def __init__(self,
thread_index,
global_net,
local_net,
initial_learning_rate,
learning_rate_input,
grad_applier,
device=None,
batch_size=None,
use_rollout=False,
one_buffer=False,
sampleR=False):
"""Initialize A3CTrainingThread class."""
assert self.action_size != -1
self.is_sil_thread = True
self.thread_idx = thread_index
self.initial_learning_rate = initial_learning_rate
self.learning_rate_input = learning_rate_input
self.local_net = local_net
self.batch_size = batch_size
self.use_rollout = use_rollout
self.one_buffer = one_buffer
self.sampleR = sampleR
logger.info("===SIL thread_index: {}===".format(self.thread_idx))
logger.info("device: {}".format(device))
logger.info("action_size: {}".format(self.action_size))
logger.info("entropy_beta: {}".format(self.entropy_beta))
logger.info("gamma: {}".format(self.gamma))
logger.info("reward_type: {}".format(self.reward_type))
logger.info("transformed_bellman: {}".format(
colored(self.transformed_bellman,
"green" if self.transformed_bellman else "red")))
logger.info("clip_norm: {}".format(self.clip_norm))
logger.info("use_grad_cam: {}".format(
colored(self.use_grad_cam,
"green" if self.use_grad_cam else "red")))
reward_clipped = True if self.reward_type == 'CLIP' else False
local_vars = self.local_net.get_vars
with tf.device(device):
critic_lr = 0.1
entropy_beta = 0
w_loss = 1.0
logger.info("sil batch_size: {}".format(self.batch_size))
logger.info("sil w_loss: {}".format(w_loss))
logger.info("sil critic_lr: {}".format(critic_lr))
logger.info("sil entropy_beta: {}".format(entropy_beta))
self.local_net.prepare_sil_loss(entropy_beta=entropy_beta,
w_loss=w_loss,
critic_lr=critic_lr)
var_refs = [v._ref() for v in local_vars()]
self.sil_gradients = tf.gradients(self.local_net.total_loss_sil,
var_refs)
global_vars = global_net.get_vars
with tf.device(device):
if self.clip_norm is not None:
self.sil_gradients, grad_norm = tf.clip_by_global_norm(
self.sil_gradients, self.clip_norm)
sil_gradients_global = list(zip(self.sil_gradients, global_vars()))
sil_gradients_local = list(zip(self.sil_gradients, local_vars()))
self.sil_apply_gradients = grad_applier.apply_gradients(
sil_gradients_global)
self.sil_apply_gradients_local = grad_applier.apply_gradients(
sil_gradients_local)
self.sync = self.local_net.sync_from(global_net)
self.episode = SILReplayMemory(self.action_size,
max_len=None,
gamma=self.gamma,
clip=reward_clipped,
height=self.local_net.in_shape[0],
width=self.local_net.in_shape[1],
phi_length=self.local_net.in_shape[2],
reward_constant=self.reward_constant)
# temp_buffer for mixing and re-sample (brown arrow in Figure 1)
# initial only when needed (A3CTBSIL & LiDER-OneBuffer does not need temp_buffer)
self.temp_buffer = None
if (self.use_rollout) and (not self.one_buffer):
self.temp_buffer = SILReplayMemory(
self.action_size,
max_len=self.batch_size * 2,
gamma=self.gamma,
clip=reward_clipped,
height=self.local_net.in_shape[0],
width=self.local_net.in_shape[1],
phi_length=self.local_net.in_shape[2],
priority=True,
reward_constant=self.reward_constant)
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()
class SILTrainingThread(CommonWorker):
"""Asynchronous Actor-Critic Training Thread Class."""
entropy_beta = 0.01
gamma = 0.99
finetune_upper_layers_only = False
transformed_bellman = False
clip_norm = 0.5
use_grad_cam = False
def __init__(self,
thread_index,
global_net,
local_net,
initial_learning_rate,
learning_rate_input,
grad_applier,
device=None,
batch_size=None,
use_rollout=False,
one_buffer=False,
sampleR=False):
"""Initialize A3CTrainingThread class."""
assert self.action_size != -1
self.is_sil_thread = True
self.thread_idx = thread_index
self.initial_learning_rate = initial_learning_rate
self.learning_rate_input = learning_rate_input
self.local_net = local_net
self.batch_size = batch_size
self.use_rollout = use_rollout
self.one_buffer = one_buffer
self.sampleR = sampleR
logger.info("===SIL thread_index: {}===".format(self.thread_idx))
logger.info("device: {}".format(device))
logger.info("action_size: {}".format(self.action_size))
logger.info("entropy_beta: {}".format(self.entropy_beta))
logger.info("gamma: {}".format(self.gamma))
logger.info("reward_type: {}".format(self.reward_type))
logger.info("transformed_bellman: {}".format(
colored(self.transformed_bellman,
"green" if self.transformed_bellman else "red")))
logger.info("clip_norm: {}".format(self.clip_norm))
logger.info("use_grad_cam: {}".format(
colored(self.use_grad_cam,
"green" if self.use_grad_cam else "red")))
reward_clipped = True if self.reward_type == 'CLIP' else False
local_vars = self.local_net.get_vars
with tf.device(device):
critic_lr = 0.1
entropy_beta = 0
w_loss = 1.0
logger.info("sil batch_size: {}".format(self.batch_size))
logger.info("sil w_loss: {}".format(w_loss))
logger.info("sil critic_lr: {}".format(critic_lr))
logger.info("sil entropy_beta: {}".format(entropy_beta))
self.local_net.prepare_sil_loss(entropy_beta=entropy_beta,
w_loss=w_loss,
critic_lr=critic_lr)
var_refs = [v._ref() for v in local_vars()]
self.sil_gradients = tf.gradients(self.local_net.total_loss_sil,
var_refs)
global_vars = global_net.get_vars
with tf.device(device):
if self.clip_norm is not None:
self.sil_gradients, grad_norm = tf.clip_by_global_norm(
self.sil_gradients, self.clip_norm)
sil_gradients_global = list(zip(self.sil_gradients, global_vars()))
sil_gradients_local = list(zip(self.sil_gradients, local_vars()))
self.sil_apply_gradients = grad_applier.apply_gradients(
sil_gradients_global)
self.sil_apply_gradients_local = grad_applier.apply_gradients(
sil_gradients_local)
self.sync = self.local_net.sync_from(global_net)
self.episode = SILReplayMemory(self.action_size,
max_len=None,
gamma=self.gamma,
clip=reward_clipped,
height=self.local_net.in_shape[0],
width=self.local_net.in_shape[1],
phi_length=self.local_net.in_shape[2],
reward_constant=self.reward_constant)
# temp_buffer for mixing and re-sample (brown arrow in Figure 1)
# initial only when needed (A3CTBSIL & LiDER-OneBuffer does not need temp_buffer)
self.temp_buffer = None
if (self.use_rollout) and (not self.one_buffer):
self.temp_buffer = SILReplayMemory(
self.action_size,
max_len=self.batch_size * 2,
gamma=self.gamma,
clip=reward_clipped,
height=self.local_net.in_shape[0],
width=self.local_net.in_shape[1],
phi_length=self.local_net.in_shape[2],
priority=True,
reward_constant=self.reward_constant)
def record_sil(self,
sil_ctr=0,
total_used=0,
num_a3c_used=0,
a3c_used_return=0,
rollout_used=0,
rollout_used_return=0,
old_used=0,
global_t=0,
mode='SIL'):
"""Record SIL."""
summary = tf.Summary()
summary.value.add(tag='{}/sil_ctr'.format(mode),
simple_value=float(sil_ctr))
summary.value.add(tag='{}/total_num_sample_used'.format(mode),
simple_value=float(total_used))
summary.value.add(tag='{}/num_a3c_used'.format(mode),
simple_value=float(num_a3c_used))
summary.value.add(tag='{}/a3c_used_return'.format(mode),
simple_value=float(a3c_used_return))
summary.value.add(tag='{}/num_rollout_used'.format(mode),
simple_value=float(rollout_used_return))
summary.value.add(tag='{}/rollout_used_return'.format(mode),
simple_value=float(rollout_used))
summary.value.add(tag='{}/num_old_used'.format(mode),
simple_value=float(old_used))
self.writer.add_summary(summary, global_t)
self.writer.flush()
def sil_train(self, sess, global_t, sil_memory, m, rollout_buffer=None):
"""Self-imitation learning process."""
# copy weights from shared to local
sess.run(self.sync)
cur_learning_rate = self._anneal_learning_rate(
global_t, self.initial_learning_rate)
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
total_used = 0
num_a3c_used = 0
num_rollout_used = 0
num_rollout_sampled = 0
num_old_used = 0
for _ in range(m):
d_batch_size, r_batch_size = 0, 0
# A3CTBSIL
if not self.use_rollout:
d_batch_size = self.batch_size
# or LiDER-OneBuffer
elif self.use_rollout and self.one_buffer:
d_batch_size = self.batch_size
# or LiDER
else:
assert rollout_buffer is not None
assert self.temp_buffer is not None
self.temp_buffer.reset()
if not self.sampleR: # otherwise, LiDER-SampleR
d_batch_size = self.batch_size
r_batch_size = self.batch_size
batch_state, batch_action, batch_returns, batch_fullstate, \
batch_rollout, batch_refresh, weights = ([] for i in range(7))
# sample from buffer D
if d_batch_size > 0 and len(sil_memory) > d_batch_size:
d_sample = sil_memory.sample(d_batch_size, beta=0.4)
d_index_list, d_batch, d_weights = d_sample
d_batch_state, d_action, d_batch_returns, \
d_batch_fullstate, d_batch_rollout, d_batch_refresh = d_batch
# update priority of sampled experiences
self.update_priorities_once(sess, sil_memory, d_index_list,
d_batch_state, d_action,
d_batch_returns)
if self.temp_buffer is not None: # when LiDER
d_batch_action = convert_onehot_to_a(d_action)
self.temp_buffer.extend_one_priority(
d_batch_state, d_batch_fullstate, d_batch_action,
d_batch_returns, d_batch_rollout, d_batch_refresh)
else: # when A3CTBSIL or LiDER-OneBuffer
batch_state.extend(d_batch_state)
batch_action.extend(d_action)
batch_returns.extend(d_batch_returns)
batch_fullstate.extend(d_batch_fullstate)
batch_rollout.extend(d_batch_rollout)
batch_refresh.extend(d_batch_refresh)
weights.extend(d_weights)
# sample from buffer R
if r_batch_size > 0 and len(rollout_buffer) > r_batch_size:
r_sample = rollout_buffer.sample(r_batch_size, beta=0.4)
r_index_list, r_batch, r_weights = r_sample
r_batch_state, r_action, r_batch_returns, \
r_batch_fullstate, r_batch_rollout, r_batch_refresh = r_batch
# update priority of sampled experiences
self.update_priorities_once(sess, rollout_buffer, r_index_list,
r_batch_state, r_action,
r_batch_returns)
if self.temp_buffer is not None: # when LiDER
r_batch_action = convert_onehot_to_a(r_action)
self.temp_buffer.extend_one_priority(
r_batch_state, r_batch_fullstate, r_batch_action,
r_batch_returns, r_batch_rollout, r_batch_refresh)
else: # when A3CTBSIL or LiDER-OneBuffer
batch_state.extend(r_batch_state)
batch_action.extend(r_action)
batch_returns.extend(r_batch_returns)
batch_fullstate.extend(r_batch_fullstate)
batch_rollout.extend(r_batch_rollout)
batch_refresh.extend(r_batch_refresh)
weights.extend(r_weights)
# LiDER only: pick 32 out of mixed
# (at the beginning the 32 could all from buffer D since rollout has no data yet)
# make sure the temp_buffer has been filled with at least size of one batch before sampling
if self.temp_buffer is not None and len(
self.temp_buffer) >= self.batch_size:
sample = self.temp_buffer.sample(self.batch_size, beta=0.4)
index_list, batch, weights = sample
# overwrite the initial empty list
batch_state, batch_action, batch_returns, \
batch_fullstate, batch_rollout, batch_refresh = batch
if self.use_rollout:
num_rollout_sampled += np.sum(batch_rollout)
# sil policy update (if one full batch is sampled)
if len(batch_state) == self.batch_size:
feed_dict = {
self.local_net.s: batch_state,
self.local_net.a_sil: batch_action,
self.local_net.returns: batch_returns,
self.local_net.weights: weights,
self.learning_rate_input: cur_learning_rate,
}
fetch = [
self.local_net.clipped_advs,
self.local_net.advs,
self.sil_apply_gradients,
self.sil_apply_gradients_local,
]
adv_clip, adv, _, _ = sess.run(fetch, feed_dict=feed_dict)
pos_idx = [i for (i, num) in enumerate(adv) if num > 0]
neg_idx = [i for (i, num) in enumerate(adv) if num <= 0]
# log number of samples used for SIL updates
total_used += len(pos_idx)
num_rollout_used += np.sum(np.take(batch_rollout, pos_idx))
num_a3c_used += (len(pos_idx) -
np.sum(np.take(batch_rollout, pos_idx)))
num_old_used += np.sum(np.take(batch_refresh, pos_idx))
# return for used rollout samples
rollout_idx = [
i for (i, num) in enumerate(batch_rollout) if num > 0
]
pos_rollout_idx = np.intersect1d(rollout_idx, pos_idx)
if len(pos_rollout_idx) > 0:
local_sil_rollout_used_return += np.sum(
np.take(adv, pos_rollout_idx))
# return for used a3c samples
a3c_idx = [
i for (i, num) in enumerate(batch_rollout) if num <= 0
]
pos_a3c_idx = np.intersect1d(a3c_idx, pos_idx)
if len(pos_a3c_idx) > 0:
local_sil_a3c_used_return += np.sum(
np.take(batch_returns, pos_a3c_idx))
local_sil_ctr += 1
local_sil_a3c_sampled += (self.batch_size * m - num_rollout_sampled)
local_sil_rollout_sampled += num_rollout_sampled
local_sil_a3c_used += num_a3c_used
local_sil_rollout_used += num_rollout_used
local_sil_old_used += num_old_used
return 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
def update_priorities_once(self, sess, memory, index_list, batch_state,
batch_action, batch_returns):
"""Self-imitation update priorities once."""
# copy weights from shared to local
sess.run(self.sync)
feed_dict = {
self.local_net.s: batch_state,
self.local_net.a_sil: batch_action,
self.local_net.returns: batch_returns,
}
fetch = self.local_net.clipped_advs
adv_clip = sess.run(fetch, feed_dict=feed_dict)
memory.set_weights(index_list, adv_clip)
def __init__(self,
thread_index,
global_net,
local_net,
initial_learning_rate,
learning_rate_input,
grad_applier,
device=None,
no_op_max=30):
"""Initialize A3CTrainingThread class."""
assert self.action_size != -1
self.is_sil_thread = False
self.is_refresh_thread = False
self.thread_idx = thread_index
self.learning_rate_input = learning_rate_input
self.local_net = local_net
self.no_op_max = no_op_max
self.override_num_noops = 0 if self.no_op_max == 0 else None
logger.info("===A3C thread_index: {}===".format(self.thread_idx))
logger.info("device: {}".format(device))
logger.info("use_sil: {}".format(
colored(self.use_sil, "green" if self.use_sil else "red")))
logger.info("local_t_max: {}".format(self.local_t_max))
logger.info("action_size: {}".format(self.action_size))
logger.info("entropy_beta: {}".format(self.entropy_beta))
logger.info("gamma: {}".format(self.gamma))
logger.info("reward_type: {}".format(self.reward_type))
logger.info("transformed_bellman: {}".format(
colored(self.transformed_bellman,
"green" if self.transformed_bellman else "red")))
logger.info("clip_norm: {}".format(self.clip_norm))
logger.info("use_grad_cam: {}".format(
colored(self.use_grad_cam,
"green" if self.use_grad_cam else "red")))
reward_clipped = True if self.reward_type == 'CLIP' else False
local_vars = self.local_net.get_vars
with tf.device(device):
self.local_net.prepare_loss(entropy_beta=self.entropy_beta,
critic_lr=0.5)
var_refs = [v._ref() for v in local_vars()]
self.gradients = tf.gradients(self.local_net.total_loss, var_refs)
global_vars = global_net.get_vars
with tf.device(device):
if self.clip_norm is not None:
self.gradients, grad_norm = tf.clip_by_global_norm(
self.gradients, self.clip_norm)
self.gradients = list(zip(self.gradients, global_vars()))
self.apply_gradients = grad_applier.apply_gradients(self.gradients)
self.sync = self.local_net.sync_from(global_net)
self.game_state = GameState(env_id=self.env_id,
display=False,
no_op_max=self.no_op_max,
human_demo=False,
episode_life=True,
override_num_noops=self.override_num_noops)
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_steps = 0
# variable controlling log output
self.prev_local_t = 0
with tf.device(device):
if self.use_grad_cam:
self.action_meaning = self.game_state.env.unwrapped \
.get_action_meanings()
self.local_net.build_grad_cam_grads()
if self.use_sil:
self.episode = SILReplayMemory(
self.action_size,
max_len=None,
gamma=self.gamma,
clip=reward_clipped,
height=self.local_net.in_shape[0],
width=self.local_net.in_shape[1],
phi_length=self.local_net.in_shape[2],
reward_constant=self.reward_constant)
class A3CTrainingThread(CommonWorker):
"""Asynchronous Actor-Critic Training Thread Class."""
log_interval = 100
perf_log_interval = 1000
local_t_max = 20
entropy_beta = 0.01
gamma = 0.99
shaping_actions = -1 # -1 all actions, 0 exclude noop
transformed_bellman = False
clip_norm = 0.5
use_grad_cam = False
use_sil = False
log_idx = 0
reward_constant = 0
def __init__(self,
thread_index,
global_net,
local_net,
initial_learning_rate,
learning_rate_input,
grad_applier,
device=None,
no_op_max=30):
"""Initialize A3CTrainingThread class."""
assert self.action_size != -1
self.is_sil_thread = False
self.is_refresh_thread = False
self.thread_idx = thread_index
self.learning_rate_input = learning_rate_input
self.local_net = local_net
self.no_op_max = no_op_max
self.override_num_noops = 0 if self.no_op_max == 0 else None
logger.info("===A3C thread_index: {}===".format(self.thread_idx))
logger.info("device: {}".format(device))
logger.info("use_sil: {}".format(
colored(self.use_sil, "green" if self.use_sil else "red")))
logger.info("local_t_max: {}".format(self.local_t_max))
logger.info("action_size: {}".format(self.action_size))
logger.info("entropy_beta: {}".format(self.entropy_beta))
logger.info("gamma: {}".format(self.gamma))
logger.info("reward_type: {}".format(self.reward_type))
logger.info("transformed_bellman: {}".format(
colored(self.transformed_bellman,
"green" if self.transformed_bellman else "red")))
logger.info("clip_norm: {}".format(self.clip_norm))
logger.info("use_grad_cam: {}".format(
colored(self.use_grad_cam,
"green" if self.use_grad_cam else "red")))
reward_clipped = True if self.reward_type == 'CLIP' else False
local_vars = self.local_net.get_vars
with tf.device(device):
self.local_net.prepare_loss(entropy_beta=self.entropy_beta,
critic_lr=0.5)
var_refs = [v._ref() for v in local_vars()]
self.gradients = tf.gradients(self.local_net.total_loss, var_refs)
global_vars = global_net.get_vars
with tf.device(device):
if self.clip_norm is not None:
self.gradients, grad_norm = tf.clip_by_global_norm(
self.gradients, self.clip_norm)
self.gradients = list(zip(self.gradients, global_vars()))
self.apply_gradients = grad_applier.apply_gradients(self.gradients)
self.sync = self.local_net.sync_from(global_net)
self.game_state = GameState(env_id=self.env_id,
display=False,
no_op_max=self.no_op_max,
human_demo=False,
episode_life=True,
override_num_noops=self.override_num_noops)
self.local_t = 0
self.initial_learning_rate = initial_learning_rate
self.episode_reward = 0
self.episode_steps = 0
# variable controlling log output
self.prev_local_t = 0
with tf.device(device):
if self.use_grad_cam:
self.action_meaning = self.game_state.env.unwrapped \
.get_action_meanings()
self.local_net.build_grad_cam_grads()
if self.use_sil:
self.episode = SILReplayMemory(
self.action_size,
max_len=None,
gamma=self.gamma,
clip=reward_clipped,
height=self.local_net.in_shape[0],
width=self.local_net.in_shape[1],
phi_length=self.local_net.in_shape[2],
reward_constant=self.reward_constant)
def train(self, sess, global_t, train_rewards):
"""Train A3C."""
states = []
fullstates = []
actions = []
rewards = []
values = []
rho = []
terminal_pseudo = False # loss of life
terminal_end = False # real terminal
# copy weights from shared to local
sess.run(self.sync)
start_local_t = self.local_t
# t_max times loop
for i in range(self.local_t_max):
state = cv2.resize(self.game_state.s_t,
self.local_net.in_shape[:-1],
interpolation=cv2.INTER_AREA)
fullstate = self.game_state.clone_full_state()
pi_, value_, logits_ = self.local_net.run_policy_and_value(
sess, state)
action = self.pick_action(logits_)
states.append(state)
fullstates.append(fullstate)
actions.append(action)
values.append(value_)
if self.thread_idx == self.log_idx \
and self.local_t % self.log_interval == 0:
log_msg1 = "lg={}".format(
np.array_str(logits_, precision=4, suppress_small=True))
log_msg2 = "pi={}".format(
np.array_str(pi_, precision=4, suppress_small=True))
log_msg3 = "V={:.4f}".format(value_)
logger.debug(log_msg1)
logger.debug(log_msg2)
logger.debug(log_msg3)
# process game
self.game_state.step(action)
# receive game result
reward = self.game_state.reward
terminal = self.game_state.terminal
self.episode_reward += reward
if self.use_sil:
# save states in episode memory
self.episode.add_item(self.game_state.s_t, fullstate, action,
reward, terminal)
if self.reward_type == 'CLIP':
reward = np.sign(reward)
rewards.append(reward)
self.local_t += 1
self.episode_steps += 1
global_t += 1
# s_t1 -> s_t
self.game_state.update()
if terminal:
terminal_pseudo = True
env = self.game_state.env
name = 'EpisodicLifeEnv'
if get_wrapper_by_name(env, name).was_real_done:
# reduce log freq
if self.thread_idx == self.log_idx:
log_msg = "train: worker={} global_t={} local_t={}".format(
self.thread_idx, global_t, self.local_t)
score_str = colored(
"score={}".format(self.episode_reward), "magenta")
steps_str = colored(
"steps={}".format(self.episode_steps), "blue")
log_msg += " {} {}".format(score_str, steps_str)
logger.debug(log_msg)
train_rewards['train'][global_t] = (self.episode_reward,
self.episode_steps)
self.record_summary(score=self.episode_reward,
steps=self.episode_steps,
episodes=None,
global_t=global_t,
mode='Train')
self.episode_reward = 0
self.episode_steps = 0
terminal_end = True
self.game_state.reset(hard_reset=False)
break
cumsum_reward = 0.0
if not terminal:
state = cv2.resize(self.game_state.s_t,
self.local_net.in_shape[:-1],
interpolation=cv2.INTER_AREA)
cumsum_reward = self.local_net.run_value(sess, state)
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
batch_state = []
batch_action = []
batch_adv = []
batch_cumsum_reward = []
# compute and accumulate gradients
for (ai, ri, si, vi) in zip(actions, rewards, states, values):
if self.transformed_bellman:
ri = np.sign(ri) * self.reward_constant + ri
cumsum_reward = transform_h(ri + self.gamma *
transform_h_inv(cumsum_reward))
else:
cumsum_reward = ri + self.gamma * cumsum_reward
advantage = cumsum_reward - vi
# convert action to one-hot vector
a = np.zeros([self.action_size])
a[ai] = 1
batch_state.append(si)
batch_action.append(a)
batch_adv.append(advantage)
batch_cumsum_reward.append(cumsum_reward)
cur_learning_rate = self._anneal_learning_rate(
global_t, self.initial_learning_rate)
feed_dict = {
self.local_net.s: batch_state,
self.local_net.a: batch_action,
self.local_net.advantage: batch_adv,
self.local_net.cumulative_reward: batch_cumsum_reward,
self.learning_rate_input: cur_learning_rate,
}
sess.run(self.apply_gradients, feed_dict=feed_dict)
t = self.local_t - self.prev_local_t
if (self.thread_idx == self.log_idx and t >= self.perf_log_interval):
self.prev_local_t += self.perf_log_interval
elapsed_time = time.time() - self.start_time
steps_per_sec = global_t / elapsed_time
logger.info("worker-{}, log_worker-{}".format(
self.thread_idx, self.log_idx))
logger.info("Performance : {} STEPS in {:.0f} sec. {:.0f}"
" STEPS/sec. {:.2f}M STEPS/hour.".format(
global_t, elapsed_time, steps_per_sec,
steps_per_sec * 3600 / 1000000.))
# return advanced local step size
diff_local_t = self.local_t - start_local_t
return diff_local_t, terminal_end, terminal_pseudo
def __init__(self, thread_index, action_size, env_id,
global_a3c, local_a3c, update_in_rollout, nstep_bc,
global_pretrained_model, local_pretrained_model,
transformed_bellman=False, no_op_max=0,
device='/cpu:0', entropy_beta=0.01, clip_norm=None,
grad_applier=None, initial_learn_rate=0.007,
learning_rate_input=None):
"""Initialize RolloutThread class."""
self.is_refresh_thread = True
self.action_size = action_size
self.thread_idx = thread_index
self.transformed_bellman = transformed_bellman
self.entropy_beta = entropy_beta
self.clip_norm = clip_norm
self.initial_learning_rate = initial_learn_rate
self.learning_rate_input = learning_rate_input
self.no_op_max = no_op_max
self.override_num_noops = 0 if self.no_op_max == 0 else None
logger.info("===REFRESH thread_index: {}===".format(self.thread_idx))
logger.info("device: {}".format(device))
logger.info("action_size: {}".format(self.action_size))
logger.info("reward_type: {}".format(self.reward_type))
logger.info("transformed_bellman: {}".format(
colored(self.transformed_bellman,
"green" if self.transformed_bellman else "red")))
logger.info("update in rollout: {}".format(
colored(update_in_rollout, "green" if update_in_rollout else "red")))
logger.info("N-step BC: {}".format(nstep_bc))
self.reward_clipped = True if self.reward_type == 'CLIP' else False
# setup local a3c
self.local_a3c = local_a3c
self.sync_a3c = self.local_a3c.sync_from(global_a3c)
with tf.device(device):
local_vars = self.local_a3c.get_vars
self.local_a3c.prepare_loss(
entropy_beta=self.entropy_beta, critic_lr=0.5)
var_refs = [v._ref() for v in local_vars()]
self.rollout_gradients = tf.gradients(self.local_a3c.total_loss, var_refs)
global_vars = global_a3c.get_vars
if self.clip_norm is not None:
self.rollout_gradients, grad_norm = tf.clip_by_global_norm(
self.rollout_gradients, self.clip_norm)
self.rollout_gradients = list(zip(self.rollout_gradients, global_vars()))
self.rollout_apply_gradients = grad_applier.apply_gradients(self.rollout_gradients)
# setup local pretrained model
self.local_pretrained = None
if nstep_bc > 0:
assert local_pretrained_model is not None
assert global_pretrained_model is not None
self.local_pretrained = local_pretrained_model
self.sync_pretrained = self.local_pretrained.sync_from(global_pretrained_model)
# setup env
self.rolloutgame = GameState(env_id=env_id, display=False,
no_op_max=0, human_demo=False, episode_life=True,
override_num_noops=0)
self.local_t = 0
self.episode_reward = 0
self.episode_steps = 0
self.action_meaning = self.rolloutgame.env.unwrapped.get_action_meanings()
assert self.local_a3c is not None
if nstep_bc > 0:
assert self.local_pretrained is not None
self.episode = SILReplayMemory(
self.action_size, max_len=None, gamma=self.gamma,
clip=self.reward_clipped,
height=self.local_a3c.in_shape[0],
width=self.local_a3c.in_shape[1],
phi_length=self.local_a3c.in_shape[2],
reward_constant=self.reward_constant)
class RefreshThread(CommonWorker):
"""Rollout Thread Class."""
advice_confidence = 0.8
gamma = 0.99
def __init__(self, thread_index, action_size, env_id,
global_a3c, local_a3c, update_in_rollout, nstep_bc,
global_pretrained_model, local_pretrained_model,
transformed_bellman=False, no_op_max=0,
device='/cpu:0', entropy_beta=0.01, clip_norm=None,
grad_applier=None, initial_learn_rate=0.007,
learning_rate_input=None):
"""Initialize RolloutThread class."""
self.is_refresh_thread = True
self.action_size = action_size
self.thread_idx = thread_index
self.transformed_bellman = transformed_bellman
self.entropy_beta = entropy_beta
self.clip_norm = clip_norm
self.initial_learning_rate = initial_learn_rate
self.learning_rate_input = learning_rate_input
self.no_op_max = no_op_max
self.override_num_noops = 0 if self.no_op_max == 0 else None
logger.info("===REFRESH thread_index: {}===".format(self.thread_idx))
logger.info("device: {}".format(device))
logger.info("action_size: {}".format(self.action_size))
logger.info("reward_type: {}".format(self.reward_type))
logger.info("transformed_bellman: {}".format(
colored(self.transformed_bellman,
"green" if self.transformed_bellman else "red")))
logger.info("update in rollout: {}".format(
colored(update_in_rollout, "green" if update_in_rollout else "red")))
logger.info("N-step BC: {}".format(nstep_bc))
self.reward_clipped = True if self.reward_type == 'CLIP' else False
# setup local a3c
self.local_a3c = local_a3c
self.sync_a3c = self.local_a3c.sync_from(global_a3c)
with tf.device(device):
local_vars = self.local_a3c.get_vars
self.local_a3c.prepare_loss(
entropy_beta=self.entropy_beta, critic_lr=0.5)
var_refs = [v._ref() for v in local_vars()]
self.rollout_gradients = tf.gradients(self.local_a3c.total_loss, var_refs)
global_vars = global_a3c.get_vars
if self.clip_norm is not None:
self.rollout_gradients, grad_norm = tf.clip_by_global_norm(
self.rollout_gradients, self.clip_norm)
self.rollout_gradients = list(zip(self.rollout_gradients, global_vars()))
self.rollout_apply_gradients = grad_applier.apply_gradients(self.rollout_gradients)
# setup local pretrained model
self.local_pretrained = None
if nstep_bc > 0:
assert local_pretrained_model is not None
assert global_pretrained_model is not None
self.local_pretrained = local_pretrained_model
self.sync_pretrained = self.local_pretrained.sync_from(global_pretrained_model)
# setup env
self.rolloutgame = GameState(env_id=env_id, display=False,
no_op_max=0, human_demo=False, episode_life=True,
override_num_noops=0)
self.local_t = 0
self.episode_reward = 0
self.episode_steps = 0
self.action_meaning = self.rolloutgame.env.unwrapped.get_action_meanings()
assert self.local_a3c is not None
if nstep_bc > 0:
assert self.local_pretrained is not None
self.episode = SILReplayMemory(
self.action_size, max_len=None, gamma=self.gamma,
clip=self.reward_clipped,
height=self.local_a3c.in_shape[0],
width=self.local_a3c.in_shape[1],
phi_length=self.local_a3c.in_shape[2],
reward_constant=self.reward_constant)
def record_rollout(self, score=0, steps=0, old_return=0, new_return=0,
global_t=0, rollout_ctr=0, rollout_added_ctr=0,
mode='Rollout', confidence=None, episodes=None):
"""Record rollout summary."""
summary = tf.Summary()
summary.value.add(tag='{}/score'.format(mode),
simple_value=float(score))
summary.value.add(tag='{}/old_return_from_s'.format(mode),
simple_value=float(old_return))
summary.value.add(tag='{}/new_return_from_s'.format(mode),
simple_value=float(new_return))
summary.value.add(tag='{}/steps'.format(mode),
simple_value=float(steps))
summary.value.add(tag='{}/all_rollout_ctr'.format(mode),
simple_value=float(rollout_ctr))
summary.value.add(tag='{}/rollout_added_ctr'.format(mode),
simple_value=float(rollout_added_ctr))
if confidence is not None:
summary.value.add(tag='{}/advice-confidence'.format(mode),
simple_value=float(confidence))
if episodes is not None:
summary.value.add(tag='{}/episodes'.format(mode),
simple_value=float(episodes))
self.writer.add_summary(summary, global_t)
self.writer.flush()
def compute_return_for_state(self, rewards, terminal):
"""Compute expected return."""
length = np.shape(rewards)[0]
returns = np.empty_like(rewards, dtype=np.float32)
if self.reward_clipped:
rewards = np.clip(rewards, -1., 1.)
else:
rewards = np.sign(rewards) * self.reward_constant + rewards
for i in reversed(range(length)):
if terminal[i]:
returns[i] = rewards[i] if self.reward_clipped else transform_h(rewards[i])
else:
if self.reward_clipped:
returns[i] = rewards[i] + self.gamma * returns[i+1]
else:
# apply transformed expected return
exp_r_t = self.gamma * transform_h_inv(returns[i+1])
returns[i] = transform_h(rewards[i] + exp_r_t)
return returns[0]
def update_a3c(self, sess, actions, states, rewards, values, global_t):
cumsum_reward = 0.0
actions.reverse()
states.reverse()
rewards.reverse()
values.reverse()
batch_state = []
batch_action = []
batch_adv = []
batch_cumsum_reward = []
# compute and accumulate gradients
for(ai, ri, si, vi) in zip(actions, rewards, states, values):
if self.transformed_bellman:
ri = np.sign(ri) * self.reward_constant + ri
cumsum_reward = transform_h(
ri + self.gamma * transform_h_inv(cumsum_reward))
else:
cumsum_reward = ri + self.gamma * cumsum_reward
advantage = cumsum_reward - vi
# convert action to one-hot vector
a = np.zeros([self.action_size])
a[ai] = 1
batch_state.append(si)
batch_action.append(a)
batch_adv.append(advantage)
batch_cumsum_reward.append(cumsum_reward)
cur_learning_rate = self._anneal_learning_rate(global_t,
self.initial_learning_rate )
feed_dict = {
self.local_a3c.s: batch_state,
self.local_a3c.a: batch_action,
self.local_a3c.advantage: batch_adv,
self.local_a3c.cumulative_reward: batch_cumsum_reward,
self.learning_rate_input: cur_learning_rate,
}
sess.run(self.rollout_apply_gradients, feed_dict=feed_dict)
return batch_adv
def rollout(self, a3c_sess, folder, pretrain_sess, global_t, past_state,
add_all_rollout, ep_max_steps, nstep_bc, update_in_rollout):
"""Perform one rollout until terminal."""
a3c_sess.run(self.sync_a3c)
if nstep_bc > 0:
pretrain_sess.run(self.sync_pretrained)
_, fs, old_a, old_return, _, _ = past_state
states = []
actions = []
rewards = []
values = []
terminals = []
confidences = []
rollout_ctr, rollout_added_ctr = 0, 0
rollout_new_return, rollout_old_return = 0, 0
terminal_pseudo = False # loss of life
terminal_end = False # real terminal
add = False
self.rolloutgame.reset(hard_reset=True)
self.rolloutgame.restore_full_state(fs)
# check if restore successful
fs_check = self.rolloutgame.clone_full_state()
assert fs_check.all() == fs.all()
del fs_check
start_local_t = self.local_t
self.rolloutgame.step(0)
# prevent rollout too long, set max_ep_steps to be lower than ALE default
# see https://github.com/openai/gym/blob/54f22cf4db2e43063093a1b15d968a57a32b6e90/gym/envs/__init__.py#L635
# but in all games tested, no rollout exceeds ep_max_steps
while ep_max_steps > 0:
state = cv2.resize(self.rolloutgame.s_t,
self.local_a3c.in_shape[:-1],
interpolation=cv2.INTER_AREA)
fullstate = self.rolloutgame.clone_full_state()
if nstep_bc > 0: # LiDER-TA or BC
model_pi = self.local_pretrained.run_policy(pretrain_sess, state)
action, confidence = self.choose_action_with_high_confidence(
model_pi, exclude_noop=False)
confidences.append(confidence) # not using "confidences" for anything
nstep_bc -= 1
else: # LiDER, refresh with current policy
pi_, _, logits_ = self.local_a3c.run_policy_and_value(a3c_sess,
state)
action = self.pick_action(logits_)
confidences.append(pi_[action])
value_ = self.local_a3c.run_value(a3c_sess, state)
values.append(value_)
states.append(state)
actions.append(action)
self.rolloutgame.step(action)
ep_max_steps -= 1
reward = self.rolloutgame.reward
terminal = self.rolloutgame.terminal
terminals.append(terminal)
self.episode_reward += reward
self.episode.add_item(self.rolloutgame.s_t, fullstate, action,
reward, terminal, from_rollout=True)
if self.reward_type == 'CLIP':
reward = np.sign(reward)
rewards.append(reward)
self.local_t += 1
self.episode_steps += 1
global_t += 1
self.rolloutgame.update()
if terminal:
terminal_pseudo = True
env = self.rolloutgame.env
name = 'EpisodicLifeEnv'
rollout_ctr += 1
terminal_end = get_wrapper_by_name(env, name).was_real_done
new_return = self.compute_return_for_state(rewards, terminals)
if not add_all_rollout:
if new_return > old_return:
add = True
else:
add = True
if add:
rollout_added_ctr += 1
rollout_new_return += new_return
rollout_old_return += old_return
# update policy immediate using a good rollout
if update_in_rollout:
batch_adv = self.update_a3c(a3c_sess, actions, states, rewards, values, global_t)
self.episode_reward = 0
self.episode_steps = 0
self.rolloutgame.reset(hard_reset=True)
break
diff_local_t = self.local_t - start_local_t
return diff_local_t, terminal_end, terminal_pseudo, rollout_ctr, \
rollout_added_ctr, add, rollout_new_return, rollout_old_return