def main():
parser = argparse.ArgumentParser(description='LUBAN runner')
register_model_args(parser)
register_game_args(parser)
params, unparsed = parser.parse_known_args(sys.argv)
params.game_variables = [('health', 101), ('sel_ammo', 301)]
params.frame_skip = 3
params.eval_time = 900
sess = tf.Session()
agent = Agent(sess, params)
agent.load("./checkpoint")
action_builder = ActionBuilder(params)
game = Game(scenario='full_deathmatch',
action_builder=action_builder,
score_variable='USER2',
freedoom=True,
screen_resolution='RES_400X225',
screen_format='CRCGCB',
use_screen_buffer=True,
use_depth_buffer=False,
labels_mapping="",
game_features="target,enemy",
mode=('SPECTATOR' if params.human_player else 'PLAYER'),
render_hud=False,
render_crosshair=True,
render_weapon=True,
freelook=params.freelook,
visible=0,
n_bots=4,
use_scripted_marines=True)
game.start(map_id=3)
game.init_bots_health(100)
n_iter = 0
last_states = []
# eval_time: seconds, 35: fps, eval_time*35: total frames
while n_iter * params.frame_skip < params.eval_time * 35:
n_iter += 1
if game.is_player_dead():
game.respawn_player()
# TODO: add reset method to network
# network.reset()
# observe the game state / select the next action
game.observe_state(params, last_states)
# single frame
screen1 = last_states[11].screen.reshape((1, 1, 60, 108, 3))
screen1 = screen1.astype(np.float32) / 255.
# two frame
screen2 = np.array([
last_states[i + 10].screen.transpose((1, 2, 0)) for i in range(2)
]).reshape((1, 2, 60, 108, 3))
screen2 = screen2.astype(np.float32) / 255.
# four frame
screen4 = np.array([
last_states[i + 8].screen.transpose((1, 2, 0)) for i in range(4)
]).reshape((1, 4, 60, 108, 3))
screen4 = screen4.astype(np.float32) / 255.
# eight frame
screen8 = np.array([
last_states[i + 4].screen.transpose((1, 2, 0)) for i in range(8)
]).reshape((1, 8, 60, 108, 3))
screen8 = screen8.astype(np.float32) / 255.
# eight frame
screen12 = np.array([
last_states[i].screen.transpose((1, 2, 0)) for i in range(12)
]).reshape((1, 12, 60, 108, 3))
screen12 = screen12.astype(np.float32) / 255.
action, output_gf = agent.choose_action(screen2)
'''
if action == 0:
action = 5
elif action == 1:
action = 7
elif action == 2:
action = 9
elif action == 3:
action = 2
elif action == 4:
action = 4
else:
action = 0
'''
print("target: %.4f, enemy: %.4f" % (output_gf[0][0], output_gf[0][1]))
#action = np.random.randint(0, 29)
print(
"health: %d\tsel_ammo: %d\taction: %d" %
(last_states[3].variables[0], last_states[3].variables[1], action))
sleep = 0.01
game.make_action(action, frame_skip=params.frame_skip, sleep=sleep)
game.close()
game.print_statistics()
grid_matrix = g.reshape((10, 10))
grid_matrix[3, 4] = 0
grid_matrix[5, 5] = 0
grid_matrix[6, 7] = 0
grid_matrix[7, 7] = 0
np.save('grid.npy', grid_matrix)
grid_matrix = np.load('grid.npy')
grid_matrix /= 2.0
g2 = copy.deepcopy(grid_matrix)
are = arena(grid_matrix, False)
# are = arena(grid_matrix,True)
a1 = Agent(0.42, .46, .25, 0, np.array([3, 4]), are)
a1.load = False
# a2 = Agent(0.2,3,4,3,np.array([5,5]),2,are)
# a2.load = True
a3 = Agent(.49, .2, .2, 2, np.array([6, 7]), are)
a3.load = False
a2 = Agent(.3, .25, .3, 0, np.array([7, 7]), are)
a2.load = False
# are.add_agents([a4,a2,a3,a1])
are.add_agents([a1, a2, a3])
g1 = are.grid_matrix
gm = []
