tuners package

Subpackages

• tuners.loggers package
  • Submodules
  • tuners.loggers.base module
    • BaseTunerObserver
      • BaseTunerObserver.on_eval_samples_updated()
      • BaseTunerObserver.on_metric_updated()
      • BaseTunerObserver.on_ministep_idx_updated()
      • BaseTunerObserver.on_parameters_updated()
      • BaseTunerObserver.on_step_idx_updated()
  • tuners.loggers.console module
    • ConsoleLogger
      • ConsoleLogger.on_ministep_idx_updated()
      • ConsoleLogger.on_parameters_updated()
      • ConsoleLogger.on_step_idx_updated()
  • tuners.loggers.json module
    • JSONLogger
      • JSONLogger.on_eval_samples_updated()
      • JSONLogger.on_metric_updated()
      • JSONLogger.on_ministep_idx_updated()
      • JSONLogger.on_parameters_updated()
      • JSONLogger.on_step_idx_updated()
      • JSONLogger.save()
  • tuners.loggers.neptune module
    • NeptuneLogger
      • NeptuneLogger.on_eval_samples_updated()
      • NeptuneLogger.on_metric_updated()
      • NeptuneLogger.on_ministep_idx_updated()
      • NeptuneLogger.on_parameters_updated()
      • NeptuneLogger.on_step_idx_updated()
    • get_proxies()
  • tuners.loggers.wandb module
    • WandBLogger
      • WandBLogger.on_eval_samples_updated()
      • WandBLogger.on_metric_updated()
      • WandBLogger.on_ministep_idx_updated()
      • WandBLogger.on_parameters_updated()
      • WandBLogger.on_step_idx_updated()
  • Module contents
• tuners.losses package
  • Submodules
  • tuners.losses.base module
    • BaseLoss
  • tuners.losses.kl module
    • KLLoss
  • Module contents

Submodules

tuners.cdpg_tuner module

class tuners.cdpg_tuner.CDPGTuner(*args, context_distribution=<disco.distributions.single_context_distribution.SingleContextDistribution object>, **kwargs)

Bases: Tuner

Contextual DPG tuning class, relying on a ContextDistribution and KLLoss().

tuners.dpg_tuner module

class tuners.dpg_tuner.DPGTuner(*args, context='', **kwargs)

Bases: CDPGTuner

DPG tuning class, a specific case of CDPG with a single, fixed, context.

tuners.tuner module

class tuners.tuner.Tuner(model, target, proposal=None, context_distribution=<disco.distributions.single_context_distribution.SingleContextDistribution object>, loss=<disco.tuners.losses.kl.KLLoss object>, features=[], track_metrics=['kl', 'tv', 'js'], track_divergence_from_base=False, **params)

Bases: object

Generic tuning class.

default_params = {'context_sampling_size': 16, 'divergence_evaluation_interval': 16, 'learning_rate': 1.41e-05, 'n_gradient_steps': 1024, 'n_samples_per_step': 1024, 'optimizer': 'Adam', 'proposal_update_metric': 'kl', 'sampling_size': 32, 'scheduler': 'constant', 'scoring_size': 64, 'warmup_steps': 12}

report_feature_moments(context, samples, proposal_log_scores, model_log_scores, target_log_scores)

tune()

Fine-tunes model distribution’s network

Fine-tunes the network of the model distribution:

• repeats n_gradient_steps tuning steps

• eventually updates the samplee according to KL divergence

Module contents