def get_demo(args):
if args.demo_memory_folder is not None:
demo_memory_folder = 'collected_demo/{}'.format(
args.demo_memory_folder)
else:
demo_memory_folder = 'collected_demo/{}'.format(
args.gym_env.replace('-', '_'))
if args.append_experiment_num is not None:
demo_memory_folder += '_' + args.append_experiment_num
if args.hostname is None:
hostname = os.uname()[1]
else:
hostname = args.hostname
prepare_dir(demo_memory_folder + '/log/{}'.format(hostname), empty=False)
prepare_dir(demo_memory_folder + '/data/{}'.format(hostname), empty=False)
episode_life = not args.not_episodic_life
datetime_collected = datetime.today().strftime('%Y%m%d_%H%M%S')
log_file = '{}.log'.format(datetime_collected)
fh = logging.FileHandler('{}/log/{}/{}'.format(demo_memory_folder,
hostname, log_file),
mode='w')
fh.setLevel(logging.DEBUG)
formatter = LogFormatter(
'%(asctime)s - %(name)s - %(levelname)s - %(message)s')
fh.setFormatter(formatter)
logger.addHandler(fh)
logging.getLogger('collect_demo').addHandler(fh)
logging.getLogger('game_state').addHandler(fh)
logging.getLogger('replay_memory').addHandler(fh)
logging.getLogger('atari_wrapper').addHandler(fh)
game_state = GameState(env_id=args.gym_env,
display=True,
human_demo=True,
episode_life=episode_life)
collect_demo = CollectDemonstration(game_state,
84,
84,
4,
args.gym_env,
folder=demo_memory_folder,
create_movie=args.create_movie,
hertz=args.hz,
skip=args.skip)
collect_demo.run_episodes(args.num_episodes,
minutes_limit=args.demo_time_limit,
demo_type=0,
log_file=log_file,
hostname=hostname)
game_state.close()
def test_collect(env_id):
from common.game_state import GameState
game_state = GameState(env_id=env_id, display=True, human_demo=True)
test_folder = "demo_samples/{}_test".format(env_id.replace('-', '_'))
prepare_dir(test_folder, empty=True)
collect_demo = CollectDemonstration(
game_state,
84, 84, 4,
env_id,
folder=test_folder, create_movie=True)
num_episodes = 1
collect_demo.run_episodes(
num_episodes,
minutes_limit=3,
demo_type=0)
def test_function():
nonlocal global_t
if args.use_transfer:
from_folder = transfer_folder.split('/')[-2]
else:
from_folder = folder.split('/')[-1]
save_folder = 'results/test_model/a3c/{}'.format(from_folder)
prepare_dir(save_folder, empty=False)
prepare_dir(save_folder + '/frames', empty=False)
# Evaluate model before training
if not stop_requested:
testing_thread.testing_model(
sess, args.eval_max_steps, global_t, save_folder,
demo_memory_cam=demo_memory_cam, demo_cam_human=demo_cam_human)
def test_function():
nonlocal global_t
if args.use_transfer:
from_folder = str(transfer_folder).split('/')[-2]
else:
from_folder = str(folder).split('/')[-1]
from_folder = pathlib.Path(from_folder)
save_folder = 'results/test_model/a3c' / from_folder
prepare_dir(str(save_folder), empty=False)
prepare_dir(str(save_folder / 'frames'), empty=False)
# Evaluate model before training
if not stop_requested:
testing_thread.testing_model(
sess, args.eval_max_steps, global_t, save_folder,
demo_memory_cam=demo_memory_cam, demo_cam_human=demo_cam_human)
def run_a3c(args):
"""
python3 run_experiment.py --gym-env=PongNoFrameskip-v4 --parallel-size=16 --initial-learn-rate=7e-4 --use-lstm --use-mnih-2015
python3 run_experiment.py --gym-env=PongNoFrameskip-v4 --parallel-size=16 --initial-learn-rate=7e-4 --use-lstm --use-mnih-2015 --use-transfer --not-transfer-fc2 --transfer-folder=<>
python3 run_experiment.py --gym-env=PongNoFrameskip-v4 --parallel-size=16 --initial-learn-rate=7e-4 --use-lstm --use-mnih-2015 --use-transfer --not-transfer-fc2 --transfer-folder=<> --load-pretrained-model --onevsall-mtl --pretrained-model-folder=<> --use-pretrained-model-as-advice --use-pretrained-model-as-reward-shaping
"""
from game_ac_network import GameACFFNetwork, GameACLSTMNetwork
from a3c_training_thread import A3CTrainingThread
if args.use_gpu:
assert args.cuda_devices != ''
os.environ['CUDA_VISIBLE_DEVICES'] = args.cuda_devices
else:
os.environ['CUDA_VISIBLE_DEVICES'] = ''
import tensorflow as tf
def log_uniform(lo, hi, rate):
log_lo = math.log(lo)
log_hi = math.log(hi)
v = log_lo * (1 - rate) + log_hi * rate
return math.exp(v)
if not os.path.exists('results/a3c'):
os.makedirs('results/a3c')
if args.folder is not None:
folder = 'results/a3c/{}_{}'.format(args.gym_env.replace('-', '_'),
args.folder)
else:
folder = 'results/a3c/{}'.format(args.gym_env.replace('-', '_'))
end_str = ''
if args.use_mnih_2015:
end_str += '_mnih2015'
if args.use_lstm:
end_str += '_lstm'
if args.unclipped_reward:
end_str += '_rawreward'
elif args.log_scale_reward:
end_str += '_logreward'
if args.transformed_bellman:
end_str += '_transformedbell'
if args.use_transfer:
end_str += '_transfer'
if args.not_transfer_conv2:
end_str += '_noconv2'
elif args.not_transfer_conv3 and args.use_mnih_2015:
end_str += '_noconv3'
elif args.not_transfer_fc1:
end_str += '_nofc1'
elif args.not_transfer_fc2:
end_str += '_nofc2'
if args.finetune_upper_layers_only:
end_str += '_tune_upperlayers'
if args.train_with_demo_num_steps > 0 or args.train_with_demo_num_epochs > 0:
end_str += '_pretrain_ina3c'
if args.use_demo_threads:
end_str += '_demothreads'
if args.load_pretrained_model:
if args.use_pretrained_model_as_advice:
end_str += '_modelasadvice'
if args.use_pretrained_model_as_reward_shaping:
end_str += '_modelasshaping'
folder += end_str
if args.append_experiment_num is not None:
folder += '_' + args.append_experiment_num
if False:
from common.util import LogFormatter
fh = logging.FileHandler('{}/a3c.log'.format(folder), mode='w')
fh.setLevel(logging.DEBUG)
formatter = LogFormatter(
'%(asctime)s - %(name)s - %(levelname)s - %(message)s')
fh.setFormatter(formatter)
logger.addHandler(fh)
demo_memory = None
num_demos = 0
max_reward = 0.
if args.load_memory or args.load_demo_cam:
if args.demo_memory_folder is not None:
demo_memory_folder = args.demo_memory_folder
else:
demo_memory_folder = 'collected_demo/{}'.format(
args.gym_env.replace('-', '_'))
if args.load_memory:
# FIXME: use new load_memory function
demo_memory, actions_ctr, max_reward = load_memory(
args.gym_env, demo_memory_folder,
imgs_normalized=True) #, create_symmetry=True)
action_freq = [
actions_ctr[a] for a in range(demo_memory[0].num_actions)
]
num_demos = len(demo_memory)
demo_memory_cam = None
if args.load_demo_cam:
demo_cam, _, total_rewards_cam, _ = load_memory(
name=None,
demo_memory_folder=demo_memory_folder,
demo_ids=args.demo_cam_id,
imgs_normalized=False)
demo_cam = demo_cam[int(args.demo_cam_id)]
demo_memory_cam = np.zeros((len(demo_cam), demo_cam.height,
demo_cam.width, demo_cam.phi_length),
dtype=np.float32)
for i in range(len(demo_cam)):
s0 = (demo_cam[i])[0]
demo_memory_cam[i] = np.copy(s0)
del demo_cam
logger.info("loaded demo {} for testing CAM".format(args.demo_cam_id))
device = "/cpu:0"
gpu_options = None
if args.use_gpu:
device = "/gpu:" + os.environ["CUDA_VISIBLE_DEVICES"]
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=args.gpu_fraction)
initial_learning_rate = args.initial_learn_rate
logger.info('Initial Learning Rate={}'.format(initial_learning_rate))
time.sleep(2)
global_t = 0
pretrain_global_t = 0
pretrain_epoch = 0
rewards = {'train': {}, 'eval': {}}
best_model_reward = -(sys.maxsize)
stop_requested = False
game_state = GameState(env_id=args.gym_env)
action_size = game_state.env.action_space.n
game_state.close()
del game_state.env
del game_state
config = tf.ConfigProto(gpu_options=gpu_options,
log_device_placement=False,
allow_soft_placement=True)
pretrained_model = None
pretrained_model_sess = None
if args.load_pretrained_model:
if args.onevsall_mtl:
from game_class_network import MTLBinaryClassNetwork as PretrainedModelNetwork
elif args.onevsall_mtl_linear:
from game_class_network import MTLMultivariateNetwork as PretrainedModelNetwork
else:
from game_class_network import MultiClassNetwork as PretrainedModelNetwork
logger.error("Not supported yet!")
assert False
if args.pretrained_model_folder is not None:
pretrained_model_folder = args.pretrained_model_folder
else:
pretrained_model_folder = '{}_classifier_use_mnih_onevsall_mtl'.format(
args.gym_env.replace('-', '_'))
PretrainedModelNetwork.use_mnih_2015 = args.use_mnih_2015
pretrained_model = PretrainedModelNetwork(action_size, -1, device)
pretrained_model_sess = tf.Session(config=config,
graph=pretrained_model.graph)
pretrained_model.load(
pretrained_model_sess,
'{}/{}_checkpoint'.format(pretrained_model_folder,
args.gym_env.replace('-', '_')))
if args.use_lstm:
GameACLSTMNetwork.use_mnih_2015 = args.use_mnih_2015
global_network = GameACLSTMNetwork(action_size, -1, device)
else:
GameACFFNetwork.use_mnih_2015 = args.use_mnih_2015
global_network = GameACFFNetwork(action_size, -1, device)
training_threads = []
learning_rate_input = tf.placeholder(tf.float32, shape=(), name="opt_lr")
grad_applier = tf.train.RMSPropOptimizer(learning_rate=learning_rate_input,
decay=args.rmsp_alpha,
epsilon=args.rmsp_epsilon)
A3CTrainingThread.log_interval = args.log_interval
A3CTrainingThread.performance_log_interval = args.performance_log_interval
A3CTrainingThread.local_t_max = args.local_t_max
A3CTrainingThread.demo_t_max = args.demo_t_max
A3CTrainingThread.use_lstm = args.use_lstm
A3CTrainingThread.action_size = action_size
A3CTrainingThread.entropy_beta = args.entropy_beta
A3CTrainingThread.demo_entropy_beta = args.demo_entropy_beta
A3CTrainingThread.gamma = args.gamma
A3CTrainingThread.use_mnih_2015 = args.use_mnih_2015
A3CTrainingThread.env_id = args.gym_env
A3CTrainingThread.finetune_upper_layers_only = args.finetune_upper_layers_only
A3CTrainingThread.transformed_bellman = args.transformed_bellman
A3CTrainingThread.clip_norm = args.grad_norm_clip
A3CTrainingThread.use_grad_cam = args.use_grad_cam
if args.unclipped_reward:
A3CTrainingThread.reward_type = "RAW"
elif args.log_scale_reward:
A3CTrainingThread.reward_type = "LOG"
else:
A3CTrainingThread.reward_type = "CLIP"
n_shapers = args.parallel_size #int(args.parallel_size * .25)
mod = args.parallel_size // n_shapers
for i in range(args.parallel_size):
is_reward_shape = False
is_advice = False
if i % mod == 0:
is_reward_shape = args.use_pretrained_model_as_reward_shaping
is_advice = args.use_pretrained_model_as_advice
training_thread = A3CTrainingThread(
i,
global_network,
initial_learning_rate,
learning_rate_input,
grad_applier,
args.max_time_step,
device=device,
pretrained_model=pretrained_model,
pretrained_model_sess=pretrained_model_sess,
advice=is_advice,
reward_shaping=is_reward_shape)
training_threads.append(training_thread)
# prepare session
sess = tf.Session(config=config)
if args.use_transfer:
if args.transfer_folder is not None:
transfer_folder = args.transfer_folder
else:
transfer_folder = 'results/pretrain_models/{}'.format(
args.gym_env.replace('-', '_'))
end_str = ''
if args.use_mnih_2015:
end_str += '_mnih2015'
end_str += '_l2beta1E-04_batchprop' #TODO: make this an argument
transfer_folder += end_str
transfer_folder += '/transfer_model'
if args.not_transfer_conv2:
transfer_var_list = [
global_network.W_conv1, global_network.b_conv1
]
elif (args.not_transfer_conv3 and args.use_mnih_2015):
transfer_var_list = [
global_network.W_conv1, global_network.b_conv1,
global_network.W_conv2, global_network.b_conv2
]
elif args.not_transfer_fc1:
transfer_var_list = [
global_network.W_conv1,
global_network.b_conv1,
global_network.W_conv2,
global_network.b_conv2,
]
if args.use_mnih_2015:
transfer_var_list += [
global_network.W_conv3, global_network.b_conv3
]
elif args.not_transfer_fc2:
transfer_var_list = [
global_network.W_conv1, global_network.b_conv1,
global_network.W_conv2, global_network.b_conv2,
global_network.W_fc1, global_network.b_fc1
]
if args.use_mnih_2015:
transfer_var_list += [
global_network.W_conv3, global_network.b_conv3
]
else:
transfer_var_list = [
global_network.W_conv1, global_network.b_conv1,
global_network.W_conv2, global_network.b_conv2,
global_network.W_fc1, global_network.b_fc1,
global_network.W_fc2, global_network.b_fc2
]
if args.use_mnih_2015:
transfer_var_list += [
global_network.W_conv3, global_network.b_conv3
]
global_network.load_transfer_model(
sess,
folder=transfer_folder,
not_transfer_fc2=args.not_transfer_fc2,
not_transfer_fc1=args.not_transfer_fc1,
not_transfer_conv3=(args.not_transfer_conv3
and args.use_mnih_2015),
not_transfer_conv2=args.not_transfer_conv2,
var_list=transfer_var_list)
def initialize_uninitialized(sess):
global_vars = tf.global_variables()
is_not_initialized = sess.run(
[tf.is_variable_initialized(var) for var in global_vars])
not_initialized_vars = [
v for (v, f) in zip(global_vars, is_not_initialized) if not f
]
if len(not_initialized_vars):
sess.run(tf.variables_initializer(not_initialized_vars))
if args.use_transfer:
initialize_uninitialized(sess)
else:
sess.run(tf.global_variables_initializer())
# summary writer for tensorboard
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(
'results/log/a3c/{}/'.format(args.gym_env.replace('-', '_')) +
folder[12:], sess.graph)
# init or load checkpoint with saver
root_saver = tf.train.Saver(max_to_keep=1)
saver = tf.train.Saver(max_to_keep=6)
best_saver = tf.train.Saver(max_to_keep=1)
checkpoint = tf.train.get_checkpoint_state(folder)
if checkpoint and checkpoint.model_checkpoint_path:
root_saver.restore(sess, checkpoint.model_checkpoint_path)
logger.info("checkpoint loaded:{}".format(
checkpoint.model_checkpoint_path))
tokens = checkpoint.model_checkpoint_path.split("-")
# set global step
global_t = int(tokens[-1])
logger.info(">>> global step set: {}".format(global_t))
# set wall time
wall_t_fname = folder + '/' + 'wall_t.' + str(global_t)
with open(wall_t_fname, 'r') as f:
wall_t = float(f.read())
with open(folder + '/pretrain_global_t', 'r') as f:
pretrain_global_t = int(f.read())
with open(folder + '/model_best/best_model_reward',
'r') as f_best_model_reward:
best_model_reward = float(f_best_model_reward.read())
rewards = pickle.load(
open(
folder + '/' + args.gym_env.replace('-', '_') +
'-a3c-rewards.pkl', 'rb'))
else:
logger.warning("Could not find old checkpoint")
# set wall time
wall_t = 0.0
prepare_dir(folder, empty=True)
prepare_dir(folder + '/model_checkpoints', empty=True)
prepare_dir(folder + '/model_best', empty=True)
prepare_dir(folder + '/frames', empty=True)
lock = threading.Lock()
test_lock = False
if global_t == 0:
test_lock = True
last_temp_global_t = global_t
ispretrain_markers = [False] * args.parallel_size
num_demo_thread = 0
ctr_demo_thread = 0
def train_function(parallel_index):
nonlocal global_t, pretrain_global_t, pretrain_epoch, \
rewards, test_lock, lock, \
last_temp_global_t, ispretrain_markers, num_demo_thread, \
ctr_demo_thread
training_thread = training_threads[parallel_index]
training_thread.set_summary_writer(summary_writer)
# set all threads as demo threads
training_thread.is_demo_thread = args.load_memory and args.use_demo_threads
if training_thread.is_demo_thread or args.train_with_demo_num_steps > 0 or args.train_with_demo_num_epochs:
training_thread.pretrain_init(demo_memory)
if global_t == 0 and (
args.train_with_demo_num_steps > 0
or args.train_with_demo_num_epochs > 0) and parallel_index < 2:
ispretrain_markers[parallel_index] = True
training_thread.replay_mem_reset()
# Pretraining with demo memory
logger.info("t_idx={} pretrain starting".format(parallel_index))
while ispretrain_markers[parallel_index]:
if stop_requested:
return
if pretrain_global_t > args.train_with_demo_num_steps and pretrain_epoch > args.train_with_demo_num_epochs:
# At end of pretraining, reset state
training_thread.replay_mem_reset()
training_thread.episode_reward = 0
training_thread.local_t = 0
if args.use_lstm:
training_thread.local_network.reset_state()
ispretrain_markers[parallel_index] = False
logger.info(
"t_idx={} pretrain ended".format(parallel_index))
break
diff_pretrain_global_t, _ = training_thread.demo_process(
sess, pretrain_global_t)
for _ in range(diff_pretrain_global_t):
pretrain_global_t += 1
if pretrain_global_t % 10000 == 0:
logger.debug(
"pretrain_global_t={}".format(pretrain_global_t))
pretrain_epoch += 1
if pretrain_epoch % 1000 == 0:
logger.debug("pretrain_epoch={}".format(pretrain_epoch))
# Waits for all threads to finish pretraining
while not stop_requested and any(ispretrain_markers):
time.sleep(0.01)
# Evaluate model before training
if not stop_requested and global_t == 0:
with lock:
if parallel_index == 0:
test_reward, test_steps, test_episodes = training_threads[
0].testing(sess,
args.eval_max_steps,
global_t,
folder,
demo_memory_cam=demo_memory_cam)
rewards['eval'][global_t] = (test_reward, test_steps,
test_episodes)
saver.save(
sess,
folder + '/model_checkpoints/' +
'{}_checkpoint'.format(args.gym_env.replace('-', '_')),
global_step=global_t)
save_best_model(test_reward)
test_lock = False
# all threads wait until evaluation finishes
while not stop_requested and test_lock:
time.sleep(0.01)
# set start_time
start_time = time.time() - wall_t
training_thread.set_start_time(start_time)
episode_end = True
use_demo_thread = False
while True:
if stop_requested:
return
if global_t >= (args.max_time_step * args.max_time_step_fraction):
return
if args.use_demo_threads and global_t < args.max_steps_threads_as_demo and episode_end and num_demo_thread < 16:
#if num_demo_thread < 2:
demo_rate = 1.0 * (args.max_steps_threads_as_demo -
global_t) / args.max_steps_threads_as_demo
if demo_rate < 0.0333:
demo_rate = 0.0333
if np.random.random() <= demo_rate and num_demo_thread < 16:
ctr_demo_thread += 1
training_thread.replay_mem_reset(D_idx=ctr_demo_thread %
num_demos)
num_demo_thread += 1
logger.info(
"idx={} as demo thread started ({}/16) rate={}".format(
parallel_index, num_demo_thread, demo_rate))
use_demo_thread = True
if use_demo_thread:
diff_global_t, episode_end = training_thread.demo_process(
sess, global_t)
if episode_end:
num_demo_thread -= 1
use_demo_thread = False
logger.info("idx={} demo thread concluded ({}/16)".format(
parallel_index, num_demo_thread))
else:
diff_global_t, episode_end = training_thread.process(
sess, global_t, rewards)
for _ in range(diff_global_t):
global_t += 1
if global_t % args.eval_freq == 0:
temp_global_t = global_t
lock.acquire()
try:
# catch multiple threads getting in at the same time
if last_temp_global_t == temp_global_t:
logger.info("Threading race problem averted!")
continue
test_lock = True
test_reward, test_steps, n_episodes = training_thread.testing(
sess,
args.eval_max_steps,
temp_global_t,
folder,
demo_memory_cam=demo_memory_cam)
rewards['eval'][temp_global_t] = (test_reward,
test_steps,
n_episodes)
if temp_global_t % (
(args.max_time_step * args.max_time_step_fraction)
// 5) == 0:
saver.save(sess,
folder + '/model_checkpoints/' +
'{}_checkpoint'.format(
args.gym_env.replace('-', '_')),
global_step=temp_global_t,
write_meta_graph=False)
if test_reward > best_model_reward:
save_best_model(test_reward)
test_lock = False
last_temp_global_t = temp_global_t
finally:
lock.release()
if global_t % (
(args.max_time_step * args.max_time_step_fraction) //
5) == 0:
saver.save(
sess,
folder + '/model_checkpoints/' +
'{}_checkpoint'.format(args.gym_env.replace('-', '_')),
global_step=global_t,
write_meta_graph=False)
# all threads wait until evaluation finishes
while not stop_requested and test_lock:
time.sleep(0.01)
def signal_handler(signal, frame):
nonlocal stop_requested
logger.info('You pressed Ctrl+C!')
stop_requested = True
if stop_requested and global_t == 0:
sys.exit(1)
def save_best_model(test_reward):
nonlocal best_model_reward
best_model_reward = test_reward
with open(folder + '/model_best/best_model_reward',
'w') as f_best_model_reward:
f_best_model_reward.write(str(best_model_reward))
best_saver.save(
sess, folder + '/model_best/' +
'{}_checkpoint'.format(args.gym_env.replace('-', '_')))
train_threads = []
for i in range(args.parallel_size):
train_threads.append(
threading.Thread(target=train_function, args=(i, )))
signal.signal(signal.SIGINT, signal_handler)
signal.signal(signal.SIGTERM, signal_handler)
# set start time
start_time = time.time() - wall_t
for t in train_threads:
t.start()
print('Press Ctrl+C to stop')
for t in train_threads:
t.join()
logger.info('Now saving data. Please wait')
# write wall time
wall_t = time.time() - start_time
wall_t_fname = folder + '/' + 'wall_t.' + str(global_t)
with open(wall_t_fname, 'w') as f:
f.write(str(wall_t))
with open(folder + '/pretrain_global_t', 'w') as f:
f.write(str(pretrain_global_t))
root_saver.save(
sess,
folder + '/{}_checkpoint_a3c'.format(args.gym_env.replace('-', '_')),
global_step=global_t)
pickle.dump(
rewards,
open(
folder + '/' + args.gym_env.replace('-', '_') + '-a3c-rewards.pkl',
'wb'), pickle.HIGHEST_PROTOCOL)
logger.info('Data saved!')
sess.close()
def strip_ansi_codes(s):
return re.sub(r'\x1b\[([0-9,A-Z]{1,2}(;[0-9]{1,2})?(;[0-9]{3})?)?[m|K]?',
'', s)
for log_num in [1]:
rl_alg = "a3c"
atari_game = "Breakout"
path = "results_kamiak/slurm_log"
log_file = "{}_{}_base_{}.log".format(rl_alg, atari_game, log_num)
folder = "results_kamiak/{}/{}NoFrameskip_v4_mnih2015_lstm_{}".format(
rl_alg, atari_game, log_num)
rewards_file = "{}NoFrameskip_v4-{}-rewards.pkl".format(atari_game, rl_alg)
prepare_dir(folder, empty=False)
rewards = {'train': {}, 'eval': {}}
with open(path + '/' + log_file, 'r') as fp:
for line in fp:
if re.search("final score", line):
line = line.rstrip()
line = line.split(" ")
global_t = int(line[5].split("=")[-1])
ave_score = float(line[8].split("=")[-1])
ave_steps = int(line[10].split("=")[-1])
num_episodes = int(line[12].split("=")[-1])
logger.debug(
"test: global_t={0:9d} score={1:.2f} steps={2} episodes={3}"
.format(global_t, ave_score, ave_steps, num_episodes))
rewards['eval'][global_t] = (ave_score, ave_steps,
def run_episodes(self, num_episodes, minutes_limit=5, demo_type=0,
model_net=None, log_file='', hostname=None):
assert hostname is not None
rewards = []
steps = []
durations = []
mem_sizes = []
episode = 0
if minutes_limit < 0:
minutes_limit = 0
while True:
replay_memory = ReplayMemory(
self.resized_w, self.resized_h,
np.random.RandomState(),
max_steps=100000,
phi_length=self.phi_length,
num_actions=self.game_state.env.action_space.n,
wrap_memory=False,
full_state_size=self.game_state.clone_full_state().shape[0])
datetime_collected = datetime.datetime.today().strftime(
'%Y%m%d_%H%M%S')
self.folder = self.main_folder / 'data'
self.folder /= hostname
self.folder /= datetime_collected
prepare_dir(self.folder, empty=True)
results = self.run(
minutes_limit=minutes_limit, episode=episode,
num_episodes=num_episodes, demo_type=demo_type,
model_net=model_net, replay_memory=replay_memory,
total_memory=np.sum(mem_sizes))
total_reward, total_steps, start_time = results[:3]
end_time, duration, mem_size = results[3:]
rewards.append(total_reward)
steps.append(total_steps)
durations.append(duration)
mem_sizes.append(mem_size)
del replay_memory
self.insert_data_to_db([(
datetime_collected, self.name,
1 if self.game_state.episode_life else 0, self._skip,
total_reward, mem_size, start_time, end_time,
str(duration), str(datetime.time(minute=minutes_limit)),
total_steps, log_file, hostname, self.hertz)])
episode += 1
if episode >= num_episodes:
break
logger.debug("steps / episode: {}".format(steps))
logger.debug("reward / episode: {}".format(rewards))
logger.debug("mean steps: {} / mean reward: {}".format(
np.mean(steps), np.mean(rewards)))
total_duration = durations[0]
logger.debug("duration / episode:")
logger.debug(" {}".format(durations[0]))
for j in range(1, len(durations)):
logger.debug(" {}".format(durations[j]))
total_duration += durations[j]
logger.debug("total duration: {}".format(total_duration))
logger.debug("mem size / episode: {}".format(mem_sizes))
logger.debug("total memory size: {}".format(np.sum(mem_sizes)))
logger.debug("total # of episodes: {}".format(num_episodes))
self.game_state.close()
self.conn.close()
def run_a3c(args):
"""Run A3C experiment."""
GYM_ENV_NAME = args.gym_env.replace('-', '_')
GAME_NAME = args.gym_env.replace('NoFrameskip-v4','')
# setup folder name and path to folder
folder = pathlib.Path(setup_folder(args, GYM_ENV_NAME))
# setup GPU (if applicable)
import tensorflow as tf
gpu_options = setup_gpu(tf, args.use_gpu, args.gpu_fraction)
######################################################
# setup default device
device = "/cpu:0"
global_t = 0
rewards = {'train': {}, 'eval': {}}
best_model_reward = -(sys.maxsize)
if args.load_pretrained_model:
class_rewards = {'class_eval': {}}
# setup logging info for analysis, see Section 4.2 of the paper
sil_dict = {
# count number of SIL updates
"sil_ctr":{},
# total number of butter D sampled during SIL
"sil_a3c_sampled":{},
# total number of buffer D samples (i.e., generated by A3C workers) used during SIL (i.e., passed max op)
"sil_a3c_used":{},
# the return of used samples for buffer D
"sil_a3c_used_return":{},
# total number of buffer R sampled during SIL
"sil_rollout_sampled":{},
# total number of buffer R samples (i.e., generated by refresher worker) used during SIL (i.e., passed max op)
"sil_rollout_used":{},
# the return of used samples for buffer R
"sil_rollout_used_return":{},
# number of old samples still used (even after refreshing)
"sil_old_used":{}
}
sil_ctr, sil_a3c_sampled, sil_a3c_used, sil_a3c_used_return = 0, 0, 0, 0
sil_rollout_sampled, sil_rollout_used, sil_rollout_used_return = 0, 0, 0
sil_old_used = 0
rollout_dict = {
# total number of rollout performed
"rollout_ctr": {},
# total number of successful rollout (i.e., Gnew > G)
"rollout_added_ctr":{},
# the return of Gnew
"rollout_new_return":{},
# the return of G
"rollout_old_return":{}
}
rollout_ctr, rollout_added_ctr = 0, 0
rollout_new_return = 0 # this records the total, avg = this / rollout_added_ctr
rollout_old_return = 0 # this records the total, avg = this / rollout_added_ctr
# setup file names
reward_fname = folder / '{}-a3c-rewards.pkl'.format(GYM_ENV_NAME)
sil_fname = folder / '{}-a3c-dict-sil.pkl'.format(GYM_ENV_NAME)
rollout_fname = folder / '{}-a3c-dict-rollout.pkl'.format(GYM_ENV_NAME)
if args.load_pretrained_model:
class_reward_fname = folder / '{}-class-rewards.pkl'.format(GYM_ENV_NAME)
sharedmem_fname = folder / '{}-sharedmem.pkl'.format(GYM_ENV_NAME)
sharedmem_params_fname = folder / '{}-sharedmem-params.pkl'.format(GYM_ENV_NAME)
sharedmem_trees_fname = folder / '{}-sharedmem-trees.pkl'.format(GYM_ENV_NAME)
rolloutmem_fname = folder / '{}-rolloutmem.pkl'.format(GYM_ENV_NAME)
rolloutmem_params_fname = folder / '{}-rolloutmem-params.pkl'.format(GYM_ENV_NAME)
rolloutmem_trees_fname = folder / '{}-rolloutmem-trees.pkl'.format(GYM_ENV_NAME)
# for removing older ckpt, save mem space
prev_ckpt_t = -1
stop_req = False
game_state = GameState(env_id=args.gym_env)
action_size = game_state.env.action_space.n
game_state.close()
del game_state.env
del game_state
input_shape = (args.input_shape, args.input_shape, 4)
#######################################################
# setup global A3C
GameACFFNetwork.use_mnih_2015 = args.use_mnih_2015
global_network = GameACFFNetwork(
action_size, -1, device, padding=args.padding,
in_shape=input_shape)
logger.info('A3C Initial Learning Rate={}'.format(args.initial_learn_rate))
# setup pretrained model
global_pretrained_model = None
local_pretrained_model = None
pretrain_graph = None
# if use pretrained model to refresh
# then must load pretrained model
# otherwise, don't load model
if args.use_lider and args.nstep_bc > 0:
assert args.load_pretrained_model, "refreshing with other policies, must load a pre-trained model (TA or BC)"
else:
assert not args.load_pretrained_model, "refreshing with the current policy, don't load pre-trained models"
if args.load_pretrained_model:
pretrain_graph, global_pretrained_model = setup_pretrained_model(tf,
args, action_size, input_shape,
device="/gpu:0" if args.use_gpu else device)
assert global_pretrained_model is not None
assert pretrain_graph is not None
time.sleep(2.0)
# setup experience memory
shared_memory = None # => this is BufferD
rollout_buffer = None # => this is BufferR
if args.use_sil:
shared_memory = SILReplayMemory(
action_size, max_len=args.memory_length, gamma=args.gamma,
clip=False if args.unclipped_reward else True,
height=input_shape[0], width=input_shape[1],
phi_length=input_shape[2], priority=args.priority_memory,
reward_constant=args.reward_constant)
if args.use_lider and not args.onebuffer:
rollout_buffer = SILReplayMemory(
action_size, max_len=args.memory_length, gamma=args.gamma,
clip=False if args.unclipped_reward else True,
height=input_shape[0], width=input_shape[1],
phi_length=input_shape[2], priority=args.priority_memory,
reward_constant=args.reward_constant)
# log memory information
shared_memory.log()
if args.use_lider and not args.onebuffer:
rollout_buffer.log()
############## Setup Thread Workers BEGIN ################
# 17 total number of threads for all experiments
assert args.parallel_size ==17, "use 17 workers for all experiments"
startIndex = 0
all_workers = []
# a3c and sil learning rate and optimizer
learning_rate_input = tf.placeholder(tf.float32, shape=(), name="opt_lr")
grad_applier = tf.train.RMSPropOptimizer(
learning_rate=learning_rate_input,
decay=args.rmsp_alpha,
epsilon=args.rmsp_epsilon)
setup_common_worker(CommonWorker, args, action_size)
# setup SIL worker
sil_worker = None
if args.use_sil:
_device = "/gpu:0" if args.use_gpu else device
sil_network = GameACFFNetwork(
action_size, startIndex, device=_device,
padding=args.padding, in_shape=input_shape)
sil_worker = SILTrainingThread(startIndex, global_network, sil_network,
args.initial_learn_rate,
learning_rate_input,
grad_applier, device=_device,
batch_size=args.batch_size,
use_rollout=args.use_lider,
one_buffer=args.onebuffer,
sampleR=args.sampleR)
all_workers.append(sil_worker)
startIndex += 1
# setup refresh worker
refresh_worker = None
if args.use_lider:
_device = "/gpu:0" if args.use_gpu else device
refresh_network = GameACFFNetwork(
action_size, startIndex, device=_device,
padding=args.padding, in_shape=input_shape)
refresh_local_pretrained_model = None
# if refreshing with other polies
if args.nstep_bc > 0:
refresh_local_pretrained_model = PretrainedModelNetwork(
pretrain_graph, action_size, startIndex,
padding=args.padding,
in_shape=input_shape, sae=False,
tied_weights=False,
use_denoising=False,
noise_factor=0.3,
loss_function='mse',
use_slv=False, device=_device)
refresh_worker = RefreshThread(
thread_index=startIndex, action_size=action_size, env_id=args.gym_env,
global_a3c=global_network, local_a3c=refresh_network,
update_in_rollout=args.update_in_rollout, nstep_bc=args.nstep_bc,
global_pretrained_model=global_pretrained_model,
local_pretrained_model=refresh_local_pretrained_model,
transformed_bellman = args.transformed_bellman,
device=_device,
entropy_beta=args.entropy_beta, clip_norm=args.grad_norm_clip,
grad_applier=grad_applier,
initial_learn_rate=args.initial_learn_rate,
learning_rate_input=learning_rate_input)
all_workers.append(refresh_worker)
startIndex += 1
# setup a3c workers
setup_a3c_worker(A3CTrainingThread, args, startIndex)
for i in range(startIndex, args.parallel_size):
local_network = GameACFFNetwork(
action_size, i, device="/cpu:0",
padding=args.padding,
in_shape=input_shape)
a3c_worker = A3CTrainingThread(
i, global_network, local_network,
args.initial_learn_rate, learning_rate_input, grad_applier,
device="/cpu:0", no_op_max=30)
all_workers.append(a3c_worker)
############## Setup Thread Workers END ################
# setup config for tensorflow
config = tf.ConfigProto(
gpu_options=gpu_options,
log_device_placement=False,
allow_soft_placement=True)
# prepare sessions
sess = tf.Session(config=config)
pretrain_sess = None
if global_pretrained_model:
pretrain_sess = tf.Session(config=config, graph=pretrain_graph)
# initial pretrained model
if pretrain_sess:
assert args.pretrained_model_folder is not None
global_pretrained_model.load(
pretrain_sess,
args.pretrained_model_folder)
sess.run(tf.global_variables_initializer())
if global_pretrained_model:
initialize_uninitialized(tf, pretrain_sess,
global_pretrained_model)
if local_pretrained_model:
initialize_uninitialized(tf, pretrain_sess,
local_pretrained_model)
# summary writer for tensorboard
summ_file = args.save_to+'log/a3c/{}/'.format(GYM_ENV_NAME) + str(folder)[58:] # str(folder)[12:]
summary_writer = tf.summary.FileWriter(summ_file, sess.graph)
# init or load checkpoint with saver
root_saver = tf.train.Saver(max_to_keep=1)
saver = tf.train.Saver(max_to_keep=1)
best_saver = tf.train.Saver(max_to_keep=1)
checkpoint = tf.train.get_checkpoint_state(str(folder)+'/model_checkpoints')
if checkpoint and checkpoint.model_checkpoint_path:
root_saver.restore(sess, checkpoint.model_checkpoint_path)
logger.info("checkpoint loaded:{}".format(
checkpoint.model_checkpoint_path))
tokens = checkpoint.model_checkpoint_path.split("-")
# set global step
global_t = int(tokens[-1])
logger.info(">>> global step set: {}".format(global_t))
tmp_t = (global_t // args.eval_freq) * args.eval_freq
logger.info(">>> tmp_t: {}".format(tmp_t))
# set wall time
wall_t = 0.
# set up reward files
best_reward_file = folder / 'model_best/best_model_reward'
with best_reward_file.open('r') as f:
best_model_reward = float(f.read())
# restore rewards
rewards = restore_dict(reward_fname, global_t)
logger.info(">>> restored: rewards")
# restore loggings
sil_dict = restore_dict(sil_fname, global_t)
sil_ctr = sil_dict['sil_ctr'][tmp_t]
sil_a3c_sampled = sil_dict['sil_a3c_sampled'][tmp_t]
sil_a3c_used = sil_dict['sil_a3c_used'][tmp_t]
sil_a3c_used_return = sil_dict['sil_a3c_used_return'][tmp_t]
sil_rollout_sampled = sil_dict['sil_rollout_sampled'][tmp_t]
sil_rollout_used = sil_dict['sil_rollout_used'][tmp_t]
sil_rollout_used_return = sil_dict['sil_rollout_used_return'][tmp_t]
sil_old_used = sil_dict['sil_old_used'][tmp_t]
logger.info(">>> restored: sil_dict")
rollout_dict = restore_dict(rollout_fname, global_t)
rollout_ctr = rollout_dict['rollout_ctr'][tmp_t]
rollout_added_ctr = rollout_dict['rollout_added_ctr'][tmp_t]
rollout_new_return = rollout_dict['rollout_new_return'][tmp_t]
rollout_old_return = rollout_dict['rollout_old_return'][tmp_t]
logger.info(">>> restored: rollout_dict")
if args.load_pretrained_model:
class_reward_file = folder / '{}-class-rewards.pkl'.format(GYM_ENV_NAME)
class_rewards = restore_dict(class_reward_file, global_t)
# restore replay buffers (if saved)
if args.checkpoint_buffer:
# restore buffer D
if args.use_sil and args.priority_memory:
shared_memory = restore_buffer(sharedmem_fname, shared_memory, global_t)
shared_memory = restore_buffer_trees(sharedmem_trees_fname, shared_memory, global_t)
shared_memory = restore_buffer_params(sharedmem_params_fname, shared_memory, global_t)
logger.info(">>> restored: shared_memory (Buffer D)")
shared_memory.log()
# restore buffer R
if args.use_lider and not args.onebuffer:
rollout_buffer = restore_buffer(rolloutmem_fname, rollout_buffer, global_t)
rollout_buffer = restore_buffer_trees(rolloutmem_trees_fname, rollout_buffer, global_t)
rollout_buffer = restore_buffer_params(rolloutmem_params_fname, rollout_buffer, global_t)
logger.info(">>> restored: rollout_buffer (Buffer R)")
rollout_buffer.log()
# if all restores okay, remove old ckpt to save storage space
prev_ckpt_t = global_t
else:
logger.warning("Could not find old checkpoint")
wall_t = 0.0
prepare_dir(folder, empty=True)
prepare_dir(folder / 'model_checkpoints', empty=True)
prepare_dir(folder / 'model_best', empty=True)
prepare_dir(folder / 'frames', empty=True)
lock = threading.Lock()
# next saving global_t
def next_t(current_t, freq):
return np.ceil((current_t + 0.00001) / freq) * freq
next_global_t = next_t(global_t, args.eval_freq)
next_save_t = next_t(
global_t, args.eval_freq*args.checkpoint_freq)
step_t = 0
def train_function(parallel_idx, th_ctr, ep_queue, net_updates):
nonlocal global_t, step_t, rewards, class_rewards, lock, \
next_save_t, next_global_t, prev_ckpt_t
nonlocal shared_memory, rollout_buffer
nonlocal sil_dict, sil_ctr, sil_a3c_sampled, sil_a3c_used, sil_a3c_used_return, \
sil_rollout_sampled, sil_rollout_used, sil_rollout_used_return, \
sil_old_used
nonlocal rollout_dict, rollout_ctr, rollout_added_ctr, \
rollout_new_return, rollout_old_return
parallel_worker = all_workers[parallel_idx]
parallel_worker.set_summary_writer(summary_writer)
with lock:
# Evaluate model before training
if not stop_req and global_t == 0 and step_t == 0:
rewards['eval'][step_t] = parallel_worker.testing(
sess, args.eval_max_steps, global_t, folder,
worker=all_workers[-1])
# testing pretrained TA or BC in game
if args.load_pretrained_model:
assert pretrain_sess is not None
assert global_pretrained_model is not None
class_rewards['class_eval'][step_t] = \
parallel_worker.test_loaded_classifier(global_t=global_t,
max_eps=50, # testing 50 episodes
sess=pretrain_sess,
worker=all_workers[-1],
model=global_pretrained_model)
# log pretrained model performance
class_eval_file = pathlib.Path(args.pretrained_model_folder[:21]+\
str(GAME_NAME)+"/"+str(GAME_NAME)+'-model-eval.txt')
class_std = np.std(class_rewards['class_eval'][step_t][-1])
class_mean = np.mean(class_rewards['class_eval'][step_t][-1])
with class_eval_file.open('w') as f:
f.write("class_mean: \n" + str(class_mean) + "\n")
f.write("class_std: \n" + str(class_std) + "\n")
f.write("class_rewards: \n" + str(class_rewards['class_eval'][step_t][-1]) + "\n")
checkpt_file = folder / 'model_checkpoints'
checkpt_file /= '{}_checkpoint'.format(GYM_ENV_NAME)
saver.save(sess, str(checkpt_file), global_step=global_t)
save_best_model(rewards['eval'][global_t][0])
# saving worker info to dicts for analysis
sil_dict['sil_ctr'][step_t] = sil_ctr
sil_dict['sil_a3c_sampled'][step_t] = sil_a3c_sampled
sil_dict['sil_a3c_used'][step_t] = sil_a3c_used
sil_dict['sil_a3c_used_return'][step_t] = sil_a3c_used_return
sil_dict['sil_rollout_sampled'][step_t] = sil_rollout_sampled
sil_dict['sil_rollout_used'][step_t] = sil_rollout_used
sil_dict['sil_rollout_used_return'][step_t] = sil_rollout_used_return
sil_dict['sil_old_used'][step_t] = sil_old_used
rollout_dict['rollout_ctr'][step_t] = rollout_ctr
rollout_dict['rollout_added_ctr'][step_t] = rollout_added_ctr
rollout_dict['rollout_new_return'][step_t] = rollout_new_return
rollout_dict['rollout_old_return'][step_t] = rollout_old_return
# dump pickle
dump_pickle([rewards, sil_dict, rollout_dict],
[reward_fname, sil_fname, rollout_fname],
global_t)
if args.load_pretrained_model:
dump_pickle([class_rewards], [class_reward_fname], global_t)
logger.info('Dump pickle at step {}'.format(global_t))
# save replay buffer (only works under priority mem)
if args.checkpoint_buffer:
if shared_memory is not None and args.priority_memory:
params = [shared_memory.buff._next_idx, shared_memory.buff._max_priority]
trees = [shared_memory.buff._it_sum._value,
shared_memory.buff._it_min._value]
dump_pickle([shared_memory.buff._storage, params, trees],
[sharedmem_fname, sharedmem_params_fname, sharedmem_trees_fname],
global_t)
logger.info('Saving shared_memory')
if rollout_buffer is not None and args.priority_memory:
params = [rollout_buffer.buff._next_idx, rollout_buffer.buff._max_priority]
trees = [rollout_buffer.buff._it_sum._value,
rollout_buffer.buff._it_min._value]
dump_pickle([rollout_buffer.buff._storage, params, trees],
[rolloutmem_fname, rolloutmem_params_fname, rolloutmem_trees_fname],
global_t)
logger.info('Saving rollout_buffer')
prev_ckpt_t = global_t
step_t = 1
# set start_time
start_time = time.time() - wall_t
parallel_worker.set_start_time(start_time)
if parallel_worker.is_sil_thread:
sil_interval = 0 # bigger number => slower SIL updates
m_repeat = 4
min_mem = args.batch_size * m_repeat
sil_train_flag = len(shared_memory) >= min_mem
while True:
if stop_req:
return
if global_t >= (args.max_time_step * args.max_time_step_fraction):
return
if parallel_worker.is_sil_thread:
# before sil starts, init local count
local_sil_ctr = 0
local_sil_a3c_sampled, local_sil_a3c_used, local_sil_a3c_used_return = 0, 0, 0
local_sil_rollout_sampled, local_sil_rollout_used, local_sil_rollout_used_return = 0, 0, 0
local_sil_old_used = 0
if net_updates.qsize() >= sil_interval \
and len(shared_memory) >= min_mem:
sil_train_flag = True
if sil_train_flag:
sil_train_flag = False
th_ctr.get()
train_out = parallel_worker.sil_train(
sess, global_t, shared_memory, m_repeat,
rollout_buffer=rollout_buffer)
local_sil_ctr, local_sil_a3c_sampled, local_sil_a3c_used, \
local_sil_a3c_used_return, \
local_sil_rollout_sampled, local_sil_rollout_used, \
local_sil_rollout_used_return, \
local_sil_old_used = train_out
th_ctr.put(1)
with net_updates.mutex:
net_updates.queue.clear()
if args.use_lider:
parallel_worker.record_sil(sil_ctr=sil_ctr,
total_used=(sil_a3c_used + sil_rollout_used),
num_a3c_used=sil_a3c_used,
a3c_used_return=sil_a3c_used_return/(sil_a3c_used+1),#add one in case divide by zero
rollout_used=sil_rollout_used,
rollout_used_return=sil_rollout_used_return/(sil_rollout_used+1),
old_used=sil_old_used,
global_t=global_t)
if sil_ctr % 200 == 0 and sil_ctr > 0:
rollout_buffsize = 0
if not args.onebuffer:
rollout_buffsize = len(rollout_buffer)
log_data = (sil_ctr, len(shared_memory),
rollout_buffsize,
sil_a3c_used+sil_rollout_used,
args.batch_size*sil_ctr,
sil_a3c_used,
sil_a3c_used_return/(sil_a3c_used+1),
sil_rollout_used,
sil_rollout_used_return/(sil_rollout_used+1),
sil_old_used)
logger.info("SIL: sil_ctr={0:}"
" sil_memory_size={1:}"
" rollout_buffer_size={2:}"
" total_sample_used={3:}/{4:}"
" a3c_used={5:}"
" a3c_used_return_avg={6:.2f}"
" rollout_used={7:}"
" rollout_used_return_avg={8:.2f}"
" old_used={9:}".format(*log_data))
else:
parallel_worker.record_sil(sil_ctr=sil_ctr,
total_used=(sil_a3c_used + sil_rollout_used),
num_a3c_used=sil_a3c_used,
rollout_used=sil_rollout_used,
global_t=global_t)
if sil_ctr % 200 == 0 and sil_ctr > 0:
log_data = (sil_ctr, sil_a3c_used+sil_rollout_used,
args.batch_size*sil_ctr,
sil_a3c_used,
len(shared_memory))
logger.info("SIL: sil_ctr={0:}"
" total_sample_used={1:}/{2:}"
" a3c_used={3:}"
" sil_memory_size={4:}".format(*log_data))
# Adding episodes to SIL memory is centralize to ensure
# sampling and updating of priorities does not become a problem
# since we add new episodes to SIL at once and during
# SIL training it is guaranteed that SIL memory is untouched.
max = args.parallel_size
while not ep_queue.empty():
data = ep_queue.get()
parallel_worker.episode.set_data(*data)
shared_memory.extend(parallel_worker.episode)
parallel_worker.episode.reset()
max -= 1
if max <= 0: # This ensures that SIL has a chance to train
break
diff_global_t = 0
# centralized rollout counting
local_rollout_ctr, local_rollout_added_ctr = 0, 0
local_rollout_new_return, local_rollout_old_return = 0, 0
elif parallel_worker.is_refresh_thread:
# before refresh starts, init local count
diff_global_t = 0
local_rollout_ctr, local_rollout_added_ctr = 0, 0
local_rollout_new_return, local_rollout_old_return = 0, 0
if len(shared_memory) >= 1:
th_ctr.get()
# randomly sample a state from buffer D
sample = shared_memory.sample_one_random()
# after sample, flip refreshed to True
# TODO: fix this so that only *succesful* refresh is flipped to True
# currently counting *all* refresh as True
assert sample[-1] == True
train_out = parallel_worker.rollout(sess, folder, pretrain_sess,
global_t, sample,
args.addall,
args.max_ep_step,
args.nstep_bc,
args.update_in_rollout)
diff_global_t, episode_end, part_end, local_rollout_ctr, \
local_rollout_added_ctr, add, local_rollout_new_return, \
local_rollout_old_return = train_out
th_ctr.put(1)
if rollout_ctr % 20 == 0 and rollout_ctr > 0:
log_msg = "ROLLOUT: rollout_ctr={} added_rollout_ct={} worker={}".format(
rollout_ctr, rollout_added_ctr, parallel_worker.thread_idx)
logger.info(log_msg)
logger.info("ROLLOUT Gnew: {}, G: {}".format(local_rollout_new_return,
local_rollout_old_return))
# should always part_end, i.e., end of episode
# and only add if new return is better (if not LiDER-AddAll)
if part_end and add:
if not args.onebuffer:
# directly put into Buffer R
rollout_buffer.extend(parallel_worker.episode)
else:
# Buffer D add sample is centralized when OneBuffer
ep_queue.put(parallel_worker.episode.get_data())
parallel_worker.episode.reset()
# centralized SIL counting
local_sil_ctr = 0
local_sil_a3c_sampled, local_sil_a3c_used, local_sil_a3c_used_return = 0, 0, 0
local_sil_rollout_sampled, local_sil_rollout_used, local_sil_rollout_used_return = 0, 0, 0
local_sil_old_used = 0
# a3c training thread worker
else:
th_ctr.get()
train_out = parallel_worker.train(sess, global_t, rewards)
diff_global_t, episode_end, part_end = train_out
th_ctr.put(1)
if args.use_sil:
net_updates.put(1)
if part_end:
ep_queue.put(parallel_worker.episode.get_data())
parallel_worker.episode.reset()
# centralized SIL counting
local_sil_ctr = 0
local_sil_a3c_sampled, local_sil_a3c_used, local_sil_a3c_used_return = 0, 0, 0
local_sil_rollout_sampled, local_sil_rollout_used, local_sil_rollout_used_return = 0, 0, 0
local_sil_old_used = 0
# centralized rollout counting
local_rollout_ctr, local_rollout_added_ctr = 0, 0
local_rollout_new_return, local_rollout_old_return = 0, 0
# ensure only one thread is updating global_t at a time
with lock:
global_t += diff_global_t
# centralize increasing count for SIL and Rollout
sil_ctr += local_sil_ctr
sil_a3c_sampled += local_sil_a3c_sampled
sil_a3c_used += local_sil_a3c_used
sil_a3c_used_return += local_sil_a3c_used_return
sil_rollout_sampled += local_sil_rollout_sampled
sil_rollout_used += local_sil_rollout_used
sil_rollout_used_return += local_sil_rollout_used_return
sil_old_used += local_sil_old_used
rollout_ctr += local_rollout_ctr
rollout_added_ctr += local_rollout_added_ctr
rollout_new_return += local_rollout_new_return
rollout_old_return += local_rollout_old_return
# if during a thread's update, global_t has reached a evaluation interval
if global_t > next_global_t:
next_global_t = next_t(global_t, args.eval_freq)
step_t = int(next_global_t - args.eval_freq)
# wait for all threads to be done before testing
while not stop_req and th_ctr.qsize() < len(all_workers):
time.sleep(0.001)
step_t = int(next_global_t - args.eval_freq)
# Evaluate for 125,000 steps
rewards['eval'][step_t] = parallel_worker.testing(
sess, args.eval_max_steps, step_t, folder,
worker=all_workers[-1])
save_best_model(rewards['eval'][step_t][0])
last_reward = rewards['eval'][step_t][0]
# saving worker info to dicts
# SIL
sil_dict['sil_ctr'][step_t] = sil_ctr
sil_dict['sil_a3c_sampled'][step_t] = sil_a3c_sampled
sil_dict['sil_a3c_used'][step_t] = sil_a3c_used
sil_dict['sil_a3c_used_return'][step_t] = sil_a3c_used_return
sil_dict['sil_rollout_sampled'][step_t] = sil_rollout_sampled
sil_dict['sil_rollout_used'][step_t] = sil_rollout_used
sil_dict['sil_rollout_used_return'][step_t] = sil_rollout_used_return
sil_dict['sil_old_used'][step_t] = sil_old_used
# ROLLOUT
rollout_dict['rollout_ctr'][step_t] = rollout_ctr
rollout_dict['rollout_added_ctr'][step_t] = rollout_added_ctr
rollout_dict['rollout_new_return'][step_t] = rollout_new_return
rollout_dict['rollout_old_return'][step_t] = rollout_old_return
# save ckpt after done with eval
if global_t > next_save_t:
next_save_t = next_t(global_t, args.eval_freq*args.checkpoint_freq)
# dump pickle
dump_pickle([rewards, sil_dict, rollout_dict],
[reward_fname, sil_fname, rollout_fname],
global_t)
if args.load_pretrained_model:
dump_pickle([class_rewards], [class_reward_fname], global_t)
logger.info('Dump pickle at step {}'.format(global_t))
# save replay buffer (only works for priority mem for now)
if args.checkpoint_buffer:
if shared_memory is not None and args.priority_memory:
params = [shared_memory.buff._next_idx, shared_memory.buff._max_priority]
trees = [shared_memory.buff._it_sum._value,
shared_memory.buff._it_min._value]
dump_pickle([shared_memory.buff._storage, params, trees],
[sharedmem_fname, sharedmem_params_fname, sharedmem_trees_fname],
global_t)
logger.info('Saved shared_memory')
if rollout_buffer is not None and args.priority_memory:
params = [rollout_buffer.buff._next_idx, rollout_buffer.buff._max_priority]
trees = [rollout_buffer.buff._it_sum._value,
rollout_buffer.buff._it_min._value]
dump_pickle([rollout_buffer.buff._storage, params, trees],
[rolloutmem_fname, rolloutmem_params_fname, rolloutmem_trees_fname],
global_t)
logger.info('Saved rollout_buffer')
# save a3c after saving buffer -- in case saving buffer OOM
# so that at least we can revert back to the previous ckpt
checkpt_file = folder / 'model_checkpoints'
checkpt_file /= '{}_checkpoint'.format(GYM_ENV_NAME)
saver.save(sess, str(checkpt_file), global_step=global_t,
write_meta_graph=False)
logger.info('Saved model ckpt')
# if everything saves okay, clean up previous ckpt to save space
remove_pickle([reward_fname, sil_fname, rollout_fname],
prev_ckpt_t)
if args.load_pretrained_model:
remove_pickle([class_reward_fname], prev_ckpt_t)
remove_pickle([sharedmem_fname, sharedmem_params_fname,
sharedmem_trees_fname],
prev_ckpt_t)
if rollout_buffer is not None and args.priority_memory:
remove_pickle([rolloutmem_fname, rolloutmem_params_fname,
rolloutmem_trees_fname],
prev_ckpt_t)
logger.info('Removed ckpt from step {}'.format(prev_ckpt_t))
prev_ckpt_t = global_t
def signal_handler(signal, frame):
nonlocal stop_req
logger.info('You pressed Ctrl+C!')
stop_req = True
if stop_req and global_t == 0:
sys.exit(1)
def save_best_model(test_reward):
nonlocal best_model_reward
if test_reward > best_model_reward:
best_model_reward = test_reward
best_reward_file = folder / 'model_best/best_model_reward'
with best_reward_file.open('w') as f:
f.write(str(best_model_reward))
best_checkpt_file = folder / 'model_best'
best_checkpt_file /= '{}_checkpoint'.format(GYM_ENV_NAME)
best_saver.save(sess, str(best_checkpt_file))
train_threads = []
th_ctr = Queue()
for i in range(args.parallel_size):
th_ctr.put(1)
episodes_queue = None
net_updates = None
if args.use_sil:
episodes_queue = Queue()
net_updates = Queue()
for i in range(args.parallel_size):
worker_thread = Thread(
target=train_function,
args=(i, th_ctr, episodes_queue, net_updates,))
train_threads.append(worker_thread)
signal.signal(signal.SIGINT, signal_handler)
signal.signal(signal.SIGTERM, signal_handler)
# set start time
start_time = time.time() - wall_t
for t in train_threads:
t.start()
print('Press Ctrl+C to stop')
for t in train_threads:
t.join()
logger.info('Now saving data. Please wait')
# write wall time
wall_t = time.time() - start_time
wall_t_fname = folder / 'wall_t.{}'.format(global_t)
with wall_t_fname.open('w') as f:
f.write(str(wall_t))
# save final model
checkpoint_file = str(folder / '{}_checkpoint_a3c'.format(GYM_ENV_NAME))
root_saver.save(sess, checkpoint_file, global_step=global_t)
dump_final_pickle([rewards, sil_dict, rollout_dict],
[reward_fname, sil_fname, rollout_fname])
logger.info('Data saved!')
# if everything saves okay & is done training (not because of pressed Ctrl+C),
# clean up previous ckpt to save space
if global_t >= (args.max_time_step * args.max_time_step_fraction):
remove_pickle([reward_fname, sil_fname, rollout_fname],
prev_ckpt_t)
if args.load_pretrained_model:
remove_pickle([class_reward_fname], prev_ckpt_t)
remove_pickle([sharedmem_fname, sharedmem_params_fname, sharedmem_trees_fname],
prev_ckpt_t)
if rollout_buffer is not None and args.priority_memory:
remove_pickle([rolloutmem_fname, rolloutmem_params_fname, rolloutmem_trees_fname],
prev_ckpt_t)
logger.info('Done training, removed ckpt from step {}'.format(prev_ckpt_t))
sess.close()
if pretrain_sess:
pretrain_sess.close()