google-deepmind/disentangled_rnns v0.1.8

v0.1.8

Repository: google-deepmind/disentangled_rnns

Tag: v0.1.8

Published: 2026-10-15T21:05:13Z

Prerelease: no

Release notes:

• Add disrnn.inverse_reparameterize_sigma, which returns the haiku parameter

for a given bottleneck sigma.

• Add bottleneck_graph, which builds a DisRNN's bottleneck graph as a

networkx DiGraph (get_bottleneck_graph), plots it (plot_bottleneck_graph), and tests it for isomorphism with a reference graph or another DisRNN (graphs_isomorphic, disrnn_isomorphic_to_graph, disrnns_isomorphic). Adds networkx, ipython and graphviz as dependencies. Plotting also needs the Graphviz dot program, which Colab has preinstalled (elsewhere see https://graphviz.org/download/).

• Add two_armed_bandits.isomorphic_to_qlearning,

isomorphic_to_policygradient, and isomorphic_to_actorcritic, which check whether a DisRNN's bottleneck graph matches that of the corresponding bandit agent.