def initialize_vizdoom(config):
game = DoomGame()
game.load_config(config)
if handleArguments().demo_mode:
game.set_window_visible(True)
else:
game.set_window_visible(False)
game.set_mode(Mode.PLAYER)
game.set_screen_format(ScreenFormat.GRAY8)
game.set_screen_resolution(ScreenResolution.RES_640X480)
game.init()
return game
def run(self):
total_step = 1
if handleArguments().load_model:
INNER_MAX_EP = MAX_EP
else:
INNER_MAX_EP = mp.cpu_count() * 3
while self.inner_ep.value < INNER_MAX_EP:
self.game.new_episode()
state = game_state(self.game)
buffer_s, buffer_a, buffer_r = [], [], []
ep_r = 0.
print("initialized:", self.name)
while True:
start = time.time()
done = False
a = self.lnet.choose_action(state)
if a in attack:
self.action_queue.put(1)
else:
self.action_queue.put(0)
r = self.game.make_action(actions[a], frame_repeat)
if self.game.is_episode_finished():
done = True
else:
s_ = game_state(self.game)
ep_r += r
buffer_a.append(a)
buffer_s.append(state)
buffer_r.append(r)
if done or total_step % UPDATE_GLOBAL_ITER == 0: # update network
# sync
optimize(opt, self.lnet, done, s_, buffer_s, buffer_a,
buffer_r, GAMMA)
buffer_s, buffer_a, buffer_r = [], [], []
if done:
self.inner_ep.value += 1
print("Inner Ep:", self.inner_ep.value)
end = time.time()
time_done = end - start
record(self.g_ep, self.g_ep_r, ep_r, self.res_queue,
self.time_queue, self.g_time, time_done,
self.name)
for lp, gp in zip(self.lnet.parameters(),
self.gnet.parameters()):
gp._grad = lp.grad
break
state = s_
total_step += 1
self.time_queue.put(None)
self.res_queue.put(None)
self.action_queue.put(None)
self.time_queue.put(None)
self.res_queue.put(None)
self.action_queue.put(None)
if __name__ == '__main__':
print("Starting A2C-Sync Agent for Vizdoom-DeadlyCorridor")
time.sleep(3)
timedelta_sum = datetime.now()
timedelta_sum -= timedelta_sum
fig, (ax1, ax2) = plt.subplots(2, 1, sharex=True)
if handleArguments().normalized_plot and not handleArguments().save_data:
runs = 3
else:
runs = 1
for i in range(runs):
starttime = datetime.now()
# load global network
if handleArguments().load_model:
model = Net(len(actions))
model = torch.load("./VIZDOOM/doom_save_model/a2c_sync_doom.pt")
model.eval()
else:
model = Net(len(actions))