def runner(env,
policy_func,
load_model_path,
timesteps_per_batch,
number_trajs,
stochastic_policy,
save=False,
reuse=False):
# Setup network
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_func("pi", ob_space, ac_space, reuse=reuse)
U.initialize()
# Prepare for rollouts
# ----------------------------------------
U.load_state(load_model_path)
obs_list = []
acs_list = []
len_list = []
ret_list = []
for _ in tqdm(range(number_trajs)):
traj = traj_1_generator(pi,
env,
timesteps_per_batch,
stochastic=stochastic_policy)
obs, acs, ep_len, ep_ret = traj['ob'], traj['ac'], traj[
'ep_len'], traj['ep_ret']
obs_list.append(obs)
acs_list.append(acs)
len_list.append(ep_len)
ret_list.append(ep_ret)
if stochastic_policy:
print('stochastic policy:')
else:
print('deterministic policy:')
if save:
filename = load_model_path.split('/')[-1] + '.' + env.spec.id
np.savez(filename,
obs=np.array(obs_list),
acs=np.array(acs_list),
lens=np.array(len_list),
rets=np.array(ret_list))
avg_len = sum(len_list) / len(len_list)
avg_ret = sum(ret_list) / len(ret_list)
print("Average length:", avg_len)
print("Average return:", avg_ret)
return avg_len, avg_ret
def learn(env,
policy_func,
dataset,
optim_batch_size=128,
max_iters=1e4,
adam_epsilon=1e-5,
optim_stepsize=3e-4,
ckpt_dir=None,
log_dir=None,
task_name=None,
verbose=False):
val_per_iter = int(max_iters / 10)
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_func("pi", ob_space,
ac_space) # Construct network for new policy
# placeholder
ob = U.get_placeholder_cached(name="ob")
ac = pi.pdtype.sample_placeholder([None])
stochastic = U.get_placeholder_cached(name="stochastic")
loss = tf.reduce_mean(tf.square(ac - pi.ac))
var_list = pi.get_trainable_variables()
adam = MpiAdam(var_list, epsilon=adam_epsilon)
lossandgrad = U.function([ob, ac, stochastic],
[loss] + [U.flatgrad(loss, var_list)])
U.initialize()
adam.sync()
logger.log("Pretraining with Behavior Cloning...")
for iter_so_far in tqdm(range(int(max_iters))):
ob_expert, ac_expert = dataset.get_next_batch(optim_batch_size,
'train')
train_loss, g = lossandgrad(ob_expert, ac_expert, True)
adam.update(g, optim_stepsize)
if verbose and iter_so_far % val_per_iter == 0:
ob_expert, ac_expert = dataset.get_next_batch(-1, 'val')
val_loss, _ = lossandgrad(ob_expert, ac_expert, True)
logger.log("Training loss: {}, Validation loss: {}".format(
train_loss, val_loss))
if ckpt_dir is None:
savedir_fname = tempfile.TemporaryDirectory().name
else:
savedir_fname = osp.join(ckpt_dir, task_name)
U.save_state(savedir_fname, var_list=pi.get_variables())
return savedir_fname
def test_runningmeanstd():
for (x1, x2, x3) in [
(np.random.randn(3), np.random.randn(4), np.random.randn(5)),
(np.random.randn(3, 2), np.random.randn(4, 2), np.random.randn(5, 2)),
]:
rms = RunningMeanStd(epsilon=0.0, shape=x1.shape[1:])
U.initialize()
x = np.concatenate([x1, x2, x3], axis=0)
ms1 = [x.mean(axis=0), x.std(axis=0)]
rms.update(x1)
rms.update(x2)
rms.update(x3)
ms2 = [rms.mean.eval(), rms.std.eval()]
assert np.allclose(ms1, ms2)
def test_dist():
np.random.seed(0)
p1, p2, p3 = (np.random.randn(3, 1), np.random.randn(4, 1),
np.random.randn(5, 1))
q1, q2, q3 = (np.random.randn(6, 1), np.random.randn(7, 1),
np.random.randn(8, 1))
# p1,p2,p3=(np.random.randn(3), np.random.randn(4), np.random.randn(5))
# q1,q2,q3=(np.random.randn(6), np.random.randn(7), np.random.randn(8))
comm = MPI.COMM_WORLD
assert comm.Get_size() == 2
if comm.Get_rank() == 0:
x1, x2, x3 = p1, p2, p3
elif comm.Get_rank() == 1:
x1, x2, x3 = q1, q2, q3
else:
assert False
rms = RunningMeanStd(epsilon=0.0, shape=(1, ))
U.initialize()
rms.update(x1)
rms.update(x2)
rms.update(x3)
bigvec = np.concatenate([p1, p2, p3, q1, q2, q3])
def checkallclose(x, y):
print(x, y)
return np.allclose(x, y)
assert checkallclose(
bigvec.mean(axis=0),
rms.mean.eval(),
)
assert checkallclose(
bigvec.std(axis=0),
rms.std.eval(),
)
def __init__(self, ob_dim, ac_dim): #pylint: disable=W0613
X = tf.placeholder(tf.float32, shape=[None, ob_dim*2+ac_dim*2+2]) # batch of observations
vtarg_n = tf.placeholder(tf.float32, shape=[None], name='vtarg')
wd_dict = {}
h1 = tf.nn.elu(dense(X, 64, "h1", weight_init=U.normc_initializer(1.0), bias_init=0, weight_loss_dict=wd_dict))
h2 = tf.nn.elu(dense(h1, 64, "h2", weight_init=U.normc_initializer(1.0), bias_init=0, weight_loss_dict=wd_dict))
vpred_n = dense(h2, 1, "hfinal", weight_init=U.normc_initializer(1.0), bias_init=0, weight_loss_dict=wd_dict)[:,0]
sample_vpred_n = vpred_n + tf.random_normal(tf.shape(vpred_n))
wd_loss = tf.get_collection("vf_losses", None)
loss = tf.reduce_mean(tf.square(vpred_n - vtarg_n)) + tf.add_n(wd_loss)
loss_sampled = tf.reduce_mean(tf.square(vpred_n - tf.stop_gradient(sample_vpred_n)))
self._predict = U.function([X], vpred_n)
optim = kfac.KfacOptimizer(learning_rate=0.001, cold_lr=0.001*(1-0.9), momentum=0.9, \
clip_kl=0.3, epsilon=0.1, stats_decay=0.95, \
async=1, kfac_update=2, cold_iter=50, \
weight_decay_dict=wd_dict, max_grad_norm=None)
vf_var_list = []
for var in tf.trainable_variables():
if "vf" in var.name:
vf_var_list.append(var)
update_op, self.q_runner = optim.minimize(loss, loss_sampled, var_list=vf_var_list)
self.do_update = U.function([X, vtarg_n], update_op) #pylint: disable=E1101
U.initialize() # Initialize uninitialized TF variables
def learn(
env,
policy_fn,
*,
timesteps_per_batch, # what to train on
max_kl,
cg_iters,
gamma,
lam, # advantage estimation
entcoeff=0.0,
cg_damping=1e-2,
vf_stepsize=3e-4,
vf_iters=3,
max_timesteps=0,
max_episodes=0,
max_iters=0, # time constraint
callback=None):
nworkers = MPI.COMM_WORLD.Get_size()
rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_fn("pi", ob_space, ac_space)
oldpi = policy_fn("oldpi", ob_space, ac_space)
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
ob = U.get_placeholder_cached(name="ob")
ac = pi.pdtype.sample_placeholder([None])
kloldnew = oldpi.pd.kl(pi.pd)
ent = pi.pd.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = entcoeff * meanent
vferr = tf.reduce_mean(tf.square(pi.vpred - ret))
ratio = tf.exp(pi.pd.logp(ac) -
oldpi.pd.logp(ac)) # advantage * pnew / pold
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy"]
dist = meankl
all_var_list = pi.get_trainable_variables()
var_list = [
v for v in all_var_list if v.name.split("/")[1].startswith("pol")
]
vf_var_list = [
v for v in all_var_list if v.name.split("/")[1].startswith("vf")
]
vfadam = MpiAdam(vf_var_list)
get_flat = U.GetFlat(var_list)
set_from_flat = U.SetFromFlat(var_list)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32,
shape=[None],
name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
sz = U.intprod(shape)
tangents.append(tf.reshape(flat_tangent[start:start + sz], shape))
start += sz
gvp = tf.add_n([
tf.reduce_sum(g * tangent)
for (g, tangent) in zipsame(klgrads, tangents)
]) #pylint: disable=E1111
fvp = U.flatgrad(gvp, var_list)
assign_old_eq_new = U.function(
[], [],
updates=[
tf.assign(oldv, newv)
for (oldv,
newv) in zipsame(oldpi.get_variables(), pi.get_variables())
])
compute_losses = U.function([ob, ac, atarg], losses)
compute_lossandgrad = U.function([ob, ac, atarg], losses +
[U.flatgrad(optimgain, var_list)])
compute_fvp = U.function([flat_tangent, ob, ac, atarg], fvp)
compute_vflossandgrad = U.function([ob, ret],
U.flatgrad(vferr, vf_var_list))
@contextmanager
def timed(msg):
if rank == 0:
print(colorize(msg, color='magenta'))
tstart = time.time()
yield
print(
colorize("done in %.3f seconds" % (time.time() - tstart),
color='magenta'))
else:
yield
def allmean(x):
assert isinstance(x, np.ndarray)
out = np.empty_like(x)
MPI.COMM_WORLD.Allreduce(x, out, op=MPI.SUM)
out /= nworkers
return out
U.initialize()
th_init = get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
set_from_flat(th_init)
vfadam.sync()
print("Init param sum", th_init.sum(), flush=True)
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi,
env,
timesteps_per_batch,
stochastic=True)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=40) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=40) # rolling buffer for episode rewards
assert sum([max_iters > 0, max_timesteps > 0, max_episodes > 0]) == 1
while True:
if callback: callback(locals(), globals())
if max_timesteps and timesteps_so_far >= max_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
logger.log("********** Iteration %i ************" % iters_so_far)
with timed("sampling"):
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, gamma, lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
ob, ac, atarg, tdlamret = seg["ob"], seg["ac"], seg["adv"], seg[
"tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()
) / atarg.std() # standardized advantage function estimate
if hasattr(pi, "ret_rms"): pi.ret_rms.update(tdlamret)
if hasattr(pi, "ob_rms"):
pi.ob_rms.update(ob) # update running mean/std for policy
args = seg["ob"], seg["ac"], atarg
fvpargs = [arr[::5] for arr in args]
def fisher_vector_product(p):
return allmean(compute_fvp(p, *fvpargs)) + cg_damping * p
assign_old_eq_new() # set old parameter values to new parameter values
with timed("computegrad"):
*lossbefore, g = compute_lossandgrad(*args)
lossbefore = allmean(np.array(lossbefore))
g = allmean(g)
if np.allclose(g, 0):
logger.log("Got zero gradient. not updating")
else:
with timed("cg"):
stepdir = cg(fisher_vector_product,
g,
cg_iters=cg_iters,
verbose=rank == 0)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(fisher_vector_product(stepdir))
lm = np.sqrt(shs / max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lm
expectedimprove = g.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = get_flat()
for _ in range(10):
thnew = thbefore + fullstep * stepsize
set_from_flat(thnew)
meanlosses = surr, kl, *_ = allmean(
np.array(compute_losses(*args)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" %
(expectedimprove, improve))
if not np.isfinite(meanlosses).all():
logger.log("Got non-finite value of losses -- bad!")
elif kl > max_kl * 1.5:
logger.log("violated KL constraint. shrinking step.")
elif improve < 0:
logger.log("surrogate didn't improve. shrinking step.")
else:
logger.log("Stepsize OK!")
break
stepsize *= .5
else:
logger.log("couldn't compute a good step")
set_from_flat(thbefore)
if nworkers > 1 and iters_so_far % 20 == 0:
paramsums = MPI.COMM_WORLD.allgather(
(thnew.sum(), vfadam.getflat().sum())) # list of tuples
assert all(
np.allclose(ps, paramsums[0]) for ps in paramsums[1:])
for (lossname, lossval) in zip(loss_names, meanlosses):
logger.record_tabular(lossname, lossval)
with timed("vf"):
for _ in range(vf_iters):
for (mbob, mbret) in dataset.iterbatches(
(seg["ob"], seg["tdlamret"]),
include_final_partial_batch=False,
batch_size=64):
g = allmean(compute_vflossandgrad(mbob, mbret))
vfadam.update(g, vf_stepsize)
logger.record_tabular("ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
lrlocal = (seg["ep_lens"], seg["ep_rets"]) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
lens, rews = map(flatten_lists, zip(*listoflrpairs))
lenbuffer.extend(lens)
rewbuffer.extend(rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - tstart)
if rank == 0:
logger.dump_tabular()
def learn(env,
q_func,
lr=5e-4,
max_timesteps=100000,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02,
train_freq=1,
batch_size=32,
print_freq=100,
checkpoint_freq=10000,
learning_starts=1000,
gamma=1.0,
target_network_update_freq=500,
prioritized_replay=False,
prioritized_replay_alpha=0.6,
prioritized_replay_beta0=0.4,
prioritized_replay_beta_iters=None,
prioritized_replay_eps=1e-6,
param_noise=False,
callback=None):
"""Train a deepq model.
Parameters
-------
env: gym.Env
environment to train on
q_func: (tf.Variable, int, str, bool) -> tf.Variable
the model that takes the following inputs:
observation_in: object
the output of observation placeholder
num_actions: int
number of actions
scope: str
reuse: bool
should be passed to outer variable scope
and returns a tensor of shape (batch_size, num_actions) with values of every action.
lr: float
learning rate for adam optimizer
max_timesteps: int
number of env steps to optimizer for
buffer_size: int
size of the replay buffer
exploration_fraction: float
fraction of entire training period over which the exploration rate is annealed
exploration_final_eps: float
final value of random action probability
train_freq: int
update the model every `train_freq` steps.
set to None to disable printing
batch_size: int
size of a batched sampled from replay buffer for training
print_freq: int
how often to print out training progress
set to None to disable printing
checkpoint_freq: int
how often to save the model. This is so that the best version is restored
at the end of the training. If you do not wish to restore the best version at
the end of the training set this variable to None.
learning_starts: int
how many steps of the model to collect transitions for before learning starts
gamma: float
discount factor
target_network_update_freq: int
update the target network every `target_network_update_freq` steps.
prioritized_replay: True
if True prioritized replay buffer will be used.
prioritized_replay_alpha: float
alpha parameter for prioritized replay buffer
prioritized_replay_beta0: float
initial value of beta for prioritized replay buffer
prioritized_replay_beta_iters: int
number of iterations over which beta will be annealed from initial value
to 1.0. If set to None equals to max_timesteps.
prioritized_replay_eps: float
epsilon to add to the TD errors when updating priorities.
callback: (locals, globals) -> None
function called at every steps with state of the algorithm.
If callback returns true training stops.
Returns
-------
act: ActWrapper
Wrapper over act function. Adds ability to save it and load it.
See header of baselines/deepq/categorical.py for details on the act function.
"""
# Create all the functions necessary to train the model
sess = tf.Session()
sess.__enter__()
# capture the shape outside the closure so that the env object is not serialized
# by cloudpickle when serializing make_obs_ph
observation_space_shape = env.observation_space.shape
def make_obs_ph(name):
return BatchInput(observation_space_shape, name=name)
action_dim = 2# env.action_space.n
act, train, update_target, debug = deepq.build_train(
make_obs_ph=make_obs_ph,
q_func=q_func,
num_actions=action_dim,
optimizer=tf.train.AdamOptimizer(learning_rate=lr),
gamma=gamma,
grad_norm_clipping=10,
param_noise=param_noise
)
act_params = {
'make_obs_ph': make_obs_ph,
'q_func': q_func,
'num_actions': action_dim,
}
act = ActWrapper(act, act_params)
# Create the replay buffer
if prioritized_replay:
replay_buffer = PrioritizedReplayBuffer(buffer_size, alpha=prioritized_replay_alpha)
if prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = max_timesteps
beta_schedule = LinearSchedule(prioritized_replay_beta_iters,
initial_p=prioritized_replay_beta0,
final_p=1.0)
else:
replay_buffer = ReplayBuffer(buffer_size)
beta_schedule = None
# Create the schedule for exploration starting from 1.
exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction * max_timesteps),
initial_p=1.0,
final_p=exploration_final_eps)
# Initialize the parameters and copy them to the target network.
U.initialize()
update_target()
episode_rewards = [0.0]
saved_mean_reward = None
obs = env.reset()
reset = True
with tempfile.TemporaryDirectory() as td:
model_saved = False
model_file = os.path.join(td, "model")
for t in range(max_timesteps):
if callback is not None:
if callback(locals(), globals()):
break
# Take action and update exploration to the newest value
kwargs = {}
if not param_noise:
update_eps = exploration.value(t)
update_param_noise_threshold = 0.
else:
update_eps = 0.
# Compute the threshold such that the KL divergence between perturbed and non-perturbed
# policy is comparable to eps-greedy exploration with eps = exploration.value(t).
# See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017
# for detailed explanation.
update_param_noise_threshold = -np.log(1. - exploration.value(t) + exploration.value(t) / float(action_dim))
kwargs['reset'] = reset
kwargs['update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
action = act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0]
env_action = action
reset = False
new_obs, rew, done, _ = env.step(env_action)
# Store transition in the replay buffer.
replay_buffer.add(obs, action, rew, new_obs, float(done))
obs = new_obs
episode_rewards[-1] += rew
if done:
obs = env.reset()
episode_rewards.append(0.0)
reset = True
if t > learning_starts and t % train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if prioritized_replay:
experience = replay_buffer.sample(batch_size, beta=beta_schedule.value(t))
(obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(batch_size)
weights, batch_idxes = np.ones_like(rewards), None
td_errors = train(obses_t, actions, rewards, obses_tp1, dones, weights)
if prioritized_replay:
new_priorities = np.abs(td_errors) + prioritized_replay_eps
replay_buffer.update_priorities(batch_idxes, new_priorities)
if t > learning_starts and t % target_network_update_freq == 0:
# Update target network periodically.
update_target()
mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1)
num_episodes = len(episode_rewards)
if done and print_freq is not None and len(episode_rewards) % print_freq == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("mean 100 episode reward", mean_100ep_reward)
logger.record_tabular("% time spent exploring", int(100 * exploration.value(t)))
logger.dump_tabular()
if (checkpoint_freq is not None and t > learning_starts and
num_episodes > 100 and t % checkpoint_freq == 0):
if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward:
if print_freq is not None:
logger.log("Saving model due to mean reward increase: {} -> {}".format(
saved_mean_reward, mean_100ep_reward))
save_state(model_file)
model_saved = True
saved_mean_reward = mean_100ep_reward
if model_saved:
if print_freq is not None:
logger.log("Restored model with mean reward: {}".format(saved_mean_reward))
load_state(model_file)
return act
def learn(
env,
policy_fn,
*,
timesteps_per_actorbatch, # timesteps per actor per update
clip_param,
entcoeff, # clipping parameter epsilon, entropy coeff
optim_epochs,
optim_stepsize,
optim_batchsize, # optimization hypers
gamma,
lam, # advantage estimation
max_timesteps=0,
max_episodes=0,
max_iters=0,
max_seconds=0, # time constraint
callback=None, # you can do anything in the callback, since it takes locals(), globals()
adam_epsilon=1e-5,
schedule='constant' # annealing for stepsize parameters (epsilon and adam)
):
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_fn("pi", ob_space,
ac_space) # Construct network for new policy
oldpi = policy_fn("oldpi", ob_space, ac_space) # Network for old policy
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
lrmult = tf.placeholder(
name='lrmult', dtype=tf.float32,
shape=[]) # learning rate multiplier, updated with schedule
clip_param = clip_param * lrmult # Annealed cliping parameter epislon
ob = U.get_placeholder_cached(name="ob")
ac = pi.pdtype.sample_placeholder([None])
kloldnew = oldpi.pd.kl(pi.pd)
ent = pi.pd.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
pol_entpen = (-entcoeff) * meanent
ratio = tf.exp(pi.pd.logp(ac) - oldpi.pd.logp(ac)) # pnew / pold
surr1 = ratio * atarg # surrogate from conservative policy iteration
surr2 = tf.clip_by_value(ratio, 1.0 - clip_param,
1.0 + clip_param) * atarg #
pol_surr = -tf.reduce_mean(tf.minimum(
surr1, surr2)) # PPO's pessimistic surrogate (L^CLIP)
vf_loss = tf.reduce_mean(tf.square(pi.vpred - ret))
total_loss = pol_surr + pol_entpen + vf_loss
losses = [pol_surr, pol_entpen, vf_loss, meankl, meanent]
loss_names = ["pol_surr", "pol_entpen", "vf_loss", "kl", "ent"]
var_list = pi.get_trainable_variables()
lossandgrad = U.function([ob, ac, atarg, ret, lrmult],
losses + [U.flatgrad(total_loss, var_list)])
adam = MpiAdam(var_list, epsilon=adam_epsilon)
assign_old_eq_new = U.function(
[], [],
updates=[
tf.assign(oldv, newv)
for (oldv,
newv) in zipsame(oldpi.get_variables(), pi.get_variables())
])
compute_losses = U.function([ob, ac, atarg, ret, lrmult], losses)
U.initialize()
adam.sync()
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi,
env,
timesteps_per_actorbatch,
stochastic=True)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=100) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=100) # rolling buffer for episode rewards
assert sum(
[max_iters > 0, max_timesteps > 0, max_episodes > 0,
max_seconds > 0]) == 1, "Only one time constraint permitted"
while True:
if callback: callback(locals(), globals())
if max_timesteps and timesteps_so_far >= max_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
elif max_seconds and time.time() - tstart >= max_seconds:
break
if schedule == 'constant':
cur_lrmult = 1.0
elif schedule == 'linear':
cur_lrmult = max(1.0 - float(timesteps_so_far) / max_timesteps, 0)
else:
raise NotImplementedError
logger.log("********** Iteration %i ************" % iters_so_far)
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, gamma, lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
ob, ac, atarg, tdlamret = seg["ob"], seg["ac"], seg["adv"], seg[
"tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()
) / atarg.std() # standardized advantage function estimate
d = Dataset(dict(ob=ob, ac=ac, atarg=atarg, vtarg=tdlamret),
shuffle=not pi.recurrent)
optim_batchsize = optim_batchsize or ob.shape[0]
if hasattr(pi, "ob_rms"):
pi.ob_rms.update(ob) # update running mean/std for policy
assign_old_eq_new() # set old parameter values to new parameter values
logger.log("Optimizing...")
logger.log(fmt_row(13, loss_names))
# Here we do a bunch of optimization epochs over the data
for _ in range(optim_epochs):
losses = [
] # list of tuples, each of which gives the loss for a minibatch
for batch in d.iterate_once(optim_batchsize):
*newlosses, g = lossandgrad(batch["ob"], batch["ac"],
batch["atarg"], batch["vtarg"],
cur_lrmult)
adam.update(g, optim_stepsize * cur_lrmult)
losses.append(newlosses)
logger.log(fmt_row(13, np.mean(losses, axis=0)))
logger.log("Evaluating losses...")
losses = []
for batch in d.iterate_once(optim_batchsize):
newlosses = compute_losses(batch["ob"], batch["ac"],
batch["atarg"], batch["vtarg"],
cur_lrmult)
losses.append(newlosses)
meanlosses, _, _ = mpi_moments(losses, axis=0)
logger.log(fmt_row(13, meanlosses))
for (lossval, name) in zipsame(meanlosses, loss_names):
logger.record_tabular("loss_" + name, lossval)
logger.record_tabular("ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
lrlocal = (seg["ep_lens"], seg["ep_rets"]) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
lens, rews = map(flatten_lists, zip(*listoflrpairs))
lenbuffer.extend(lens)
rewbuffer.extend(rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - tstart)
if MPI.COMM_WORLD.Get_rank() == 0:
logger.dump_tabular()
def learn(env,
policy,
vf,
gamma,
lam,
timesteps_per_batch,
num_timesteps,
animate=False,
callback=None,
desired_kl=0.002):
obfilter = ZFilter(env.observation_space.shape)
max_pathlength = env.spec.timestep_limit
stepsize = tf.Variable(initial_value=np.float32(np.array(0.03)),
name='stepsize')
inputs, loss, loss_sampled = policy.update_info
optim = kfac.KfacOptimizer(learning_rate=stepsize, cold_lr=stepsize*(1-0.9), momentum=0.9, kfac_update=2,\
epsilon=1e-2, stats_decay=0.99, async=1, cold_iter=1,
weight_decay_dict=policy.wd_dict, max_grad_norm=None)
pi_var_list = []
for var in tf.trainable_variables():
if "pi" in var.name:
pi_var_list.append(var)
update_op, q_runner = optim.minimize(loss,
loss_sampled,
var_list=pi_var_list)
do_update = U.function(inputs, update_op)
U.initialize()
# start queue runners
enqueue_threads = []
coord = tf.train.Coordinator()
for qr in [q_runner, vf.q_runner]:
assert (qr != None)
enqueue_threads.extend(
qr.create_threads(tf.get_default_session(),
coord=coord,
start=True))
i = 0
timesteps_so_far = 0
while True:
if timesteps_so_far > num_timesteps:
break
logger.log("********** Iteration %i ************" % i)
# Collect paths until we have enough timesteps
timesteps_this_batch = 0
paths = []
while True:
path = rollout(env,
policy,
max_pathlength,
animate=(len(paths) == 0 and (i % 10 == 0)
and animate),
obfilter=obfilter)
paths.append(path)
n = pathlength(path)
timesteps_this_batch += n
timesteps_so_far += n
if timesteps_this_batch > timesteps_per_batch:
break
# Estimate advantage function
vtargs = []
advs = []
for path in paths:
rew_t = path["reward"]
return_t = common.discount(rew_t, gamma)
vtargs.append(return_t)
vpred_t = vf.predict(path)
vpred_t = np.append(vpred_t,
0.0 if path["terminated"] else vpred_t[-1])
delta_t = rew_t + gamma * vpred_t[1:] - vpred_t[:-1]
adv_t = common.discount(delta_t, gamma * lam)
advs.append(adv_t)
# Update value function
vf.fit(paths, vtargs)
# Build arrays for policy update
ob_no = np.concatenate([path["observation"] for path in paths])
action_na = np.concatenate([path["action"] for path in paths])
oldac_dist = np.concatenate([path["action_dist"] for path in paths])
adv_n = np.concatenate(advs)
standardized_adv_n = (adv_n - adv_n.mean()) / (adv_n.std() + 1e-8)
# Policy update
do_update(ob_no, action_na, standardized_adv_n)
min_stepsize = np.float32(1e-8)
max_stepsize = np.float32(1e0)
# Adjust stepsize
kl = policy.compute_kl(ob_no, oldac_dist)
if kl > desired_kl * 2:
logger.log("kl too high")
tf.assign(stepsize, tf.maximum(min_stepsize,
stepsize / 1.5)).eval()
elif kl < desired_kl / 2:
logger.log("kl too low")
tf.assign(stepsize, tf.minimum(max_stepsize,
stepsize * 1.5)).eval()
else:
logger.log("kl just right!")
logger.record_tabular(
"EpRewMean", np.mean([path["reward"].sum() for path in paths]))
logger.record_tabular(
"EpRewSEM",
np.std([
path["reward"].sum() / np.sqrt(len(paths)) for path in paths
]))
logger.record_tabular("EpLenMean",
np.mean([pathlength(path) for path in paths]))
logger.record_tabular("KL", kl)
if callback:
callback()
logger.dump_tabular()
i += 1
coord.request_stop()
coord.join(enqueue_threads)
def __init__(self, ob_dim, ac_dim):
# Here we'll construct a bunch of expressions, which will be used in two places:
# (1) When sampling actions
# (2) When computing loss functions, for the policy update
# Variables specific to (1) have the word "sampled" in them,
# whereas variables specific to (2) have the word "old" in them
ob_no = tf.placeholder(tf.float32, shape=[None, ob_dim * 2],
name="ob") # batch of observations
oldac_na = tf.placeholder(
tf.float32, shape=[None, ac_dim],
name="ac") # batch of actions previous actions
oldac_dist = tf.placeholder(
tf.float32, shape=[None, ac_dim * 2], name="oldac_dist"
) # batch of actions previous action distributions
adv_n = tf.placeholder(tf.float32, shape=[None],
name="adv") # advantage function estimate
wd_dict = {}
h1 = tf.nn.tanh(
dense(ob_no,
64,
"h1",
weight_init=U.normc_initializer(1.0),
bias_init=0.0,
weight_loss_dict=wd_dict))
h2 = tf.nn.tanh(
dense(h1,
64,
"h2",
weight_init=U.normc_initializer(1.0),
bias_init=0.0,
weight_loss_dict=wd_dict))
mean_na = dense(h2,
ac_dim,
"mean",
weight_init=U.normc_initializer(0.1),
bias_init=0.0,
weight_loss_dict=wd_dict) # Mean control output
self.wd_dict = wd_dict
self.logstd_1a = logstd_1a = tf.get_variable(
"logstd", [ac_dim], tf.float32,
tf.zeros_initializer()) # Variance on outputs
logstd_1a = tf.expand_dims(logstd_1a, 0)
std_1a = tf.exp(logstd_1a)
std_na = tf.tile(std_1a, [tf.shape(mean_na)[0], 1])
ac_dist = tf.concat([
tf.reshape(mean_na, [-1, ac_dim]),
tf.reshape(std_na, [-1, ac_dim])
], 1)
sampled_ac_na = tf.random_normal(
tf.shape(ac_dist[:, ac_dim:])
) * ac_dist[:,
ac_dim:] + ac_dist[:, :
ac_dim] # This is the sampled action we'll perform.
logprobsampled_n = -tf.reduce_sum(tf.log(
ac_dist[:, ac_dim:]), axis=1) - 0.5 * tf.log(
2.0 * np.pi) * ac_dim - 0.5 * tf.reduce_sum(
tf.square(ac_dist[:, :ac_dim] - sampled_ac_na) /
(tf.square(ac_dist[:, ac_dim:])),
axis=1) # Logprob of sampled action
logprob_n = -tf.reduce_sum(
tf.log(ac_dist[:, ac_dim:]), axis=1
) - 0.5 * tf.log(2.0 * np.pi) * ac_dim - 0.5 * tf.reduce_sum(
tf.square(ac_dist[:, :ac_dim] - oldac_na) /
(tf.square(ac_dist[:, ac_dim:])),
axis=1
) # Logprob of previous actions under CURRENT policy (whereas oldlogprob_n is under OLD policy)
kl = tf.reduce_mean(kl_div(oldac_dist, ac_dist, ac_dim))
#kl = .5 * tf.reduce_mean(tf.square(logprob_n - oldlogprob_n)) # Approximation of KL divergence between old policy used to generate actions, and new policy used to compute logprob_n
surr = -tf.reduce_mean(
adv_n * logprob_n
) # Loss function that we'll differentiate to get the policy gradient
surr_sampled = -tf.reduce_mean(logprob_n) # Sampled loss of the policy
self._act = U.function([ob_no],
[sampled_ac_na, ac_dist, logprobsampled_n
]) # Generate a new action and its logprob
#self.compute_kl = U.function([ob_no, oldac_na, oldlogprob_n], kl) # Compute (approximate) KL divergence between old policy and new policy
self.compute_kl = U.function([ob_no, oldac_dist], kl)
self.update_info = (
(ob_no, oldac_na, adv_n), surr, surr_sampled
) # Input and output variables needed for computing loss
U.initialize() # Initialize uninitialized TF variables
env = gym.make("CartPole-v0")
# Create all the functions necessary to train the model
act, train, update_target, debug = deepq.build_train(
make_obs_ph=lambda name: BatchInput(env.observation_space.shape, name=name),
q_func=model,
num_actions=env.action_space.n,
optimizer=tf.train.AdamOptimizer(learning_rate=5e-4),
)
# Create the replay buffer
replay_buffer = ReplayBuffer(50000)
# Create the schedule for exploration starting from 1 (every action is random) down to
# 0.02 (98% of actions are selected according to values predicted by the model).
exploration = LinearSchedule(schedule_timesteps=10000, initial_p=1.0, final_p=0.02)
# Initialize the parameters and copy them to the target network.
U.initialize()
update_target()
episode_rewards = [0.0]
obs = env.reset()
for t in itertools.count():
# Take action and update exploration to the newest value
action = act(obs[None], update_eps=exploration.value(t))[0]
new_obs, rew, done, _ = env.step(action)
# Store transition in the replay buffer.
replay_buffer.add(obs, action, rew, new_obs, float(done))
obs = new_obs
episode_rewards[-1] += rew
if done:
obs = env.reset()