def retro_analyze_trials(predir):
'''Retro-analyze all trial level datas.'''
logger.info('Retro-analyzing trials from file')
from slm_lab.experiment.control import Trial
filenames = ps.filter_(os.listdir(predir),
lambda filename: filename.endswith('_trial_df.csv'))
for idx, filename in enumerate(filenames):
filepath = f'{predir}/{filename}'
prepath = filepath.replace('_trial_df.csv', '')
spec, info_space = util.prepath_to_spec_info_space(prepath)
trial_index, _ = util.prepath_to_idxs(prepath)
trial = Trial(spec, info_space)
trial.session_data_dict = session_data_dict_from_file(
predir, trial_index, ps.get(info_space, 'ckpt'))
# zip only at the last
zip = (idx == len(filenames) - 1)
trial_fitness_df = analysis.analyze_trial(trial, zip)
# write trial_data that was written from ray search
trial_data_filepath = filepath.replace('_trial_df.csv',
'_trial_data.json')
if os.path.exists(trial_data_filepath):
fitness_vec = trial_fitness_df.iloc[0].to_dict()
fitness = analysis.calc_fitness(trial_fitness_df)
trial_data = util.read(trial_data_filepath)
trial_data.update({
**fitness_vec,
'fitness': fitness,
'trial_index': trial_index,
})
util.write(trial_data, trial_data_filepath)
def run_trial(experiment, config):
trial_index = config.pop('trial_index')
spec = spec_from_config(experiment, config)
info_space = deepcopy(experiment.info_space)
info_space.set('trial', trial_index)
trial_fitness_df = experiment.init_trial_and_run(spec, info_space)
fitness_vec = trial_fitness_df.iloc[0].to_dict()
fitness = analysis.calc_fitness(trial_fitness_df)
trial_data = {**config, **fitness_vec, 'fitness': fitness, 'trial_index': trial_index}
prepath = util.get_prepath(spec, info_space, unit='trial')
util.write(trial_data, f'{prepath}_trial_data.json')
return trial_data
def run_trial(experiment, config):
trial_index = config.pop('trial_index')
spec = spec_from_config(experiment, config)
info_space = deepcopy(experiment.info_space)
info_space.set('trial', trial_index)
trial_fitness_df = experiment.init_trial_and_run(spec, info_space)
fitness_vec = trial_fitness_df.iloc[0].to_dict()
fitness = analysis.calc_fitness(trial_fitness_df)
trial_data = {
**config, **fitness_vec, 'fitness': fitness, 'trial_index': trial_index,
}
prepath = analysis.get_prepath(spec, info_space, unit='trial')
util.write(trial_data, f'{prepath}_trial_data.json')
return trial_data
def lab_trial(config, reporter):
'''Trainable method to run a trial given ray config and reporter'''
trial_index = config.pop('trial_index')
spec = self.spec_from_config(config)
info_space = deepcopy(self.experiment.info_space)
info_space.set('trial', trial_index)
trial_fitness_df = self.experiment.init_trial_and_run(
spec, info_space)
fitness_vec = trial_fitness_df.iloc[0].to_dict()
fitness = analysis.calc_fitness(trial_fitness_df)
trial_index = trial_fitness_df.index[0]
trial_data = {
**config,
**fitness_vec,
'fitness': fitness,
'trial_index': trial_index,
}
done = True
# TODO timesteps = episode len or total_t from space_clock
# call reporter from inside trial/session loop
reporter(timesteps_total=-1, done=done, info=trial_data)