def evaluate_all(sim, env, pop, print_log, save_vxa_every, run_directory, run_name, max_eval_time=120,
time_to_try_again=10, save_lineages=False):
"""Evaluate all individuals of the population in VoxCad.
Parameters
----------
sim : Sim
Configures parameters of the Voxelyze simulation.
env : Env
Configures parameters of the Voxelyze environment.
pop : Population
This provides the individuals to evaluate.
print_log : PrintLog()
For logging with time stamps
save_vxa_every : int
Which generations to save information about individual SoftBots
run_directory : string
Where to save
run_name : string
Experiment name for files
max_eval_time : int
How long to run physical simulation per ind in pop
time_to_try_again : int
How long to wait until relaunching remaining unevaluated (crashed?) simulations
save_lineages : bool
Save the vxa of every ancestor of the surviving individual
"""
start_time = time.time()
num_evaluated_this_gen = 0
ids_to_analyze = []
controller_evolution = hasattr(pop[0].genotype, "controller")
for ind in pop:
# set environmental parameters defined in the controller
if controller_evolution:
controller = ind.genotype.controller
env.temp_amp = env.temp_base + controller.temp_amplitude
env.period = controller.temp_period
env.cte = controller.muscles_cte
# insert individual in the environment if obstacles have been enabled
if env.obstacles:
for name, details in ind.genotype.to_phenotype_mapping.items():
if details["env_kws"] is None:
env.insert_individual(details)
# write the phenotype of a SoftBot to a file so that VoxCad can access for sim.
ind.md5 = write_voxelyze_file(sim, env, ind, run_directory, run_name)
# don't evaluate if invalid
if not ind.phenotype.is_valid():
for rank, goal in pop.objective_dict.items():
if goal["name"] != "age":
setattr(ind, goal["name"], goal["worst_value"])
print_log.message("Skipping invalid individual")
# don't evaluate if identical phenotype has already been evaluated
elif env.actuation_variance == 0 and ind.md5 in pop.already_evaluated:
for rank, goal in pop.objective_dict.items():
if goal["tag"] is not None:
setattr(ind, goal["name"], pop.already_evaluated[ind.md5][rank])
# print_log.message("Individual already evaluated: cached fitness is {}".format(ind.fitness))
if pop.gen % save_vxa_every == 0 and save_vxa_every > 0:
sub.call("cp " + run_directory + "/voxelyzeFiles/" + run_name + "--id_%05i.vxa" % ind.id +
" " + run_directory + "/Gen_%04i/" % pop.gen + run_name +
"--Gen_%04i--fit_%.08f--id_%05i.vxa" % (pop.gen, ind.fitness, ind.id), shell=True)
# otherwise evaluate with voxelyze
else:
num_evaluated_this_gen += 1
pop.total_evaluations += 1
ids_to_analyze += [ind.id]
sub.Popen("./voxelyze -f " + run_directory + "/voxelyzeFiles/" + run_name + "--id_%05i.vxa" % ind.id,
shell=True)
print_log.message("Launched {0} voxelyze calls, out of {1} individuals".format(num_evaluated_this_gen, len(pop)))
num_evals_finished = 0
all_done = False
already_analyzed_ids = []
redo_attempts = 1
fitness_eval_start_time = time.time()
while not all_done:
time_waiting_for_fitness = time.time() - fitness_eval_start_time
# this protects against getting stuck when voxelyze doesn't return a fitness value
# (diverging simulations, crash, error reading .vxa)
if time_waiting_for_fitness > pop.pop_size * max_eval_time:
# TODO ** WARNING: This could in fact alter the sim and undermine the reproducibility **
all_done = False # something bad with this individual, probably sim diverged
break
if time_waiting_for_fitness > pop.pop_size * time_to_try_again * redo_attempts:
# try to redo any simulations that crashed
redo_attempts += 1
non_analyzed_ids = [idx for idx in ids_to_analyze if idx not in already_analyzed_ids]
print "Rerunning voxelyze for: ", non_analyzed_ids
for idx in non_analyzed_ids:
sub.Popen("./voxelyze -f " + run_directory + "/voxelyzeFiles/" + run_name + "--id_%05i.vxa" % idx,
shell=True)
# check to see if all are finished
all_done = True
for ind in pop:
if ind.phenotype.is_valid() and ind.fitness == pop.objective_dict[0]["worst_value"]:
all_done = False
# check for any fitness files that are present
ls_check = sub.check_output(["ls", run_directory + "/fitnessFiles/"])
# duplicated ids issue: may be due to entering here two times for the same fitness file found in the directory.
if ls_check:
# ls_check = random.choice(ls_check.split()) # doesn't accomplish anything and undermines reproducibility
ls_check = ls_check.split()[0]
if "softbotsOutput--id_" in ls_check:
this_id = int(ls_check.split("_")[1].split(".")[0])
if this_id in already_analyzed_ids:
# workaround to avoid any duplicated ids when restarting sims
print_log.message("Duplicate voxelyze results found from THIS gen with id {}".format(this_id))
sub.call("rm " + run_directory + "/fitnessFiles/" + ls_check, shell=True)
elif this_id in pop.all_evaluated_individuals_ids:
print_log.message("Duplicate voxelyze results found from PREVIOUS gen with id {}".format(this_id))
sub.call("rm " + run_directory + "/fitnessFiles/" + ls_check, shell=True)
else:
num_evals_finished += 1
already_analyzed_ids.append(this_id)
ind_filename = run_directory + "/fitnessFiles/" + ls_check
objective_values_dict = read_voxlyze_results(pop, print_log, ind_filename)
if env.novelty_based:
centroids = read_voxelyze_centroids(pop, print_log, ind_filename)
print_log.message("{0} fit = {1} ({2} / {3})".format(ls_check, objective_values_dict[0],
num_evals_finished,
num_evaluated_this_gen))
# now that we've read the fitness file, we can remove it
sub.call("rm " + run_directory + "/fitnessFiles/" + ls_check, shell=True)
# assign the values to the corresponding individual
for ind in pop:
if ind.id == this_id:
for rank, details in pop.objective_dict.items():
if objective_values_dict[rank] is not None:
setattr(ind, details["name"], objective_values_dict[rank])
if env.novelty_based:
setattr(ind, "trajectory", from_centroids_to_trajectory(centroids))
else:
# for network in ind.genotype:
# for name in network.output_node_names:
# if name == details["output_node_name"]:
# print "here!"
# # apply the specified function to the specified output node
# state = network.graph.node[name]["state"]
# setattr(ind, details["name"], details["node_func"](state))
for name, details_phenotype in ind.genotype.to_phenotype_mapping.items():
if name == details["output_node_name"]:
state = details_phenotype["state"]
setattr(ind, details["name"], details["node_func"](state))
pop.already_evaluated[ind.md5] = [getattr(ind, details["name"])
for rank, details in
pop.objective_dict.items()]
pop.all_evaluated_individuals_ids += [this_id]
# update the run statistics and file management
if ind.fitness > pop.best_fit_so_far:
pop.best_fit_so_far = ind.fitness
sub.call("cp " + run_directory + "/voxelyzeFiles/" + run_name + "--id_%05i.vxa" %
ind.id + " " + run_directory + "/bestSoFar/fitOnly/" + run_name +
"--Gen_%04i--fit_%.08f--id_%05i.vxa" %
(pop.gen, ind.fitness, ind.id), shell=True)
if save_lineages:
sub.call("cp " + run_directory + "/voxelyzeFiles/" + run_name + "--id_%05i.vxa" %
ind.id + " " + run_directory + "/ancestors/", shell=True)
if pop.gen % save_vxa_every == 0 and save_vxa_every > 0:
sub.call("mv " + run_directory + "/voxelyzeFiles/" + run_name + "--id_%05i.vxa" %
ind.id + " " + run_directory + "/Gen_%04i/" % pop.gen +
run_name + "--Gen_%04i--fit_%.08f--id_%05i.vxa" %
(pop.gen, ind.fitness, ind.id), shell=True)
else:
sub.call("rm " + run_directory + "/voxelyzeFiles/" + run_name + "--id_%05i.vxa" %
ind.id, shell=True)
break
# wait a second and try again
else:
time.sleep(0.5)
else:
time.sleep(0.5)
if not all_done:
print_log.message("WARNING: Couldn't get a fitness value in time for some individuals. "
"The min fitness was assigned for these individuals")
print_log.message("\nAll Voxelyze evals finished in {} seconds".format(time.time() - start_time))
print_log.message("num_evaluated_this_gen: {0}".format(num_evaluated_this_gen))
print_log.message("total_evaluations: {}".format(pop.total_evaluations))
def evaluate_all(sim,
env,
pop,
print_log,
save_vxa_every,
run_directory,
run_name,
max_eval_time=60,
time_to_try_again=10,
save_lineages=False,
batch_size=None):
"""Evaluate all individuals of the population in VoxCad.
Parameters
----------
sim : Sim
Configures parameters of the Voxelyze simulation.
env : Env
Configures parameters of the Voxelyze environment.
pop : Population
This provides the individuals to evaluate.
print_log : PrintLog()
For logging with time stamps
save_vxa_every : int
Which generations to save information about individual SoftBots
run_directory : string
Where to save
run_name : string
Experiment name for files
max_eval_time : int
How long to run physical simulation per ind in pop
time_to_try_again : int
How long to wait until relaunching remaining unevaluated (crashed?) simulations
save_lineages : bool
Save the vxa of every ancestor of the surviving individual
batch_size : int
How many voxelyze processes start at once
"""
start_time = time.time()
num_evaluated_this_gen = 0
ids_to_analyze = []
for n, ind in enumerate(pop):
# write the phenotype of a SoftBot to a file so that VoxCad can access for sim.
ind.md5 = write_voxelyze_file(sim, env, ind, run_directory, run_name)
# don't evaluate if invalid
if not ind.phenotype.is_valid():
for rank, goal in pop.objective_dict.items():
if goal["name"] != "age":
setattr(ind, goal["name"], goal["worst_value"])
print_log.message("Skipping invalid individual")
# don't evaluate if identical phenotype has already been evaluated
elif env.actuation_variance == 0 and ind.md5 in pop.already_evaluated:
if not ind.previously_aggregated: # don't overwrite the fitness for aggregations of multiple trials
for rank, goal in pop.objective_dict.items():
if goal["tag"] is not None:
setattr(ind, goal["name"],
pop.already_evaluated[ind.md5][rank])
# print_log.message("Individual already evaluated: cached fitness is {}".format(ind.fitness))
if pop.gen % save_vxa_every == 0 and save_vxa_every > 0:
sub.call("cp " + run_directory + "/voxelyzeFiles/" + run_name +
"--id_%05i.vxa" % ind.id + " " + run_directory +
"/Gen_%04i/" % pop.gen + run_name +
"--Gen_%04i--fit_%.08f--id_%05i.vxa" %
(pop.gen, ind.fitness, ind.id),
shell=True)
# ##############################################################################################################
# # todo: there might be cases (generalize this!) in which we don't want to waste time evaluating in simulation
# elif pop.learning_trials > 0 and \
# get_depths_of_material_from_shell(
# ind.genotype.to_phenotype_mapping['material']['state'], mat=8)[4] > 0:
# for rank, goal in pop.objective_dict.items():
# if goal["name"] != "age":
# setattr(ind, goal["name"], goal["default_value"])
# ##############################################################################################################
# otherwise evaluate with voxelyze
else:
num_evaluated_this_gen += 1
pop.total_evaluations += 1
ids_to_analyze += [ind.id]
sub.Popen("./voxelyze -f " + run_directory + "/voxelyzeFiles/" +
run_name + "--id_%05i.vxa" % ind.id,
shell=True)
# # pause if using mini-batches
# batch_start_time = time.time()
# if batch_size < pop.pop_size:
# if (n+1) % batch_size == 0:
# done_with_batch = False
# while not done_with_batch:
# ls_check = sub.check_output(["ls", run_directory + "/fitnessFiles/"])
# if len(ls_check.split()) == n+1:
# done_with_batch = True
#
# elif time.time() - batch_start_time > time_to_try_again*batch_size:
# done_with_batch = True # resend this robot later
#
# else:
# time.sleep(0.5)
print_log.message(
"Launched {0} voxelyze calls, out of {1} individuals".format(
num_evaluated_this_gen, len(pop)))
num_evals_finished = 0
all_done = False
already_analyzed_ids = []
redo_attempts = 1
fitness_eval_start_time = time.time()
if num_evaluated_this_gen == 0:
all_done = True
# if batch_size < pop.pop_size:
# fitness_eval_start_time = batch_start_time # use timer from just before batch started
while not all_done:
time_waiting_for_fitness = time.time() - fitness_eval_start_time
# this protects against getting stuck when voxelyze doesn't return a fitness value
# (diverging simulations, crash, error reading .vxa)
if time_waiting_for_fitness > pop.pop_size * max_eval_time:
# TODO ** WARNING: This could in fact alter the sim and undermine the reproducibility **
all_done = False # something bad with this individual, probably sim diverged
break
if time_waiting_for_fitness > pop.pop_size * time_to_try_again * redo_attempts:
# try to redo any simulations that crashed
redo_attempts += 1
non_analyzed_ids = [
idx for idx in ids_to_analyze
if idx not in already_analyzed_ids
]
print "Rerunning voxelyze for: ", non_analyzed_ids
for idx in non_analyzed_ids:
sub.Popen("./voxelyze -f " + run_directory +
"/voxelyzeFiles/" + run_name + "--id_%05i.vxa" % idx,
shell=True)
# check to see if all are finished
all_done = True
for ind in pop:
if ind.phenotype.is_valid(
) and ind.fitness == pop.objective_dict[0]["worst_value"]:
all_done = False
# check for any fitness files that are present
ls_check = sub.check_output(["ls", run_directory + "/fitnessFiles/"])
# duplicated ids issue: may be due to entering here two times for the same fitness file found in the directory.
if ls_check:
# ls_check = random.choice(ls_check.split()) # doesn't accomplish anything and undermines reproducibility
ls_check = ls_check.split()[0]
if "softbotsOutput--id_" in ls_check:
this_id = int(ls_check.split("_")[1].split(".")[0])
if this_id in already_analyzed_ids:
# workaround to avoid any duplicated ids when restarting sims
print_log.message(
"Duplicate voxelyze results found from THIS gen with id {}"
.format(this_id))
sub.call("rm " + run_directory + "/fitnessFiles/" +
ls_check,
shell=True)
elif this_id in pop.all_evaluated_individuals_ids and pop.learning_trials <= 1:
print_log.message(
"Duplicate voxelyze results found from PREVIOUS gen with id {}"
.format(this_id))
sub.call("rm " + run_directory + "/fitnessFiles/" +
ls_check,
shell=True)
else:
num_evals_finished += 1
already_analyzed_ids.append(this_id)
ind_filename = run_directory + "/fitnessFiles/" + ls_check
objective_values_dict = read_voxlyze_results(
pop, print_log, ind_filename)
print_log.message("{0} fit = {1} ({2} / {3})".format(
ls_check, objective_values_dict[0], num_evals_finished,
num_evaluated_this_gen))
# now that we've read the fitness file, we can remove it
sub.call("rm " + run_directory + "/fitnessFiles/" +
ls_check,
shell=True)
# assign the values to the corresponding individual
for ind in pop:
if ind.id == this_id:
for rank, details in pop.objective_dict.items():
if objective_values_dict[rank] is not None:
this_value = objective_values_dict[rank]
if details["meta_func"] is not None:
this_value = details["meta_func"](
this_value, ind)
setattr(ind, details["name"], this_value)
else:
# for network in ind.genotype:
# for name in network.output_node_names:
# if name == details["output_node_name"]:
# print "here!"
# # apply the specified function to the specified output node
# state = network.graph.node[name]["state"]
# setattr(ind, details["name"], details["node_func"](state))
for name, details_phenotype in ind.genotype.to_phenotype_mapping.items(
):
if name == details["output_node_name"]:
state = details_phenotype["state"]
setattr(
ind, details["name"],
details["node_func"](state))
pop.already_evaluated[ind.md5] = [
getattr(ind, details["name"]) for rank, details
in pop.objective_dict.items()
]
pop.all_evaluated_individuals_ids += [this_id]
# update the run statistics and file management
new_run_champ = False
if pop.objective_dict[0]["maximize"]:
if ind.fitness > pop.best_fit_so_far:
new_run_champ = True
elif ind.fitness < pop.best_fit_so_far:
new_run_champ = True
if new_run_champ:
pop.best_fit_so_far = ind.fitness
sub.call(
"cp " + run_directory + "/voxelyzeFiles/" +
run_name + "--id_%05i.vxa" % ind.id + " " +
run_directory + "/bestSoFar/fitOnly/" +
run_name +
"--Gen_%04i--fit_%.08f--id_%05i.vxa" %
(pop.gen, ind.fitness, ind.id),
shell=True)
if save_lineages:
sub.call("cp " + run_directory +
"/voxelyzeFiles/" + run_name +
"--id_%05i.vxa" % ind.id + " " +
run_directory + "/ancestors/",
shell=True)
if pop.gen % save_vxa_every == 0 and save_vxa_every > 0:
sub.call(
"mv " + run_directory + "/voxelyzeFiles/" +
run_name + "--id_%05i.vxa" % ind.id + " " +
run_directory + "/Gen_%04i/" % pop.gen +
run_name +
"--Gen_%04i--fit_%.08f--id_%05i.vxa" %
(pop.gen, ind.fitness, ind.id),
shell=True)
else:
sub.call("rm " + run_directory +
"/voxelyzeFiles/" + run_name +
"--id_%05i.vxa" % ind.id,
shell=True)
break
# wait a second and try again
else:
time.sleep(0.5)
else:
time.sleep(0.5)
if not all_done:
print_log.message(
"WARNING: Couldn't get a fitness value in time for some individuals. "
"The min fitness was assigned for these individuals")
print_log.message(
"\nAll Voxelyze evals finished in {} seconds".format(time.time() -
start_time))
print_log.message(
"num_evaluated_this_gen: {0}".format(num_evaluated_this_gen))
print_log.message("total_evaluations: {}".format(pop.total_evaluations))
