Major changes:
- trainer.py: Checkpoints saved every N steps + after each phase completion.
Phase-level checkpoints (phase1_complete.pt, phase2_complete.pt, etc.)
enable resuming from any phase.
- main.py: --resume-phase N loads phase N-1 checkpoint and skips completed phases.
--sinkhorn-batch overrides Sinkhorn batch size (for OOM on smaller GPUs).
--checkpoint-dir and --checkpoint-every for checkpoint control.
- config.yaml: CIFAR-10 sinkhorn.batch_size reduced 128→32 for T4 16GB VRAM.
Pool batches increased 2500→10000 to compensate (same total pool entries).
torch.cuda.empty_cache() called after pool building to free Sinkhorn memory."
Update SKILL.md with DataLoader batch size mismatch bug and shared state pitfall
- compute_velocity now flattens (N,C,H,W) → (N,D) before geomloss calls,
reshapes gradients back to original shape after. Fixes MNIST/CIFAR crash.
- compute_sinkhorn_divergence also handles image tensors.
- TrajectoryPool.finalize() pre-concatenates after pool building (O(1) sampling
instead of O(pool_size) per step)."
Add SKILL.md — paper reproduction skill with lessons from NSGF++ implementation