def main(args, thread_num=0):
print(thread_num)
# settings
alfred_dataset_path = '../data/json_2.1.0/train'
constants.DATA_SAVE_PATH = args.save_path
print("Force Unsave Data: %s" % str(args.force_unsave))
# Set up data structure to track dataset balance and use for selecting next parameters.
# In actively gathering data, we will try to maximize entropy for each (e.g., uniform spread of goals,
# uniform spread over patient objects, uniform recipient objects, and uniform scenes).
succ_traj = pd.DataFrame(
columns=["goal", "pickup", "movable", "receptacle", "scene"])
# objects-to-scene and scene-to-objects database
for scene_type, ids in constants.SCENE_TYPE.items():
for id in ids:
obj_json_file = os.path.join('layouts',
'FloorPlan%d-objects.json' % id)
with open(obj_json_file, 'r') as of:
scene_objs = json.load(of)
id_str = str(id)
scene_id_to_objs[id_str] = scene_objs
for obj in scene_objs:
if obj not in obj_to_scene_ids:
obj_to_scene_ids[obj] = set()
obj_to_scene_ids[obj].add(id_str)
# scene-goal database
for g in constants.GOALS:
for st in constants.GOALS_VALID[g]:
scenes_for_goal[g].extend(
[str(s) for s in constants.SCENE_TYPE[st]])
scenes_for_goal[g] = set(scenes_for_goal[g])
# scene-type database
for st in constants.SCENE_TYPE:
for s in constants.SCENE_TYPE[st]:
scene_to_type[str(s)] = st
# pre-populate counts in this structure using saved trajectories path.
succ_traj, full_traj = load_successes_from_disk(args.save_path, succ_traj,
args.just_examine,
args.repeats_per_cond)
if args.just_examine:
print_successes(succ_traj)
return
print(succ_traj.groupby('goal').count())
# pre-populate failed trajectories.
fail_traj = load_fails_from_disk(args.save_path)
print("Loaded %d known failed tuples" % len(fail_traj))
# create env and agent
env = ThorEnv(x_display='0.%d' % (thread_num % 2))
game_state = TaskGameStateFullKnowledge(env)
agent = DeterministicPlannerAgent(thread_id=0, game_state=game_state)
errors = {
} # map from error strings to counts, to be shown after every failure.
goal_candidates = constants.GOALS[:]
pickup_candidates = list(set().union(*[
constants.VAL_RECEPTACLE_OBJECTS[
obj] # Union objects that can be placed.
for obj in constants.VAL_RECEPTACLE_OBJECTS
]))
pickup_candidates = [
p for p in pickup_candidates
if constants.OBJ_PARENTS[p] in obj_to_scene_ids
]
movable_candidates = list(
set(constants.MOVABLE_RECEPTACLES).intersection(
obj_to_scene_ids.keys()))
receptacle_candidates = [obj for obj in constants.VAL_RECEPTACLE_OBJECTS
if obj not in constants.MOVABLE_RECEPTACLES and obj in obj_to_scene_ids] + \
[obj for obj in constants.VAL_ACTION_OBJECTS["Toggleable"]
if obj in obj_to_scene_ids]
# toaster isn't interesting in terms of producing linguistic diversity
receptacle_candidates.remove('Toaster')
receptacle_candidates.sort()
scene_candidates = list(scene_id_to_objs.keys())
n_until_load_successes = args.async_load_every_n_samples
print_successes(succ_traj)
task_sampler = sample_task_params(succ_traj, full_traj, fail_traj,
goal_candidates, pickup_candidates,
movable_candidates,
receptacle_candidates, scene_candidates)
# main generation loop
# keeps trying out new task tuples as trajectories either fail or suceed
while True:
# for _ in range(20):
for ii, json_path in enumerate(
glob.iglob(os.path.join(alfred_dataset_path, "**",
"traj_data.json"),
recursive=True)):
# if ii % args.num_threads == thread_num:
# if ii == 5:
sampled_task = json_path.split('/')[-3].split('-')
# sampled_task = next(task_sampler)
# print("===============")
# print(ii, json_path)
print(sampled_task) # DEBUG
# print("===============")
if sampled_task is None:
sys.exit(
"No valid tuples left to sample (all are known to fail or already have %d trajectories"
% args.repeats_per_cond)
gtype, pickup_obj, movable_obj, receptacle_obj, sampled_scene = sampled_task
sampled_scene = int(sampled_scene)
print("sampled tuple: " + str((gtype, pickup_obj, movable_obj,
receptacle_obj, sampled_scene)))
tries_remaining = args.trials_before_fail
# only try to get the number of trajectories left to make this tuple full.
target_remaining = args.repeats_per_cond - len(
succ_traj.loc[(succ_traj['goal'] == gtype)
& (succ_traj['pickup'] == pickup_obj) &
(succ_traj['movable'] == movable_obj) &
(succ_traj['receptacle'] == receptacle_obj) &
(succ_traj['scene'] == str(sampled_scene))])
num_place_fails = 0 # count of errors related to placement failure for no valid positions.
# continue until we're (out of tries + have never succeeded) or (have gathered the target number of instances)
while num_place_fails > args.trials_before_fail or target_remaining > 0:
# environment setup
constants.pddl_goal_type = gtype
print("PDDLGoalType: " + constants.pddl_goal_type)
task_id = create_dirs(gtype, pickup_obj, movable_obj,
receptacle_obj, sampled_scene)
# setup data dictionary
setup_data_dict()
constants.data_dict['task_id'] = task_id
constants.data_dict['task_type'] = constants.pddl_goal_type
constants.data_dict['dataset_params'][
'video_frame_rate'] = constants.VIDEO_FRAME_RATE
# plan & execute
try:
# if True:
# Agent reset to new scene.
constraint_objs = {
'repeat': [(
constants.OBJ_PARENTS[
pickup_obj], # Generate multiple parent objs.
np.random.randint(
2 if gtype == "pick_two_obj_and_place" else 1,
constants.PICKUP_REPEAT_MAX + 1))],
'sparse':
[(receptacle_obj.replace('Basin', ''),
num_place_fails * constants.RECEPTACLE_SPARSE_POINTS)
]
}
if movable_obj != "None":
constraint_objs['repeat'].append(
(movable_obj,
np.random.randint(1, constants.PICKUP_REPEAT_MAX +
1)))
for obj_type in scene_id_to_objs[str(sampled_scene)]:
if (obj_type in pickup_candidates and
obj_type != constants.OBJ_PARENTS[pickup_obj]
and obj_type != movable_obj):
constraint_objs['repeat'].append(
(obj_type,
np.random.randint(
1,
constants.MAX_NUM_OF_OBJ_INSTANCES + 1)))
if gtype in goal_to_invalid_receptacle:
constraint_objs['empty'] = [
(r.replace('Basin', ''), num_place_fails *
constants.RECEPTACLE_EMPTY_POINTS)
for r in goal_to_invalid_receptacle[gtype]
]
constraint_objs['seton'] = []
if gtype == 'look_at_obj_in_light':
constraint_objs['seton'].append(
(receptacle_obj, False))
if num_place_fails > 0:
print(
"Failed %d placements in the past; increased free point constraints: "
% num_place_fails + str(constraint_objs))
scene_info = {
'scene_num': sampled_scene,
'random_seed': random.randint(0, 2**32)
}
info = agent.reset(scene=scene_info, objs=constraint_objs)
# Problem initialization with given constraints.
task_objs = {'pickup': pickup_obj}
if movable_obj != "None":
task_objs['mrecep'] = movable_obj
if gtype == "look_at_obj_in_light":
task_objs['toggle'] = receptacle_obj
else:
task_objs['receptacle'] = receptacle_obj
agent.setup_problem({'info': info},
scene=scene_info,
objs=task_objs)
# Now that objects are in their initial places, record them.
object_poses = [{
'objectName':
obj['name'].split('(Clone)')[0],
'position':
obj['position'],
'rotation':
obj['rotation']
} for obj in env.last_event.metadata['objects']
if obj['pickupable']]
dirty_and_empty = gtype == 'pick_clean_then_place_in_recep'
object_toggles = [{
'objectType': o,
'stateChange': 'toggleable',
'isToggled': v
} for o, v in constraint_objs['seton']]
constants.data_dict['scene']['object_poses'] = object_poses
constants.data_dict['scene'][
'dirty_and_empty'] = dirty_and_empty
constants.data_dict['scene'][
'object_toggles'] = object_toggles
# Pre-restore the scene to cause objects to "jitter" like they will when the episode is replayed
# based on stored object and toggle info. This should put objects closer to the final positions they'll
# be inlay at inference time (e.g., mugs fallen and broken, knives fallen over, etc.).
print("Performing reset via thor_env API")
env.reset(sampled_scene)
print("Performing restore via thor_env API")
env.restore_scene(object_poses, object_toggles,
dirty_and_empty)
event = env.step(
dict(constants.data_dict['scene']['init_action']))
terminal = False
while not terminal and agent.current_frame_count <= constants.MAX_EPISODE_LENGTH:
action_dict = agent.get_action(None)
agent.step(action_dict)
reward, terminal = agent.get_reward()
dump_data_dict()
save_video()
# else:
except Exception as e:
import traceback
traceback.print_exc()
print("Error: " + repr(e))
print("Invalid Task: skipping...")
if args.debug:
print(traceback.format_exc())
deleted = delete_save(args.in_parallel)
if not deleted: # another thread is filling this task successfully, so leave it alone.
target_remaining = 0 # stop trying to do this task.
else:
if str(
e
) == "API Action Failed: No valid positions to place object found":
# Try increasing the space available on sparse and empty flagged objects.
num_place_fails += 1
tries_remaining -= 1
else: # generic error
tries_remaining -= 1
estr = str(e)
if len(estr) > 120:
estr = estr[:120]
if estr not in errors:
errors[estr] = 0
errors[estr] += 1
print("%%%%%%%%%%")
es = sum([errors[er] for er in errors])
print("\terrors (%d):" % es)
for er, v in sorted(errors.items(),
key=lambda kv: kv[1],
reverse=True):
if v / es < 0.01: # stop showing below 1% of errors.
break
print("\t(%.2f) (%d)\t%s" % (v / es, v, er))
print("%%%%%%%%%%")
continue
if args.force_unsave:
delete_save(args.in_parallel)
# add to save structure.
succ_traj = succ_traj.append(
{
"goal": gtype,
"movable": movable_obj,
"pickup": pickup_obj,
"receptacle": receptacle_obj,
"scene": str(sampled_scene)
},
ignore_index=True)
target_remaining -= 1
tries_remaining += args.trials_before_fail # on success, add more tries for future successes
# if this combination resulted in a certain number of failures with no successes, flag it as not possible.
if tries_remaining == 0 and target_remaining == args.repeats_per_cond:
new_fails = [(gtype, pickup_obj, movable_obj, receptacle_obj,
str(sampled_scene))]
fail_traj = load_fails_from_disk(args.save_path,
to_write=new_fails)
print("%%%%%%%%%%")
print("failures (%d)" % len(fail_traj))
# print("\t" + "\n\t".join([str(ft) for ft in fail_traj]))
print("%%%%%%%%%%")
# if this combination gave us the repeats we wanted, note it as filled.
if target_remaining == 0:
full_traj.add((gtype, pickup_obj, movable_obj, receptacle_obj,
sampled_scene))
# if we're sharing with other processes, reload successes from disk to update local copy with others' additions.
if args.in_parallel:
if n_until_load_successes > 0:
n_until_load_successes -= 1
else:
print(
"Reloading trajectories from disk because of parallel processes..."
)
succ_traj = pd.DataFrame(
columns=succ_traj.columns) # Drop all rows.
succ_traj, full_traj = load_successes_from_disk(
args.save_path, succ_traj, False,
args.repeats_per_cond)
print("... Loaded %d trajectories" % len(succ_traj.index))
n_until_load_successes = args.async_load_every_n_samples
print_successes(succ_traj)
task_sampler = sample_task_params(
succ_traj, full_traj, fail_traj, goal_candidates,
pickup_candidates, movable_candidates,
receptacle_candidates, scene_candidates)
print(
"... Created fresh instance of sample_task_params generator"
)
class Thor(threading.Thread):
def __init__(self, queue, train_eval="train"):
Thread.__init__(self)
self.action_queue = queue
self.mask_rcnn = None
self.env = None
self.train_eval = train_eval
self.controller_type = "oracle"
def run(self):
while True:
action, reset, task_file = self.action_queue.get()
try:
if reset:
self.reset(task_file)
else:
self.step(action)
finally:
self.action_queue.task_done()
def init_env(self, config):
self.config = config
screen_height = config['env']['thor']['screen_height']
screen_width = config['env']['thor']['screen_width']
smooth_nav = config['env']['thor']['smooth_nav']
save_frames_to_disk = config['env']['thor']['save_frames_to_disk']
if not self.env:
self.env = ThorEnv(player_screen_height=screen_height,
player_screen_width=screen_width,
smooth_nav=smooth_nav,
save_frames_to_disk=save_frames_to_disk)
self.controller_type = self.config['controller']['type']
self._done = False
self._res = ()
self._feedback = ""
self.expert = HandCodedThorAgent(self.env, max_steps=200)
self.prev_command = ""
self.load_mask_rcnn()
def load_mask_rcnn(self):
# load pretrained MaskRCNN model if required
if 'mrcnn' in self.config['controller'][
'type'] and not self.mask_rcnn:
model_path = os.path.join(
os.environ['ALFRED_ROOT'],
self.config['mask_rcnn']['pretrained_model_path'])
self.mask_rcnn = load_pretrained_model(model_path)
def set_task(self, task_file):
self.task_file = task_file
self.traj_root = os.path.dirname(task_file)
with open(task_file, 'r') as f:
self.traj_data = json.load(f)
def reset(self, task_file):
assert self.env
assert self.controller_type
self.set_task(task_file)
# scene setup
scene_num = self.traj_data['scene']['scene_num']
object_poses = self.traj_data['scene']['object_poses']
dirty_and_empty = self.traj_data['scene']['dirty_and_empty']
object_toggles = self.traj_data['scene']['object_toggles']
scene_name = 'FloorPlan%d' % scene_num
self.env.reset(scene_name)
self.env.restore_scene(object_poses, object_toggles,
dirty_and_empty)
# recording
save_frames_path = self.config['env']['thor']['save_frames_path']
self.env.save_frames_path = os.path.join(
save_frames_path, self.traj_root.replace('../', ''))
# initialize to start position
self.env.step(dict(
self.traj_data['scene']['init_action'])) # print goal instr
task_desc = get_templated_task_desc(self.traj_data)
print("Task: %s" % task_desc)
# print("Task: %s" % (self.traj_data['turk_annotations']['anns'][0]['task_desc']))
# setup task for reward
class args:
pass
args.reward_config = os.path.join(os.environ['ALFRED_ROOT'],
'agents/config/rewards.json')
self.env.set_task(self.traj_data, args, reward_type='dense')
# set controller
self.controller_type = self.config['controller']['type']
self.goal_desc_human_anns_prob = self.config['env'][
'goal_desc_human_anns_prob']
load_receps = self.config['controller']['load_receps']
debug = self.config['controller']['debug']
if self.controller_type == 'oracle':
self.controller = OracleAgent(
self.env,
self.traj_data,
self.traj_root,
load_receps=load_receps,
debug=debug,
goal_desc_human_anns_prob=self.goal_desc_human_anns_prob)
elif self.controller_type == 'oracle_astar':
self.controller = OracleAStarAgent(
self.env,
self.traj_data,
self.traj_root,
load_receps=load_receps,
debug=debug,
goal_desc_human_anns_prob=self.goal_desc_human_anns_prob)
elif self.controller_type == 'mrcnn':
self.controller = MaskRCNNAgent(
self.env,
self.traj_data,
self.traj_root,
pretrained_model=self.mask_rcnn,
load_receps=load_receps,
debug=debug,
goal_desc_human_anns_prob=self.goal_desc_human_anns_prob,
save_detections_to_disk=self.env.save_frames_to_disk,
save_detections_path=self.env.save_frames_path)
elif self.controller_type == 'mrcnn_astar':
self.controller = MaskRCNNAStarAgent(
self.env,
self.traj_data,
self.traj_root,
pretrained_model=self.mask_rcnn,
load_receps=load_receps,
debug=debug,
goal_desc_human_anns_prob=self.goal_desc_human_anns_prob,
save_detections_to_disk=self.env.save_frames_to_disk,
save_detections_path=self.env.save_frames_path)
else:
raise NotImplementedError()
# zero steps
self.steps = 0
# reset expert state
self.expert.reset(task_file)
self.prev_command = ""
# return intro text
self._feedback = self.controller.feedback
self._res = self.get_info()
return self._feedback
def step(self, action):
if not self._done:
# take action
self.prev_command = str(action)
self._feedback = self.controller.step(action)
self._res = self.get_info()
if self.env.save_frames_to_disk:
self.record_action(action)
self.steps += 1
def get_results(self):
return self._res
def record_action(self, action):
txt_file = os.path.join(self.env.save_frames_path, 'action.txt')
with open(txt_file, 'a+') as f:
f.write("%s\r\n" % str(action))
def get_info(self):
won = self.env.get_goal_satisfied()
pcs = self.env.get_goal_conditions_met()
goal_condition_success_rate = pcs[0] / float(pcs[1])
acs = self.controller.get_admissible_commands()
# expert action
if self.train_eval == "train":
game_state = {
'admissible_commands': acs,
'feedback': self._feedback,
'won': won
}
expert_actions = ["look"]
try:
if not self.prev_command:
self.expert.observe(game_state['feedback'])
else:
next_action = self.expert.act(game_state, 0, won,
self.prev_command)
if next_action in acs:
expert_actions = [next_action]
except HandCodedAgentTimeout:
print("Expert Timeout")
except Exception as e:
print(e)
traceback.print_exc()
else:
expert_actions = []
training_method = self.config["general"]["training_method"]
if training_method == "dqn":
max_nb_steps_per_episode = self.config["rl"]["training"][
"max_nb_steps_per_episode"]
elif training_method == "dagger":
max_nb_steps_per_episode = self.config["dagger"]["training"][
"max_nb_steps_per_episode"]
else:
raise NotImplementedError
self._done = won or self.steps > max_nb_steps_per_episode
return (self._feedback, self._done, acs, won,
goal_condition_success_rate, expert_actions)
def get_last_frame(self):
return self.env.last_event.frame[:, :, ::-1]
def get_exploration_frames(self):
return self.controller.get_exploration_frames()