google-deepmind/disentangled_rnns v0.1.7

v0.1.7

Repository: google-deepmind/disentangled_rnns

Tag: v0.1.7

Published: 2026-09-26T14:05:54Z

Prerelease: no

Release notes:

• Validate penalty fields, max_latent_value > 0, and unknown attribute

assignments in DisRnnConfig, MultisubjectDisRnnConfig, and DisRnnWNeuralActivityConfig.

• Default trainable_param_names to () instead of None and gate parameter

path printing behind verbose=False in checkpoint_utils.get_optimizer_with_frozen_params.

• Apply learned choice bottleneck multipliers in plotting.compute_choice_rule.
• Fix "All Trials" title string iteration bug and observation_names typing

in plotting.compute_update_rules.

• Unify SubjEmb labels to 1-based indexing, add sort_latents support across

update, choice, and neural-activity rule plots so latent numbering matches plot_bottlenecks, and scope matplotlib.rcParams overrides in plotting.

• Return figure objects in neuro_disrnn.plot_choice_rule and

neuro_disrnn.plot_update_rules, remove the unused plt.subplots allocation in neuro_disrnn.plot_neural_activity_rules, and compute 2D color scale limits across all four (choice, reward) conditions.

• Fix double normalization of decisions prior to applying lapse in

pclicks.drift_diffusion_model.