30k, with one change: the lr is held constant at 1e-4 after warmup instead of decaying on a cosine.
Warmup is unchanged and is not part of that choice -- the trainer always applies warmup_steps = total_steps * 0.05 (1,500 here) and wraps whichever main scheduler is selected, so this run still ramps 0 -> 1e-4 over 1,500 steps and then flattens.
4 GPU x batch 16 x grad-accum 3 = global batch 192, 30k steps, 192x320 single camera. Weights only; checkpoints/state/ is not published.
- Downloads last month
- 9
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support