Log-reward uniformisation, logging and new hyperparams for temperature, Z-training, validation.#54

Merged

julienroyd merged 55 commits intotrunkfrom julien-pr1

Mar 22, 2023

+367-206

Commits on Feb 20, 2023

more comments and slightly better architecture adaptation
bengioe
committed

Commits on Mar 1, 2023

fix: removed np.int (deprecated)
julienroyd
committed
feat: added option hps.temperature_sample_dist='constant' which correspond to self.temperature_dist_params=<fixed_value>
julienroyd
committed
fix: added case for self.temperature_sample_dist == 'constant' in encode_conditional_information()
julienroyd
committed
feat: added hyperparams 'objectives' and 'num_thermometer_dim'
julienroyd
committed
fix: use 'objectives' hyperparam for validation set construction
julienroyd
committed
fix: added security-check to thermometer and avoid passing both vmin=vmax=0 when beta=0
julienroyd
committed
feat: added a logger to run()
julienroyd
committed
fix: isolating sensitive imports (cvxopt) that are tricky to install on Mac M1 chip
julienroyd
committed
(fix): fixed encode_conditional_information() for constant beta (was passing full thermometer instead of empty one)
julienroyd
committed
feat: added hyperparameter 'Z_params'
julienroyd
committed
fix: removed batch_size*2 for training sampling_iterator
julienroyd
committed
(minor): added prints
julienroyd
committed
feat: added argument 'valid_random_action_prob' and moved random_action_prob to be a argument of graph_sampler.sample_from_model()
julienroyd
committed
Added hyperparameters:
julienroyd
committed
docs: added hyperparmeters in printing and in logs as json file
julienroyd
committed
chore: moved config saving from GFNTrainer.run() to SEHMOOFragTrainer.setup() to be caught by Determined agents as well
julienroyd
committed
feat: added git_hash to saved hyperparams
julienroyd
committed
feat: added 'hostname' in hyperparameters for cluster troubleshooting
julienroyd
committed
feat: use log-rewards for numerical stability (forgotten commit from PR39)
julienroyd
committed
docs: renamed functions to logrewards (to avoid TrajectoryBalance calling task.cond_info_to_reward() for tasks that wouldn't have been updated to using logrewards rather than rewards)
julienroyd
committed
fix: convert logrewards back into regular rewards for logging
julienroyd
committed
Merge branch 'fix_frag_env' into julien-general-pr
julienroyd
committed
chore: pep8
julienroyd
committed
Added support for self.temperature_sample_dist='constant' to other tasks:
julienroyd
committed

Commits on Mar 3, 2023

Commits on Mar 15, 2023

fix: clamp rewards before taking into log-space (rather than add a small epsilon)
julienroyd
committed

Commits on Mar 20, 2023