num_configurations, num_ace_runs, num_policies = policies_memmap[
'policies'].shape
parameters_dtype = policies_memmap['parameters'].dtype
performance_dtype = np.dtype([('parameters', parameters_dtype),
('results', float,
(num_ace_runs, num_policies,
args.num_evaluation_runs))])
# Create the memmapped array of results to be populated in parallel:
performance_memmap_path = str(experiment_path /
'{}_performance.npy'.format(args.objective))
performance_memmap = np.lib.format.open_memmap(
performance_memmap_path,
shape=(num_configurations, ),
dtype=performance_dtype,
mode='w+')
# Evaluate the learned policies in parallel:
with tqdm_joblib(
tqdm(total=args.num_evaluation_runs * num_ace_runs *
num_configurations * num_policies)) as progress_bar:
Parallel(n_jobs=args.num_cpus, verbose=0)(
delayed(evaluate_policies)(policies_memmap, performance_memmap,
evaluation_run_num, ace_run_num,
config_num, policy_num, random_seed)
for evaluation_run_num, random_seed in enumerate(random_seeds)
for ace_run_num in range(num_ace_runs)
for config_num in range(num_configurations)
for policy_num in range(num_policies))
mode='w+')
# Create the memmapped array of performance results for the learned policies:
performance_dtype = np.dtype([
('parameters', parameters_dtype),
('results', float, (num_runs, num_policies, args.num_evaluation_runs)),
('results_excursions', float, (num_runs, num_policies, num_test_eval))
])
performance_memmap_path = str(output_dir / 'performance.npy')
if os.path.isfile(performance_memmap_path):
performance_memmap = np.lib.format.open_memmap(performance_memmap_path,
mode='r+')
else:
performance_memmap = np.lib.format.open_memmap(performance_memmap_path,
shape=(len(
args.parameters), ),
dtype=performance_dtype,
mode='w+')
# Run ACE for each configuration in parallel:
with utils.tqdm_joblib(
tqdm(total=num_runs * len(args.parameters),
smoothing=0)) as progress_bar:
Parallel(n_jobs=args.num_cpus, verbose=0)(
delayed(
run_ace)(experience_memmap, policies_memmap,
performance_memmap, run_num, config_num, parameters,
random_seed, experience_memmap_test, num_test_eval)
for config_num, parameters in enumerate(args.parameters)
for run_num, random_seed in enumerate(random_seeds))
transition_dtype = np.dtype([('s_t', float,
env.observation_space.shape)])
experience_memmap_path = str(output_dir / 'experience_test.npy')
if os.path.isfile(experience_memmap_path):
experience_memmap = np.lib.format.open_memmap(
experience_memmap_path, mode='r+')
else:
experience_memmap = np.lib.format.open_memmap(
experience_memmap_path,
shape=(args.num_runs, args.num_timesteps),
dtype=transition_dtype,
mode='w+')
# Generate the experience in parallel:
with utils.tqdm_joblib(tqdm(total=args.num_runs)) as progress_bar:
Parallel(n_jobs=args.num_cpus, verbose=0)(
delayed(generate_experience_test)(experience_memmap, run_num,
random_seed)
for run_num, random_seed in enumerate(random_seeds))
else:
utils.save_args_to_file(args, output_dir / 'experience.args')
# Create the memmapped structured array of experience to be populated in parallel:
if args.environment == 'pw':
env = puddleworld()
else:
env = gym.make(
args.environment
).unwrapped # Make a dummy env to get shape info for observations.
transition_dtype = np.dtype([