def __init__(self, var_list, *, beta1=0.9, beta2=0.999, epsilon=1e-08, scale_grad_by_procs=True, comm=None):
self.var_list = var_list
self.beta1 = beta1
self.beta2 = beta2
self.epsilon = epsilon
self.scale_grad_by_procs = scale_grad_by_procs
size = sum(U.numel(v) for v in var_list)
self.m = np.zeros(size, 'float32')
self.v = np.zeros(size, 'float32')
self.t = 0
self.setfromflat = U.SetFromFlat(var_list)
self.getflat = U.GetFlat(var_list)
self.comm = MPI.COMM_WORLD if comm is None and MPI is not None else comm
def __init__(self,
var_list,
*,
beta1=0.9,
beta2=0.999,
epsilon=1e-08,
scale_grad_by_procs=True,
comm=None,
sess=None):
"""
A parallel MPI implementation of the Adam optimizer for TensorFlow
https://arxiv.org/abs/1412.6980
:param var_list: ([TensorFlow Tensor]) the variables
:param beta1: (float) Adam beta1 parameter
:param beta2: (float) Adam beta1 parameter
:param epsilon: (float) to help with preventing arithmetic issues
:param scale_grad_by_procs: (bool) if the scaling should be done by processes
:param comm: (MPI Communicators) if None, mpi4py.MPI.COMM_WORLD
:param sess: (TensorFlow Session) if None, tf.get_default_session()
"""
self.var_list = var_list
self.beta1 = beta1
self.beta2 = beta2
self.epsilon = epsilon
self.scale_grad_by_procs = scale_grad_by_procs
size = sum(tf_utils.numel(v) for v in var_list)
# Exponential moving average of gradient values
# "first moment estimate" m in the paper
self.exp_avg = np.zeros(size, 'float32')
# Exponential moving average of squared gradient values
# "second raw moment estimate" v in the paper
self.exp_avg_sq = np.zeros(size, 'float32')
self.step = 0
self.setfromflat = tf_utils.SetFromFlat(var_list, sess=sess)
self.getflat = tf_utils.GetFlat(var_list, sess=sess)
self.comm = mpi4py.MPI.COMM_WORLD if comm is None else comm
def learn(
env,
policy_func,
discriminator,
expert_dataset,
embedding_z,
pretrained,
pretrained_weight,
*,
g_step,
d_step,
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,
d_stepsize=3e-4,
vf_iters=3,
max_timesteps=0,
max_episodes=0,
max_iters=0, # time constraint
callback=None,
save_per_iter=100,
ckpt_dir=None,
log_dir=None,
load_model_path=None,
task_name=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_func("pi",
ob_space,
ac_space,
reuse=(pretrained_weight != None))
oldpi = policy_func("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 = U.mean(kloldnew)
meanent = U.mean(ent)
entbonus = entcoeff * meanent
vferr = U.mean(tf.square(pi.vpred - ret))
ratio = tf.exp(pi.pd.logp(ac) -
oldpi.pd.logp(ac)) # advantage * pnew / pold
surrgain = U.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")
]
d_adam = MpiAdam(discriminator.get_trainable_variables())
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(
[U.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
writer = U.FileWriter(log_dir)
U.initialize()
th_init = get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
set_from_flat(th_init)
d_adam.sync()
vfadam.sync()
print("Init param sum", th_init.sum(), flush=True)
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi,
env,
discriminator,
embedding=embedding_z,
timesteps_per_batch=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
true_rewbuffer = deque(maxlen=40)
assert sum([max_iters > 0, max_timesteps > 0, max_episodes > 0]) == 1
g_loss_stats = stats(loss_names)
d_loss_stats = stats(discriminator.loss_name)
ep_stats = stats(["True_rewards", "Rewards", "Episode_length"])
# if provide pretrained weight
if pretrained_weight is not None:
U.load_state(pretrained_weight, var_list=pi.get_variables())
# if provieded model path
if load_model_path is not None:
U.load_state(load_model_path)
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
# Save model
if iters_so_far % save_per_iter == 0 and ckpt_dir is not None:
U.save_state(os.path.join(ckpt_dir, task_name),
counter=iters_so_far)
logger.log("********** Iteration %i ************" % iters_so_far)
def fisher_vector_product(p):
return allmean(compute_fvp(p, *fvpargs)) + cg_damping * p
# ------------------ Update G ------------------
logger.log("Optimizing Policy...")
for _ in range(g_step):
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, "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]
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:])
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=128):
if hasattr(pi, "ob_rms"):
pi.ob_rms.update(
mbob) # update running mean/std for policy
g = allmean(compute_vflossandgrad(mbob, mbret))
vfadam.update(g, vf_stepsize)
g_losses = meanlosses
for (lossname, lossval) in zip(loss_names, meanlosses):
logger.record_tabular(lossname, lossval)
logger.record_tabular("ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
# ------------------ Update D ------------------
logger.log("Optimizing Discriminator...")
logger.log(fmt_row(13, discriminator.loss_name))
ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob))
batch_size = len(ob) // d_step
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
for ob_batch, ac_batch in dataset.iterbatches(
(ob, ac), include_final_partial_batch=False,
batch_size=batch_size):
ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob_batch))
# update running mean/std for discriminator
if hasattr(discriminator, "obs_rms"):
discriminator.obs_rms.update(
np.concatenate((ob_batch, ob_expert), 0))
*newlosses, g = discriminator.lossandgrad(ob_batch, ac_batch,
ob_expert, ac_expert)
d_adam.update(allmean(g), d_stepsize)
d_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
lrlocal = (seg["ep_lens"], seg["ep_rets"], seg["ep_true_rets"]
) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
lens, rews, true_rets = map(flatten_lists, zip(*listoflrpairs))
true_rewbuffer.extend(true_rets)
lenbuffer.extend(lens)
rewbuffer.extend(rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpTrueRewMean", np.mean(true_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()
g_loss_stats.add_all_summary(writer, g_losses, iters_so_far)
d_loss_stats.add_all_summary(writer, np.mean(d_losses, axis=0),
iters_so_far)
ep_stats.add_all_summary(writer, [
np.mean(true_rewbuffer),
np.mean(rewbuffer),
np.mean(lenbuffer)
], iters_so_far)
def learn(
env,
policy_func,
discriminator,
expert_dataset,
timesteps_per_batch,
*,
g_step,
d_step, # 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,
d_stepsize=3e-4,
schedule='constant', # annealing for stepsize parameters (epsilon and adam)
save_per_iter=100,
ckpt_dir=None,
task="train",
sample_stochastic=True,
load_model_path=None,
task_name=None,
max_sample_traj=1500):
nworkers = MPI.COMM_WORLD.Get_size()
rank = MPI.COMM_WORLD.Get_rank()
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_func("pi", ob_space,
ac_space) # Construct network for new policy
oldpi = policy_func("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 = U.mean(kloldnew)
meanent = U.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 = U.clip(ratio, 1.0 - clip_param, 1.0 + clip_param) * atarg #
pol_surr = -U.mean(tf.minimum(
surr1, surr2)) # PPO's pessimistic surrogate (L^CLIP)
vf_loss = U.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)])
d_adam = MpiAdam(discriminator.get_trainable_variables())
adam = MpiAdam(var_list, epsilon=adam_epsilon)
get_flat = U.GetFlat(var_list)
set_from_flat = U.SetFromFlat(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, ret, lrmult], losses)
U.initialize()
th_init = get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
set_from_flat(th_init)
d_adam.sync()
adam.sync()
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
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi,
env,
discriminator,
timesteps_per_batch,
stochastic=True)
traj_gen = traj_episode_generator(pi,
env,
timesteps_per_batch,
stochastic=sample_stochastic)
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
true_rewbuffer = deque(maxlen=100)
assert sum(
[max_iters > 0, max_timesteps > 0, max_episodes > 0,
max_seconds > 0]) == 1, "Only one time constraint permitted"
if task == 'sample_trajectory':
# not elegant, i know :(
sample_trajectory(load_model_path, max_sample_traj, traj_gen,
task_name, sample_stochastic)
sys.exit()
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
# Save model
if iters_so_far % save_per_iter == 0 and ckpt_dir is not None:
U.save_state(os.path.join(ckpt_dir, task_name),
counter=iters_so_far)
logger.log("********** Iteration %i ************" % iters_so_far)
for _ in range(g_step):
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)
# ------------------ Update D ------------------
logger.log("Optimizing Discriminator...")
logger.log(fmt_row(13, discriminator.loss_name))
ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob))
batch_size = len(ob) // d_step
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob))
batch_size = len(ob) // d_step
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
for ob_batch, ac_batch in dataset.iterbatches(
(ob, ac), include_final_partial_batch=False,
batch_size=batch_size):
ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob_batch))
# update running mean/std for discriminator
if hasattr(discriminator, "obs_rms"):
discriminator.obs_rms.update(
np.concatenate((ob_batch, ob_expert), 0))
*newlosses, g = discriminator.lossandgrad(ob_batch, ac_batch,
ob_expert, ac_expert)
d_adam.update(allmean(g), d_stepsize)
d_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
# ----------------- logger --------------------
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"], seg["ep_true_rets"]
) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
lens, rews, true_rews = map(flatten_lists, zip(*listoflrpairs))
lenbuffer.extend(lens)
rewbuffer.extend(rews)
true_rewbuffer.extend(true_rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpTrueRewMean", np.mean(true_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()