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 else comm
def learn(env,
policy_func,
reward_giver,
reward_guidance,
expert_dataset,
rank,
pretrained,
pretrained_weight,
*,
g_step,
d_step,
entcoeff,
save_per_iter,
ckpt_dir,
log_dir,
timesteps_per_batch,
task_name,
gamma,
lam,
algo,
max_kl,
cg_iters,
cg_damping=1e-2,
vf_stepsize=3e-4,
d_stepsize=1e-4,
vf_iters=3,
max_timesteps=0,
max_episodes=0,
max_iters=0,
loss_percent=0.0,
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
policy = build_policy(env, 'mlp', value_network='copy')
ob = observation_placeholder(ob_space)
with tf.variable_scope('pi'):
pi = policy(observ_placeholder=ob)
with tf.variable_scope('oldpi'):
oldpi = policy(observ_placeholder=ob)
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
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.vf - 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 = get_trainable_variables('pi')
# var_list = [v for v in all_var_list if v.name.startswith("pi/pol") or v.name.startswith("pi/logstd")]
# vf_var_list = [v for v in all_var_list if v.name.startswith("pi/vff")]
var_list = get_pi_trainable_variables("pi")
vf_var_list = get_vf_trainable_variables("pi")
# assert len(var_list) == len(vf_var_list) + 1
d_adam = MpiAdam(reward_giver.get_trainable_variables())
guidance_adam = MpiAdam(reward_guidance.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([
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(get_variables('oldpi'), get_variables('pi'))
])
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)
d_adam.sync()
guidance_adam.sync()
vfadam.sync()
if rank == 0:
print("Init param sum", th_init.sum(), flush=True)
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi,
env,
reward_giver,
reward_guidance,
timesteps_per_batch,
stochastic=True,
algo=algo,
loss_percent=loss_percent)
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(reward_giver.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())
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 rank == 0 and iters_so_far % save_per_iter == 0 and ckpt_dir is not None:
fname = os.path.join(ckpt_dir, task_name)
os.makedirs(os.path.dirname(fname), exist_ok=True)
saver = tf.train.Saver()
saver.save(tf.get_default_session(), fname)
logger.log("********** Iteration %i ************" % iters_so_far)
# global flag_render
# if iters_so_far > 0 and iters_so_far % 10 ==0:
# flag_render = True
# else:
# flag_render = False
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__()
print('rewards', seg['rew'])
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, reward_giver.loss_name))
ob_expert, ac_expert = expert_dataset.get_next_batch(
batch_size=len(ob))
batch_size = 128
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
with timed("Discriminator"):
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(
batch_size=batch_size)
# update running mean/std for reward_giver
if hasattr(reward_giver, "obs_rms"):
reward_giver.obs_rms.update(
np.concatenate((ob_batch, ob_expert), 0))
*newlosses, g = reward_giver.lossandgrad(ob_batch, ob_expert)
d_adam.update(allmean(g), d_stepsize)
d_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
# ------------------ Update Guidance ------------
logger.log("Optimizing Guidance...")
logger.log(fmt_row(13, reward_guidance.loss_name))
batch_size = 128
guidance_losses = [
] # list of tuples, each of which gives the loss for a minibatch
with timed("Guidance"):
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(
batch_size=batch_size)
idx_condition = process_expert(ob_expert, ac_expert)
pick_idx = (idx_condition >= loss_percent)
# pick_idx = idx_condition
ob_expert_p = ob_expert[pick_idx]
ac_expert_p = ac_expert[pick_idx]
ac_batch_p = []
for each_ob in ob_expert_p:
tmp_ac, _, _, _ = pi.step(each_ob, stochastic=True)
ac_batch_p.append(tmp_ac)
# update running mean/std for reward_giver
if hasattr(reward_guidance, "obs_rms"):
reward_guidance.obs_rms.update(ob_expert_p)
# reward_guidance.train(expert_s=ob_batch_p, agent_a=ac_batch_p, expert_a=ac_expert_p)
*newlosses, g = reward_guidance.lossandgrad(
ob_expert_p, ac_batch_p, ac_expert_p)
guidance_adam.update(allmean(g), d_stepsize)
guidance_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(guidance_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) * g_step
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 __init__(self,
observations,
action_space,
latent,
optimizer=None,
sess=None,
train=True,
beta=1.0,
l2=0.,
lr=0.001,
init_scale=0.01,
init_bias=0.0,
trainable_variance=True,
trainable_bias=True,
init_logstd=0.,
scope_name="pi",
clip=None,
state_dependent_variance=True,
**tensors):
"""
Parameters:
----------
env RL environment
observations tensorflow placeholder in which the observations will be fed
latent latent state from which policy distribution parameters should be inferred
sess tensorflow session to run calculations in (if None, default session is used)
**tensors tensorflow tensors for additional attributes such as state or mask
"""
self.X = observations
self.state = tf.constant([])
self.initial_state = None
self.__dict__.update(tensors)
latent = tf.layers.flatten(latent)
self.action_space = action_space
self.pdtype = make_pdtype(action_space)
self.pd, self.pi = self.pdtype.pdfromlatent(
latent,
init_scale=init_scale,
init_bias=init_bias,
trainable_variance=trainable_variance,
trainable_bias=trainable_bias,
init_logstd=init_logstd,
clip=clip,
beta=beta) # init_bias=0.0
self.stochastic = tf.placeholder(dtype=tf.bool, shape=())
self.action = tf_util.switch(self.stochastic, self.pd.sample(),
self.pd.mode())
self.neglogp = self.pd.neglogp(self.action)
if beta == 1.0:
self.prob = tf.nn.softmax(self.pd.flatparam())
else:
self.prob = boltzmann(self.pd.flatparam(), beta=beta)
if optimizer is None:
self.optimizer = tf.train.AdamOptimizer(learning_rate=lr)
else:
self.optimizer = optimizer
self.sess = sess or tf.get_default_session()
self.vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope=scope_name)
try:
self.action_ph = tf.placeholder(tf.int64, [None],
name='targets_placeholder')
self.action_selected = action_selected = tf.one_hot(
self.action_ph, self.action_space.n)
#out = tf.reduce_sum(tf.reduce_sum(tf.log(self.logits+1e-5)*action_selected, axis=1))
out = tf.reduce_mean(
tf.log(tf.reduce_sum(self.prob * action_selected, axis=1)))
gradients = tf.gradients(out, self.vars)
except:
self.action_ph = tf.placeholder(dtype=tf.float32,
shape=(None, ) +
action_space.shape,
name='targets_placeholder')
gradients = tf.gradients(-self.pd.neglogp(self.action_ph),
self.vars)
self.cont = cont = not isinstance(self.action_space, Discrete)
flat_grad = tf_util.GetFlat(gradients).op
self.compute_gradients = tf_util.function(
inputs=[self.X, self.action_ph], outputs=[flat_grad])
'''self.compute_cont_gradients = tf_util.function(
inputs=[self.X, self.action_ph],
outputs=tf.gradients(-self.pd.neglogp(self.action_ph), self.vars)
)'''
self.debug = tf_util.function(
inputs=[self.X, self.action_ph],
outputs=[gradients, self.prob, self.action_ph])
self.set_from_flat = tf_util.SetFromFlat(self.vars)
if self.cont:
total_error = tf.reduce_sum(
tf.square(self.action_ph -
tf.reduce_mean(self.action_ph, axis=0)),
axis=0)
unexplained_error = tf.reduce_sum(tf.square(self.action_ph -
self.pd.mean),
axis=0)
R_squared = 1 - (unexplained_error / total_error)
self.accuracy = accuracy = R_squared
else:
self.accuracy = accuracy = tf.reduce_mean(
tf.cast(tf.math.equal(self.pd.mode(), self.action_ph),
tf.float32))
self.entropy = entropy = tf.reduce_mean(self.pd.entropy())
if train:
self.gamma = l2
self._build_train(
cont=cont, state_dependent_variance=state_dependent_variance)
self.pdf = tf.exp(self.pd.logp(self.action_ph))
def learn(
*,
network,
env,
eval_policy,
total_timesteps,
timesteps_per_batch=1024, # what to train on
max_kl=0.001,
cg_iters=10,
gamma=0.99,
lam=1.0, # advantage estimation
seed=None,
ent_coef=0.0,
cg_damping=1e-2,
vf_stepsize=3e-4,
vf_iters=3,
max_episodes=0,
max_iters=0, # time constraint
callback=None,
load_path=None,
checkpoint_path_in=None,
checkpoint_dir_out=None,
checkpoint_freq=100, # In iterations!!,
from_iter=0,
eval_episodes=20,
**network_kwargs):
'''
learn a policy function with TRPO algorithm
Parameters:
----------
network neural network to learn. Can be either string ('mlp', 'cnn', 'lstm', 'lnlstm' for basic types)
or function that takes input placeholder and returns tuple (output, None) for feedforward nets
or (output, (state_placeholder, state_output, mask_placeholder)) for recurrent nets
env environment (one of the gym environments or wrapped via baselines.common.vec_env.VecEnv-type class
timesteps_per_batch timesteps per gradient estimation batch
max_kl max KL divergence between old policy and new policy ( KL(pi_old || pi) )
ent_coef coefficient of policy entropy term in the optimization objective
cg_iters number of iterations of conjugate gradient algorithm
cg_damping conjugate gradient damping
vf_stepsize learning rate for adam optimizer used to optimie value function loss
vf_iters number of iterations of value function optimization iterations per each policy optimization step
total_timesteps max number of timesteps
max_episodes max number of episodes
max_iters maximum number of policy optimization iterations
callback function to be called with (locals(), globals()) each policy optimization step
load_path str, path to load the model from (default: None, i.e. no model is loaded)
**network_kwargs keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
Returns:
-------
learnt model
'''
nworkers = MPI.COMM_WORLD.Get_size()
rank = MPI.COMM_WORLD.Get_rank()
cpus_per_worker = 1
U.get_session(
config=tf.ConfigProto(allow_soft_placement=True,
inter_op_parallelism_threads=cpus_per_worker,
intra_op_parallelism_threads=cpus_per_worker))
policy = build_policy(env, network, value_network='copy', **network_kwargs)
set_global_seeds(seed)
np.set_printoptions(precision=3)
# Setup losses and stuff
# ----------------------------------------
# ob_space = Box(low=-np.inf, high=np.inf, shape=(env.observation_space.n,))
ob_space = env.observation_space
ac_space = env.action_space
ob = observation_placeholder(ob_space)
with tf.variable_scope("pi"):
pi = policy(observ_placeholder=ob)
with tf.variable_scope("oldpi"):
oldpi = policy(observ_placeholder=ob)
# Loading checkpoint
if checkpoint_path_in is not None and os.path.isfile(checkpoint_path_in):
pi.load(checkpoint_path_in)
logger.log('Loaded policy weights from %s' % checkpoint_path_in)
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
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 = ent_coef * meanent
vferr = tf.reduce_mean(tf.square(pi.vf - 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 = get_trainable_variables("pi")
# 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")]
var_list = get_pi_trainable_variables("pi")
vf_var_list = get_vf_trainable_variables("pi")
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(get_variables("oldpi"), get_variables("pi"))
])
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()
# s = env.reset()
# start = time.time()
# for i in range(10000):
# pi.step(s, stochastic=True)
# duration = time.time() - start
# print(duration)
# return
if load_path is not None:
pi.load(load_path)
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,
gamma=gamma)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
iters_eval = 0
all_logs = []
best_rew = -np.inf
tstart = time.time()
lenbuffer = deque(maxlen=40) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=40) # rolling buffer for episode rewards
online_scores = []
offline_scores = []
if sum([max_iters > 0, total_timesteps > 0, max_episodes > 0]) == 0:
# noththing to be done
return pi
assert sum([max_iters > 0, total_timesteps > 0, max_episodes > 0]) < 2, \
'out of max_iters, total_timesteps, and max_episodes only one should be specified'
while True:
if callback: callback(locals(), globals())
if total_timesteps and timesteps_so_far >= total_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)
if iters_so_far % checkpoint_freq == 0 and checkpoint_dir_out is not None:
if not os.path.exists(checkpoint_dir_out):
os.makedirs(checkpoint_dir_out)
pi.save(
os.path.join(checkpoint_dir_out,
'checkpoint_%d' % iters_so_far))
logger.log('Saved policy weights as %s' % os.path.join(
checkpoint_dir_out, 'checkpoint_%d.npy' % iters_so_far))
def pi_wrapper(ob):
ac, vpred, _, _ = pi.step(ob, stochastic=True)
return ac
rew, _, logs, disc_rets, num_stops, avg_damages = eval_policy(
pi=pi_wrapper, n_episodes=eval_episodes, verbose=True)
offline_scores.append(
[np.mean(disc_rets),
np.mean(num_stops),
np.mean(avg_damages)])
np.save(os.path.join(checkpoint_dir_out, 'offline_scores.npy'),
offline_scores)
for log in logs:
log['iter'] = iters_eval
all_logs = all_logs + logs
iters_eval += 1
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)
ep_rew_mean = np.mean(rewbuffer)
online_scores.append(ep_rew_mean)
np.save(os.path.join(checkpoint_dir_out, 'online_scores.npy'),
online_scores)
# Saving best
if iters_so_far % checkpoint_freq == 0 and ep_rew_mean > best_rew and checkpoint_dir_out is not None:
pi.save(os.path.join(checkpoint_dir_out, 'best'))
best_rew = ep_rew_mean
logger.log('Saved policy weights as %s' %
os.path.join(checkpoint_dir_out, 'best.npy'))
if rank == 0:
logger.dump_tabular()
return pi
def __init__(self, env, observations, latent, estimate_q=False, vf_latent=None,
sess=None,trainable_variance=True,init_logstd=0, clip=None, **tensors):
"""
Parameters:
----------
env RL environment
observations tensorflow placeholder in which the observations will be fed
latent latent state from which policy distribution parameters should be inferred
vf_latent latent state from which value function should be inferred (if None, then latent is used)
sess tensorflow session to run calculations in (if None, default session is used)
**tensors tensorflow tensors for additional attributes such as state or mask
"""
self.X = observations
self.state = tf.constant([])
self.initial_state = None
self.__dict__.update(tensors)
vf_latent = vf_latent if vf_latent is not None else latent
vf_latent = tf.layers.flatten(vf_latent)
latent = tf.layers.flatten(latent)
self.pdtype = make_pdtype(env.action_space)
self.pd, self.pi = self.pdtype.pdfromlatent(latent, init_scale=0.01,
trainable_variance=trainable_variance,
init_logstd=init_logstd,
clip=clip)
self.stochastic = tf.placeholder(dtype=tf.bool, shape=())
self.action = tf_util.switch(self.stochastic, self.pd.sample(), self.pd.mode())
self.neglogp = self.pd.neglogp(self.action)
self.logits=tf.nn.softmax(self.pd.flatparam())
self.sess = sess
self.prob = tf.nn.softmax(self.pd.flatparam())
#out = tf.reduce_mean(tf.log(tf.reduce_sum(self.prob * action_selected, axis=1)))
self.vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="pi/pi")
if len(self.vars) == 0:
self.vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope="pi")
self.set_from_flat = tf_util.SetFromFlat(self.vars)
try:
self.action_ph = tf.placeholder(tf.int64, [None], name='targets_placeholder')
self.action_selected = action_selected = tf.one_hot(self.action_ph, env.action_space.n)
#out = tf.reduce_sum(tf.reduce_sum(tf.log(self.logits+1e-5)*action_selected, axis=1))
out = tf.reduce_mean(tf.log(tf.reduce_sum(self.prob*action_selected, axis=1)))
gradients = tf.gradients(out, self.vars)
except:
self.action_ph = tf.placeholder(dtype=tf.float32, shape=(None,) + env.action_space.shape,
name='targets_placeholder')
gradients = tf.gradients(-self.pd.neglogp(self.action_ph), self.vars)
#gradients = tf.gradients(out, self.vars)
if gradients[0] is not None:
flat_grad = tf_util.GetFlat(gradients).op
self.compute_gradients = tf_util.function(
inputs=[self.X, self.action_ph],
outputs=[flat_grad]
)
if estimate_q:
assert isinstance(env.action_space, gym.spaces.Discrete)
self.q = fc(vf_latent, 'q', env.action_space.n)
self.vf = self.q
else:
self.vf = fc(vf_latent, 'vf', 1)
self.vf = self.vf[:,0]