Fix RAFT Training Non-Convergence
Budget: $10 – $30 USD
My custom PyTorch implementation of the RAFT optical-flow model stubbornly refuses to converge. I have already verified the model architecture layer by layer, so the issue is likely hiding elsewhere—in the data pipeline, loss computation, hyper-parameter schedule, or gradient handling.
I will share the full repository, sample datasets, and current training logs. Your job is to trace the exact reason training plateaus, implement the necessary code-level or configuration fixes, and show a clean training run that converges to expected benchmarks.
Deliverables:
• Annotated code or pull request with the fixes
• Brief report summarising what went wrong and how you resolved it
• Screenshot or log of a successful training curve reaching the target loss
A solid grasp of PyTorch internals, mixed-precision training, and RAFT’s iterative refinement logic is essential. If this sounds routine for you, let’s get the model learning again.
I will share the full repository, sample datasets, and current training logs. Your job is to trace the exact reason training plateaus, implement the necessary code-level or configuration fixes, and show a clean training run that converges to expected benchmarks.
Deliverables:
• Annotated code or pull request with the fixes
• Brief report summarising what went wrong and how you resolved it
• Screenshot or log of a successful training curve reaching the target loss
A solid grasp of PyTorch internals, mixed-precision training, and RAFT’s iterative refinement logic is essential. If this sounds routine for you, let’s get the model learning again.