def measure(name, iters=5000, **settings):
print(name)
for k, v in settings.items():
print("\t{}: {}".format(k, v))
# Vizdoom wrapper
doom_wrapper = VizdoomWrapper(**settings)
start = time()
for _ in trange(iters, leave=False):
current_img, current_misc = doom_wrapper.get_current_state()
action_index = randint(0, doom_wrapper.actions_num - 1)
doom_wrapper.make_action(action_index)
if doom_wrapper.is_terminal():
doom_wrapper.reset()
end = time()
wrapper_t = (end - start)
# Vanilla vizdoom:
doom = vzd.DoomGame()
if "scenarios_path" not in settings:
scenarios_path = vzd.__path__[0] + "/scenarios"
else:
scenarios_path = settings["scenarios_path"]
config_file = scenarios_path + "/" + settings["config_file"]
doom.load_config(config_file)
doom.set_window_visible(False)
doom.set_screen_format(vzd.ScreenFormat.GRAY8)
doom.set_screen_resolution(vzd.ScreenResolution.RES_160X120)
doom.init()
actions = [
list(a)
for a in it.product([0, 1],
repeat=len(doom.get_available_game_variables()))
]
start = time()
frame_skip = settings["frame_skip"]
for _ in trange(iters, leave=False):
if doom.is_episode_finished():
doom.new_episode()
doom.make_action(choice(actions), frame_skip)
end = time()
vanilla_t = end - start
print(green("\twrapper: {:0.2f} steps/s".format(iters / wrapper_t)))
print(
green("\twrapper: {:0.2f} s/1000 steps".format(wrapper_t / iters *
1000)))
print(blue("\tvanilla: {:0.2f} steps/s".format(iters / vanilla_t)))
print(
blue("\tvanilla: {:0.2f} s/1000 steps\n".format(vanilla_t / iters *
1000)))
def _print_train_log(self, scores, overall_start_time, last_log_time, steps):
current_time = time.time()
mean_score = np.mean(scores)
score_std = np.std(scores)
min_score = np.min(scores)
max_score = np.max(scores)
elapsed_time = time.time() - overall_start_time
global_steps = self._global_steps_counter.get()
local_steps_per_sec = steps / (current_time - last_log_time)
global_steps_per_sec = global_steps / elapsed_time
global_mil_steps_per_hour = global_steps_per_sec * 3600 / 1000000.0
log(
"TRAIN: {}(GlobalSteps), {} episodes, mean: {}, min: {}, max: {}, "
"\nLocalSpd: {:.0f} STEPS/s GlobalSpd: "
"{} STEPS/s, {:.2f}M STEPS/hour, total elapsed time: {}".format(
global_steps,
len(scores),
green("{:0.3f}±{:0.2f}".format(mean_score, score_std)),
red("{:0.3f}".format(min_score)),
blue("{:0.3f}".format(max_score)),
local_steps_per_sec,
blue("{:.0f}".format(
global_steps_per_sec)),
global_mil_steps_per_hour,
sec_to_str(elapsed_time)
))
def test(self, episodes_num=None, deterministic=True):
if episodes_num is None:
episodes_num = self.test_episodes_per_epoch
test_start_time = time.time()
test_rewards = []
test_actions = []
test_frameskips = []
for _ in trange(episodes_num, desc="Testing", file=sys.stdout,
leave=False, disable=not self.enable_progress_bar):
total_reward, actions, frameskips, _ = self.run_episode(deterministic=deterministic, return_stats=True)
test_rewards.append(total_reward)
test_actions += actions
test_frameskips += frameskips
self.doom_wrapper.reset()
if self.local_network.has_state():
self.local_network.reset_state()
test_end_time = time.time()
test_duration = test_end_time - test_start_time
min_score = np.min(test_rewards)
max_score = np.max(test_rewards)
mean_score = np.mean(test_rewards)
score_std = np.std(test_rewards)
log(
"TEST: mean: {}, min: {}, max: {}, test time: {}".format(
green("{:0.3f}±{:0.2f}".format(mean_score, score_std)),
red("{:0.3f}".format(min_score)),
blue("{:0.3f}".format(max_score)),
sec_to_str(test_duration)))
return test_rewards, test_actions, test_frameskips
def print_epoch_log(prefix, scores, steps, epoch_time):
mean_score = np.mean(scores)
score_std = np.std(scores)
min_score = np.min(scores)
max_score = np.max(scores)
episodes = len(scores)
steps_per_sec = steps / epoch_time
mil_steps_per_hour = steps_per_sec * 3600 / 1000000.0
log("{}: Episodes: {}, mean: {}, min: {}, max: {}, "
" Speed: {:.0f} STEPS/s, {:.2f}M STEPS/hour, time: {}".format(
prefix, episodes,
green("{:0.3f}±{:0.2f}".format(mean_score, score_std)),
red("{:0.3f}".format(min_score)),
blue("{:0.3f}".format(max_score)), steps_per_sec,
mil_steps_per_hour, sec_to_str(epoch_time)))
def train_async(q_learning, settings):
proto_vizdoom = VizdoomWrapper(noinit=True, **settings)
actions_num = proto_vizdoom.actions_num
misc_len = proto_vizdoom.misc_len
img_shape = proto_vizdoom.img_shape
del proto_vizdoom
# TODO target global network
# This global step counts gradient applications not performed actions.
global_train_step = tf.Variable(0, trainable=False, name="global_step")
global_learning_rate = tf.train.polynomial_decay(
name="larning_rate",
learning_rate=settings["initial_learning_rate"],
end_learning_rate=settings["final_learning_rate"],
decay_steps=settings["learning_rate_decay_steps"],
global_step=global_train_step)
optimizer = ClippingRMSPropOptimizer(learning_rate=global_learning_rate,
**settings["rmsprop"])
learners = []
network_class = eval(settings["network_type"])
global_network = network_class(actions_num=actions_num,
misc_len=misc_len,
img_shape=img_shape,
**settings)
global_steps_counter = ThreadsafeCounter()
if q_learning:
global_target_network = network_class(thread="global_target",
actions_num=actions_num,
misc_len=misc_len,
img_shape=img_shape,
**settings)
global_network.prepare_unfreeze_op(global_target_network)
unfreeze_thread = min(1, settings["threads_num"] - 1)
for i in range(settings["threads_num"]):
learner = ADQNLearner(thread_index=i,
global_network=global_network,
unfreeze_thread=i == unfreeze_thread,
global_target_network=global_target_network,
optimizer=optimizer,
learning_rate=global_learning_rate,
global_steps_counter=global_steps_counter,
**settings)
learners.append(learner)
else:
for i in range(settings["threads_num"]):
learner = A3CLearner(thread_index=i,
global_network=global_network,
optimizer=optimizer,
learning_rate=global_learning_rate,
global_steps_counter=global_steps_counter,
**settings)
learners.append(learner)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
session = tf.Session(config=config)
log("Initializing variables...")
session.run(tf.global_variables_initializer())
log("Initialization finished.\n")
if q_learning:
session.run(global_network.ops.unfreeze)
log(green("Starting training.\n"))
for l in learners:
l.run_training(session)
for l in learners:
l.join()
def train(self, session):
# Prefill replay memory:
for _ in trange(self.replay_memory.capacity, desc="Filling replay memory",
leave=False, disable=not self.enable_progress_bar, file=sys.stdout):
if self.doom_wrapper.is_terminal():
self.doom_wrapper.reset()
s1 = self.doom_wrapper.get_current_state()
action_frameskip_index = randint(0, self.actions_num * len(self.frameskips) - 1)
action_index, frameskip = self.get_action_and_frameskip(action_frameskip_index)
reward = self.doom_wrapper.make_action(action_index, frameskip)
terminal = self.doom_wrapper.is_terminal()
s2 = self.doom_wrapper.get_current_state()
self.replay_memory.add_transition(s1, action_frameskip_index, s2, reward, terminal)
overall_start_time = time()
self.network.update_target_network(session)
log(green("Starting training.\n"))
while self._epoch <= self._epochs:
self.doom_wrapper.reset()
train_scores = []
test_scores = []
train_start_time = time()
for _ in trange(self.train_steps_per_epoch, desc="Training, epoch {}".format(self._epoch),
leave=False, disable=not self.enable_progress_bar, file=sys.stdout):
self.steps += 1
s1 = self.doom_wrapper.get_current_state()
if random() <= self.get_current_epsilon():
action_frameskip_index = randint(0, self.actions_num*len(self.frameskips) - 1)
action_index, frameskip = self.get_action_and_frameskip(action_frameskip_index)
else:
action_frameskip_index = self.network.get_action(session, s1)
action_index, frameskip = self.get_action_and_frameskip(action_frameskip_index)
reward = self.doom_wrapper.make_action(action_index, frameskip)
terminal = self.doom_wrapper.is_terminal()
s2 = self.doom_wrapper.get_current_state()
self.replay_memory.add_transition(s1, action_frameskip_index, s2, reward, terminal)
if self.steps % self.update_pattern[0] == 0:
for _ in range(self.update_pattern[1]):
self.network.train_batch(session, self.replay_memory.get_sample())
if terminal:
train_scores.append(self.doom_wrapper.get_total_reward())
self.doom_wrapper.reset()
if self.steps % self.frozen_steps == 0:
self.network.update_target_network(session)
train_time = time() - train_start_time
log("Epoch {}".format(self._epoch))
log("Training steps: {}, epsilon: {}".format(self.steps, self.get_current_epsilon()))
self.print_epoch_log("TRAIN", train_scores, self.train_steps_per_epoch, train_time)
test_start_time = time()
test_steps = 0
# TESTING
for _ in trange(self.test_episodes_per_epoch, desc="Testing, epoch {}".format(self._epoch),
leave=False, disable=not self.enable_progress_bar, file=sys.stdout):
self.doom_wrapper.reset()
while not self.doom_wrapper.is_terminal():
test_steps += 1
state = self.doom_wrapper.get_current_state()
action_frameskip_index = self.network.get_action(session, state)
action_index, frameskip = self.get_action_and_frameskip(action_frameskip_index)
self.doom_wrapper.make_action(action_index, frameskip)
test_scores.append(self.doom_wrapper.get_total_reward())
test_time = time() - test_start_time
self.print_epoch_log("TEST", test_scores, test_steps, test_time)
if self.write_summaries:
log("Writing summaries.")
train_summary = session.run(self._summaries, {self.scores_placeholder: train_scores})
self._train_writer.add_summary(train_summary, self.steps)
if self._run_tests:
test_summary = session.run(self._summaries, {self.scores_placeholder: test_scores})
self._test_writer.add_summary(test_summary, self.steps)
# Save model
if self._epoch % self.save_interval == 0:
savedir = os.path.dirname(self._model_savefile)
if not os.path.exists(savedir):
log("Creating directory: {}".format(savedir))
os.makedirs(savedir)
log("Saving model to: {}".format(self._model_savefile))
saver = tf.train.Saver()
saver.save(session, self._model_savefile)
overall_time = time() - overall_start_time
log("Total elapsed time: {}\n".format(sec_to_str(overall_time)))
self._epoch += 1