Release v0.4

👉 Corresponds to pypi release 0.4.1

New modules:

• CustomGoalAchievementLoss: allow users to define custom loss_f functions and pass it to the constructor

Updated modules:

• L2GoalAchievementLoss: now inherits from CustomGoalAchievementLoss
• NearestNeighborInterventionSelector: now has a loss_f method (given in constructor) allowing custom distance function
• BaseIM and IMFlowGoalGenerator have been updated to also allow custom distance_fn function
• PushPerturbationGenerator: now the push is sampled on the hyper-rectangle of side magnitude

Updated util functions:

• nearest_neighbors:
  • Y is now replaced by a vector called target
  • takes a loss_f function as input (allowing custom distance function)
  • distance is now computed by vmaping the loss_f function over X
• is_stable: now check that the error |y(t)–yfinal| ≤ settling_threshold for all t in time_window (similar to matlab stepinfo function)

API changes:

• in all pipeline, logs is now a DictTree that is created and saved in the pipeline, similar to history (removing dependencies to exputils package)

Bug Fixes:

• IMFlowGoalGenerator IM_vals and IM_grads has been reincluded into the logs
• PiecewiseSetConstantIntervention: removed hardplus when setting w and c
• NoisePerturbationGenerator, PushPerturbationGenerator, WallPerturbationGenerator: perturbations are now scaled by trajectories_extent = jnp.nanmax(trajectories[.., 1:], -1) - jnp.nanmin(trajectories[..., 1:], -1) to avoid NaN values, don't take into account the position at t0 when scaling (often big jumps between t0 and t1), and when needed trajectories_extent is set to 1 (to avoid division by zero)

Tests: updated with the new changes

Examples: updated with the new changes

v0.3

AutoDiscJax v0.3

• New modules:

  • ClampModule: new util module to allow clamping of module outputs
  • UniformGoalGenerator: module to sample goals uniformly between a low and high value
  • RandomInterventionSelector: module to randomly select an intervention among history of prior interventions
  • NullRolloutStatisticsEncoder: module to allow users to not compute any statistics on the system outputs
  • PushPerturbationGenerator: new module to generate push perturbation parameters (separated from the NoisePerturbationGenerator)

• Updated modules:

  • Module: each autodiscjax module now returns a log_data object

  • BaseGenerator and BaseOptimizer now inherit from the ClampModule module

  • big refacto of optimizers + BaseGCInterventionOptimizer + LearningProgressIM modules to allow save of local optim runs in imgep_experiment_pipeline

  • EAOptimizer now has a noise scheduler

  • GRNRollout: removed the vmap of grn_step (no needed anymore)

  • GRNRolloutStatisticsEncoder: added the is_converging statistic

  • WallPerturbationGenerator:

    • takes a sigma parameter allowing to tune the perpendicular vs radial repulsive forces exerted by the wall
    • positions wall based on distance travelled (instead of time travelled)

  • NoisePerturbationGenerator: added possibility to generate noise for w and c parameters

  • L2GoalAchievementLoss: added sqrt (before was squared L2) and adaptive scaling by current reached goal space extent

  • NearestNeighborInterventionSelector: added adaptive scaling by current reached goal space extent

• New util functions:

  • create_modules.py: bunch of util functions to instantiate autodiscjax modules from a given config
  • append_to_log: util function to append the module's log_data outputs to log

• Updated util functions:

  • is_converging: now looks if the signal amplitude has decreased (when compared to a prior phase of the signal)

• Bug Fixes:

  • HypercubeGoalGenerator: fix hypercube center calculation
  • GRNRolloutStatisticsEncoder: stats were mixed, fixed

• API changes:

  • big refacto of optimizers + BaseGCInterventionOptimizer + LearningProgressIM modules to allow save of local optim runs in imgep_experiment_pipeline
  • imgep_experiment_pipeline:
    • now saves local optim runs in history
  • rs_experiment_pipeline: added random search experiment pipeline
  • imgep_evaluation_pipeline was renamed into robustness_evaluation_pipeline (and run_imgep_evaluation into run_robustness_tests), and was modified with:
    • vmap vectorization of modules
    • logging
    • perturbation generator now takes raw system outputs as input (instead of just ys as this is more general)

• jit decorators where removed/added to optimize compute time

• Tests: renamed, updated and novel tests to test the different modules

• Examples: updated with the new modules/pipeline

• TODOs for next release:

  • docs
  • pip package
  • SGDOptimizer: deal with NaN values
  • IMFlowGoalGenerator bug fixes:
    • saving IM_vals into the logs can crash the code
    • can return NaN values

AutoDiscJax v0.2

• New modules:

  • BaseGoalAchievementLoss: base module for goal-achievement losses
  • BaseOptimizer: base module for optimizers
  • OpenESOptimizer: sgd optimizer where grad is not the exact value but estimated with ES (Salimans et al., 2017)
  • EAOptimizer: simple ea optimizer
  • BaseIM: base module for computing IM values and grad
  • LearningProgressIM: learning-progress IM variant
  • IMFlowGoalGenerator: imgep module for sampling new goals by flowing previous goals along IM gradients
  • BasePerturbationGenerator: base module for generating perturbations during evaluation (dependent of experiment's system_output_library)
  • NoisePerturbationGenerator: submodule for generating noise perturbation params, given experiment's grn trajectories
  • WallPerturbationGenerator: submodule for generating wall perturbation params, given experiment's grn trajectories
  • PiecewiseWallCollisionIntervention: grn intervention module to apply walls

• New util functions in misc.py :

  • flat_top_gaussian
  • sigmoid
  • calc_segment_intersection
  • wall_sticky_collision
  • wall_elastic_collision
  • calc_perpendicular_wall_distance
  • calc_radial_wall_distance
  • wall_force_field_collision

• Bug Fixes:

  • Clamp intervention parameters in optimizers
  • NearestNeighborInterventionSelector allow k>1

• API changes:

  • BaseGCInterventionOptimizer: imgep module now allows several optimizer variants (where optimizer is a BaseOptimizer object)
  • grn interventions __call__() function now takes y, y_, w, w_, c, c_, t_ as arguments
  • GRNRollout output pytree keys are now ys, ws, cs, ts instead of y, w, c, times
  • imgep_experiment_pipeline:
    • goal_generator now takes the target_goal_embedding_library as input
    • saving modules is optional
  • imgep_evaluation_pipeline:
    • removed the goal_embedding_encoder module from the pipeline
    • perturbation_generator now takes the experiment_system_output_library as input
    • saving modules is optional

• Tests:

  • test_grn_modules
  • test_imgep_modules

• Examples:

  • new example python scripts to run imgep experiment and evaluation pipelines with different possible configurations
  • new jupyter notebooks to analyze the results