def play_train(self, continues_attack=False, verbose=False):
is_attack = False
state_last = None
self.safe_action(C._NO_OP, 0, [])
self.safe_action(C._MOVE_CAMERA, 0, [C.base_camera_pos])
self._gases = U.find_initial_gases(self.obs)
while True:
self.safe_action(C._MOVE_CAMERA, 0, [C.base_camera_pos])
if self.policy_flag and (not self.is_end):
non_image_feature = self.mapping_source_to_mini_by_rule(
self.get_the_input())
#print('non_image_feature.shape:', non_image_feature.shape)
#print('non_image_feature:', non_image_feature)
image_feature = U.get_simple_map_data(self.obs)
#print('image_feature.shape:', image_feature.shape)
#print('image_feature:', image_feature)
latent_image_feature, mu, logvar = self.encode_obs(
image_feature)
#print('latent_image_feature.shape:', latent_image_feature.shape)
#print('latent_image_feature:', latent_image_feature)
feature = np.concatenate(
[non_image_feature, latent_image_feature], axis=-1)
#print('feature.shape:', feature.shape)
#print('feature:', feature)
#state_now = feature
reward_last = 0
state_now, action, v_preds = self.get_action(
feature, reward_last)
# print(ProtossAction(action).name)
self.mini_step(action)
if state_last is not None:
if 0:
print('state_last:', state_last, ', action_last:',
action_last, ', state_now:', state_now)
v_preds_next = self.net.policy.get_values(state_now)
v_preds_next = self.get_values(v_preds_next)
reward = 0
self.local_buffer.append(state_last, action_last,
state_now, reward, v_preds,
v_preds_next)
state_last = state_now
action_last = action
self.policy_flag = False
if self.is_end:
if self.rl_training:
self.local_buffer.rewards[-1] += 1 * self.result[
'reward'] # self.result['win']
#print(self.local_buffer.rewards)
self.global_buffer.add(self.local_buffer)
#print("add %d buffer!" % (len(self.local_buffer.rewards)))
break
def sample(self, verbose=False, use_image=False):
is_attack = False
state_last = None
random_generated_int = random.randint(0, 2**31 - 1)
filename = self.extract_save_dir + "/" + str(
random_generated_int) + ".npz"
recording_obs = []
recording_img = []
recording_action = []
np.random.seed(random_generated_int)
tf.set_random_seed(random_generated_int)
self.safe_action(C._NO_OP, 0, [])
self.safe_action(C._MOVE_CAMERA, 0, [C.base_camera_pos])
self._gases = U.find_initial_gases(self.obs)
while True:
self.safe_action(C._MOVE_CAMERA, 0, [C.base_camera_pos])
if self.policy_flag and (not self.is_end):
state_now = self.mapping_source_to_mini_by_rule(
self.get_the_input())
recording_obs.append(state_now)
if use_image:
recording_img.append(U.get_simple_map_data(self.obs))
action, v_preds = self.net.policy.get_action(state_now,
verbose=False)
recording_action.append(action)
self.mini_step(action)
if state_last is not None:
if False:
print('state_last:', state_last, ', action_last:',
action_last, ', state_now:', state_now)
v_preds_next = self.net.policy.get_values(state_now)
v_preds_next = self.get_values(v_preds_next)
reward = 0
self.local_buffer.append(state_last, action_last,
state_now, reward, v_preds,
v_preds_next)
state_last = state_now
action_last = action
self.policy_flag = False
if self.is_end:
if True:
#note this will not consider the minerals larger than 256!
recording_obs = np.array(recording_obs, dtype=np.uint16)
recording_action = np.array(recording_action,
dtype=np.uint8)
if not use_image:
np.savez_compressed(filename,
obs=recording_obs,
action=recording_action)
else:
recording_img = np.array(recording_img,
dtype=np.float16)
np.savez_compressed(filename,
obs=recording_obs,
img=recording_img,
action=recording_action)
break
def sample(self, verbose=False, use_image=True):
is_attack = False
state_last = None
random_generated_int = random.randint(0, 2**31 - 1)
filename = self.extract_save_dir + "/" + str(
random_generated_int) + ".npz"
recording_obs = []
recording_img = []
recording_action = []
recording_reward = []
np.random.seed(random_generated_int)
tf.set_random_seed(random_generated_int)
self.safe_action(C._NO_OP, 0, [])
self.safe_action(C._MOVE_CAMERA, 0, [C.base_camera_pos])
self._gases = U.find_initial_gases(self.obs)
while True:
self.safe_action(C._MOVE_CAMERA, 0, [C.base_camera_pos])
if self.policy_flag and (not self.is_end):
non_image_feature = self.mapping_source_to_mini_by_rule(
self.get_the_input())
#print('non_image_feature.shape:', non_image_feature.shape)
#print('non_image_feature:', non_image_feature)
image_feature = U.get_simple_map_data(self.obs)
#print('image_feature.shape:', image_feature.shape)
#print('image_feature:', image_feature)
latent_image_feature, mu, logvar = self.encode_obs(
image_feature)
#print('latent_image_feature.shape:', latent_image_feature.shape)
#print('latent_image_feature:', latent_image_feature)
feature = np.concatenate(
[non_image_feature, latent_image_feature], axis=-1)
#print('feature.shape:', feature.shape)
#print('feature:', feature)
#state_now = feature
reward_last = 0
state_now, action, v_preds = self.get_action(
feature, reward_last)
# print(ProtossAction(action).name)
self.mini_step(action)
if state_last is not None:
if False:
print('state_last:', state_last, ', action_last:',
action_last, ', state_now:', state_now)
v_preds_next = self.net.policy.get_values(state_now)
v_preds_next = self.get_values(v_preds_next)
reward = 0
recording_obs.append(non_image_feature)
recording_img.append(image_feature)
recording_action.append(action)
recording_reward.append(reward)
#self.local_buffer.append(state_last, action_last, state_now, reward, v_preds, v_preds_next)
state_last = state_now
action_last = action
self.policy_flag = False
if self.is_end:
if True:
# consider the win/loss, to 0(not end), 1(loss), 2(draw), 3(win)
recording_reward[-1] = (1 * self.result['reward'] + 2)
if recording_reward[-1] != 0:
print("result is:", recording_reward[-1])
recording_obs = np.array(recording_obs, dtype=np.uint16)
recording_action = np.array(recording_action,
dtype=np.uint8)
recording_reward = np.array(recording_reward,
dtype=np.uint8)
recording_img = np.array(recording_img, dtype=np.float16)
np.savez_compressed(filename,
obs=recording_obs,
img=recording_img,
action=recording_action,
reward=recording_reward)
break