def start(self):
_features = features.features_from_game_info(
self.controller.game_info())
i = 0
while i < self.info.game_duration_loops:
i += self.skip
self.controller.step(self.step_mul)
obs = self.controller.observe()
try:
agent_obs = _features.transform_obs(obs)
except:
pass
if obs.player_result:
self._state = StepType.LAST
discount = 0
else:
discount = self.discount
self._episode_steps += self.step_mul
step = TimeStep(step_type=self._state,
reward=0,
discount=discount,
observation=agent_obs)
self.agent.step(step, obs.actions)
if obs.player_result:
break
self._state = StepType.MID
self.save_data()
def start(self):
_features = features.Features(self.controller.game_info())
while True:
self.controller.step(self.step_mul)
obs = self.controller.observe()
agent_obs = _features.transform_obs(obs.observation)
if obs.player_result: # Episide over.
self._state = StepType.LAST
discount = 0
else:
discount = self.discount
self._episode_steps += self.step_mul
step = TimeStep(step_type=self._state,
reward=0,
discount=discount,
observation=agent_obs)
self.agent.step(step, obs.actions)
if obs.player_result:
break
self._state = StepType.MID
def test_one_input(self):
d = 48
# These shapes are what is actually returned from environment
dummy_obs = {
"screen": np.zeros((16, d, d), dtype="int32"),
"minimap": np.zeros((7, d, d), dtype="int32"),
"available_actions": np.arange(10)
}
dummy_ts = TimeStep(StepType.MID, 0.0, 0.0, dummy_obs)
p = ObsProcesser()
assert p.process_one_input(dummy_ts)["screen_numeric"].shape == (
ObsProcesser.N_SCREEN_CHANNELS, d, d)
assert p.process_one_input(dummy_ts)["minimap_numeric"].shape == (
ObsProcesser.N_MINIMAP_CHANNELS, d, d)
n_screen_scalar_features = len(
[k for k in SCREEN_FEATURES if k.type == FeatureType.SCALAR])
total_screen_dim = n_screen_scalar_features + 3 + 1 # binary flags + visibility_flag
assert total_screen_dim == ObsProcesser.N_SCREEN_CHANNELS
n_screen_minimap_features = len(
[k for k in MINIMAP_FEATURES if k.type == FeatureType.SCALAR])
total_minimap_dim = n_screen_minimap_features + 3 + 1
assert total_minimap_dim == ObsProcesser.N_MINIMAP_CHANNELS
def start(self, replay_file_path):
self.openReplay(replay_file_path)
_features = features.Features(self.controller.game_info())
while True:
self.controller.step(self.step_mul)
obs = self.controller.observe()
agent_obs = _features.transform_obs(obs.observation)
if obs.player_result:
self._state = StepType.LAST
discount = 0
else:
discount = self.discount
self._episode_steps += self.step_mul
step = TimeStep(step_type=self._state,
reward=0,
discount=discount,
observation=agent_obs)
self.agent.step(step, obs.actions, self._state == StepType.LAST)
if obs.player_result:
break
self._state = StepType.MID
self.closeReplay()
def start(self):
_features = features.Features(self.controller.game_info())
frames = random.sample(
np.arange(self.info.game_duration_loops).tolist(),
self.info.game_duration_loops)
# frames = frames[0 : min(self.frames_per_game, self.info.game_duration_loops)]
step_mul = 10
frames = frames[0:int(self.info.game_duration_loops) // step_mul]
frames.sort()
last_frame = 0
i = 0
# for frame in frames:
skips = step_mul
while i < self.info.game_duration_loops:
# skips = frame - last_frame
# last_frame = frame
i += skips
self.controller.step(skips)
obs = self.controller.observe()
agent_obs = _features.transform_obs(obs.observation)
if obs.player_result: # Episode over.
self._state = StepType.LAST
discount = 0
else:
discount = self.discount
self._episode_steps += skips
step = TimeStep(step_type=self._state,
reward=0,
discount=discount,
observation=agent_obs)
self.agent.step(step, obs.actions, self.info, _features)
if obs.player_result:
break
self._state = StepType.MID
print("Saving data")
#print(self.agent.states)
pickle.dump({
"info": self.info,
"state": self.agent.states
}, open("data/" + self.replay_file_name + ".p", "wb"))
print("Data successfully saved")
self.agent.states = []
print("Data flushed")
print("Done")
def convert_step(timestep: TimeStep) -> Tuple[Any, float, bool, Dict]:
obs = timestep.observation["feature_screen"][SCREEN_FEATURES.player_relative.index]
obs = obs.view(type=np.ndarray) # Get a standard ndarray view instead of pysc2's subclass (NamedNumpyArray)
# Reshape from (84, 84) to (84, 84, 1). '...' is for slicing higher-dimensional data structures and means
# insert as many full slices (:) to extend the multi-dimensional slice to all dimensions.
# Ref: https://stackoverflow.com/questions/118370/how-do-you-use-the-ellipsis-slicing-syntax-in-python
obs = obs[..., np.newaxis]
done = timestep.last()
info = {}
return obs, timestep.reward, done, info
def train():
env_args = dict(
map_name=FLAGS.map_name,
step_mul=FLAGS.step_mul,
game_steps_per_episode=0,
screen_size_px=(FLAGS.resolution,) * 2,
minimap_size_px=(FLAGS.resolution,) * 2,
visualize=FLAGS.visualize
)
max_frames = FLAGS.frames * 1e6
total_frames = 0
vis = visdom.Visdom()
vis.close(env=FLAGS.expirement_name, win=None)
envs = SC2ProcVec([partial(SC2TorchEnv, env_args) for i in range(FLAGS.n_envs)])
print(f"Starting {FLAGS.n_envs} workers")
try:
agent = A2CAgent(screen_width=FLAGS.resolution,
screen_height=FLAGS.resolution,
expirement_name=FLAGS.expirement_name,
learning_rate=FLAGS.learning_rate,
num_processes=FLAGS.n_envs,
value_coef=FLAGS.value_weight,
entropy_coef=FLAGS.entropy_weight,
continue_training=FLAGS.continue_training,
horizon=FLAGS.horizon)
num_processes = FLAGS.n_envs
horizon = FLAGS.horizon
timesteps = envs.reset()
agent.reset()
while total_frames * num_processes <= max_frames:
total_frames += 1
step = total_frames % horizon
agent.finish_step()
actions = [agent.step(step, p, TimeStep(*t))
for p, t in enumerate(timesteps.reshape(num_processes, 4))]
if step == 0:
agent.rollout()
agent.reset()
timesteps = envs.step(actions)
except KeyboardInterrupt:
pass
finally:
envs.close()
print(f"Training done after {total_frames} steps")
def step(self, ep, step, pb_obs, agent_obs, agent_actions):
"""
Puts the given observations in the queue.
:param int ep: the episode that this observation was made.
:param int step: the episode time-step in which this observation was made.
:param ResponseObservation pb_obs: the observation in protobuf form.
:param TimeStep agent_obs: the observation in pysc2 features form.
:param list[FunctionCall] agent_actions: list of actions executed by the agent between the previous observation
and the current observation.
:return:
"""
if self._ignore_replay:
return
# checks new episode
if step == 0:
self._total_eps += 1
# force step type in observation
agent_obs = TimeStep(StepType.FIRST, agent_obs.reward, agent_obs.discount, agent_obs.observation)
# put sample in queue and wait for ack
self._samples_queue.put((agent_obs, agent_actions, step == 0, self._total_steps))
self._samples_queue.join()
self._total_steps += 1
def start(self):
_features = features.features_from_game_info(
self.controller.game_info())
_features.init_camera(
features.Dimensions(self.screen_size_px, self.minimap_size_px),
point.Point(*self.map_size), self.camera_width)
while True:
self.controller.step(self.step_mul)
obs = self.controller.observe()
try:
agent_obs = _features.transform_obs(obs)
except:
pass
#screenpoint = (42, 42)
#screenpoint = point.Point(*screenpoint)
if (len(obs.actions) == 0):
continue
#else:
# print(obs.actions)
#if obs.observation.game_loop in config.actions:
#func = config.actions[o.game_loop](obs)
#_features.reverse_action(obs.actions[1])
#action = _features.transform_action(obs.observation, actions.FUNCTIONS.move_camera([42,42]))
#self.controller.act(action)
#self.assertEqual(actions.FUNCTIONS.move_camera.id, func.function)
#s2clientprotocol_dot_common__pb2._POINT2D
#actions.FUNCTIONS.move_camera(screenpoint)
#remote_controller.RemoteController.act(actions.move_camera(actions.FUNCTIONS.move_camera,['FEATURES'], screenpoint))
#action_observer_camera_move = (sc_pb.ActionObserverCameraMove(world_pos = screenpoint))
#sc_pb.RequestObserverAction
#screenpoint.assign_to(action_observer_camera_move.world_pos)
#self.controller.act(sc_pb.ActionObserverCameraMove(world_pos=screenpoint))
#sc_pb.RequestObserverAction(actions=[sc_pb.ObserverAction(player_perspective=sc_pb.ActionObserverPlayerPerspective(player_id=2))])
#obsAction = self.controller.act(sc_pb.RequestObserverAction(actions=[sc_pb.ObserverAction(player_perspective=sc_pb.ActionObserverPlayerPerspective(player_id=2))]))# [sc_pb.ActionObserverCameraMove(distance=50)]))
#screenpoint.assign_to(obsAction.camera_move.world_pos)
#remote_controller.RemoteController.actions
if obs.player_result: # Episide over.
self._state = StepType.LAST
discount = 0
else:
discount = self.discount
#if (_features.reverse_action(obs.actions[0]).function == actions.FUNCTIONS.select_rect.id):
agent_obs = _features.transform_obs(obs)
#self._episode_steps += self.step_mul
step = TimeStep(step_type=self._state,
reward=0,
discount=discount,
observation=agent_obs)
acts = []
for action in obs.actions:
for num in self.agent.action_dict.keys():
if (format(
_features.reverse_action(action).function) == num):
acts.append(_features.reverse_action(action))
break
data = self.controller.data()
self.agent.step(step, self.info, acts)
#offset = self.agent.step(step, self.info)
#print(_features.reverse_action(obs.actions[0]))
#print ("+")
#print(offset)
if obs.player_result:
break
self._state = StepType.MID
def get_random_trajectory(self):
function_dict = {}
for _FUNCTION in actions._FUNCTIONS:
#print(_FUNCTION)
function_dict[_FUNCTION.ability_id] = _FUNCTION.name
race_list = ['Terran', 'Zerg', 'Protoss']
"""How many agent steps the agent has been trained for."""
run_config = run_configs.get()
sc2_proc = run_config.start()
controller = sc2_proc.controller
#print ("source: {}".format(source))
#root_path = '/media/kimbring2/Steam/StarCraftII/Replays/4.8.2.71663-20190123_035823-1'
root_path = self.source
file_list = glob.glob(root_path + '*.*')
#print ("file_list: {}".format(file_list))
for i in range(0, 500):
#print("i: " + str(i))
replay_file_path = random.choice(file_list)
#print ("replay_file_path: {}".format(replay_file_path))
#replay_file_path = root_path + '0a0f62052fe4311368910ad38c662bf979e292b86ad02b49b41a87013e58c432.SC2Replay'
#replay_file_path = root_path + '/0a1b09abc9e98f4e0c3921ae0a427c27e97c2bbdcf34f50df18dc41cea3f3249.SC2Replay'
#replay_file_path_2 = root_path + '/0a01d32e9a98e1596b88bc2cdec7752249b22aca774e3305dae2e93efef34be3.SC2Replay'
#replay_file_path_0 = human_data
#print ("replay_file_path: {}".format(replay_file_path))
try:
replay_data = run_config.replay_data(replay_file_path)
ping = controller.ping()
info = controller.replay_info(replay_data)
print("ping: " + str(ping))
print("replay_info: " + str(info))
player0_race = info.player_info[0].player_info.race_actual
player0_mmr = info.player_info[0].player_mmr
player0_apm = info.player_info[0].player_apm
player0_result = info.player_info[0].player_result.result
print("player0_race: " + str(player0_race))
print("player0_mmr: " + str(player0_mmr))
print("player0_apm: " + str(player0_apm))
print("player0_result: " + str(player0_result))
home_race = race_list.index(self.home_race_name) + 1
if (home_race == player0_race):
print("player0_race pass")
else:
print("player0_race fail")
continue
if (player0_mmr >= self.replay_filter):
print("player0_mmr pass ")
else:
print("player0_mmr fail")
continue
player1_race = info.player_info[0].player_info.race_actual
player1_mmr = info.player_info[0].player_mmr
player1_apm = info.player_info[0].player_apm
player1_result = info.player_info[0].player_result.result
print("player1_race: " + str(player1_race))
print("player1_mmr: " + str(player1_mmr))
print("player1_apm: " + str(player1_apm))
print("player1_result: " + str(player1_result))
away_race = race_list.index(self.away_race_name) + 1
if (away_race == player1_race):
print("player1_race pass ")
else:
print("player1_race fail ")
continue
if (player1_mmr >= self.replay_filter):
print("player1_mmr pass ")
else:
print("player1_mmr fail")
continue
screen_size_px = (128, 128)
minimap_size_px = (64, 64)
player_id = 1
discount = 1.
step_mul = 8
screen_size_px = point.Point(*screen_size_px)
minimap_size_px = point.Point(*minimap_size_px)
interface = sc_pb.InterfaceOptions(
raw=False,
score=True,
feature_layer=sc_pb.SpatialCameraSetup(width=24))
screen_size_px.assign_to(interface.feature_layer.resolution)
minimap_size_px.assign_to(
interface.feature_layer.minimap_resolution)
map_data = None
if info.local_map_path:
map_data = run_config.map_data(info.local_map_path)
_episode_length = info.game_duration_loops
_episode_steps = 0
controller.start_replay(
sc_pb.RequestStartReplay(replay_data=replay_data,
map_data=map_data,
options=interface,
observed_player_id=player_id))
_state = StepType.FIRST
if (info.HasField("error")
or info.base_build != ping.base_build
or # different game version
info.game_duration_loops < 1000 or
len(info.player_info) != 2):
# Probably corrupt, or just not interesting.
print("error")
continue
feature_screen_size = 128
feature_minimap_size = 64
rgb_screen_size = None
rgb_minimap_size = None
action_space = None
use_feature_units = True
agent_interface_format = sc2_env.parse_agent_interface_format(
feature_screen=feature_screen_size,
feature_minimap=feature_minimap_size,
rgb_screen=rgb_screen_size,
rgb_minimap=rgb_minimap_size,
action_space=action_space,
use_feature_units=use_feature_units)
_features = features.features_from_game_info(
controller.game_info())
build_info = []
build_name = []
replay_step = 0
print("True loop")
while True:
replay_step += 1
print("replay_step: " + str(replay_step))
controller.step(step_mul)
obs = controller.observe()
self.home_trajectory.append(obs)
if (len(obs.actions) != 0):
action = (obs.actions)[0]
action_spatial = action.action_feature_layer
unit_command = action_spatial.unit_command
ability_id = unit_command.ability_id
function_name = function_dict[ability_id]
if (function_name != 'build_queue'):
function_name_parse = function_name.split('_')
function_name_first = function_name_parse[0]
#print("function_name_first: " + str(function_name_first))
if (function_name_first == 'Build'
or function_name_first == 'Train'):
unit_name = function_name_parse[1]
unit_info = int(
units_new.get_unit_type(
self.home_race_name, unit_name))
#print("unit_name: " + str(unit_name))
#print("unit_info: " + str(unit_info))
#print("function_name_parse[1]: " + str(function_name_parse[1]))
build_name.append(unit_name)
build_info.append(unit_info)
if obs.player_result: # Episide over.
_state = StepType.LAST
discount = 0
else:
discount = discount
_episode_steps += step_mul
agent_obs = _features.transform_obs(obs)
step = TimeStep(step_type=_state,
reward=0,
discount=discount,
observation=agent_obs)
score_cumulative = agent_obs['score_cumulative']
score_cumulative_dict = {}
score_cumulative_dict['score'] = score_cumulative.score
score_cumulative_dict[
'idle_production_time'] = score_cumulative.idle_production_time
score_cumulative_dict[
'idle_worker_time'] = score_cumulative.idle_worker_time
score_cumulative_dict[
'total_value_units'] = score_cumulative.total_value_units
score_cumulative_dict[
'total_value_structures'] = score_cumulative.total_value_structures
score_cumulative_dict[
'killed_value_units'] = score_cumulative.killed_value_units
score_cumulative_dict[
'killed_value_structures'] = score_cumulative.killed_value_structures
score_cumulative_dict[
'collected_minerals'] = score_cumulative.collected_minerals
score_cumulative_dict[
'collected_vespene'] = score_cumulative.collected_vespene
score_cumulative_dict[
'collection_rate_minerals'] = score_cumulative.collection_rate_minerals
score_cumulative_dict[
'collection_rate_vespene'] = score_cumulative.collection_rate_vespene
score_cumulative_dict[
'spent_minerals'] = score_cumulative.spent_minerals
score_cumulative_dict[
'spent_vespene'] = score_cumulative.spent_vespene
if obs.player_result:
break
_state = StepType.MID
self.home_BO = build_info
self.away_BU = score_cumulative_dict
break
except:
continue
build_name.append(unit_name)
build_info.append(unit_info)
if obs.player_result: # Episide over.
_state = StepType.LAST
discount = 0
else:
discount = discount
_episode_steps += step_mul
agent_obs = _features.transform_obs(obs)
#print("agent_obs['feature_units']: " + str(agent_obs['feature_units']))
step = TimeStep(step_type=_state,
reward=0,
discount=discount,
observation=agent_obs)
score_cumulative = agent_obs['score_cumulative']
score_cumulative_dict = {}
score_cumulative_dict['score'] = score_cumulative.score
score_cumulative_dict[
'idle_production_time'] = score_cumulative.idle_production_time
score_cumulative_dict[
'idle_worker_time'] = score_cumulative.idle_worker_time
score_cumulative_dict[
'total_value_units'] = score_cumulative.total_value_units
score_cumulative_dict[
'total_value_structures'] = score_cumulative.total_value_structures
score_cumulative_dict[
'killed_value_units'] = score_cumulative.killed_value_units
def start(self):
print("Hello we are in Start")
step_mul = 1
trainingDataPath = 'C:\\Users\\Charlie\\training_data\\4101\\'
_features = features.features_from_game_info(
self.controller.game_info(), use_camera_position=True)
#print("world_tl_to_world_camera_rel: {}\n\nworld_to_feature_screen_px: {}\n\nworld_to_world_tl: {}".format(_features._world_tl_to_world_camera_rel,
# _features._world_to_feature_screen_px,
# _features._world_to_world_tl))
# _features.init_camera(features.Dimensions(self.screen_size_px, self.minimap_size_px),
# point.Point(*const.WorldSize()),
# self.camera_width)
packageCounter = 0
fileName = trainingDataPath + self.replay_file_name + "/" + str(
packageCounter) + '.csv'
npFileName = trainingDataPath + self.replay_file_name + "/" + str(
packageCounter) + '.npy'
npFileNameComp = trainingDataPath + self.replay_file_name + "/" + str(
packageCounter)
dirname = os.path.dirname(fileName)
if not os.path.exists(dirname):
os.makedirs(dirname)
# keyboard = Controller()
# time.sleep(1)
# keyboard.press(str(self.player_id))
# time.sleep(0.5)
# keyboard.release(str(self.player_id))
while True:
#Takes one step through the replay
self.controller.step(step_mul)
#Converts visual data into abstract data
obs = self.controller.observe()
if obs.player_result: # Episide over.
self._state = StepType.LAST
print("Episode Over")
break
discount = 0
else:
discount = self.discount
if (len(obs.actions) == 0):
continue
agent_obs = _features.transform_obs(obs)
step = TimeStep(step_type=self._state,
reward=0,
discount=discount,
observation=agent_obs)
for action in obs.actions:
for num in self.agent.action_dict.keys():
# If action is worth recording
if (int(_features.reverse_action(action).function) == num):
# Check if the action is on a Micro Unit
if (const.IsMicroUnit(agent_obs.single_select)
or const.IsMicroUnit(agent_obs.multi_select)):
# Record action
#print(_features._world_tl_to_world_camera_rel.offset)
#self.agent.states.append(self.agent.step(step, self.info, _features.reverse_action(action)))
state = self.agent.step(
step, self.info,
_features.reverse_action(action))
if state != 0:
npFileNameComp = trainingDataPath + self.replay_file_name + "/" + str(
packageCounter)
np.savez_compressed(
npFileNameComp,
action=translate_outputs_to_NN(
state["action"][0]),
feature_layers=np.moveaxis(
(np.array(state["feature_layers"])), 0,
2))
packageCounter += 1
break
#print("%s: %s" % (len(agent_obs.multi_select), units.Zerg(agent_obs.multi_select[0][0])))
#print(action)
#print(units.Zerg(agent_obs.single_select[0][0]))
#self.agent.step(step, self.info, acts)
#print(_features.reverse_action(obs.actions[0]))
#print ("+")
#print(offset)
#screenpoint = (84, 84)
#screenpoint = point.Point(*screenpoint)
if obs.player_result:
os.remove(replay_file_path)
print("Game Ended, File Removed")
break
self._state = StepType.MID