def atari_learn(env, num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
dqn.learn(
env,
q_func=atari_model,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
# learning_starts=50000,
learning_starts=50,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10)
env.close()
def lander_learn(
env,
session,
num_timesteps,
# YOUR OWN CODE
seed,
doubleQ=True,
exp_name='doubleQ',
schedule='PiecewiseSchedule',
rew_file='lander_test.pk1'):
# optimizer = lander_optimizer()
# stopping_criterion = lander_stopping_criterion(num_timesteps)
# exploration_schedule = lander_exploration_schedule(num_timesteps)
dqn.learn(
env=env,
session=session,
exploration=lander_exploration_schedule(num_timesteps, schedule),
stopping_criterion=lander_stopping_criterion(num_timesteps),
double_q=doubleQ,
# YOUR OWN CODE
rew_file=rew_file,
seed=seed,
env_name='LunarLander-v2',
exp_name=exp_name,
**lander_kwargs())
env.close()
def learn(env, session, args):
if args.env == 'PongNoFrameskip-v4':
lr_schedule = ConstantSchedule(1e-4)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
limit = max(int(args.num_steps / 2), 2e6)
exploration_schedule = PiecewiseSchedule([
(0, 1.00),
(1e6, 0.10),
(limit, 0.01),
],
outside_value=0.01)
dqn.learn(env=env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10,
double_q=args.double_q,
logdir=args.logdir,
max_steps=args.num_steps)
elif args.env == 'CartPole-v0':
lr_schedule = ConstantSchedule(5e-4)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
exploration_schedule = PiecewiseSchedule([
(0, 1.00),
(5e4, 0.10),
(1e5, 0.02),
],
outside_value=0.02)
dqn.learn(env=env,
q_func=cartpole_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
replay_buffer_size=10000,
batch_size=100,
gamma=0.99,
learning_starts=1000,
learning_freq=4,
frame_history_len=1,
target_update_freq=500,
grad_norm_clipping=10,
double_q=args.double_q,
logdir=args.logdir,
max_steps=args.num_steps,
cartpole=True)
else:
raise ValueError(args.env)
env.close()
def atari_learn(env, session, num_timesteps):
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule(
[(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier)],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdadeltaOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule([(0, 1.0), (1e6, 0.1),
(num_iterations / 2, 0.01)],
outside_value=0.01)
dqn.learn(env=env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10,
double_q=True)
env.close()
def atari_learn(
env,
session,
num_timesteps,
# YOUR OWN CODE
seed,
doubleQ=True,
exp_name='doubleQ',
rew_file='ram_test.pk1'):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 0.2),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
dqn.learn(
env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=1,
target_update_freq=10000,
grad_norm_clipping=10,
# YOUR OWN CODE
double_q=doubleQ,
rew_file=rew_file,
seed=seed,
env_name='Pong-ram-v0',
exp_name=exp_name)
env.close()
def atari_learn(env,
session,
num_timesteps,
double_q,
explore,
env_name,
ex2=ex2,
coef=coef):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
# therefore, the exploration gradually decrease
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
# TO-DO: Pay attention to arg here, double_q
dqn.learn(env=env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10,
double_q=double_q,
rew_file='./pkl/' + env_name + '_' +
time.strftime("%d-%m-%Y_%H-%M-%S") + '.pkl',
explore=explore,
ex2=ex2,
coef=coef)
env.close()
def atari_learn(env,
session,
num_timesteps,
lr_multiplier):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(
constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule
)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule(
[
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
], outside_value=0.01
)
if not (os.path.exists('data')):
os.makedirs('data')
logdir = os.path.join('data', 'PongNoFrameskip-v4')
if not (os.path.exists(logdir)):
os.makedirs(logdir)
dqn.learn(
env=env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10,
double_q=False,
# rew_file='%s_lr_%s.pkl' % (os.path.join(logdir, time.strftime("%d-%m-%Y_%H-%M-%S")), str(lr_multiplier))
rew_file='%s_lr_%s.pkl' % (os.path.join(logdir, time.strftime("%d-%m-%Y_%H-%M-%S")), 'vanilla')
)
env.close()
def atari_learn(env,
session,
args,
num_timesteps):
logdir = os.path.join('data', args.exp_name)
#if not(os.path.exists(logdir)):
#os.makedirs(logdir)
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(
constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule
)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule(
[
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
], outside_value=0.01
)
dqn.learn(
env=env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=args.gamma,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10,
double_q=args.double_q,
logdir=logdir
)
env.close()
def atari_learn(env, session, num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
if (t % 10000 == 0):
print("get_total_steps:" +
str(get_wrapper_by_name(env, "Monitor").get_total_steps()) +
", t:" + str(t) + ", num_timesteps:" + str(num_timesteps))
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
exploration_schedule2 = PiecewiseSchedule([
(0, 1.0),
(2e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
dqn.learn(
env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule, #pipaek
stopping_criterion=stopping_criterion,
#replay_buffer_size=1000000,
replay_buffer_size=2000000, #pipaek
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10)
env.close()
def cartpole_learn(env, session, num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
# lr_multiplier = 1.0
# lr_multiplier = 0.1
# lr_schedule = PiecewiseSchedule([
# (0, 1e-4 * lr_multiplier),
# (num_iterations / 2, 1e-5 * lr_multiplier),
# ],
# outside_value=5e-5 * lr_multiplier)
lr_schedule = InverseSchedule(initial_p=0.1, gamma=0.6)
optimizer = dqn.OptimizerSpec(
constructor=tf.train.GradientDescentOptimizer,
# constructor=tf.train.AdamOptimizer,
# kwargs=dict(epsilon=1e-4),
kwargs=dict(),
# constructor=tf.train.RMSPropOptimizer,
# kwargs=dict(epsilon=1e-1),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule(
[
(0, 1.0),
# (0.2 * num_timesteps, 0.9),
# (0.5 * num_timesteps, 0.5),
(0.1 * num_timesteps, 0.1),
],
outside_value=0.01)
dqn.learn(
env,
q_func=cartpole_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=100000,
batch_size=256,
gamma=0.99,
learning_starts=2000,
learning_freq=1,
frame_history_len=4,
target_update_freq=1000,
grad_norm_clipping=1000,
)
env.close()
def lander_learn(env, session, num_timesteps, seed):
optimizer = lander_optimizer()
stopping_criterion = lander_stopping_criterion(num_timesteps)
exploration_schedule = lander_exploration_schedule(num_timesteps)
dqn.learn(env=env,
session=session,
exploration=lander_exploration_schedule(num_timesteps),
stopping_criterion=lander_stopping_criterion(num_timesteps),
double_q=True,
**lander_kwargs())
env.close()
def arm_learn(env,
session,
scope_name,
num_timesteps,
spec_file=None,
exp_dir=None):
# # This is just a rough estimate
# num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_timesteps / 40, 1e-4 * lr_multiplier),
(num_timesteps / 8, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(t):
return t >= num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(num_timesteps / 20, 0.3),
(num_timesteps / 10, 0.1),
(num_timesteps / 2, 0.01),
],
outside_value=0.01)
dqn.learn(env,
q_func=arm_model,
optimizer_spec=optimizer,
session=session,
scope_name=scope_name,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=2000,
learning_freq=1,
frame_history_len=1,
target_update_freq=500,
grad_norm_clipping=10,
log_every_n_steps=500,
spec_file=spec_file,
exp_dir=exp_dir)
ep_rew = env.get_episode_rewards()
ep_len = env.get_episode_lengths()
return ep_rew, ep_len
def atari_learn(env, session, num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 0.2),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
# Initialize Logging Dir
data_path = osp.join(osp.dirname(osp.realpath(__file__)), 'data')
if not (osp.exists(data_path)):
os.makedirs(data_path)
logdir = 'dqn_' + env.spec.id + '_' + time.strftime("%d-%m-%Y_%H-%M-%S")
logdir = osp.join(data_path, logdir)
dqn.learn(env=env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=1,
target_update_freq=10000,
grad_norm_clipping=10,
logdir=logdir)
env.close()
def atari_learn(env, session, num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
if REW_FILE == 'LinearSchedule':
exploration_schedule = LinearSchedule(num_iterations,
final_p=0.01,
initial_p=1.0)
elif REW_FILE == 'ConstantSchedule':
exploration_schedule = ConstantSchedule(0.05)
else:
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(num_iterations / 5, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
dqn.learn(env=env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=100000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10,
double_q=True,
rew_file=REW_FILE)
env.close()
def lander_learn(env, session, discount, num_timesteps, batch_size, double):
optimizer = lander_optimizer()
stopping_criterion = lander_stopping_criterion(num_timesteps)
exploration_schedule = lander_exploration_schedule(num_timesteps)
dqn.learn(env=env,
session=session,
exploration=lander_exploration_schedule(num_timesteps),
stopping_criterion=lander_stopping_criterion(num_timesteps),
batch_size=batch_size,
gamma=discount,
double_q=double,
**lander_kwargs())
env.close()
def atari_learn(env, session, discount, num_timesteps, batch_size, double,
target_update_freq, **kwargs):
# [Mehran Shakerinava] change end
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
dqn.learn(
env=env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
grad_norm_clipping=10,
# [Mehran Shakerinava] change begin
target_update_freq=target_update_freq,
batch_size=batch_size,
gamma=discount,
double_q=double
# [Mehran Shakerinava] change end
)
env.close()
def knapsack_learn(env,
session,
num_timesteps,
lr_multiplier=1.0,
target_update_freq=10000,
exp_name='Knapsack_DQN',
boltzmann_exploration=False):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
dqn.learn(env,
q_func=knapsack_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=512,
gamma=1,
learning_starts=5000,
learning_freq=4,
frame_history_len=4,
target_update_freq=target_update_freq,
grad_norm_clipping=10,
exp_name=exp_name,
boltzmann_exploration=boltzmann_exploration)
env.close()
def main():
args = get_args()
env = make_atari_env(args.env, args.seed)
benchmark_env = make_atari_env(args.env, args.seed + 1)
optimizer = tf.train.AdamOptimizer(learning_rate=args.lr, epsilon=1e-4)
exploration_schedule = utils.PiecewiseSchedule(
[
(0, 1.0),
(args.prepopulate, 1.0),
(args.prepopulate + args.explore_time, args.final_eps),
],
outside_value=args.final_eps,
)
if not args.legacy:
assert args.train_freq == 4 # Training frequency is undefined for DQN(lambda)
replay_memory = make_replay_memory(args.return_est, args.mem_size,
args.history_len, args.discount,
args.cache_size, args.block_size,
args.priority)
else:
assert args.cache_size == 80000 # Cache-related args are undefined for legacy DQN
assert args.priority == 0.0
assert args.block_size == 100
replay_memory = make_legacy_replay_memory(args.return_est,
args.mem_size,
args.history_len,
args.discount)
with utils.make_session(args.seed) as session:
dqn.learn(
session,
env,
benchmark_env,
atari_cnn,
replay_memory,
optimizer,
exploration_schedule,
args.timesteps,
args.batch_size,
args.prepopulate,
args.update_freq,
train_freq=args.train_freq,
grad_clip=args.grad_clip,
log_every_n_steps=1000,
)
env.close()
def atari_learn(env,
session,
num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(
constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule
)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule(
[
(0, 0.2),
(1e6, 0.1),
(num_iterations / 2, 0.01),
], outside_value=0.01
)
dqn.learn(
env,
q_func=atari_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=1,
target_update_freq=10000,
grad_norm_clipping=10
)
env.close()
def atari_learn(env,
num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
LEARNING_RATE = 5e-5
lr_multiplier = 3.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(
constructor=optim.Adam,
kwargs=dict(lr=LEARNING_RATE, eps=1e-4)
)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule(
[
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
], outside_value=0.01
)
dqn.learn(
env,
q_func=atari_model,
optimizer_spec=optimizer,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10,
num_target_values = 10
)
env.close()
def lander_learn(env, session, num_timesteps, seed):
optimizer = lander_optimizer()
stopping_criterion = lander_stopping_criterion(num_timesteps)
exploration_schedule = lander_exploration_schedule(num_timesteps)
dqn.learn(env=env,
session=session,
exploration=lander_exploration_schedule(num_timesteps),
stopping_criterion=lander_stopping_criterion(num_timesteps),
double_q=True,
save_name='lander_ddqn_replay500000_target5000',
save=True,
test=False,
**lander_kwargs())
env.close()
def atari_learn(env, num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return env.get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule(
[
# (0, 1.0),
# (1e6, 0.1),
(0, 0.4),
(5e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
lr_schedule = dict(milestones=[num_iterations / 2], gamma=0.5)
dqn.learn(
env=env,
lr_schedule=lr_schedule,
load_path='model/step_2400000.pth.tar',
# load_path=None,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=700000,
# replay_buffer_size=70000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
# learning_starts=50,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10,
double_q=True)
env.close()
def lander_learn(env, session, num_timesteps, seed, double_q, explore):
optimizer = lander_optimizer()
stopping_criterion = lander_stopping_criterion(num_timesteps)
exploration_schedule = lander_exploration_schedule(num_timesteps)
dqn.learn(
env=env,
session=session,
exploration=lander_exploration_schedule(num_timesteps),
stopping_criterion=lander_stopping_criterion(num_timesteps),
# double_q=True,
double_q=double_q,
rew_file='./pkl/lander_' + time.strftime("%d-%m-%Y_%H-%M-%S") + '.pkl',
explore=explore,
**lander_kwargs())
env.close()
def atari_learn(env, session, num_timesteps, result_dir):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
lander_optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs={},
lr_schedule=ConstantSchedule(1e-3))
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 1),
(num_timesteps * 0.1, 0.02),
],
outside_value=0.02)
dqn.learn(
env=env,
q_func=lander_model,
optimizer_spec=lander_optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=50000,
batch_size=32,
gamma=1,
learning_starts=1000,
learning_freq=1,
frame_history_len=1,
target_update_freq=3000,
grad_norm_clipping=10,
lander=True,
rew_file=osp.join(result_dir, 'episode_rewards.pkl'),
)
env.close()
def lander_learn(env, session, num_timesteps, seed):
# Initialize Logging Dir
data_path = osp.join(osp.dirname(osp.realpath(__file__)), 'data')
if not (osp.exists(data_path)):
os.makedirs(data_path)
logdir = 'dqn_' + env.spec.id + '_' + time.strftime("%d-%m-%Y_%H-%M-%S")
logdir = osp.join(data_path, logdir)
dqn.learn(env=env,
session=session,
exploration=lander_exploration_schedule(num_timesteps),
stopping_criterion=lander_stopping_criterion(num_timesteps),
double_q=False,
logdir=logdir,
**lander_kwargs())
env.close()
def lander_learn(env, session, seed, exp_name, num_timesteps, double_q,
replay_buffer_size):
optimizer = lander_optimizer()
stopping_criterion = lander_stopping_criterion(num_timesteps)
exploration_schedule = lander_exploration_schedule(num_timesteps)
dqn.learn(env=env,
session=session,
exp_name=exp_name,
seed=seed,
exploration=lander_exploration_schedule(num_timesteps),
stopping_criterion=lander_stopping_criterion(num_timesteps),
double_q=double_q,
replay_buffer_size=replay_buffer_size,
**lander_kwargs())
env.close()
def atari_learn(env, session, num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
dqn.learn(
env,
q_func=atari_model, #just the neural network as defined above
optimizer_spec=
optimizer, #just a named tuple containing the grad alg, lr_shedule etc..
session=session, #The tf session
exploration=exploration_schedule, #epsilon greedy schedule
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000, #Replay buffer size... size of what?
batch_size=32, #Gradient ascent batch size I guess..
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10)
env.close()
def run(worker_id, monitor, args):
task_id, env_code, agent = args
env = gym.make(env_code)
_, reward_history, _ = dqn.learn(env,
agent,
monitor=monitor,
worker_id=worker_id)
env.close()
return reward_history
def atari_learn(env, session, num_timesteps, model, double_q, logdir):
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_timesteps / 10, 1e-4 * lr_multiplier),
(num_timesteps / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return get_wrapper_by_name(
env, "Monitor").get_total_steps() >= 4 * num_timesteps
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(1e6, 0.1),
(num_timesteps / 2, 0.01),
],
outside_value=0.01)
dqn.learn(env=env,
q_func=globals()["atari_model_" + model],
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10,
double_q=double_q,
rew_file=None,
logdir=logdir)
env.close()
def knapsack_learn(env, num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
return False
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
dqn.learn(env,
q_func=knapsack_model,
nn_size=3,
n_hidden_units=128,
exploration=exploration_schedule,
stopping_criterion=stopping_criterion,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
target_update_freq=10000,
grad_norm_clipping=10,
double_DQN=True,
n_steps_ahead=3)
env.close()
def game_learn(env, session, num_timesteps):
# This is just a rough estimate
num_iterations = float(num_timesteps) / 4.0
lr_multiplier = 1.0
lr_schedule = PiecewiseSchedule([
(0, 1e-4 * lr_multiplier),
(num_iterations / 10, 1e-4 * lr_multiplier),
(num_iterations / 2, 5e-5 * lr_multiplier),
],
outside_value=5e-5 * lr_multiplier)
optimizer = dqn.OptimizerSpec(constructor=tf.train.AdamOptimizer,
kwargs=dict(epsilon=1e-4),
lr_schedule=lr_schedule)
def stopping_criterion(env, t):
# notice that here t is the number of steps of the wrapped env,
# which is different from the number of steps in the underlying env
pass
exploration_schedule = PiecewiseSchedule([
(0, 1.0),
(1e6, 0.1),
(num_iterations / 2, 0.01),
],
outside_value=0.01)
dqn.learn(env,
q_func=cnn_model,
optimizer_spec=optimizer,
session=session,
exploration=exploration_schedule,
stopping_criterion=None,
replay_buffer_size=1000000,
batch_size=32,
gamma=0.99,
learning_starts=50000,
learning_freq=4,
frame_history_len=4,
target_update_freq=10000,
grad_norm_clipping=10)
env.close()