def get_player(viz=False, train=False, dumpdir=None):
# pl = None
# if ENV_NAME == "MultiRoomEnv":
# print "An awful hack for the moment"
# pl = nmt.NonMarkovEnvironment()
# else:
pl = GymEnv(ENV_NAME, dumpdir=dumpdir)
# print "HAS ATTR:{}".format(hasattr(pl, "experiment_id"))
if EXPERIMENT_ID != None and hasattr(pl.gymenv, "set_experiment_id"):
pl.gymenv.set_experiment_id(EXPERIMENT_ID)
pl = MapPlayerState(pl, EXPERIMENT_MODEL.get_screen_processor())
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
if hasattr(EXPERIMENT_MODEL, "get_history_processor"):
pl = HistoryFramePlayer(pl, FRAME_HISTORY,
EXPERIMENT_MODEL.get_history_processor())
else:
pl = HistoryFramePlayer(
pl,
FRAME_HISTORY,
)
if not train:
pl = PreventStuckPlayer(pl, 30, 1)
pl = LimitLengthPlayer(pl, 40000)
return pl
def get_player(viz=False, train=False, dumpdir=None):
if PC_METHOD and train:
pl = GymEnv(ENV_NAME, dumpdir=dumpdir, pc_method=PC_METHOD)
else:
pl = GymEnv(ENV_NAME, dumpdir=dumpdir)
def resize(img):
return cv2.resize(img, IMAGE_SIZE)
def grey(img):
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img = resize(img)
img = img[:, :, np.newaxis]
return img
pl = MapPlayerState(pl, grey)
#show_images(pl.current_state())
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
if not train:
pl = HistoryFramePlayer(pl, FRAME_HISTORY)
pl = PreventStuckPlayer(pl, 30, 1)
pl = LimitLengthPlayer(pl, 40000)
return pl
def get_player(viz=False, train=False, dumpdir=None):
#TODO: (Next Plan)idea1 use CNN as features of our density model
#TODO: idea1.5 clear counter in some intermeidate points
#TODO: (on EXP now)idea2 time increasing with psuedo reward. IF the pseudo reward is less than a threshold (e.g.,0.01) for most of the states, increase the pseudo reward.
#TODO: (on EXP now)Not decrease Explore Factor after several epochs. The exp results show not enough exploration afterwards. But the scores are remained greatly.
#TODO: (Read more papers)idea2.5: Intuition from people. Exploration and Exploitation modes. Remember the good rewards and turn into Exploitation modes, explore other possibilities.
#TODO: (Done)Evaluate with policy probability
if PC_METHOD and train:
pl = GymEnv(ENV_NAME, dumpdir=dumpdir, pc_method=PC_METHOD, pc_mult=PC_MULT, pc_thre=PC_THRE, pc_time=PC_TIME, feature=FEATURE, pc_action=PC_ACTION, pc_downsample_value=PC_DOWNSAMPLE_VALUE, pc_clean=PC_CLEAN, UCB1=UCB1)
else:
pl = GymEnv(ENV_NAME, dumpdir=dumpdir)
def resize(img):
return cv2.resize(img, IMAGE_SIZE)
def grey(img):
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img = resize(img)
img = img[:, :, np.newaxis]
return img
pl = MapPlayerState(pl, grey)
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
pl = HistoryFramePlayer(pl, FRAME_HISTORY)
#if not train:
# pl = PreventStuckPlayer(pl, 30, 1)
pl = LimitLengthPlayer(pl, 40000)
return pl
def get_player(viz=False, train=False, dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir)
global NUM_ACTIONS
global IMAGE_SHAPE3
global FRAME_HISTORY
NUM_ACTIONS = pl.get_action_space().num_actions()
def resize(img):
return cv2.resize(img, IMAGE_SIZE)
def grey(img):
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img = resize(img)
img = img[:, :, np.newaxis] / 255.0
return img
if OBJECT_METHOD == 'swap_input_combine': #1
# swap the input with combined object image
FRAME_HISTORY = 4
IMAGE_SHAPE3 = IMAGE_SIZE + (FRAME_HISTORY,)
pl = ObjectSensitivePlayer(pl, TEMPLATE_MATCHER, OBJECT_METHOD, resize)
#pl = HistoryFramePlayer(pl, FRAME_HISTORY)
#show_images(pl.current_state())
if OBJECT_METHOD == 'add_input_combine': #2
# add the input with combined object image for each history
FRAME_HISTORY = 4
IMAGE_SHAPE3 = IMAGE_SIZE + (FRAME_HISTORY * 2,)
pl = MapPlayerState(pl, grey)
pl = ObjectSensitivePlayer(pl, TEMPLATE_MATCHER, OBJECT_METHOD, resize)
#pl = HistoryFramePlayer(pl, FRAME_HISTORY)
#show_images(pl.current_state())
if OBJECT_METHOD == 'add_input_separate': #3
# For the current state, add the object images
# (obj1_his1, obj2_his1, cur_his1, obj1_his2, ... )
# For each image, use the grey scale image, and resize it to 84 * 84
FRAME_HISTORY = 4
IMAGE_SHAPE3 = IMAGE_SIZE + (FRAME_HISTORY * (len(TEMPLATE_MATCHER.index2obj)+1),)
pl = MapPlayerState(pl, grey)
pl = ObjectSensitivePlayer(pl, TEMPLATE_MATCHER, OBJECT_METHOD, resize)
#show_images(pl.current_state())
if OBJECT_METHOD == 'swap_input_separate': #4
# swap the input images with object images
# (obj1_his1, obj2_his1, obj1_his2, obj2_his2, ... )
# TODO: If we need to add walls
# Interesting thing is we don't have any wall info here
FRAME_HISTORY = 4
IMAGE_SHAPE3 = IMAGE_SIZE + (FRAME_HISTORY * len(TEMPLATE_MATCHER.index2obj),)
pl = ObjectSensitivePlayer(pl, TEMPLATE_MATCHER, OBJECT_METHOD, resize)
if not train:
pl = HistoryFramePlayer(pl, FRAME_HISTORY)
pl = PreventStuckPlayer(pl, 30, 1)
#show_images(pl.current_state())
pl = LimitLengthPlayer(pl, 40000)
#show_images(pl.current_state())
#exit()
return pl
def get_player(dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir, auto_restart=False)
pl = MapPlayerState(pl, grey)
pl = MapPlayerState(pl, resize)
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
pl = HistoryFramePlayer(pl, FRAME_HISTORY)
#show_images(pl.current_state())
return pl
def get_player(dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir, auto_restart=False)
pl = MapPlayerState(pl, lambda img: cv2.resize(img, IMAGE_SIZE[::-1]))
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
pl = HistoryFramePlayer(pl, FRAME_HISTORY)
return pl
def get_player(viz=False, train=False, dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir)
def func(img):
return cv2.resize(img, IMAGE_SIZE[::-1])
pl = MapPlayerState(pl, func)
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
if not train:
pl = HistoryFramePlayer(pl, FRAME_HISTORY)
pl = PreventStuckPlayer(pl, 30, 1)
pl = LimitLengthPlayer(pl, 40000)
return pl
def get_player(viz=False, train=False, dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir)
def func(img):
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
#img = cv2.resize(img, IMAGE_SIZE[::])
return img
pl = MapPlayerState(pl, func)
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
pl = LimitLengthPlayer(pl, 40000)
return pl
def get_player(viz=False, train=False, dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir)
#def func(img):
# return cv2.resize(img, IMAGE_SIZE[::-1])
#pl = MapPlayerState(pl, func)
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
if not train:
pass
pl = HistoryFramePlayer(pl, FRAME_HISTORY)
#pl = PreventStuckPlayer(pl, 30, 1) #TODO: I think we don't need this in freeway. Is any bug in this code? didn't see repeated actions.
pl = LimitLengthPlayer(pl, 40000)
return pl
def get_player(viz=False, train=False, dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir)
def func(img):
return cv2.resize(
img, IMAGE_SIZE[::-1]
) #TODO: Do we really need to resize here? Check the original paper.
pl = MapPlayerState(pl, func)
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
if not train: # When testing
pl = HistoryFramePlayer(pl, FRAME_HISTORY)
#pl = PreventStuckPlayer(pl, 30, 1) #TODO: Need to know the start button. Is it different for each game?
pl = LimitLengthPlayer(pl, 30000) # 500s
return pl
def get_player(dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir, auto_restart=False)
def resize(img):
return cv2.resize(img, IMAGE_SIZE)
def grey(img):
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img = resize(img)
img = img[:, :, np.newaxis]
return img
pl = MapPlayerState(pl, grey)
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
pl = HistoryFramePlayer(pl, FRAME_HISTORY)
return pl
def get_player(dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir, force=True)
pl = MapPlayerState(pl, EXPERIMENT_MODEL.get_screen_processor())
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
if hasattr(EXPERIMENT_MODEL, "get_history_processor"):
pl = HistoryFramePlayer(pl, FRAME_HISTORY,
EXPERIMENT_MODEL.get_history_processor())
else:
pl = HistoryFramePlayer(
pl,
FRAME_HISTORY,
)
if not train:
pl = PreventStuckPlayer(pl, 30, 1)
pl = LimitLengthPlayer(pl, 40000)
return pl
def get_player(viz=False, train=False, dumpdir=None):
pl = GymEnv(ENV_NAME, dumpdir=dumpdir)
#TODO: Preprocessing goes here
def func(img):
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
#img = img[:,:,0]
#img = cv2.resize(img, IMAGE_SIZE[::])
return img
pl = MapPlayerState(pl, func)
global NUM_ACTIONS
NUM_ACTIONS = pl.get_action_space().num_actions()
if not train:
pass
#pl = HistoryFramePlayer(pl, FRAME_HISTORY)
#pl = PreventStuckPlayer(pl, 30, 1)
pl = LimitLengthPlayer(pl, 40000)
return pl