def main():
"""Run DQN until the environment throws an exception."""
base_path = "results/rainbow/6/"
env = make_env(stack=False, scale_rew=False, render=None, monitor=base_path + "train_monitor",
episodic_life=True)
# I think the env itself allows Backtracking
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.8
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n, gym_space_vectorizer(env.observation_space),
min_val=-200, max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4)
saver = tf.train.Saver(name="rainbow")
sess.run(tf.global_variables_initializer())
saver.save(sess, base_path + "training", global_step=0)
try:
dqn.train(num_steps=2_000_000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000, 0.5, 0.4, epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=64,
min_buffer_size=20000,
handle_ep=handle_ep) # in seconds
except KeyboardInterrupt:
print("keyboard interrupt")
print("finishing")
saver.save(sess, base_path + "final", global_step=2_000_000)
def main():
env = AllowBacktracking(make_env(stack=False, scale_rew=False))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-421,
max_val=421))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4)
sess.run(tf.global_variables_initializer())
dqn.train(num_steps=2000000,
player=player,
replay_buffer=PrioritizedReplayBuffer(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=64,
batch_size=32,
min_buffer_size=25000)
def main():
"""Run DQN until the environment throws an exception."""
env = make(game='SonicTheHedgehog-Genesis', state='GreenHillZone.Act1')
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4)
sess.run(tf.global_variables_initializer())
dqn.train(
num_steps=2000000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=32,
min_buffer_size=20000)
def main():
"""Run DQN until the environment throws an exception."""
env = AllowBacktracking(make_env(stack=False, scale_rew=False))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4)
sess.run(tf.global_variables_initializer())
dqn.train(
num_steps=2000000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=StochasticMaxStochasticDeltaDeletionPRB(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=32,
min_buffer_size=20000)
def main():
"""Run DQN until the environment throws an exception."""
env = make(game='SonicTheHedgehog-Genesis', state='GreenHillZone.Act1')
env = AllowBacktracking(make_local_env(env, stack=False, scale_rew=False))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4)
sess.run(tf.global_variables_initializer())
dqn.train(num_steps=num_steps, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000, 0.5, 0.4, epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=32,
min_buffer_size=20000)
print(tf.trainable_variables())
save_path='/home/noob/retro-noob/rainbow/params/params'
utils.save_state(save_path+'_tf_saver')
with tf.variable_scope('model'):
params = tf.trainable_variables()
ps = sess.run(params)
joblib.dump(ps, save_path + '_joblib')
def main():
"""Run DQN until the environment throws an exception."""
env = AllowBacktracking(make_env(stack=False, scale_rew=False))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
"""
Create a TF Op that optimizes the objective.
Args:
learning_rate: the Adam learning rate.
epsilon: the Adam epsilon.
"""
optimize = dqn.optimize(learning_rate=6.25e-5, epsilon=1.5e-4)
sess.run(tf.global_variables_initializer())
"""
Run an automated training loop.
This is meant to provide a convenient way to run a
standard training loop without any modifications.
You may get more flexibility by writing your own
training loop.
Args:
num_steps: the number of timesteps to run.
player: the Player for gathering experience.
replay_buffer: the ReplayBuffer for experience.
optimize_op: a TF Op to optimize the model.
train_interval: timesteps per training step.
target_interval: number of timesteps between
target network updates.
batch_size: the size of experience mini-batches.
min_buffer_size: minimum replay buffer size
before training is performed.
tf_schedules: a sequence of TFSchedules that are
updated with the number of steps taken.
handle_ep: called with information about every
completed episode.
timeout: if set, this is a number of seconds
after which the training loop should exit.
"""
dqn.train(
num_steps=1000000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=32,
min_buffer_size=20000)
def main():
"""Run DQN until the environment throws an exception."""
env = AllowBacktracking(make_env(stack=False, scale_rew=False))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
# Other exploration schedules
#eps_decay_sched = LinearTFSchedule(50000, 1.0, 0.01)
#player = NStepPlayer(BatchedPlayer(env, EpsGreedyQNetwork(dqn.online_net, 0.1)), 3)
#player = NStepPlayer(BatchedPlayer(env, EpsGreedyQNetwork(dqn.online_net, TFScheduleValue(sess, eps_decay_sched))), 3)
#player = NStepPlayer(BatchedPlayer(env, SonicEpsGreedyQNetwork(dqn.online_net, TFScheduleValue(sess, eps_decay_sched))), 3)
optimize = dqn.optimize(learning_rate=1e-4)
sess.run(tf.global_variables_initializer())
reward_hist = []
total_steps = 0
def _handle_ep(steps, rew, env_rewards):
nonlocal total_steps
total_steps += steps
reward_hist.append(rew)
if total_steps % 10 == 0:
print('%d episodes, %d steps: mean of last 100 episodes=%f' %
(len(reward_hist), total_steps,
sum(reward_hist[-100:]) / len(reward_hist[-100:])))
dqn.train(
num_steps=2000000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=32,
min_buffer_size=20000,
tf_schedules=[eps_decay_sched],
handle_ep=_handle_ep,
restore_path='./pretrained_model',
save_interval=None,
)
def main():
"""Run DQN until the environment throws an exception."""
env_fns, env_names = create_envs()
env = BatchedFrameStack(batched_gym_env(env_fns),
num_images=4,
concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4) # Use ADAM
sess.run(tf.global_variables_initializer())
reward_hist = []
total_steps = 0
def _handle_ep(steps, rew, env_rewards):
nonlocal total_steps
total_steps += steps
reward_hist.append(rew)
if total_steps % 1 == 0:
print('%d episodes, %d steps: mean of last 100 episodes=%f' %
(len(reward_hist), total_steps,
sum(reward_hist[-100:]) / len(reward_hist[-100:])))
dqn.train(
num_steps=
2000000000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=32,
min_buffer_size=20000,
handle_ep=_handle_ep,
num_envs=len(env_fns),
save_interval=10,
)
def main():
"""Run DQN until the environment throws an exception."""
# "results/rainbow/2/videos/6"
env = make_env(stack=False,
scale_rew=False,
render=20,
monitor=None,
timelimit=False)
# env = AllowBacktracking(make_env(stack=False, scale_rew=False))
# TODO we might not want to allow backtracking, it kinda hurts in mario
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
config.gpu_options.per_process_gpu_memory_fraction = 0.6
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
# TODO rebuild the online_net form the saved model
# type <anyrl.models.dqn_dist.NatureDistQNetwork object at ???>
# important methods
#
model = dqn.online_net
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
with tf.device("/cpu"):
# sess.run(tf.global_variables_initializer())
vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
try:
for i in tqdm(range(100000)):
trajectories = player.play()
for trajectori in trajectories:
trajectori
pass
except KeyboardInterrupt:
env.close()
def main():
"""Run DQN until the environment throws an exception."""
env = AllowBacktracking(make_env(stack=False, scale_rew=False))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(sess, "/root/compo/model.ckpt")
#print('model restored')
replay_buffer = pickle.load(
gzip.open('/root/compo/replay_buffer.p.gz', 'rb'))
replay_buffer.alpha = 0.2
replay_buffer.beta = 0.4
replay_buffer.capacity = 100000
restore_ppo2_weights(sess)
dqn.train(
num_steps=2000000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=
replay_buffer, #PrioritizedReplayBuffer(500000, 0.5, 0.4, epsilon=0.1),
optimize_op=optimize,
train_interval=4,
target_interval=8192,
batch_size=32,
min_buffer_size=20000)
def make_net(name):
return rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200)
def main():
"""Run DQN until the environment throws an exception."""
print('creating env')
env = AllowBacktracking(make_env(stack=False, scale_rew=False))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
print('starting tf session')
with tf.Session(config=config) as sess:
print('creating agent')
online_net, target_net = rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(
env.observation_space),
min_val=-200,
max_val=200)
dqn = DQN(online_net, target_net)
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4)
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
train_steps = 5000
print('training steps:', train_steps)
for j in range(1):
print(j)
start = time.time()
dqn.train(
num_steps=
train_steps, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=32,
min_buffer_size=10000)
end = time.time()
print(end - start)
print('done training')
print('save nn')
save_path = saver.save(sess, "saved_models/rainbow5.ckpt")
print("Model saved in path: %s" % save_path)
tvars = tf.trainable_variables()
tvars_vals = sess.run(tvars)
#for var, val in zip(tvars, tvars_vals):
# print(var.name, val[0])
#print(tvars_vals[0][-5:])
#print('stepping')
#obs = env.reset()
#online_net.step(obs, obs)
'''
def main():
"""
Entry-point for the program.
"""
args = _parse_args()
# batched env = creates gym env, not sure what batched means
# make_single_env = GrayscaleEnv > DownsampleEnv
# GrayscaleEnv = turns RGB into grayscale
# DownsampleEnv = down samples observation by N times where N is the specified variable (e.g. 2x smaller)
env = batched_gym_env([partial(make_single_env, args.game)] * args.workers)
env_test = make_single_env(args.game)
#make_single_env(args.game)
print('OBSSSS', env_test.observation_space)
#env = CustomWrapper(args.game)
# Using BatchedFrameStack with concat=False is more
# memory efficient than other stacking options.
env = BatchedFrameStack(env, num_images=4, concat=False)
with tf.Session() as sess:
def make_net(name):
return rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200)
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = BatchedPlayer(env,
EpsGreedyQNetwork(dqn.online_net, args.epsilon))
optimize = dqn.optimize(learning_rate=args.lr)
sess.run(tf.global_variables_initializer())
reward_hist = []
total_steps = 0
def _handle_ep(steps, rew):
nonlocal total_steps
total_steps += steps
reward_hist.append(rew)
if len(reward_hist) == REWARD_HISTORY:
print('%d steps: mean=%f' %
(total_steps, sum(reward_hist) / len(reward_hist)))
reward_hist.clear()
dqn.train(num_steps=int(1e7),
player=player,
replay_buffer=UniformReplayBuffer(args.buffer_size),
optimize_op=optimize,
target_interval=args.target_interval,
batch_size=args.batch_size,
min_buffer_size=args.min_buffer_size,
handle_ep=_handle_ep)
env.close()
def train(batched_env,
env_count=1,
batch_size_multiplier=32,
num_steps=2000000,
pretrained_model='artifacts/model/model.cpkt',
output_dir='artifacts/model',
use_schedules=True):
"""
Trains on a batched_env using anyrl-py's dqn and rainbow model.
env_count: The number of envs in batched_env
batch_size_multiplier: batch_size of the dqn train call will be env_count * batch_size_multiplier
num_steps: The number of steps to run training for
pretrained_model: Load tf weights from this model file
output_dir: Save tf weights to this file
use_schedules: Enables the tf_schedules for the train call. Schedules require internet access, so don't include on
retro-contest evaluation server
"""
env = CollisionMapWrapper(batched_env)
env = BatchedResizeImageWrapper(env)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
scheduled_saver = ScheduledSaver(save_interval=10000,
save_dir=output_dir)
print('Outputting trained model to', output_dir)
# Reporting uses BatchedPlayer to get _total_rewards
batched_player = BatchedPlayer(env, dqn.online_net)
player = NStepPlayer(batched_player, 3)
optimize = dqn.optimize(learning_rate=1e-4)
if pretrained_model is None:
print('Initializing with random weights')
sess.run(tf.global_variables_initializer())
else:
print('Loading pre-trained model from', pretrained_model)
scheduled_saver.saver.restore(sess, pretrained_model)
print('Beginning Training, steps', num_steps)
tf_schedules = []
if (use_schedules):
tf_schedules = [
scheduled_saver,
LosswiseSchedule(num_steps, batched_player),
LoadingBar(num_steps)
]
print(env_count * batch_size_multiplier)
dqn.train(
num_steps=num_steps,
player=player,
replay_buffer=PrioritizedReplayBuffer(300000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=env_count,
target_interval=8192,
batch_size=env_count * batch_size_multiplier,
min_buffer_size=max(4500, env_count * batch_size_multiplier),
# min_buffer_size=60,
tf_schedules=tf_schedules,
handle_ep=print)
scheduled_saver.save(sess)
#print(env.action_space.n)
#StackedBox(84,84,1)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
print('starting tf session')
with tf.Session(config=config) as sess:
print('creating agent')
online_net, target_net = rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(
env.observation_space),
min_val=-200,
max_val=200)
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
saver.restore(sess, "saved_models/rainbow4.ckpt")
print('model loaded')
# RESET
env.reset_start()
obs = env.reset_wait()
#print(obs1[0][0].shape)
def main():
"""Run DQN until the environment throws an exception."""
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
comm = MPI.COMM_WORLD
# Use MPI for parallel evaluation
rank = comm.Get_rank()
size = comm.Get_size()
env_fns, env_names = create_eval_envs()
env = AllowBacktracking(env_fns[rank](stack=False, scale_rew=False))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4)
sess.run(tf.global_variables_initializer())
reward_hist = []
total_steps = 0
def _handle_ep(steps, rew, env_rewards):
nonlocal total_steps
total_steps += steps
reward_hist.append(rew)
if total_steps % 1 == 0:
avg_score = sum(reward_hist[-100:]) / len(reward_hist[-100:])
# Global Score
global_score = np.zeros(1)
local_score = np.array(avg_score)
print("Local Score for " + env_names[rank] + " at episode " +
str(len(reward_hist)) + " with timesteps: " +
str(total_steps) + ": " + str(local_score))
comm.Allreduce(local_score, global_score, op=MPI.SUM)
global_score /= size
if rank == 0:
print("Global Average Score at episode: " +
str(len(reward_hist)) + ": " + str(global_score))
dqn.train(
num_steps=2000000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=32,
min_buffer_size=20000,
handle_ep=_handle_ep,
save_interval=None,
restore_path=
'./checkpoints_rainbow/model-10' # Model to be evaluated
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--restore',
'-restore',
action='store_true',
help='restore from checkpoint file')
parser.add_argument('--record',
'-record',
action='store_true',
help='record bk2 movies')
args = parser.parse_args()
"""Run DQN until the environment throws an exception."""
env = AllowBacktracking(
make_env(stack=False, scale_rew=False, record=args.record))
env = BatchedFrameStack(BatchedGymEnv([[env]]), num_images=4, concat=False)
checkpoint_dir = os.path.join(os.getcwd(), 'results')
results_dir = os.path.join(os.getcwd(), 'results',
time.strftime("%d-%m-%Y_%H-%M-%S"))
if not os.path.exists(results_dir):
os.makedirs(results_dir)
summary_writer = tf.summary.FileWriter(results_dir)
# TODO
# env = wrappers.Monitor(env, results_dir, force=True)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # pylint: disable=E1101
with tf.Session(config=config) as sess:
dqn = DQN(*rainbow_models(sess,
env.action_space.n,
gym_space_vectorizer(env.observation_space),
min_val=-200,
max_val=200))
saver = tf.train.Saver()
if args.restore:
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
if latest_checkpoint:
print("Loading model checkpoint {} ...\n".format(
latest_checkpoint))
saver.restore(sess, latest_checkpoint)
else:
print("Checkpoint not found")
player = NStepPlayer(BatchedPlayer(env, dqn.online_net), 3)
optimize = dqn.optimize(learning_rate=1e-4)
sess.run(tf.global_variables_initializer())
reward_hist = []
total_steps = 0
# runs with every completed episode
def _handle_ep(steps, rew):
nonlocal total_steps
total_steps += steps
reward_hist.append(rew)
summary_reward = tf.Summary()
summary_reward.value.add(tag='global/reward', simple_value=rew)
summary_writer.add_summary(summary_reward, global_step=total_steps)
print('save model')
saver.save(sess=sess,
save_path=checkpoint_dir + '/model',
global_step=total_steps)
if len(reward_hist) == REWARD_HISTORY:
print('%d steps: mean=%f' %
(total_steps, sum(reward_hist) / len(reward_hist)))
summary_meanreward = tf.Summary()
summary_meanreward.value.add(tag='global/mean_reward',
simple_value=sum(reward_hist) /
len(reward_hist))
summary_writer.add_summary(summary_meanreward,
global_step=total_steps)
reward_hist.clear()
dqn.train(
num_steps=7000000, # Make sure an exception arrives before we stop.
player=player,
replay_buffer=PrioritizedReplayBuffer(500000,
0.5,
0.4,
epsilon=0.1),
optimize_op=optimize,
train_interval=1,
target_interval=8192,
batch_size=32,
min_buffer_size=20000,
handle_ep=_handle_ep)
