def run(self, max_hours_runtime=29, max_gens=3000, num_random_individuals=1, num_env_cycles=0,
directory="tests_data", name="TestRun", max_eval_time=60, time_to_try_again=30, checkpoint_every=100,
save_vxa_every=100, save_pareto=False, save_nets=False, save_lineages=False,
continued_from_checkpoint=False, batch_size=None, update_survivors_age=True, max_fitness=None):
if self.autosuspended:
sub.call("rm %s/AUTOSUSPENDED" % directory, shell=True)
self.autosuspended = False
self.max_gens = max_gens # can add additional gens through checkpointing
print_log = PrintLog()
print_log.add_timer("evaluation")
self.start_time = print_log.timers["start"] # sync start time with logging
if not continued_from_checkpoint: # generation zero
self.directory = directory
self.name = name
self.num_random_inds = num_random_individuals
self.num_env_cycles = num_env_cycles
initialize_folders(self.pop, self.directory, self.name, save_nets, save_lineages=save_lineages)
sub.call("touch {}/RUNNING".format(self.directory), shell=True)
while self.pop.gen < max_gens:
if self.pop.gen % checkpoint_every == 0:
print_log.message("Saving checkpoint at generation {0}".format(self.pop.gen + 1), timer_name="start")
self.save_checkpoint(self.directory, self.pop.gen)
if self.elapsed_time(units="h") > max_hours_runtime or self.pop.best_fit_so_far == max_fitness:
self.autosuspended = True
print_log.message("Autosuspending at generation {0}".format(self.pop.gen + 1), timer_name="start")
self.save_checkpoint(self.directory, self.pop.gen)
sub.call("touch {0}/AUTOSUSPENDED && rm {0}/RUNNING".format(self.directory), shell=True)
break
print_log.message("Creating folders structure for this generation")
make_gen_directories(self.pop, self.directory, save_vxa_every, save_nets)
# mutation
print_log.message("Mutation starts")
new_solutions = self.make_children()
new_children = [self.encoding(t) for t in new_solutions]
print_log.message("Mutation ends: successfully generated %d new children." % (len(new_children)))
# replace old pop
print_log.message("Replacing population with new designs")
self.pop.individuals = []
for mat in new_children:
self.pop.add_random_individual()
for old_net in self.pop[-1].genotype.networks:
if old_net.output_node_names[0] == "material": # TODO: make this dynamic
old_net.values = mat
self.pop[-1].genotype.express()
print_log.message("New population size is %d" % len(self.pop))
# evaluate fitness
print_log.message("Starting fitness evaluation", timer_name="start")
print_log.reset_timer("evaluation")
self.update_env()
self.evaluate(self.sim, self.env[self.curr_env_idx], self.pop, print_log, save_vxa_every,
self.directory, self.name, max_eval_time, time_to_try_again, save_lineages, batch_size)
print_log.message("Fitness evaluation finished", timer_name="evaluation") # record total eval time in log
if self.solver is not None: # update solver
self.solver.tell([ind.fitness for ind in self.pop])
# print population to stdout and save all individual data
print_log.message("Saving statistics")
write_gen_stats(self.pop, self.directory, self.name, save_vxa_every, save_pareto, save_nets,
save_lineages=save_lineages)
self.pop.gen += 1
if not self.autosuspended: # print end of run stats
print_log.message("Finished {0} generations".format(self.pop.gen + 1))
print_log.message("DONE!", timer_name="start")
sub.call("touch {0}/RUN_FINISHED && rm {0}/RUNNING".format(self.directory), shell=True)
def run(self,
max_hours_runtime=29,
max_gens=3000,
num_random_individuals=1,
num_env_cycles=0,
directory="tests_data",
name="TestRun",
max_eval_time=60,
time_to_try_again=10,
checkpoint_every=100,
save_vxa_every=100,
save_pareto=True,
save_nets=False,
save_lineages=True,
continued_from_checkpoint=False):
if self.autosuspended:
sub.call("rm %s/AUTOSUSPENDED" % directory, shell=True)
self.autosuspended = False
self.max_gens = max_gens # can add additional gens through checkpointing
print_log = PrintLog()
print_log.add_timer("evaluation")
self.start_time = print_log.timers[
"start"] # sync start time with logging
# sub.call("clear", shell=True)
if not continued_from_checkpoint: # generation zero
self.directory = directory
self.name = name
self.num_random_inds = num_random_individuals
self.num_env_cycles = num_env_cycles
# global fitness_list
# fitness_list += [self.pop.best_fit_so_far]
global RUN_NAME
global RUN_DIR
RUN_NAME = name
RUN_DIR = directory
initialize_folders(self.pop,
self.directory,
self.name,
save_nets,
save_lineages=save_lineages)
make_gen_directories(self.pop, self.directory, save_vxa_every,
save_nets)
sub.call("touch {}/RUNNING".format(self.directory), shell=True)
self.evaluate(self.sim, self.env[self.curr_env_idx], self.pop,
print_log, save_vxa_every, self.directory, self.name,
max_eval_time, time_to_try_again, save_lineages)
self.select(
self.pop, directory, name
) # only produces dominated_by stats, no selection happening (population not replaced)
write_gen_stats(self.pop,
self.directory,
self.name,
save_vxa_every,
save_pareto,
save_nets,
save_lineages=save_lineages)
while self.pop.gen < max_gens:
if self.pop.gen % checkpoint_every == 0:
print_log.message(
"Saving checkpoint at generation {0}".format(self.pop.gen +
1),
timer_name="start")
self.save_checkpoint(self.directory, self.pop.gen)
if self.elapsed_time(units="h") > max_hours_runtime:
self.autosuspended = True
print_log.message(
"Autosuspending at generation {0}".format(self.pop.gen +
1),
timer_name="start")
self.save_checkpoint(self.directory, self.pop.gen)
sub.call("touch {0}/AUTOSUSPENDED && rm {0}/RUNNING".format(
self.directory),
shell=True)
break
self.pop.gen += 1
print_log.message("Creating folders structure for this generation")
make_gen_directories(self.pop, self.directory, save_vxa_every,
save_nets)
# update ages
self.pop.update_ages()
# mutation
print_log.message("Mutation starts")
new_children = self.mutate(self.pop, print_log=print_log)
print_log.message(
"Mutation ends: successfully generated %d new children." %
(len(new_children)))
# combine children and parents for selection
print_log.message("Now creating new population")
self.pop.append(new_children)
for _ in range(self.num_random_inds):
print_log.message("Random individual added to population")
self.pop.add_random_individual()
print_log.message("New population size is %d" % len(self.pop))
# evaluate fitness
print_log.message("Starting fitness evaluation",
timer_name="start")
print_log.reset_timer("evaluation")
self.update_env()
self.evaluate(self.sim, self.env[self.curr_env_idx], self.pop,
print_log, save_vxa_every, self.directory, self.name,
max_eval_time, time_to_try_again, save_lineages)
print_log.message(
"Fitness evaluation finished",
timer_name="evaluation") # record total eval time in log
# perform selection by pareto fronts
new_population = self.select(self.pop, directory, name)
# adding individuals if the pop is too small
if len(self.pop) < self.pop.pop_size:
for _ in range(self.pop.pop_size - len(self.pop)):
print_log.message(
"Random individual added to population because too many were deleted"
)
self.pop.add_random_individual()
print_log.message("New population size is %d" % len(self.pop))
# print population to stdout and save all individual data
print_log.message("Saving statistics")
write_gen_stats(self.pop,
self.directory,
self.name,
save_vxa_every,
save_pareto,
save_nets,
save_lineages=save_lineages)
# replace population with selection
self.pop.individuals = new_population
print_log.message("Population size reduced to %d" % len(self.pop))
if not self.autosuspended: # print end of run stats
print_log.message("Finished {0} generations".format(self.pop.gen +
1))
print_log.message("DONE!", timer_name="start")
sub.call("touch {0}/RUN_FINISHED && rm {0}/RUNNING".format(
self.directory),
shell=True)
def run(self, max_hours_runtime=29, max_gens=350, num_random_individuals=1, num_env_cycles=0,
directory="tests_data", name="TestRun",
max_eval_time=60, time_to_try_again=30,
checkpoint_every=100, save_vxa_every=100, save_pareto=True,
save_nets=False, save_lineages=False, continued_from_checkpoint=False,
batch_size=None, update_survivors_age=True,
max_fitness=None):
if self.autosuspended:
sub.call("rm run_1/AUTOSUSPENDED" % directory, shell=True) #sub.call("rm %s/AUTOSUSPENDED" % directory, shell=True)
self.autosuspended = False
self.max_gens = max_gens # can add additional gens through checkpointing
#Class that init a dictionary that contain important times
print_log = PrintLog() #here you have the default times: start and last_call
print_log.add_timer("evaluation") #add "evaluation" time in the dict list
self.start_time = print_log.timers["start"] # sync start time with logging
# sub.call("clear", shell=True)
#If you are starting from generation zero
if not continued_from_checkpoint: # generation zero
self.directory = directory #the name of directoty given to this run. example: run_(seed)
self.name = name #the name given to this run. exemplo "Swimmers"
self.num_random_inds = num_random_individuals #parameter about the number of new random individuals added in each generations. typical example = 1
self.num_env_cycles = num_env_cycles #parameter to change environment. typical example = 1 (same environment always)
#Function used to create the folders where the simulations outputs will be saved:
initialize_folders(self.pop, self.directory, self.name, save_nets, save_lineages=save_lineages)
#saves networks and vxa file for each generation if save_vxa_every and save_nets = True
make_gen_directories(self.pop, self.directory, save_vxa_every, save_nets)
#Create RUNNING file
sub.call("touch {}/RUNNING".format(self.directory), shell=True)
#Evaluate fitness of all individuals of the population in Voxelyze
#In evaluation, the functions read_voxlyze_results AND write_voxelyze_file are used
#So it is the place where the contact with Voxelyze is done.
self.evaluate(self.sim, self.env[self.curr_env_idx], self.pop, print_log, save_vxa_every, self.directory,
self.name, max_eval_time, time_to_try_again, save_lineages, batch_size)
self.select(self.pop) # only produces dominated_by stats, no selection happening (population not replaced)
write_gen_stats(self.pop, self.directory, self.name, save_vxa_every, save_pareto, save_nets,
save_lineages=save_lineages)
while self.pop.gen < max_gens:
if self.pop.gen % checkpoint_every == 0: #CHECK_POINT_EVERY = 1 saves always
print_log.message("Saving checkpoint at generation {0}".format(self.pop.gen+1), timer_name="start")
self.save_checkpoint(self.directory, self.pop.gen)
if self.elapsed_time(units="h") > max_hours_runtime or self.pop.best_fit_so_far == max_fitness:
self.autosuspended = True
print_log.message("Autosuspending at generation {0}".format(self.pop.gen+1), timer_name="start")
self.save_checkpoint(self.directory, self.pop.gen)
sub.call("touch {0}/AUTOSUSPENDED && rm {0}/RUNNING".format(self.directory), shell=True)
break
self.pop.gen += 1
print_log.message("Creating folders structure for this generation")
make_gen_directories(self.pop, self.directory, save_vxa_every, save_nets)
# update ages
self.pop.update_ages(update_survivors_age)
# mutation
print_log.message("Mutation starts")
new_children = self.mutate(self.pop, print_log=print_log)
print_log.message("Mutation ends: successfully generated %d new children." % (len(new_children)))
# combine children and parents for selection
print_log.message("Now creating new population")
self.pop.append(new_children) #Add the new children to population
for _ in range(self.num_random_inds):
print_log.message("Random individual added to population")
self.pop.add_random_individual() #add new random individual to population
print_log.message("New population size is %d" % len(self.pop))
# evaluate fitness
print_log.message("Starting fitness evaluation", timer_name="start")
print_log.reset_timer("evaluation")
self.update_env()
self.evaluate(self.sim, self.env[self.curr_env_idx], self.pop, print_log, save_vxa_every, self.directory,
self.name, max_eval_time, time_to_try_again, save_lineages, batch_size)
print_log.message("Fitness evaluation finished", timer_name="evaluation") # record total eval time in log
# perform selection by pareto fronts
new_population = self.select(self.pop)
# print population to stdout and save all individual data
print_log.message("Saving statistics")
write_gen_stats(self.pop, self.directory, self.name, save_vxa_every, save_pareto, save_nets,
save_lineages=save_lineages)
# replace population with selection
self.pop.individuals = new_population
print_log.message("Population size reduced to %d" % len(self.pop))
if not self.autosuspended: # print end of run stats
print_log.message("Finished {0} generations".format(self.pop.gen + 1))
print_log.message("DONE!", timer_name="start")
sub.call("touch {0}/RUN_FINISHED && rm {0}/RUNNING".format(self.directory), shell=True)