def create_motion(args, task, oracle_func=None):
if args['motion'] == 'rnn':
if (args['warmstart_dict'] is not None) and os.path.isfile(
args['warmstart_dict']):
with open(args['warmstart_dict'], 'r') as f:
trainer_args = json.load(f)
else:
trainer_args = args
common.process_observation_shape(
'rnn',
trainer_args['resolution_level'],
segmentation_input=trainer_args['segment_input'],
depth_input=trainer_args['depth_input'],
history_frame_len=1,
target_mask_input=trainer_args['target_mask_input'])
import zmq_train
trainer = zmq_train.create_zmq_trainer('a3c', 'rnn', trainer_args)
model_file = args['warmstart']
if model_file is not None:
trainer.load(model_file)
trainer.eval()
motion = RNNMotion(task,
trainer,
pass_target=args['multi_target'],
term_measure=args['terminate_measure'],
oracle_func=oracle_func)
elif args['motion'] == 'random':
motion = RandomMotion(task,
None,
term_measure=args['terminate_measure'],
oracle_func=oracle_func)
elif args['motion'] == 'fake':
motion = FakeMotion(task,
None,
term_measure=args['terminate_measure'],
oracle_func=oracle_func)
else: # mixture motion
mixture_dict_file = args['mixture_motion_dict']
try:
with open(mixture_dict_file, 'r') as f:
arg_dict = json.load(f)
except Exception as e:
print('Invalid Mixture Motion Dict!! file = <{}>'.format(
mixture_dict_file))
raise e
trainer_dict, pass_tar_dict, obs_mode_dict = create_mixture_motion_trainer_dict(
arg_dict)
motion = MixMotion(task,
trainer_dict,
pass_tar_dict,
term_measure=args['terminate_measure'],
obs_mode=obs_mode_dict,
oracle_func=oracle_func)
common.ensure_object_targets(args['object_target'])
if ('force_oracle_done' in args) and args['force_oracle_done']:
motion.set_force_oracle_done(True)
return motion
def train(args=None, warmstart=None):
# Process Observation Shape
common.process_observation_shape(model='rnn',
resolution_level=args['resolution_level'],
segmentation_input=args['segment_input'],
depth_input=args['depth_input'],
target_mask_input=args['target_mask_input'],
history_frame_len=1)
args['logger'] = utils.MyLogger(args['log_dir'], True, keep_file_handler=not args['append_file'])
name = 'ipc://@whatever' + args['job_name']
name2 = 'ipc://@whatever' + args['job_name'] + '2'
n_proc = args['n_proc']
config = create_zmq_config(args)
procs = [ZMQSimulator(k, name, name2, config) for k in range(n_proc)]
[k.start() for k in procs]
ensure_proc_terminate(procs)
trainer = create_zmq_trainer(args['algo'], model='rnn', args=args)
if warmstart is not None:
if os.path.exists(warmstart):
print('Warmstarting from <{}> ...'.format(warmstart))
trainer.load(warmstart)
else:
save_dir = args['save_dir']
print('Warmstarting from save_dir <{}> with version <{}> ...'.format(save_dir, warmstart))
trainer.load(save_dir, warmstart)
master = ZMQMaster(name, name2, trainer=trainer, config=args)
try:
# both loops must be running
print('Start Iterations ....')
send_thread = threading.Thread(target=master.send_loop, daemon=True)
send_thread.start()
master.recv_loop()
print('Done!')
trainer.save(args['save_dir'], version='final')
except KeyboardInterrupt:
trainer.save_all(args['save_dir'], version='interrupt')
raise
def create_mixture_motion_trainer_dict(arg_dict):
import zmq_train
trainer_dict = dict()
pass_tar_dict = dict()
obs_mode_dict = dict(
) # segment_input, depth_signal=True, target_mask_signal=False, joint_visual_signal=False
loaded_model = dict()
for target in all_allowed_targets:
assert target in arg_dict, '[MixtureMotion] Invalid <arg_dict>! Key=<{}> does not exist!'.format(
target)
args = arg_dict[target]
model_file = args['warmstart']
assert (model_file is not None) and os.path.exists(model_file), \
'[MixtureMotion] model file <{}> for target <{}> does not exist!!'.format(model_file, target)
if model_file in loaded_model:
trainer_dict[target] = trainer_dict[loaded_model[model_file]]
pass_tar_dict[target] = pass_tar_dict[loaded_model[model_file]]
obs_mode_dict[target] = obs_mode_dict[loaded_model[model_file]]
continue
common.process_observation_shape(
'rnn',
args['resolution_level'],
segmentation_input=args['segment_input'],
depth_input=args['depth_input'],
history_frame_len=1,
target_mask_input=args['target_mask_input'])
# ensure object target
__backup_CFG = common.CFG.copy()
common.ensure_object_targets(args['object_target'])
trainer = zmq_train.create_zmq_trainer('a3c', 'rnn', args)
common.CFG = __backup_CFG # backup
# load model
trainer.load(model_file)
trainer.eval()
loaded_model[model_file] = target
trainer_dict[target] = trainer
pass_tar_dict[target] = args['multi_target']
obs_mode_dict[target] = dict(
segment_input=(args['segment_input'] != 'none'),
depth_signal=args['depth_input'],
target_mask_signal=args['target_mask_input'],
joint_visual_signal=(args['segment_input'] == 'joint'))
return trainer_dict, pass_tar_dict, obs_mode_dict
def evaluate(args):
elap = time.time()
# Do not need to log detailed computation stats
common.debugger = utils.FakeLogger()
# ensure observation shape
common.process_observation_shape(
'rnn',
args['resolution'],
segmentation_input=args['segmentation_input'],
depth_input=args['depth_input'],
history_frame_len=1,
target_mask_input=args['target_mask_input'])
fixed_target = args['fixed_target']
if (fixed_target is not None) and (fixed_target != 'any-room') and (
fixed_target != 'any-object'):
assert fixed_target in common.all_target_instructions, 'invalid fixed target <{}>'.format(
fixed_target)
__backup_CFG = common.CFG.copy()
if fixed_target == 'any-room':
common.ensure_object_targets(False)
if args['hardness'] is not None:
print('>>>> Hardness = {}'.format(args['hardness']))
if args['max_birthplace_steps'] is not None:
print('>>>> Max BirthPlace Steps = {}'.format(
args['max_birthplace_steps']))
set_seed(args['seed'])
task = common.create_env(args['house'],
task_name=args['task_name'],
false_rate=args['false_rate'],
hardness=args['hardness'],
max_birthplace_steps=args['max_birthplace_steps'],
success_measure=args['success_measure'],
depth_input=args['depth_input'],
target_mask_input=args['target_mask_input'],
segment_input=args['segmentation_input'],
genRoomTypeMap=False,
cacheAllTarget=args['multi_target'],
render_device=args['render_gpu'],
use_discrete_action=True,
include_object_target=args['object_target']
and (fixed_target != 'any-room'),
include_outdoor_target=args['outdoor_target'],
discrete_angle=True,
min_birthplace_grids=args['min_birthplace_grids'])
if (fixed_target is not None) and (fixed_target != 'any-room') and (
fixed_target != 'any-object'):
task.reset_target(fixed_target)
if fixed_target == 'any-room':
common.CFG = __backup_CFG
common.ensure_object_targets(True)
# create semantic classifier
if args['semantic_dir'] is not None:
assert os.path.exists(
args['semantic_dir']
), '[Error] Semantic Dir <{}> not exists!'.format(args['semantic_dir'])
assert not args[
'object_target'], '[ERROR] currently do not support --object-target!'
print('Loading Semantic Oracle from dir <{}>...'.format(
args['semantic_dir']))
if args['semantic_gpu'] is None:
args['semantic_gpu'] = common.get_gpus_for_rendering()[0]
oracle = SemanticOracle(model_dir=args['semantic_dir'],
model_device=args['semantic_gpu'],
include_object=args['object_target'])
oracle_func = OracleFunction(
oracle,
threshold=args['semantic_threshold'],
filter_steps=args['semantic_filter_steps'])
else:
oracle_func = None
# create motion
motion = create_motion(args, task, oracle_func)
logger = utils.MyLogger(args['log_dir'], True)
logger.print('Start Evaluating ...')
episode_success = []
episode_good = []
episode_stats = []
t = 0
seed = args['seed']
max_episode_len = args['max_episode_len']
plan_req = args['plan_dist_iters'] if 'plan_dist_iters' in args else None
for it in range(args['max_iters']):
cur_infos = []
motion.reset()
set_seed(seed + it + 1) # reset seed
if plan_req is not None:
while True:
task.reset(target=fixed_target)
m = len(task.get_optimal_plan())
if (m in plan_req) and plan_req[m] > 0:
break
plan_req[m] -= 1
else:
task.reset(target=fixed_target)
info = task.info
episode_success.append(0)
episode_good.append(0)
cur_stats = dict(best_dist=info['dist'],
success=0,
good=0,
reward=0,
target=task.get_current_target(),
meters=task.info['meters'],
optstep=task.info['optsteps'],
length=max_episode_len,
images=None)
if hasattr(task.house, "_id"):
cur_stats['world_id'] = task.house._id
store_history = args['store_history']
if store_history:
cur_infos.append(proc_info(task.info))
if args['temperature'] is not None:
ep_data = motion.run(task.get_current_target(),
max_episode_len,
temperature=args['temperature'])
else:
ep_data = motion.run(task.get_current_target(), max_episode_len)
for dat in ep_data:
info = dat[4]
if store_history:
cur_infos.append(proc_info(info))
cur_dist = info['dist']
if cur_dist == 0:
cur_stats['good'] += 1
episode_good[-1] = 1
if cur_dist < cur_stats['best_dist']:
cur_stats['best_dist'] = cur_dist
episode_step = len(ep_data)
if ep_data[-1][3]: # done
if ep_data[-1][2] > 5: # magic number:
episode_success[-1] = 1
cur_stats['success'] = 1
cur_stats['length'] = episode_step # store length
if store_history:
cur_stats['infos'] = cur_infos
episode_stats.append(cur_stats)
dur = time.time() - elap
logger.print('Episode#%d, Elapsed = %.3f min' % (it + 1, dur / 60))
if args['multi_target']:
logger.print(' ---> Target Room = {}'.format(cur_stats['target']))
logger.print(' ---> Total Samples = {}'.format(t))
logger.print(' ---> Success = %d (rate = %.3f)' %
(cur_stats['success'], np.mean(episode_success)))
logger.print(
' ---> Times of Reaching Target Room = %d (rate = %.3f)' %
(cur_stats['good'], np.mean(episode_good)))
logger.print(' ---> Best Distance = %d' % cur_stats['best_dist'])
logger.print(' ---> Birth-place Distance = %d' % cur_stats['optstep'])
logger.print('######## Final Stats ###########')
logger.print('Success Rate = %.3f' % np.mean(episode_success))
logger.print(
'> Avg Ep-Length per Success = %.3f' %
np.mean([s['length'] for s in episode_stats if s['success'] > 0]))
logger.print(
'> Avg Birth-Meters per Success = %.3f' %
np.mean([s['meters'] for s in episode_stats if s['success'] > 0]))
logger.print('Reaching Target Rate = %.3f' % np.mean(episode_good))
logger.print('> Avg Ep-Length per Target Reach = %.3f' %
np.mean([s['length']
for s in episode_stats if s['good'] > 0]))
logger.print('> Avg Birth-Meters per Target Reach = %.3f' %
np.mean([s['meters']
for s in episode_stats if s['good'] > 0]))
if args['multi_target']:
all_targets = list(set([s['target'] for s in episode_stats]))
for tar in all_targets:
n = sum([1.0 for s in episode_stats if s['target'] == tar])
succ = [
float(s['success'] > 0) for s in episode_stats
if s['target'] == tar
]
good = [
float(s['good'] > 0) for s in episode_stats
if s['target'] == tar
]
length = [s['length'] for s in episode_stats if s['target'] == tar]
meters = [s['meters'] for s in episode_stats if s['target'] == tar]
good_len = np.mean([l for l, g in zip(length, good) if g > 0.5])
succ_len = np.mean([l for l, s in zip(length, succ) if s > 0.5])
good_mts = np.mean([l for l, g in zip(meters, good) if g > 0.5])
succ_mts = np.mean([l for l, s in zip(meters, succ) if s > 0.5])
logger.print(
'>>>>> Multi-Target <%s>: Rate = %.3f (n=%d), Good = %.3f (AvgLen=%.3f; Mts=%.3f), Succ = %.3f (AvgLen=%.3f; Mts=%.3f)'
% (tar, n / len(episode_stats), n, np.mean(good), good_len,
good_mts, np.mean(succ), succ_len, succ_mts))
return episode_stats
def evaluate(args, data_saver=None):
args['segment_input'] = args['segmentation_input']
backup_rate = args['backup_rate']
elap = time.time()
# Do not need to log detailed computation stats
common.debugger = utils.FakeLogger()
# ensure observation shape
common.process_observation_shape(
'rnn',
args['resolution'],
args['segmentation_input'],
args['depth_input'],
target_mask_input=args['target_mask_input'])
fixed_target = args['fixed_target']
if (fixed_target is not None) and (fixed_target != 'any-room') and (
fixed_target != 'any-object'):
assert fixed_target in common.n_target_instructions, 'invalid fixed target <{}>'.format(
fixed_target)
__backup_CFG = common.CFG.copy()
if fixed_target == 'any-room':
common.ensure_object_targets(False)
if args['hardness'] is not None:
print('>>>> Hardness = {}'.format(args['hardness']))
if args['max_birthplace_steps'] is not None:
print('>>>> Max BirthPlace Steps = {}'.format(
args['max_birthplace_steps']))
set_seed(args['seed'])
task = common.create_env(args['house'],
task_name=args['task_name'],
false_rate=args['false_rate'],
hardness=args['hardness'],
max_birthplace_steps=args['max_birthplace_steps'],
success_measure=args['success_measure'],
depth_input=args['depth_input'],
target_mask_input=args['target_mask_input'],
segment_input=args['segmentation_input'],
genRoomTypeMap=False,
cacheAllTarget=args['multi_target'],
render_device=args['render_gpu'],
use_discrete_action=True,
include_object_target=args['object_target']
and (fixed_target != 'any-room'),
include_outdoor_target=args['outdoor_target'],
discrete_angle=True,
min_birthplace_grids=args['min_birthplace_grids'])
if (fixed_target is not None) and (fixed_target != 'any-room') and (
fixed_target != 'any-object'):
task.reset_target(fixed_target)
if fixed_target == 'any-room':
common.CFG = __backup_CFG
common.ensure_object_targets(True)
# logger
logger = utils.MyLogger(args['log_dir'], True)
logger.print('Start Evaluating ...')
# create semantic classifier
if args['semantic_dir'] is not None:
assert os.path.exists(
args['semantic_dir']
), '[Error] Semantic Dir <{}> not exists!'.format(args['semantic_dir'])
assert not args[
'object_target'], '[ERROR] currently do not support --object-target!'
print('Loading Semantic Oracle from dir <{}>...'.format(
args['semantic_dir']))
if args['semantic_gpu'] is None:
args['semantic_gpu'] = common.get_gpus_for_rendering()[0]
oracle = SemanticOracle(model_dir=args['semantic_dir'],
model_device=args['semantic_gpu'],
include_object=args['object_target'])
oracle_func = OracleFunction(
oracle,
threshold=args['semantic_threshold'],
filter_steps=args['semantic_filter_steps'],
batched_size=args['semantic_batch_size'])
else:
oracle_func = None
# create motion
motion = create_motion(args, task, oracle_func=oracle_func)
if args['motion'] == 'random':
motion.set_skilled_rate(args['random_motion_skill'])
flag_interrupt = args['interruptive_motion']
# create planner
graph = None
max_motion_steps = args['n_exp_steps']
if (args['planner'] == None) or (args['planner'] == 'void'):
graph = VoidPlanner(motion)
elif args['planner'] == 'oracle':
graph = OraclePlanner(motion)
elif args['planner'] == 'rnn':
#assert False, 'Currently only support Graph-planner'
graph = RNNPlanner(motion,
args['planner_units'],
args['planner_filename'],
oracle_func=oracle_func)
else:
graph = GraphPlanner(motion)
if not args['outdoor_target']:
graph.add_excluded_target('outdoor')
filename = args['planner_filename']
if filename == 'None': filename = None
if filename is not None:
logger.print(' > Loading Graph from file = <{}>'.format(filename))
with open(filename, 'rb') as f:
_params = pickle.load(f)
graph.set_parameters(_params)
# hack
if args['planner_obs_noise'] is not None:
graph.set_param(-1, args['planner_obs_noise']) # default 0.95
episode_success = []
episode_good = []
episode_stats = []
t = 0
seed = args['seed']
max_episode_len = args['max_episode_len']
plan_req = args['plan_dist_iters'] if 'plan_dist_iters' in args else None
####################
accu_plan_time = 0
accu_exe_time = 0
accu_mask_time = 0
####################
for it in range(args['max_iters']):
if (it > 0) and (backup_rate > 0) and (it % backup_rate
== 0) and (data_saver
is not None):
data_saver.save(episode_stats, ep_id=it)
cur_infos = []
motion.reset()
set_seed(seed + it + 1) # reset seed
if plan_req is not None:
while True:
task.reset(target=fixed_target)
m = len(task.get_optimal_plan())
if (m in plan_req) and plan_req[m] > 0:
break
plan_req[m] -= 1
else:
task.reset(target=fixed_target)
info = task.info
episode_success.append(0)
episode_good.append(0)
task_target = task.get_current_target()
cur_stats = dict(best_dist=info['dist'],
success=0,
good=0,
reward=0,
target=task_target,
plan=[],
meters=task.info['meters'],
optstep=task.info['optsteps'],
length=max_episode_len,
images=None)
if hasattr(task.house, "_id"):
cur_stats['world_id'] = task.house._id
store_history = args['store_history']
if store_history:
cur_infos.append(proc_info(task.info))
episode_step = 0
# reset planner
if graph is not None:
graph.reset()
while episode_step < max_episode_len:
if flag_interrupt and motion.is_interrupt():
graph_target = task.get_current_target()
else:
# TODO #####################
tt = time.time()
mask_feat = oracle_func.get(
task
) if oracle_func is not None else task.get_feature_mask()
accu_mask_time += time.time() - tt
tt = time.time()
graph_target = graph.plan(mask_feat, task_target)
accu_plan_time += time.time() - tt
################################
graph_target_id = common.target_instruction_dict[graph_target]
allowed_steps = min(max_episode_len - episode_step,
max_motion_steps)
###############
# TODO
tt = time.time()
motion_data = motion.run(graph_target, allowed_steps)
accu_exe_time += time.time() - tt
cur_stats['plan'].append(
(graph_target, len(motion_data),
(motion_data[-1][0][graph_target_id] > 0)))
# store stats
for dat in motion_data:
info = dat[4]
if store_history:
cur_infos.append(proc_info(info))
cur_dist = info['dist']
if cur_dist == 0:
cur_stats['good'] += 1
episode_good[-1] = 1
if cur_dist < cur_stats['best_dist']:
cur_stats['best_dist'] = cur_dist
# update graph
## TODO ############
tt = time.time()
graph.observe(motion_data, graph_target)
accu_plan_time += time.time() - tt
episode_step += len(motion_data)
# check done
if motion_data[-1][3]:
if motion_data[-1][2] > 5: # magic number
episode_success[-1] = 1
cur_stats['success'] = 1
break
cur_stats['length'] = episode_step # store length
if store_history:
cur_stats['infos'] = cur_infos
episode_stats.append(cur_stats)
dur = time.time() - elap
logger.print('Episode#%d, Elapsed = %.3f min' % (it + 1, dur / 60))
#TODO #################
logger.print(' >>> Mask Time = %.4f min' % (accu_mask_time / 60))
logger.print(' >>> Plan Time = %.4f min' % (accu_plan_time / 60))
logger.print(' >>> Motion Time = %.4f min' % (accu_exe_time / 60))
if args['multi_target']:
logger.print(' ---> Target Room = {}'.format(cur_stats['target']))
logger.print(' ---> Total Samples = {}'.format(t))
logger.print(' ---> Success = %d (rate = %.3f)' %
(cur_stats['success'], np.mean(episode_success)))
logger.print(
' ---> Times of Reaching Target Room = %d (rate = %.3f)' %
(cur_stats['good'], np.mean(episode_good)))
logger.print(' ---> Best Distance = %d' % cur_stats['best_dist'])
logger.print(' ---> Birth-place Meters = %.4f (optstep = %d)' %
(cur_stats['meters'], cur_stats['optstep']))
logger.print(' ---> Planner Results = {}'.format(cur_stats['plan']))
logger.print('######## Final Stats ###########')
logger.print('Success Rate = %.3f' % np.mean(episode_success))
logger.print(
'> Avg Ep-Length per Success = %.3f' %
np.mean([s['length'] for s in episode_stats if s['success'] > 0]))
logger.print(
'> Avg Birth-Meters per Success = %.3f' %
np.mean([s['meters'] for s in episode_stats if s['success'] > 0]))
logger.print('Reaching Target Rate = %.3f' % np.mean(episode_good))
logger.print('> Avg Ep-Length per Target Reach = %.3f' %
np.mean([s['length']
for s in episode_stats if s['good'] > 0]))
logger.print('> Avg Birth-Meters per Target Reach = %.3f' %
np.mean([s['meters']
for s in episode_stats if s['good'] > 0]))
if args['multi_target']:
all_targets = list(set([s['target'] for s in episode_stats]))
for tar in all_targets:
n = sum([1.0 for s in episode_stats if s['target'] == tar])
succ = [
float(s['success'] > 0) for s in episode_stats
if s['target'] == tar
]
good = [
float(s['good'] > 0) for s in episode_stats
if s['target'] == tar
]
length = [s['length'] for s in episode_stats if s['target'] == tar]
meters = [s['meters'] for s in episode_stats if s['target'] == tar]
good_len = np.mean([l for l, g in zip(length, good) if g > 0.5])
succ_len = np.mean([l for l, s in zip(length, succ) if s > 0.5])
good_mts = np.mean([l for l, g in zip(meters, good) if g > 0.5])
succ_mts = np.mean([l for l, s in zip(meters, succ) if s > 0.5])
logger.print(
'>>>>> Multi-Target <%s>: Rate = %.3f (n=%d), Good = %.3f (AvgLen=%.3f; Mts=%.3f), Succ = %.3f (AvgLen=%.3f; Mts=%.3f)'
% (tar, n / len(episode_stats), n, np.mean(good), good_len,
good_mts, np.mean(succ), succ_len, succ_mts))
return episode_stats