def test_dump_and_load_results(self):
result_data = random_result_data(size = 10)
br = BenchmarkResult(result_dict=result_data)
br.dump("./results")
br_loaded = BenchmarkResult.load("./results")
loaded_dict = br_loaded.get_result_dict()
self.assertEqual(result_data, loaded_dict)
def run(self, viewer=None, maintain_history=False, checkpoint_every=None):
results = []
histories = {}
for idx, bmark_conf in enumerate(self.configs_to_run):
self.logger.info(
"Running config idx {} being {}/{}: Scenario {} of set \"{}\" for behavior \"{}\""
.format(bmark_conf.config_idx, idx,
len(self.benchmark_configs) - 1,
bmark_conf.scenario_idx, bmark_conf.scenario_set_name,
bmark_conf.behavior_config.behavior_name))
bmark_conf = copy.deepcopy(
bmark_conf) if self._deepcopy else bmark_conf
result_dict, scenario_history = self._run_benchmark_config(
bmark_conf, viewer, maintain_history)
results.append(result_dict)
histories[bmark_conf.config_idx] = scenario_history
if self.log_eval_avg_every and (idx +
1) % self.log_eval_avg_every == 0:
self._log_eval_average(results, self.configs_to_run)
if checkpoint_every and (idx + 1) % checkpoint_every == 0:
intermediate_result = BenchmarkResult(results, \
self.configs_to_run[0:idx+1], histories=histories)
checkpoint_file = os.path.join(self.checkpoint_dir,
self.get_checkpoint_file_name())
intermediate_result.dump(checkpoint_file,
dump_configs=True,
dump_histories=maintain_history)
self.logger.info("Saved checkpoint {}".format(checkpoint_file))
benchmark_result = BenchmarkResult(results,
self.configs_to_run,
histories=histories)
self.existing_benchmark_result.extend(benchmark_result)
return self.existing_benchmark_result
def _log_eval_average(self, result_dct_list, configs):
bresult = BenchmarkResult(result_dct_list, configs)
df = bresult.get_data_frame()
grouped = df.apply(pd.to_numeric, errors='ignore').groupby(bresult.get_evaluation_groups()).mean()[
self._evaluation_criteria()]
self.logger.info("\n------------------- Current Evaluation Results ---------------------- \n Num. Results:{}\n {} \n \
---------------------------------------------------------------------".format(len(result_dct_list),
grouped.to_string()))
def __init__(self,
benchmark_database=None,
evaluators=None,
terminal_when=None,
behaviors=None,
behavior_configs=None,
num_scenarios=None,
benchmark_configs=None,
scenario_generation=None,
logger_name=None,
log_eval_avg_every=None,
checkpoint_dir=None,
merge_existing=False,
deepcopy=True):
self.benchmark_database = benchmark_database
self.scenario_generation = scenario_generation
self.evaluators = evaluators if isinstance(evaluators, EvaluationConfig) else EvaluationConfig(evaluators)
self.terminal_when = terminal_when or []
if behaviors:
self.behavior_configs = BehaviorConfig.configs_from_dict(behaviors)
else:
self.behavior_configs = behavior_configs or {}
if benchmark_configs:
self.benchmark_configs = benchmark_configs
elif benchmark_database:
self.benchmark_configs = \
self._create_configurations_from_database(num_scenarios)
elif scenario_generation:
self.benchmark_configs = \
self._create_configurations_from_scenario_generation(num_scenarios)
else:
self.benchmark_configs = [] # to be compatible when benchmark runner is used in ray actor
self.logger = logging.getLogger(logger_name or "BenchmarkRunner")
self.logger.setLevel(logging.DEBUG)
self.logger.info("Total number of {} configs to run".format(len(self.benchmark_configs)))
self.configs_to_run = self.benchmark_configs
self._deepcopy = deepcopy
self.checkpoint_dir = checkpoint_dir or "checkpoints"
self.existing_benchmark_result = BenchmarkResult(file_name= \
BenchmarkRunner.get_merged_result_filename(self.checkpoint_dir))
if not os.path.exists(self.checkpoint_dir):
os.makedirs(self.checkpoint_dir)
if merge_existing:
self.existing_benchmark_result = \
BenchmarkRunner.merge_checkpoint_benchmark_results(checkpoint_dir)
self.logger.info("Merged {} processed configs in folder {}". \
format(len(self.existing_benchmark_result.get_benchmark_configs()), checkpoint_dir))
self.configs_to_run = self.get_configs_to_run(self.benchmark_configs, \
self.existing_benchmark_result)
self.logger.info("Remaining number of {} configs to run".format(len(self.configs_to_run)))
self.exceptions_caught = []
self.log_eval_avg_every = log_eval_avg_every
def _log_eval_average(self, result_dct_list, configs):
bresult = BenchmarkResult(result_dct_list, configs)
df = bresult.get_data_frame()
for eval_group in bresult.get_evaluation_groups():
if eval_group not in df.columns:
df[eval_group] = np.nan
df.fillna(-1, inplace=True)
grouped = df.apply(pd.to_numeric, errors='ignore').groupby(bresult.get_evaluation_groups()).mean()
self.logger.info("\n------------------- Current Evaluation Results ---------------------- \n Num. Results:{}\n {} \n \
---------------------------------------------------------------------".format(len(result_dct_list),
grouped.to_string()))
def test_dump_and_load_histories_one(self):
result_num = 2
result_data = random_result_data(size = result_num)
histories = random_history_data(result_num, 20)
br = BenchmarkResult(result_dict=result_data, histories=histories)
br.dump("./results_with_history", dump_histories=True, max_mb_per_file = 2)
br_loaded = BenchmarkResult.load("./results_with_history")
br_loaded.load_histories(config_idx_list = list(histories.keys()))
loaded_histories = br_loaded.get_histories()
self.assertEqual(histories, loaded_histories)
loaded_dict = br_loaded.get_result_dict()
self.assertEqual(result_data, loaded_dict)
def merge_checkpoint_benchmark_results(checkpoint_dir):
checkpoint_files = glob.glob(os.path.join(checkpoint_dir,
"**/*.ckpnt"),
recursive=True)
merged_result = BenchmarkResult()
# merge all checkpoints with new results
for checkpoint_file in checkpoint_files:
logging.info("Loading checkpoint {}".format(
os.path.abspath(checkpoint_file)))
next_result = BenchmarkResult.load(os.path.abspath(checkpoint_file), \
load_configs=True, load_histories=True)
merged_result.extend(next_result)
# dump merged result
if len(merged_result.get_result_dict()) > 0:
logging.info("Dumping merged result")
merged_result_filename = os.path.join(checkpoint_dir,
"merged_results.ckpnt")
merged_result.dump(merged_result_filename, \
dump_configs=True, dump_histories=True)
# delete checkpoints
for checkpoint_file in checkpoint_files:
if checkpoint_file == merged_result_filename:
continue
os.remove(checkpoint_file)
logging.info(
"Removed old checkpoint file {}".format(checkpoint_file))
return merged_result
def dump():
ckpt_dir = [
os.path.join('/home/ekumar/master_thesis/code/hythe-src/checkpoints/',
ck) for ck in
os.listdir("/home/ekumar/master_thesis/code/hythe-src/checkpoints/")
]
eval_criteria = {"goal_reached": lambda x: x}
print(ckpt_dir)
demos_dir = os.path.join('/home/ekumar/demos/')
list_of_demos = []
for cdir in ckpt_dir:
print(f"Extracting result {cdir}")
result = BenchmarkResult.load_results(cdir)
democ = DemonstrationCollector()
democ._collection_result = result
democ._directory = demos_dir
demos = democ.ProcessCollectionResult(eval_criteria)
list_of_demos.extend(demos)
# make the demonstrations dir in the exp root
os.makedirs(
"/home/ekumar/output/experiments/exp_c76fc949-e95f-4774-91ba-6bec575ada37/demonstrations/generated_demonstrations"
)
to_pickle(
list_of_demos,
"/home/ekumar/output/experiments/exp_c76fc949-e95f-4774-91ba-6bec575ada37/demonstrations/generated_demonstrations",
"demonstrations")
collector = DemonstrationCollector.load(
"/home/ekumar/output/experiments/exp_c76fc949-e95f-4774-91ba-6bec575ada37/demonstrations/generated_demonstrations"
)
print("Total demonstations found:",
len(collector.GetDemonstrationExperiences()))
return
def get_configs_to_run(benchmark_configs, existing_benchmark_result):
existing_inds = existing_benchmark_result.get_benchmark_config_indices()
required_inds = BenchmarkResult(benchmark_configs=benchmark_configs).get_benchmark_config_indices()
missing_inds = list(set(required_inds) - set(existing_inds))
filtered_configs = filter(lambda bc : bc.config_idx in missing_inds, benchmark_configs)
return list(filtered_configs)
def dummy_benchmark_results():
br = [
{
"config_idx": 24,
"collision": False,
"rss": True,
"behavior": "test1",
"scen_idx": 12
},
{
"config_idx": 1,
"collision": True,
"rss": False,
"behavior": "test",
"scen_idx": 10
},
{
"config_idx": 500,
"collision": True,
"rss": True,
"behavior": "test",
"scen_idx": 112
},
]
behaviorc = BehaviorConfig("behavior_name", "behavior")
benchmarkc = BenchmarkConfig(0, behaviorc, "test scenario", 0,
"scenario_set_name")
return BenchmarkResult(result_dict=br, benchmark_configs=[benchmarkc])
def run_benchmark_config(self, config_idx, **kwargs):
for idx, bmark_conf in enumerate(self.benchmark_configs):
if bmark_conf.config_idx == config_idx:
result_dict, scenario_history = self._run_benchmark_config(copy.deepcopy(bmark_conf), **kwargs)
return BenchmarkResult(result_dict, [bmark_conf], histories={config_idx : scenario_history})
self.logger.error("Config idx {} not found in benchmark configs. Skipping...".format(config_idx))
return
def __init__(self,
benchmark_database=None,
evaluators=None,
terminal_when=None,
behaviors=None,
behavior_configs=None,
num_scenarios=None,
benchmark_configs=None,
logger_name=None,
log_eval_avg_every=None,
checkpoint_dir=None,
merge_existing=False):
self.benchmark_database = benchmark_database
self.evaluators = evaluators or {}
self.terminal_when = terminal_when or []
if behaviors:
self.behavior_configs = BehaviorConfig.configs_from_dict(behaviors)
else:
self.behavior_configs = behavior_configs or {}
self.benchmark_configs = benchmark_configs or \
self._create_configurations(num_scenarios)
self.logger = logging.getLogger(logger_name or "BenchmarkRunner")
self.logger.setLevel(logging.DEBUG)
self.logger.info("Total number of {} configs to run".format(len(self.benchmark_configs)))
self.existing_benchmark_result = BenchmarkResult()
self.configs_to_run = self.benchmark_configs
self.checkpoint_dir = checkpoint_dir or "checkpoints"
if not os.path.exists(self.checkpoint_dir):
os.makedirs(self.checkpoint_dir)
if merge_existing:
self.existing_benchmark_result = \
BenchmarkRunner.merge_checkpoint_benchmark_results(checkpoint_dir)
self.logger.info("Merged {} processed configs in folder {}". \
format(len(self.existing_benchmark_result.get_benchmark_configs()), checkpoint_dir))
self.configs_to_run = self.get_configs_to_run(self.benchmark_configs, \
self.existing_benchmark_result)
self.logger.info("Remaining number of {} configs to run".format(len(self.configs_to_run)))
self.exceptions_caught = []
self.log_eval_avg_every = log_eval_avg_every
def test_make_scenarios_congruent(self):
brst = BenchmarkResult(result_dict=dummy_benchmark_results(),
benchmark_configs=None)
analyzer = BenchmarkAnalyzer(benchmark_result=brst)
# scenarios 5 and 12 in all lists
congruent_list = analyzer.make_scenarios_congruent(
configs_idx_lists=[[3, 2, 500, 11], [35, 1, 41],
[500, 41, 12, 121]])
self.assertEqual(congruent_list[0], [11, 2])
def test_dump_and_partial_load(self):
result_num = 100
confs, result_data, histories = random_benchmark_conf_data(result_num, 2000000, hist_size=1500000)
br = BenchmarkResult(result_dict=result_data,
benchmark_configs=confs, histories=histories)
br.dump("./results_all", dump_configs=True, dump_histories=True, max_mb_per_file = 5)
br_loaded = BenchmarkResult.load("./results_all")
loaded_dict = br_loaded.get_result_dict()
self.assertEqual(br.get_result_dict(), loaded_dict)
loaded_configs_idx = list(range(10, 20))
processed_files = br_loaded.load_benchmark_configs(config_idx_list = loaded_configs_idx)
loaded_confs = br_loaded.get_benchmark_configs()
self.assertEqual(len(loaded_confs), 10)
# 2mb per conf, max 5 mb per file -> 2 confs per file -> 10/2 = 5files
self.assertEqual(len(processed_files), 5)
for conf_idx in loaded_configs_idx:
self.assertEqual(br_loaded.get_benchmark_config(conf_idx), confs[conf_idx])
loaded_configs_idx = list(range(10, 27))
processed_files = br_loaded.load_histories(config_idx_list = loaded_configs_idx)
loaded_histories = br_loaded.get_histories()
self.assertEqual(len(loaded_histories), 18) # one more as specified since it was in the last file
# 1.5mb per history, max 5 mb per file -> 3 confs per file -> 17/3 = 6files
self.assertEqual(len(processed_files), 6)
for conf_idx in loaded_configs_idx:
self.assertEqual(br_loaded.get_history(conf_idx), histories[conf_idx])
def load(directory):
collector = DemonstrationCollector()
collection_result_fullname = os.path.join(
directory, DemonstrationCollector.collection_result_filename())
if os.path.exists(collection_result_fullname):
collector._collection_result = BenchmarkResult.load(
collection_result_fullname)
demonstration_fullname = os.path.join(
directory, DemonstrationCollector.demonstrations_filename())
if os.path.exists(demonstration_fullname):
collector._demonstrations = from_pickle(
directory, DemonstrationCollector.demonstrations_filename())
collector._directory = directory
return collector
def test_find_config_idx(self):
brst = BenchmarkResult(result_dict=dummy_benchmark_results(),
benchmark_configs=None)
analyzer = BenchmarkAnalyzer(benchmark_result=brst)
configs_found = analyzer.find_configs({"collision": lambda x: x})
self.assertEqual(configs_found, [1, 500, 11])
configs_found = analyzer.find_configs({"collision": lambda x: not x})
self.assertEqual(configs_found, [24, 2, 41, 3, 12, 121, 35, 42])
configs_found = analyzer.find_configs({"metric1": lambda x: x == 0.1})
self.assertEqual(configs_found, [24, 11])
configs_found = analyzer.find_configs({
"collision": lambda x: not x,
"metric1": lambda x: x == 0.1
})
self.assertEqual(configs_found, [24])
configs_found = analyzer.find_configs({
"collision": lambda x: x,
"metric1": lambda x: x > 1
})
self.assertEqual(configs_found, [1, 500])
configs_found = analyzer.find_configs({
"collision":
lambda x: x,
"behavior":
lambda x: x == "test2"
})
self.assertEqual(configs_found, [])
configs_found = analyzer.find_configs(scenario_idx_list=[4, 10, 7])
self.assertEqual(configs_found, [121, 1, 42])
configs_found = analyzer.find_configs({"collision": lambda x: not x},
scenario_idx_list=[4, 10, 7])
self.assertEqual(configs_found, [121, 42])
configs_found = analyzer.find_configs(
{"collision": lambda x: x},
scenarios_as_in_configs=[24, 11, 121, 1])
self.assertEqual(configs_found, [11, 1])
configs_found = analyzer.find_configs(scenario_idx_list=[4, 10, 7],
in_configs=[1, 42])
self.assertEqual(configs_found, [1, 42])
def test_dump_and_load_benchmark_configs(self):
result_num = 100
confs, result_data, _ = random_benchmark_conf_data(result_num, 2000000)
br = BenchmarkResult(result_dict=result_data, benchmark_configs=confs)
br.dump("./results_with_confs", dump_configs=True, max_mb_per_file = 5)
br_loaded = BenchmarkResult.load("./results_with_confs")
br_loaded.load_benchmark_configs(config_idx_list = list(range(0, result_num)))
loaded_confs = br_loaded.get_benchmark_configs()
self.assertEqual(confs, loaded_confs)
loaded_dict = br_loaded.get_result_dict()
self.assertEqual(br.get_result_dict(), loaded_dict)
def run(self, viewer=None, maintain_history=False, checkpoint_every=None):
results_tmp = ray.get([
actor.run.remote(viewer, maintain_history, checkpoint_every)
for actor in self.actors
])
result_dict = []
benchmark_configs = []
histories = {}
for result_tmp in results_tmp:
result_dict.extend(result_tmp.get_result_dict())
benchmark_configs.extend(result_tmp.get_benchmark_configs())
histories.update(result_tmp.get_histories())
benchmark_result = BenchmarkResult(result_dict,
benchmark_configs,
histories=histories)
self.existing_benchmark_result.extend(benchmark_result)
return self.existing_benchmark_result
def merge_checkpoint_benchmark_results(checkpoint_dir):
checkpoint_files = glob.glob(os.path.join(checkpoint_dir, "**/*.ckpnt"), recursive=True)
merged_result_filename = BenchmarkRunner.get_merged_result_filename(checkpoint_dir)
if os.path.exists(merged_result_filename):
merged_result = BenchmarkResult.load_results(filename=merged_result_filename)
else:
merged_result = BenchmarkResult(file_name=merged_result_filename)
# merge all checkpoints with new results
for checkpoint_file in checkpoint_files:
loaded_result = BenchmarkResult.load(os.path.abspath(checkpoint_file))
merged_result.extend(loaded_result, file_level=True)
logging.info("Extending with checkpoint {}".format(checkpoint_file))
# delete checkpoints
for checkpoint_file in checkpoint_files:
if "merged_result" in checkpoint_file:
continue
os.remove(checkpoint_file)
logging.info("Removed old checkpoint file {}".format(checkpoint_file))
return merged_result
def test_database_run_and_analyze(self):
dbs = DatabaseSerializer(test_scenarios=2, test_world_steps=3, num_serialize_scenarios=2)
# to find database files
cwd = os.getcwd()
if not debug:
os.chdir("../benchmark_database/")
else:
os.chdir("bazel-bin/bark/benchmark/tests/py_benchmark_process_tests.runfiles/benchmark_database")
dbs.process("data/database1")
local_release_filename = dbs.release(version="test")
db = BenchmarkDatabase(database_root=local_release_filename)
evaluators = {"success" : "EvaluatorGoalReached", "collision" : "EvaluatorCollisionEgoAgent",
"max_steps": "EvaluatorStepCount"}
terminal_when = {"collision" :lambda x: x, "max_steps": lambda x : x>2}
params = ParameterServer() # only for evaluated agents not passed to scenario!
behaviors_tested = {"IDM": BehaviorIDMClassic(params), "Const" : BehaviorConstantAcceleration(params)}
benchmark_runner = BenchmarkRunnerMP(benchmark_database=db,
evaluators=evaluators,
terminal_when=terminal_when,
behaviors=behaviors_tested,
log_eval_avg_every=2)
result = benchmark_runner.run(maintain_history=True)
result.dump(os.path.join("./benchmark_results"), dump_configs=True, \
dump_histories=True, max_mb_per_file=1)
result_loaded = BenchmarkResult.load(os.path.join("./benchmark_results"))
result_loaded.load_histories()
result_loaded.load_benchmark_configs()
params2 = ParameterServer()
fig = plt.figure(figsize=[10, 10])
viewer = MPViewer(
params=params2,
center=[5112, 5165],
y_length = 120,
enforce_y_length=True,
axis = fig.gca())
analyzer = BenchmarkAnalyzer(benchmark_result=result_loaded)
configs = analyzer.find_configs(criteria={"behavior": lambda x: x=="IDM", "success": lambda x : not x})
configs_const = analyzer.find_configs(criteria={"behavior": lambda x: x=="Const", "success": lambda x : not x})
os.chdir(cwd)
#analyzer.visualize(configs_idx_list = configs,
# viewer = viewer, real_time_factor=10, fontsize=12)
plt.close(fig)
fig, (ax1, ax2) = plt.subplots(1, 2)
viewer1 = MPViewer(
params=params2,
center=[5112, 5165],
y_length = 120,
enforce_y_length=True,
axis = ax1)
viewer2 = MPViewer(
params=params2,
center=[5112, 5165],
y_length = 120,
enforce_y_length=True,
axis = ax2)
analyzer.visualize(configs_idx_list = [configs[1:3], configs_const[1:3]],
viewer = [viewer1, viewer2], viewer_names=["IDM", "ConstVelocity"], real_time_factor=10, fontsize=12)
class BenchmarkRunner:
def __init__(self,
benchmark_database=None,
evaluators=None,
terminal_when=None,
behaviors=None,
behavior_configs=None,
num_scenarios=None,
benchmark_configs=None,
logger_name=None,
log_eval_avg_every=None,
checkpoint_dir=None,
merge_existing=False,
deepcopy=True):
self.benchmark_database = benchmark_database
self.evaluators = evaluators or {}
self.terminal_when = terminal_when or []
if behaviors:
self.behavior_configs = BehaviorConfig.configs_from_dict(behaviors)
else:
self.behavior_configs = behavior_configs or {}
self.benchmark_configs = benchmark_configs or \
self._create_configurations(num_scenarios)
self.logger = logging.getLogger(logger_name or "BenchmarkRunner")
self.logger.setLevel(logging.DEBUG)
self.logger.info("Total number of {} configs to run".format(
len(self.benchmark_configs)))
self.existing_benchmark_result = BenchmarkResult()
self.configs_to_run = self.benchmark_configs
self._deepcopy = deepcopy
self.checkpoint_dir = checkpoint_dir or "checkpoints"
if not os.path.exists(self.checkpoint_dir):
os.makedirs(self.checkpoint_dir)
if merge_existing:
self.existing_benchmark_result = \
BenchmarkRunner.merge_checkpoint_benchmark_results(checkpoint_dir)
self.logger.info("Merged {} processed configs in folder {}". \
format(len(self.existing_benchmark_result.get_benchmark_configs()), checkpoint_dir))
self.configs_to_run = self.get_configs_to_run(self.benchmark_configs, \
self.existing_benchmark_result)
self.logger.info("Remaining number of {} configs to run".format(
len(self.configs_to_run)))
self.exceptions_caught = []
self.log_eval_avg_every = log_eval_avg_every
def get_checkpoint_file_name(self):
return "benchmark_runner.ckpnt"
def clear_checkpoint_dir(self):
files = glob.glob(os.path.join(self.checkpoint_dir, "*.ckpnt"))
for f in files:
os.remove(f)
@staticmethod
def merge_checkpoint_benchmark_results(checkpoint_dir):
checkpoint_files = glob.glob(os.path.join(checkpoint_dir,
"**/*.ckpnt"),
recursive=True)
merged_result = BenchmarkResult()
# merge all checkpoints with new results
for checkpoint_file in checkpoint_files:
logging.info("Loading checkpoint {}".format(
os.path.abspath(checkpoint_file)))
next_result = BenchmarkResult.load(os.path.abspath(checkpoint_file), \
load_configs=True, load_histories=True)
merged_result.extend(next_result)
# dump merged result
if len(merged_result.get_result_dict()) > 0:
logging.info("Dumping merged result")
merged_result_filename = os.path.join(checkpoint_dir,
"merged_results.ckpnt")
merged_result.dump(merged_result_filename, \
dump_configs=True, dump_histories=True)
# delete checkpoints
for checkpoint_file in checkpoint_files:
if checkpoint_file == merged_result_filename:
continue
os.remove(checkpoint_file)
logging.info(
"Removed old checkpoint file {}".format(checkpoint_file))
return merged_result
@staticmethod
def get_configs_to_run(benchmark_configs, existing_benchmark_result):
existing_inds = existing_benchmark_result.get_benchmark_config_indices(
)
required_inds = BenchmarkResult(benchmark_configs=benchmark_configs
).get_benchmark_config_indices()
missing_inds = list(set(required_inds) - set(existing_inds))
filtered_configs = filter(lambda bc: bc.config_idx in missing_inds,
benchmark_configs)
return list(filtered_configs)
def _create_configurations(self, num_scenarios=None):
benchmark_configs = []
for behavior_config in self.behavior_configs:
# run over all scenario generators from benchmark database
for scenario_generator, scenario_set_name, scenario_set_param_desc in self.benchmark_database:
for scenario, scenario_idx in scenario_generator:
if num_scenarios and scenario_idx >= num_scenarios:
break
benchmark_config = \
BenchmarkConfig(
len(benchmark_configs),
behavior_config,
scenario,
scenario_idx,
scenario_set_name,
scenario_set_param_desc
)
benchmark_configs.append(benchmark_config)
return benchmark_configs
def run(self, viewer=None, maintain_history=False, checkpoint_every=None):
results = []
histories = {}
for idx, bmark_conf in enumerate(self.configs_to_run):
self.logger.info(
"Running config idx {} being {}/{}: Scenario {} of set \"{}\" for behavior \"{}\""
.format(bmark_conf.config_idx, idx,
len(self.benchmark_configs) - 1,
bmark_conf.scenario_idx, bmark_conf.scenario_set_name,
bmark_conf.behavior_config.behavior_name))
bmark_conf = copy.deepcopy(
bmark_conf) if self._deepcopy else bmark_conf
result_dict, scenario_history = self._run_benchmark_config(
bmark_conf, viewer, maintain_history)
results.append(result_dict)
histories[bmark_conf.config_idx] = scenario_history
if self.log_eval_avg_every and (idx +
1) % self.log_eval_avg_every == 0:
self._log_eval_average(results, self.configs_to_run)
if checkpoint_every and (idx + 1) % checkpoint_every == 0:
intermediate_result = BenchmarkResult(results, \
self.configs_to_run[0:idx+1], histories=histories)
checkpoint_file = os.path.join(self.checkpoint_dir,
self.get_checkpoint_file_name())
intermediate_result.dump(checkpoint_file,
dump_configs=True,
dump_histories=maintain_history)
self.logger.info("Saved checkpoint {}".format(checkpoint_file))
benchmark_result = BenchmarkResult(results,
self.configs_to_run,
histories=histories)
self.existing_benchmark_result.extend(benchmark_result)
return self.existing_benchmark_result
def run_benchmark_config(self, config_idx, **kwargs):
for idx, bmark_conf in enumerate(self.benchmark_configs):
if bmark_conf.config_idx == config_idx:
bmark_conf = copy.deepcopy(
bmark_conf) if self._deepcopy else bmark_conf
result_dict, scenario_history = self._run_benchmark_config(
bmark_conf, **kwargs)
return BenchmarkResult(
result_dict, [bmark_conf],
histories={config_idx: scenario_history})
self.logger.error(
"Config idx {} not found in benchmark configs. Skipping...".format(
config_idx))
return
def _run_benchmark_config(self,
benchmark_config,
viewer=None,
maintain_history=False):
scenario = benchmark_config.scenario
behavior = benchmark_config.behavior_config.behavior
parameter_server = ParameterServer(json=scenario._json_params)
scenario_history = []
step = 0
try:
world = scenario.GetWorldState()
except Exception as e:
self.logger.error(
"For config-idx {}, Exception thrown in scenario.GetWorldState: {}"
.format(benchmark_config.config_idx, e))
self._append_exception(benchmark_config, e)
return {
**benchmark_config.as_dict(), "step": step,
"Terminal": "exception_raised"
}
# if behavior is not None (None specifies that also the default model can be evalauted)
if behavior:
world.agents[scenario._eval_agent_ids[0]].behavior_model = behavior
if maintain_history:
self._append_to_scenario_history(scenario_history, world, scenario)
self._reset_evaluators(world, scenario._eval_agent_ids)
step_time = parameter_server["Simulation"]["StepTime", "", 0.2]
if not isinstance(step_time, float):
step_time = 0.2
terminal = False
terminal_why = None
while not terminal:
try:
evaluation_dict = self._get_evalution_dict(world)
except Exception as e:
self.logger.error(
"For config-idx {}, Exception thrown in evaluation: {}".
format(benchmark_config.config_idx, e))
terminal_why = "exception_raised"
self._append_exception(benchmark_config, e)
evaluation_dict = {}
break
terminal, terminal_why = self._is_terminal(evaluation_dict)
if not terminal:
if viewer:
viewer.drawWorld(
world,
scenario._eval_agent_ids,
scenario_idx=benchmark_config.scenario_idx)
viewer.show(block=False)
time.sleep(step_time)
viewer.clear()
try:
world.Step(step_time)
except Exception as e:
self.logger.error(
"For config-idx {}, Exception thrown in world.Step: {}"
.format(benchmark_config.config_idx, e))
terminal_why = "exception_raised"
self._append_exception(benchmark_config, e)
break
if maintain_history:
self._append_to_scenario_history(scenario_history, world,
scenario)
step += 1
dct = {
**benchmark_config.as_dict(), "step": step,
**evaluation_dict, "Terminal": terminal_why
}
return dct, scenario_history
def _append_to_scenario_history(self, scenario_history, world, scenario):
scenario = Scenario(agent_list=list(world.agents.values()),
map_file_name=scenario.map_file_name,
eval_agent_ids=scenario.eval_agent_ids,
json_params=scenario.json_params)
scenario_history.append(scenario.copy())
def _append_exception(self, benchmark_config, exception):
self.exceptions_caught.append((benchmark_config.config_idx, exception))
def _reset_evaluators(self, world, eval_agent_ids):
for evaluator_name, evaluator_params in self.evaluators.items():
evaluator_bark = None
if isinstance(evaluator_params, str):
try:
evaluator_bark = eval(
"{}(eval_agent_ids[0])".format(evaluator_params))
except:
evaluator_bark = eval("{}()".format(evaluator_params))
elif isinstance(evaluator_params, dict):
evaluator_bark = eval(
"{}(agent_id=eval_agent_ids[0], **evaluator_params['params'])"
.format(evaluator_params["type"]))
else:
raise ValueError
world.AddEvaluator(evaluator_name, evaluator_bark)
def _evaluation_criteria(self):
bark_evals = [eval_crit for eval_crit, _ in self.evaluators.items()]
bark_evals.append("step")
return bark_evals
def _get_evalution_dict(self, world):
return world.Evaluate()
def _is_terminal(self, evaluation_dict):
terminal = False
terminal_why = []
for evaluator_name, function in self.terminal_when.items():
if function(evaluation_dict[evaluator_name]):
terminal = True
terminal_why.append(evaluator_name)
return terminal, terminal_why
def _log_eval_average(self, result_dct_list, configs):
bresult = BenchmarkResult(result_dct_list, configs)
df = bresult.get_data_frame()
for eval_group in bresult.get_evaluation_groups():
if eval_group not in df.columns:
df[eval_group] = np.nan
df.fillna(-1, inplace=True)
grouped = df.apply(pd.to_numeric, errors='ignore').groupby(
bresult.get_evaluation_groups()).mean()[
self._evaluation_criteria()]
self.logger.info(
"\n------------------- Current Evaluation Results ---------------------- \n Num. Results:{}\n {} \n \
---------------------------------------------------------------------".format(
len(result_dct_list), grouped.to_string()))
def run(self, viewer=None, maintain_history=False, checkpoint_every=None):
last_results = []
last_histories = {}
last_run_configs = []
results = []
checkpoint_file = os.path.abspath(os.path.join(self.checkpoint_dir, self.get_checkpoint_file_name()))
last_result_file = os.path.abspath(os.path.join(self.checkpoint_dir, "tmp_{}".format(self.get_checkpoint_file_name())))
checkpoint_result = BenchmarkResult(file_name=checkpoint_file)
for idx, bmark_conf in enumerate(self.configs_to_run):
self.logger.info("Running config idx {} being {}/{}: Scenario {} of set \"{}\" for behavior \"{}\"".format(
bmark_conf.config_idx, idx, len(self.benchmark_configs) - 1, bmark_conf.scenario_idx,
bmark_conf.scenario_set_name, bmark_conf.behavior_config.behavior_name))
bmark_conf = copy.deepcopy(bmark_conf) if self._deepcopy else bmark_conf
result_dict, scenario_history = self._run_benchmark_config(bmark_conf, viewer,
maintain_history)
results.append(result_dict)
last_results.append(result_dict)
last_histories[bmark_conf.config_idx] = scenario_history
last_run_configs.append(bmark_conf)
if self.log_eval_avg_every and (idx + 1) % self.log_eval_avg_every == 0:
self._log_eval_average(results, self.configs_to_run)
if checkpoint_every and (idx+1) % checkpoint_every == 0:
# append results since last checkpoint
last_benchmark_result = BenchmarkResult(result_dict=last_results, file_name = last_result_file, \
benchmark_configs=last_run_configs, histories=last_histories)
last_benchmark_result.dump(last_result_file, dump_configs=True, dump_histories=maintain_history, append=False)
checkpoint_result.extend(benchmark_result=last_benchmark_result, file_level=True)
self.logger.info("Extended checkpoint {} with last result.".format(checkpoint_file))
last_histories.clear()
last_run_configs.clear()
last_results.clear()
# append results of last run
last_benchmark_result = BenchmarkResult(result_dict=last_results, file_name = last_result_file, \
benchmark_configs=last_run_configs, histories=last_histories)
last_benchmark_result.dump(last_result_file, dump_configs=True, dump_histories=maintain_history, append=False)
checkpoint_result.extend(benchmark_result=last_benchmark_result, file_level=True)
os.remove(last_result_file)
self.logger.info("Extended checkpoint {} with final result.".format(checkpoint_file))
checkpoint_result.extend(benchmark_result=self.existing_benchmark_result, file_level=True)
return checkpoint_result
def test_extend_from_file(self):
try:
os.remove("./br1")
os.remove("./br2")
os.remove("./br3")
except:
pass
result_num = 100
confs, result_data, histories1 = random_benchmark_conf_data(result_num, 2000000, hist_size=1500000, offset=0)
br1 = BenchmarkResult(result_dict=result_data,
benchmark_configs=confs, histories=histories1)
br1.dump("./br1", dump_histories=True, dump_configs=True)
br1_df = br1.get_data_frame().copy()
result_num = 30
confs2, result_data2, histories2 = random_benchmark_conf_data(result_num, 2000000, hist_size=1500000, offset=200)
br2 = BenchmarkResult(result_dict=result_data2,
benchmark_configs=confs2, histories=histories2)
br2.dump(filename="./br2", dump_histories=True, dump_configs=True)
result_num = 10
confs3, result_data3, histories3 = random_benchmark_conf_data(result_num, 2000000, hist_size=1500000, offset=400)
br3 = BenchmarkResult(result_dict=result_data3,
benchmark_configs=confs3, histories=histories3)
br3.dump(filename="./br3", dump_histories=True, dump_configs=True)
br1.extend(benchmark_result=br2, file_level=True)
br1.extend(benchmark_result=br3, file_level=True)
br_loaded = BenchmarkResult.load("./br1", load_histories=True, load_configs=True)
df_desired = br1_df
df_desired = pd.concat([df_desired, br2.get_data_frame()])
df_desired = pd.concat([df_desired, br3.get_data_frame()])
self.assertEqual(len(br_loaded.get_data_frame().index), len(df_desired.index))
extended_confs = br_loaded.get_benchmark_configs()
self.assertEqual(len(extended_confs), 140)
extended_histories = br_loaded.get_histories()
self.assertEqual(len(extended_histories), 140)
extended_histories = histories1
extended_histories.update(histories2)
extended_histories.update(histories3)
for bc in extended_confs:
self.assertEqual(br_loaded.get_history(bc.config_idx), extended_histories[bc.config_idx])