def learn_graph(args):
elap = time.time()
# Do not need to log detailed computation stats
common.debugger = utils.FakeLogger()
common.ensure_object_targets(True)
set_seed(args['seed'])
task = common.create_env(args['house'], task_name=args['task_name'], false_rate=args['false_rate'],
success_measure=args['success_measure'],
depth_input=args['depth_input'],
target_mask_input=args['target_mask_input'],
segment_input=args['segmentation_input'],
cacheAllTarget=True,
render_device=args['render_gpu'],
use_discrete_action=True,
include_object_target=True,
include_outdoor_target=True,
discrete_angle=True)
# create motion
__graph_warmstart = args['warmstart']
args['warmstart'] = args['motion_warmstart']
motion = create_motion(args, task)
# create graph
args['warmstart'] = __graph_warmstart
graph = GraphPlanner(motion)
# logger
logger = utils.MyLogger(args['save_dir'], True)
logger.print("> Training Mode = {}".format(args['training_mode']))
logger.print("> Graph Eps = {}".format(args['graph_eps']))
logger.print("> N_Trials = {}".format(args['n_trials']))
logger.print("> Max Exploration Steps = {}".format(args['max_exp_steps']))
# Graph Building
logger.print('Start Graph Building ...')
if args['warmstart'] is not None:
filename = args['warmstart']
logger.print(' >>> Loading Pre-Trained Graph from {}'.format(filename))
with open(filename, 'rb') as file:
g_params = pickle.load(file)
graph.set_parameters(g_params)
train_mode = args['training_mode']
if train_mode in ['mle', 'joint']:
graph.learn(n_trial=args['n_trials'], max_allowed_steps=args['max_exp_steps'], eps=args['graph_eps'], logger=logger)
if train_mode in ['evolution', 'joint']:
graph.evolve() # TODO: not implemented yet
logger.print('######## Final Stats ###########')
graph._show_prior_room(logger=logger)
graph._show_prior_object(logger=logger)
return graph
def train(
args=None,
houseID=0,
reward_type='indicator',
success_measure='center',
multi_target=False,
include_object_target=False,
algo='pg',
model_name='cnn', # NOTE: optional: model_name='rnn'
iters=2000000,
report_rate=20,
save_rate=1000,
eval_range=200,
log_dir='./temp',
save_dir='./_model_',
warmstart=None,
log_debug_info=True):
if 'scheduler' in args:
scheduler = args['scheduler']
else:
scheduler = None
if args is None:
args = common.create_default_args(algo)
hardness = args['hardness']
max_birthplace_steps = args['max_birthplace_steps']
if hardness is not None:
print('>>> Hardness Level = {}'.format(hardness))
if max_birthplace_steps is not None:
print('>>>> Max BirthPlace Steps = {}'.format(max_birthplace_steps))
env = common.create_env(houseID,
task_name=args['task_name'],
false_rate=args['false_rate'],
reward_type=reward_type,
hardness=hardness,
max_birthplace_steps=max_birthplace_steps,
success_measure=success_measure,
segment_input=args['segment_input'],
depth_input=args['depth_input'],
render_device=args['render_gpu'],
cacheAllTarget=args['multi_target'],
use_discrete_action=('dpg' not in algo),
include_object_target=include_object_target)
trainer = common.create_trainer(algo, model_name, args)
logger = utils.MyLogger(log_dir, True)
if multi_target:
assert hasattr(trainer, 'set_target')
if warmstart is not None:
if os.path.exists(warmstart):
logger.print('Warmstarting from <{}> ...'.format(warmstart))
trainer.load(warmstart)
else:
logger.print(
'Warmstarting from save_dir <{}> with version <{}> ...'.format(
save_dir, warmstart))
trainer.load(save_dir, warmstart)
logger.print('Start Training')
if log_debug_info:
common.debugger = utils.MyLogger(log_dir, True, 'full_logs.txt')
else:
common.debugger = utils.FakeLogger()
episode_rewards = [0.0]
episode_success = [0.0]
episode_length = [0.0]
episode_targets = ['kitchen']
trainer.reset_agent()
if multi_target:
obs = env.reset()
target_room = env.info['target_room']
trainer.set_target(target_room)
episode_targets[-1] = target_room
else:
env.reset(target='kitchen')
assert not np.any(np.isnan(obs)), 'nan detected in the observation!'
obs = obs.transpose([1, 0, 2])
logger.print('Observation Shape = {}'.format(obs.shape))
episode_step = 0
t = 0
best_res = -1e50
elap = time.time()
update_times = 0
print('Starting iterations...')
try:
while (len(episode_rewards) <= iters):
idx = trainer.process_observation(obs)
# get action
if scheduler is not None:
noise_level = scheduler.value(len(episode_rewards) - 1)
action = trainer.action(noise_level)
else:
action = trainer.action()
#proc_action = [np.exp(a) for a in action]
# environment step
obs, rew, done, info = env.step(action)
assert not np.any(
np.isnan(obs)), 'nan detected in the observation!'
obs = obs.transpose([1, 0, 2])
episode_step += 1
episode_length[-1] += 1
terminal = (episode_step >= args['episode_len'])
# collect experience
trainer.process_experience(idx, action, rew, done, terminal, info)
episode_rewards[-1] += rew
if rew > 5: # magic number
episode_success[-1] = 1.0
if done or terminal:
trainer.reset_agent()
if multi_target:
obs = env.reset()
target_room = env.info['target_room']
trainer.set_target(target_room)
episode_targets.append(target_room)
else:
obs = env.reset(target='kitchen')
assert not np.any(
np.isnan(obs)), 'nan detected in the observation!'
obs = obs.transpose([1, 0, 2])
episode_step = 0
episode_rewards.append(0)
episode_success.append(0)
episode_length.append(0)
# update all trainers
trainer.preupdate()
stats = trainer.update()
if stats is not None:
update_times += 1
if common.debugger is not None:
common.debugger.print(
'>>>>>> Update#{} Finished!!!'.format(update_times),
False)
# save results
if ((done or terminal) and (len(episode_rewards) % save_rate == 0)) or\
(len(episode_rewards) > iters):
trainer.save(save_dir)
logger.print(
'Successfully Saved to <{}>'.format(save_dir + '/' +
trainer.name + '.pkl'))
if np.mean(episode_rewards[-eval_range:]) > best_res:
best_res = np.mean(episode_rewards[-eval_range:])
trainer.save(save_dir, "best")
# display training output
if ((update_times % report_rate == 0) and (algo != 'pg') and (stats is not None)) or \
((update_times == 0) and (algo != 'pg') and (len(episode_rewards) % 100 == 0) and (done or terminal)) or \
((algo == 'pg') and (stats is not None)):
logger.print(
'Episode#%d, Updates=%d, Time Elapsed = %.3f min' %
(len(episode_rewards), update_times,
(time.time() - elap) / 60))
logger.print('-> Total Samples: %d' % t)
logger.print('-> Avg Episode Length: %.4f' %
(t / len(episode_rewards)))
if stats is not None:
for k in stats:
logger.print(' >> %s = %.4f' % (k, stats[k]))
logger.print(' >> Reward = %.4f' %
np.mean(episode_rewards[-eval_range:]))
logger.print(' >> Success Rate = %.4f' %
np.mean(episode_success[-eval_range:]))
if multi_target:
ep_rew = episode_rewards[-eval_range:]
ep_suc = episode_success[-eval_range:]
ep_tar = episode_targets[-eval_range:]
ep_len = episode_length[-eval_range:]
total_n = len(ep_rew)
tar_stats = dict()
for k, r, s, l in zip(ep_tar, ep_rew, ep_suc, ep_len):
if k not in tar_stats:
tar_stats[k] = [0.0, 0.0, 0.0, 0.0]
tar_stats[k][0] += 1
tar_stats[k][1] += r
tar_stats[k][2] += s
tar_stats[k][3] += l
for k in tar_stats.keys():
n, r, s, l = tar_stats[k]
logger.print(
' --> Multi-Room<%s> Freq = %.4f, Rew = %.4f, Succ = %.4f (AvgLen = %.3f)'
% (k, n / total_n, r / n, s / n, l / n))
print('----> Data Loading Time = %.4f min' %
(time_counter[-1] / 60))
print('----> GPU Data Transfer Time = %.4f min' %
(time_counter[0] / 60))
print('----> Training Time = %.4f min' %
(time_counter[1] / 60))
print('----> Target Net Update Time = %.4f min' %
(time_counter[2] / 60))
t += 1
except KeyboardInterrupt:
print('Keyboard Interrupt!!!!!!')
trainer.save(save_dir, "final")
with open(save_dir + '/final_training_stats.pkl', 'wb') as f:
pickle.dump([
episode_rewards, episode_success, episode_targets, episode_length
], f)
def evaluate_aux_pred(house,
seed=0,
iters=1000,
max_episode_len=10,
algo='a3c',
model_name='rnn',
model_file=None,
log_dir='./log/eval',
store_history=False,
use_batch_norm=True,
rnn_units=None,
rnn_layers=None,
rnn_cell=None,
multi_target=True,
use_target_gating=False,
segmentation_input='none',
depth_input=False,
resolution='normal'):
# TODO: currently do not support this
assert False, 'Aux Prediction Not Supported!'
# Do not need to log detailed computation stats
assert algo in ['a3c', 'nop']
flag_run_random_policy = (algo == 'nop')
common.debugger = utils.FakeLogger()
args = common.create_default_args(algo,
model=model_name,
use_batch_norm=use_batch_norm,
replay_buffer_size=50,
episode_len=max_episode_len,
rnn_units=rnn_units,
rnn_layers=rnn_layers,
rnn_cell=rnn_cell,
segmentation_input=segmentation_input,
resolution_level=resolution,
depth_input=depth_input,
history_frame_len=1)
# TODO: add code for evaluation aux-task (concept learning)
args['multi_target'] = multi_target
args['target_gating'] = use_target_gating
args['aux_task'] = True
import zmq_train
set_seed(seed)
env = common.create_env(house,
hardness=1e-8,
success_measure='stay',
depth_input=depth_input,
segment_input=args['segment_input'],
genRoomTypeMap=True,
cacheAllTarget=True,
use_discrete_action=True)
trainer = zmq_train.create_zmq_trainer(algo, model_name, args)
if model_file is not None:
trainer.load(model_file)
trainer.eval() # evaluation mode
logger = utils.MyLogger(log_dir, True)
logger.print('Start Evaluating Auxiliary Task ...')
logger.print(
' --> Episode (Left) Turning Steps = {}'.format(max_episode_len))
episode_err = []
episode_succ = []
episode_good = []
episode_rews = []
episode_stats = []
elap = time.time()
for it in range(iters):
trainer.reset_agent()
set_seed(seed + it + 1) # reset seed
obs = env.reset() if multi_target else env.reset(
target=env.get_current_target())
target_id = common.target_instruction_dict[env.get_current_target()]
if multi_target and hasattr(trainer, 'set_target'):
trainer.set_target(env.get_current_target())
cur_infos = []
if store_history:
cur_infos.append(proc_info(env.info))
# cur_images.append(env.render(renderMapLoc=env.cam_info['loc'], display=False))
if model_name != 'rnn': obs = obs.transpose([1, 0, 2])
episode_succ.append(0)
episode_err.append(0)
episode_good.append(0)
cur_rew = []
cur_pred = []
if flag_run_random_policy:
predefined_aux_pred = common.all_aux_predictions[random.choice(
common.all_target_instructions)]
for _st in range(max_episode_len):
# get action
if flag_run_random_policy:
aux_pred = predefined_aux_pred
else:
if multi_target:
_, _, aux_prob = trainer.action(obs,
return_numpy=True,
target=[[target_id]],
return_aux_pred=True,
return_aux_logprob=False)
else:
_, _, aux_prob = trainer.action(obs,
return_numpy=True,
return_aux_pred=True,
return_aux_logprob=False)
aux_prob = aux_prob.squeeze() # [n_pred]
aux_pred = int(
np.argmax(aux_prob)
) # greedy action, takes the output with the maximum confidence
aux_rew = trainer.get_aux_task_reward(
aux_pred, env.get_current_room_pred_mask())
cur_rew.append(aux_rew)
cur_pred.append(common.all_aux_prediction_list[aux_pred])
if aux_rew < 0:
episode_err[-1] += 1
if aux_rew >= 0.9: # currently a hack
episode_succ[-1] += 1
if aux_rew > 0:
episode_good[-1] += 1
action = 5 # Left Rotation
# environment step
obs, rew, done, info = env.step(action)
if store_history:
cur_infos.append(proc_info(info))
cur_infos[-1]['aux_pred'] = cur_pred
#cur_images.append(env.render(renderMapLoc=env.cam_info['loc'], display=False))
if model_name != 'rnn': obs = obs.transpose([1, 0, 2])
if episode_err[-1] > 0:
episode_succ[-1] = 0
room_mask = env.get_current_room_pred_mask()
cur_room_types = []
for i in range(common.n_aux_predictions):
if (room_mask & (1 << i)) > 0:
cur_room_types.append(common.all_aux_prediction_list[i])
cur_stats = dict(err=episode_err[-1],
good=episode_good[-1],
succ=episode_succ[-1],
rew=cur_rew,
err_rate=episode_err[-1] / max_episode_len,
good_rate=episode_good[-1] / max_episode_len,
succ_rate=episode_succ[-1] / max_episode_len,
target=env.get_current_target(),
mask=room_mask,
room_types=cur_room_types,
length=max_episode_len)
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))
logger.print(' ---> Target Room = {}'.format(cur_stats['target']))
logger.print(' ---> Aux Rew = {}'.format(cur_rew))
if (episode_succ[-1] > 0) and (episode_err[-1] == 0):
logger.print(' >>>> Success!')
elif episode_err[-1] == 0:
logger.print(' >>>> Good!')
else:
logger.print(' >>>> Failed!')
logger.print(
" ---> Indep. Prediction: Succ Rate = %.3f, Good Rate = %.3f, Err Rate = %.3f"
% (episode_succ[-1] * 100.0 / max_episode_len,
episode_good[-1] * 100.0 / max_episode_len,
episode_err[-1] * 100.0 / max_episode_len))
logger.print(
" > Accu. Succ = %.3f, Good = %.3f, Fail = %.3f" %
(float(np.mean([float(s == max_episode_len)
for s in episode_succ])) * 100.0,
float(np.mean([float(e == 0) for e in episode_err])) * 100,
float(np.mean([float(e > 0) for e in episode_err])) * 100))
logger.print(
" > Accu. Rate: Succ Rate = %.3f, Good Rate = %.3f, Fail Rate = %.3f"
% (float(np.mean([s / max_episode_len
for s in episode_succ])) * 100.0,
float(np.mean([g / max_episode_len
for g in episode_good])) * 100,
float(np.mean([e / max_episode_len
for e in episode_err])) * 100))
return episode_stats
def evaluate(house,
seed=0,
render_device=None,
iters=1000,
max_episode_len=1000,
task_name='roomnav',
false_rate=0.0,
hardness=None,
max_birthplace_steps=None,
success_measure='center',
multi_target=False,
fixed_target=None,
algo='nop',
model_name='cnn',
model_file=None,
log_dir='./log/eval',
store_history=False,
use_batch_norm=True,
rnn_units=None,
rnn_layers=None,
rnn_cell=None,
use_action_gating=False,
use_residual_critic=False,
use_target_gating=False,
segmentation_input='none',
depth_input=False,
target_mask_input=False,
resolution='normal',
history_len=4,
include_object_target=False,
include_outdoor_target=True,
aux_task=False,
no_skip_connect=False,
feed_forward=False,
greedy_execution=False,
greedy_aux_pred=False):
assert not aux_task, 'Do not support Aux-Task now!'
elap = time.time()
# Do not need to log detailed computation stats
common.debugger = utils.FakeLogger()
args = common.create_default_args(algo,
model=model_name,
use_batch_norm=use_batch_norm,
replay_buffer_size=50,
episode_len=max_episode_len,
rnn_units=rnn_units,
rnn_layers=rnn_layers,
rnn_cell=rnn_cell,
segmentation_input=segmentation_input,
resolution_level=resolution,
depth_input=depth_input,
target_mask_input=target_mask_input,
history_frame_len=history_len)
args['action_gating'] = use_action_gating
args['residual_critic'] = use_residual_critic
args['multi_target'] = multi_target
args['object_target'] = include_object_target
args['target_gating'] = use_target_gating
args['aux_task'] = aux_task
args['no_skip_connect'] = no_skip_connect
args['feed_forward'] = feed_forward
if (fixed_target is not None) and (fixed_target
not in ['any-room', '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 hardness is not None:
print('>>>> Hardness = {}'.format(hardness))
if max_birthplace_steps is not None:
print('>>>> Max BirthPlace Steps = {}'.format(max_birthplace_steps))
set_seed(seed)
env = common.create_env(house,
task_name=task_name,
false_rate=false_rate,
hardness=hardness,
max_birthplace_steps=max_birthplace_steps,
success_measure=success_measure,
depth_input=depth_input,
target_mask_input=target_mask_input,
segment_input=args['segment_input'],
genRoomTypeMap=aux_task,
cacheAllTarget=multi_target,
render_device=render_device,
use_discrete_action=('dpg' not in algo),
include_object_target=include_object_target
and (fixed_target != 'any-room'),
include_outdoor_target=include_outdoor_target,
discrete_angle=True)
if (fixed_target is not None) and (fixed_target != 'any-room') and (
fixed_target != 'any-object'):
env.reset_target(fixed_target)
if fixed_target == 'any-room':
common.CFG = __backup_CFG
common.ensure_object_targets(True)
# create model
if model_name == 'rnn':
import zmq_train
trainer = zmq_train.create_zmq_trainer(algo, model_name, args)
else:
trainer = common.create_trainer(algo, model_name, args)
if model_file is not None:
trainer.load(model_file)
trainer.eval() # evaluation mode
if greedy_execution and hasattr(trainer, 'set_greedy_execution'):
trainer.set_greedy_execution()
else:
print('[Eval] WARNING!!! Greedy Policy Execution NOT Available!!!')
greedy_execution = False
if greedy_aux_pred and hasattr(trainer, 'set_greedy_aux_prediction'):
trainer.set_greedy_aux_prediction()
else:
print(
'[Eval] WARNING!!! Greedy Execution of Auxiliary Task NOT Available!!!'
)
greedy_aux_pred = False
if aux_task: assert trainer.is_rnn() # only rnn support aux_task
#flag_random_reset_target = multi_target and (fixed_target is None)
logger = utils.MyLogger(log_dir, True)
logger.print('Start Evaluating ...')
episode_success = []
episode_good = []
episode_stats = []
t = 0
for it in range(iters):
cur_infos = []
trainer.reset_agent()
set_seed(seed + it + 1) # reset seed
obs = env.reset(target=fixed_target)
#if multi_target and (fixed_target is not None) and (fixed_target != 'kitchen'):
# # TODO: Currently a hacky solution
# env.reset(target=fixed_target)
# if house < 0: # multi-house env
# obs = env.reset(reset_target=False, keep_world=True)
# else:
# obs = env.reset(reset_target=False)
#else:
# # TODO: Only support multi-target + fixed kitchen; or fixed-target (kitchen)
# obs = env.reset(reset_target=flag_random_reset_target)
target_id = common.target_instruction_dict[env.get_current_target()]
if multi_target and hasattr(trainer, 'set_target'):
trainer.set_target(env.get_current_target())
if store_history:
cur_infos.append(proc_info(env.info))
#cur_images.append(env.render(renderMapLoc=env.cam_info['loc'], display=False))
if model_name != 'rnn': obs = obs.transpose([1, 0, 2])
episode_success.append(0)
episode_good.append(0)
cur_stats = dict(best_dist=1e50,
success=0,
good=0,
reward=0,
target=env.get_current_target(),
meters=env.info['meters'],
optstep=env.info['optsteps'],
length=max_episode_len,
images=None)
if aux_task:
cur_stats['aux_pred_rew'] = 0
cur_stats['aux_pred_err'] = 0
if hasattr(env.house, "_id"):
cur_stats['world_id'] = env.house._id
episode_step = 0
for _st in range(max_episode_len):
# get action
if trainer.is_rnn():
idx = 0
if multi_target:
if aux_task:
action, _, aux_pred = trainer.action(
obs,
return_numpy=True,
target=[[target_id]],
return_aux_pred=True)
else:
action, _ = trainer.action(obs,
return_numpy=True,
target=[[target_id]])
else:
if aux_task:
action, _, aux_pred = trainer.action(
obs, return_numpy=True, return_aux_pred=True)
else:
action, _ = trainer.action(obs, return_numpy=True)
action = action.squeeze()
if greedy_execution:
action = int(np.argmax(action))
else:
action = int(action)
if aux_task:
aux_pred = aux_pred.squeeze()
if greedy_aux_pred:
aux_pred = int(np.argmax(aux_pred))
else:
aux_pred = int(aux_pred)
aux_rew = trainer.get_aux_task_reward(
aux_pred, env.get_current_room_pred_mask())
cur_stats['aux_pred_rew'] += aux_rew
if aux_rew < 0: cur_stats['aux_pred_err'] += 1
else:
idx = trainer.process_observation(obs)
action = trainer.action(
None if greedy_execution else 1.0) # use gumbel noise
# environment step
obs, rew, done, info = env.step(action)
if store_history:
cur_infos.append(proc_info(info))
#cur_images.append(env.render(renderMapLoc=env.cam_info['loc'], display=False))
if model_name != 'rnn': obs = obs.transpose([1, 0, 2])
cur_dist = info['dist']
if cur_dist == 0:
cur_stats['good'] += 1
episode_good[-1] = 1
t += 1
if cur_dist < cur_stats['best_dist']:
cur_stats['best_dist'] = cur_dist
episode_step += 1
# collect experience
trainer.process_experience(idx, action, rew, done,
(_st + 1 >= max_episode_len), info)
if done:
if rew > 5: # magic number:
episode_success[-1] = 1
cur_stats['success'] = 1
cur_stats['length'] = episode_step
if aux_task:
cur_stats['aux_pred_err'] /= episode_step
cur_stats['aux_pred_rew'] /= episode_step
break
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 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'])
if aux_task:
logger.print(
' >>>>>> Aux-Task: Avg Rew = %.4f, Avg Err = %.4f' %
(cur_stats['aux_pred_rew'], cur_stats['aux_pred_err']))
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 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))
if aux_task:
logger.print(
' -->>> Auxiliary-Task: Mean Episode Avg Rew = %.6f, Mean Episode Avg Err = %.6f'
% (np.mean([float(s['aux_pred_rew']) for s in episode_stats]),
np.mean([float(s['aux_pred_err']) for s in episode_stats])))
return episode_stats
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 __init__(self,
k=0,
task_name='roomnav',
false_rate=0.0,
reward_type='indicator',
reward_silence=0,
success_measure='see',
multi_target=True,
include_object_target=True,
fixed_target=None,
aux_task=False,
hardness=None,
max_birthplace_steps=None,
min_birthplace_grids=0,
curriculum_schedule=None,
segment_input='none',
depth_input=False,
target_mask_input=False,
cache_supervision=False,
include_outdoor_target=True,
mask_feature_dim=None,
max_steps=-1,
device=0):
assert k >= 0
init_birthplace = max_birthplace_steps if curriculum_schedule is None else curriculum_schedule[
0]
self.env = common.create_env(
k,
task_name=task_name,
false_rate=false_rate,
reward_type=reward_type,
hardness=hardness,
max_birthplace_steps=init_birthplace,
success_measure=success_measure,
segment_input=segment_input,
depth_input=depth_input,
target_mask_input=target_mask_input,
max_steps=max_steps,
render_device=device,
genRoomTypeMap=aux_task,
cacheAllTarget=multi_target,
include_object_target=include_object_target,
use_discrete_action=True, # assume A3C with discrete actions
reward_silence=reward_silence,
#curriculum_schedule=curriculum_schedule,
cache_supervision=cache_supervision,
include_outdoor_target=include_outdoor_target,
min_birthplace_grids=min_birthplace_grids)
self.obs = self.env.reset() if multi_target else self.env.reset(
target='kitchen')
self.done = False
self.multi_target = multi_target
self.fixed_target = fixed_target
self.aux_task = aux_task
self.supervision = cache_supervision
self.mask_feature_dim = mask_feature_dim
self._mask_feature = self._get_mask_feature(
) if mask_feature_dim is not None else None
self._sup_act = self.env.info[
'supervision'] if self.supervision else None
if self.aux_task:
#self._aux_target = self.env.get_current_room_pred_mask() TODO: Currently do not support aux room pred
assert False, 'Aux Room Prediction Currently Not Supported!'
self._target = common.target_instruction_dict[
self.env.get_current_target()]
def learn_controller(args):
elap = time.time()
# Do not need to log detailed computation stats
common.debugger = utils.FakeLogger()
if args['object_target']:
common.ensure_object_targets()
set_seed(args['seed'])
task = common.create_env(args['house'],
task_name=args['task_name'],
false_rate=args['false_rate'],
success_measure=args['success_measure'],
depth_input=args['depth_input'],
target_mask_input=args['target_mask_input'],
segment_input=args['segmentation_input'],
cacheAllTarget=True,
render_device=args['render_gpu'],
use_discrete_action=True,
include_object_target=args['object_target'],
include_outdoor_target=args['outdoor_target'],
discrete_angle=True)
# create motion
__controller_warmstart = args['warmstart']
args['warmstart'] = args['motion_warmstart']
motion = create_motion(args, task)
args['warmstart'] = __controller_warmstart
# logger
logger = utils.MyLogger(args['save_dir'], True)
logger.print("> Planner Units = {}".format(args['units']))
logger.print("> Max Planner Steps = {}".format(args['max_planner_steps']))
logger.print("> Max Exploration Steps = {}".format(args['max_exp_steps']))
logger.print("> Reward = {} & {}".format(args['time_penalty'],
args['success_reward']))
# Planner Learning
logger.print('Start RNN Planner Learning ...')
planner = RNNPlanner(motion, args['units'], args['warmstart'])
fixed_target = None
if args['only_eval_room']:
fixed_target = 'any-room'
elif args['only_eval_object']:
fixed_target = 'any-object'
train_stats, eval_stats = \
planner.learn(args['iters'], args['max_episode_len'],
target=fixed_target,
motion_steps=args['max_exp_steps'],
planner_steps=args['max_planner_steps'],
batch_size=args['batch_size'],
lrate=args['lrate'], grad_clip=args['grad_clip'],
weight_decay=args['weight_decay'], gamma=args['gamma'],
entropy_penalty=args['entropy_penalty'],
save_dir=args['save_dir'],
report_rate=5, eval_rate=20, save_rate=100,
logger=logger, seed=args['seed'])
logger.print('######## Done ###########')
filename = args['save_dir']
if filename[-1] != '/': filename = filename + '/'
filename = filename + 'train_stats.pkl'
with open(filename, 'wb') as f:
pickle.dump([train_stats, eval_stats], f)
logger.print(' --> Training Stats Saved to <{}>!'.format(filename))
return planner
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