Tutorials#
Tutorials may demonstrate a method on a composed model, a same-run hook composition, or a multi-stage pipeline. These terms follow the Composition contract contract. Existing notebook hand-offs are migration candidates for named TensorDict artifacts and declared pipeline stages.
TorchRL PPO Action Probing
Set up a TorchRL PPO agent and use tdhook to probe action representations.
Chess Value Saliency
Compute attribution maps for chess model predictions using saliency methods.
Concept Attribution (LRP + RelMax)
Build a striped concept and visualize concept-conditioned relevance on natural images.
Chess Dimension Estimation
Estimate channel-wise intrinsic dimension from chess activations and compare layers, heads, and opening splits.