tsn-affinity

How-to guides

The guides on this page walk you through task-oriented recipes. Pick the one closest to what you want to accomplish, then drill in.

Recipes

- :material-tune:{ .lg .middle } **[Tune sparsity](/tsn-affinity/guides/sparsity-tuning.html)** --- Choose `keep_ratio`, opt embeddings in or out, decide whether to quantize weights between tasks. - :material-vector-combine:{ .lg .middle } **[Pick a routing mode](/tsn-affinity/guides/routing-modes.html)** --- Switch between `action`, `latent`, and `hybrid` affinity routing and tune the threshold or margin. - :material-gamepad-variant:{ .lg .middle } **[Run Atari experiments](/tsn-affinity/guides/atari.html)** --- Collect random-policy trajectories, train a TSN strategy on them, and compute human-normalized scores. - :material-robot:{ .lg .middle } **[Add a new environment](/tsn-affinity/guides/new-environment.html)** --- Register a custom environment with `TaskRegistry` and plug it into the affinity router. - :material-tools:{ .lg .middle } **[Train your own model](/tsn-affinity/guides/custom-training.html)** --- Wire the public API into a custom training loop with your own data.

Conventions used in the guides