def learn(self,
total_timesteps,
callback=None,
log_interval=100,
tb_log_name="A2C",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
self.learning_rate_schedule = Scheduler(
initial_value=self.learning_rate,
n_values=total_timesteps,
schedule=self.lr_schedule)
t_start = time.time()
callback.on_training_start(locals(), globals())
for update in range(1, total_timesteps // self.n_batch + 1):
callback.on_rollout_start()
# true_reward is the reward without discount
rollout = self.runner.run(callback)
# unpack
obs, states, rewards, masks, actions, values, ep_infos, true_reward = rollout
callback.on_rollout_end()
# Early stopping due to the callback
if not self.runner.continue_training:
break
self.ep_info_buf.extend(ep_infos)
_, value_loss, policy_entropy = self._train_step(
obs, states, rewards, masks, actions, values,
self.num_timesteps // self.n_batch, writer)
n_seconds = time.time() - t_start
fps = int((update * self.n_batch) / n_seconds)
if writer is not None:
total_episode_reward_logger(
self.episode_reward,
true_reward.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)), writer,
self.num_timesteps)
if self.verbose >= 1 and (update % log_interval == 0
or update == 1):
explained_var = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps",
self.num_timesteps)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy",
float(policy_entropy))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance",
float(explained_var))
if len(self.ep_info_buf) > 0 and len(
self.ep_info_buf[0]) > 0:
logger.logkv(
'ep_reward_mean',
safe_mean([
ep_info['r'] for ep_info in self.ep_info_buf
]))
logger.logkv(
'ep_len_mean',
safe_mean([
ep_info['l'] for ep_info in self.ep_info_buf
]))
logger.dump_tabular()
callback.on_training_end()
return self
def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="MDPO",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
print("got seed {}, sgd_steps {}".format(seed, self.sgd_steps))
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
with self.sess.as_default():
callback.on_training_start(locals(), globals())
seg_gen = traj_segment_generator(self.old_policy, self.env, self.timesteps_per_batch,
reward_giver=self.reward_giver,
gail=self.using_gail, mdal=self.using_mdal, neural=self.neural,
action_space=self.action_space, gamma=self.gamma, callback=callback)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
t_start = time.time()
len_buffer = deque(maxlen=40) # rolling buffer for episode lengths
reward_buffer = deque(maxlen=40) # rolling buffer for episode rewards
self.episode_reward = np.zeros((self.n_envs,))
self.outer_learning_rate = get_schedule_fn(3e-4)
self.cliprange_vf = get_schedule_fn(0.2)
true_reward_buffer = None
if self.using_gail or self.using_mdal:
true_reward_buffer = deque(maxlen=40)
# Initialize dataloader
batchsize = self.timesteps_per_batch // self.d_step
self.expert_dataset.init_dataloader(batchsize)
# Stats not used for now
# TODO: replace with normal tb logging
# g_loss_stats = Stats(loss_names)
# d_loss_stats = Stats(reward_giver.loss_name)
# ep_stats = Stats(["True_rewards", "Rewards", "Episode_length"])
while True:
# if callback is not None:
# # Only stop training if return value is False, not when it is None. This is for backwards
# # compatibility with callbacks that have no return statement.
# if callback(locals(), globals()) is False:
# break
if total_timesteps and timesteps_so_far >= total_timesteps:
break
logger.log("********** Iteration %i ************" % iters_so_far)
#def fisher_vector_product(vec):
# return self.allmean(self.compute_fvp(vec, *fvpargs, sess=self.sess)) + self.cg_damping * vec
# ------------------ Update G ------------------
# logger.log("Optimizing Policy...")
# g_step = 1 when not using GAIL
mean_losses = None
vpredbefore = None
tdlamret = None
observation = None
action = None
seg = None
for k in range(self.g_step):
with self.timed("sampling"):
seg = seg_gen.__next__()
if not seg.get('continue_training', True): # pytype: disable=attribute-error
break
add_vtarg_and_adv(seg, self.gamma, self.lam)
if self.using_mdal:
policy_successor_features = add_successor_features(seg, self.gamma,
is_action_features=self.is_action_features)
else:
policy_successor_features = add_successor_features(seg, self.gamma)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
observation, action = seg["observations"], seg["actions"]
atarg, tdlamret = seg["adv"], seg["tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before update
atarg = (atarg - atarg.mean()) / atarg.std() # standardized advantage function estimate
# true_rew is the reward without discount
if writer is not None:
self.episode_reward = total_episode_reward_logger(self.episode_reward,
seg["true_rewards"].reshape(
(self.n_envs, -1)),
seg["dones"].reshape((self.n_envs, -1)),
writer, self.num_timesteps)
n_updates = int(total_timesteps / self.timesteps_per_batch)
lr_now = np.float32(1.0 - (iters_so_far - 1.0) / n_updates)
outer_lr_now = self.outer_learning_rate(1.0 - (iters_so_far - 1.0) / n_updates)
clip_now = self.cliprange_vf(1.0 - (iters_so_far - 1.0) / n_updates)
args = seg["observations"], seg["observations"], seg["actions"], atarg
# Subsampling: see p40-42 of John Schulman thesis
# http://joschu.net/docs/thesis.pdf
#fvpargs = [arr[::5] for arr in args]
with self.timed("computegrad"):
steps = self.num_timesteps + (k + 1) * (seg["total_timestep"] / self.g_step)
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata() if self.full_tensorboard_log else None
# run loss backprop with summary, and save the metadata (memory, compute time, ...)
if writer is not None:
summary, grad, *lossbefore = self.compute_lossandgrad(*args, tdlamret,
lr_now, seg["vpred"],
seg["observations"],
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
if self.full_tensorboard_log:
writer.add_run_metadata(run_metadata, 'step%d' % steps)
writer.add_summary(summary, steps)
else:
_, grad, *lossbefore = self.compute_lossandgrad(*args, tdlamret,
lr_now, seg["vpred"],
seg["observations"],
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
td_map = {self.policy_pi.obs_ph: seg["observations"],
self.old_policy.obs_ph: seg["observations"],
self.closed_policy.obs_ph: seg["observations"],
self.action: seg["actions"], self.atarg: atarg, self.ret: tdlamret,
self.learning_rate_ph: lr_now, self.outer_learning_rate_ph: outer_lr_now,
self.vtarg: seg["vpred"]}
for _ in range(int(self.sgd_steps)):
_ = self.sess.run(self._train, td_map)
#if self.method == "closed-KL":
# _ = self.sess.run(self._train_policy, td_map)
if np.allclose(grad, 0):
logger.log("Got zero gradient. not updating")
else:
for _ in range(1):
mean_losses = surr, kl_loss, *_ = self.allmean(
np.array(self.compute_losses(*args, lr_now, seg["vpred"], sess=self.sess)))
with self.timed("vf"):
for _ in range(self.vf_iters):
# NOTE: for recurrent policies, use shuffle=False?
for (mbob, mbret, mbval) in dataset.iterbatches((seg["observations"], seg["tdlamret"], seg["vpred"]),
include_final_partial_batch=False,
batch_size=128,
shuffle=True):
grad = self.allmean(self.compute_vflossandgrad(mbob, mbob, mbret, mbval, clip_now, sess=self.sess))
self.vfadam.update(grad, outer_lr_now) #self.vf_stepsize)
if iters_so_far % 1 == 0:
# print("updating theta now")
self.assign_old_eq_new(sess=self.sess)
for (loss_name, loss_val) in zip(self.loss_names, mean_losses):
logger.record_tabular(loss_name, loss_val)
logger.record_tabular("explained_variance_tdlam_before",
explained_variance(vpredbefore, tdlamret))
if self.using_gail:
# ------------------ Update D ------------------
logger.log("Optimizing Discriminator...")
logger.log(fmt_row(13, self.reward_giver.loss_name))
assert len(observation) == self.timesteps_per_batch
batch_size = self.timesteps_per_batch // self.d_step
# NOTE: uses only the last g step for observation
d_losses = [] # list of tuples, each of which gives the loss for a minibatch
# NOTE: for recurrent policies, use shuffle=False?
for ob_batch, ac_batch in dataset.iterbatches((observation, action),
include_final_partial_batch=False,
batch_size=batch_size,
shuffle=True):
ob_expert, ac_expert = self.expert_dataset.get_next_batch()
# update running mean/std for reward_giver
if self.reward_giver.normalize:
self.reward_giver.obs_rms.update(np.concatenate((ob_batch, ob_expert), 0))
# Reshape actions if needed when using discrete actions
if isinstance(self.action_space, gym.spaces.Discrete):
if len(ac_batch.shape) == 2:
ac_batch = ac_batch[:, 0]
if len(ac_expert.shape) == 2:
ac_expert = ac_expert[:, 0]
*newlosses, grad = self.reward_giver.lossandgrad(ob_batch, ac_batch, ob_expert, ac_expert)
self.d_adam.update(self.allmean(grad), self.d_stepsize)
d_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
elif self.using_mdal:
batch_sampling = True
if self.neural:
if batch_sampling:
batch_size = self.timesteps_per_batch // self.d_step
# NOTE: uses only the last g step for observation
d_losses = [] # list of tuples, each of which gives the loss for a minibatch
# NOTE: for recurrent policies, use shuffle=False?
for ob_batch, ac_batch in dataset.iterbatches((observation, action),
include_final_partial_batch=False,
batch_size=batch_size,
shuffle=True):
# ob_batch, ac_batch, gamma_batch = np.array(batch_buffer['obs']), np.array(
# batch_buffer['acs']), np.array(batch_buffer['gammas'])
gamma_batch = np.ones((ob_batch.shape[0]))
ob_expert, ac_expert = self.expert_dataset.get_next_batch()
gamma_expert = np.ones((ob_expert.shape[0]))
# ob_expert, ac_expert, gamma_expert = np.concatenate(self.expert_dataset.ep_obs),\
# np.concatenate(self.expert_dataset.ep_acs),\
# np.concatenate(self.expert_dataset.ep_gammas)
# update running mean/std for reward_giver
if self.reward_giver.normalize:
self.reward_giver.obs_rms.update(np.concatenate((ob_batch, ob_expert), 0))
# Reshape actions if needed when using discrete actions
if isinstance(self.action_space, gym.spaces.Discrete):
if len(ac_batch.shape) == 2:
ac_batch = ac_batch[:, 0]
if len(ac_expert.shape) == 2:
ac_expert = ac_expert[:, 0]
ob_reg_expert, ac_reg_expert = np.array(ob_expert), np.array(ac_expert)
# while True:
# if ob_reg_expert.shape[0] == ob_batch.shape[0] and ac_reg_expert.shape[0] == \
# ac_batch.shape[0]:
# break
# ob_reg_expert, ac_reg_expert = self.expert_dataset.get_next_batch()
# ob_reg_expert, ac_reg_expert = np.array(ob_reg_expert), np.array(ac_reg_expert)
alpha = np.random.uniform(0.0, 1.0, size=(ob_reg_expert.shape[0], 1))
ob_mix_batch = alpha * ob_batch[:ob_reg_expert.shape[0]] + (1 - alpha) * ob_reg_expert
ac_mix_batch = alpha * ac_batch[:ac_reg_expert.shape[0]] + (1 - alpha) * ac_reg_expert
with self.sess.as_default():
# self.reward_giver.train(ob_batch, ac_batch, np.expand_dims(gamma_batch, axis=1),
# ob_expert, ac_expert, np.expand_dims(gamma_expert, axis=1))
*newlosses, grad = self.reward_giver.lossandgrad(
ob_batch, ac_batch, np.expand_dims(gamma_batch, axis=1),
ob_expert, ac_expert, np.expand_dims(gamma_expert, axis=1),
ob_mix_batch, ac_mix_batch)
self.d_adam.update(self.allmean(grad), self.d_stepsize)
else:
# assert len(observation) == self.timesteps_per_batch
# Comment out if you want only the latest rewards:
obs_batch, acs_batch, gammas_batch = seg['obs_batch'], seg['acs_batch'], seg['gammas_batch']
batch_successor_features = seg['successor_features_batch']
if self.reward_giver.normalize:
ob_reg_batch, ac_reg_batch = observation, action
ob_expert, _ = self.expert_dataset.get_next_batch()
self.reward_giver.obs_rms.update(np.concatenate((ob_reg_batch, ob_expert), 0))
# self.reward_giver.obs_rms.update(
# np.array(batch_successor_features)[:, :self.observation_space.shape[0]])
for idx, (ob_batch, ac_batch, gamma_batch) in enumerate(
zip(obs_batch, acs_batch, gammas_batch)):
rand_traj = np.random.randint(self.expert_dataset.num_traj)
ob_expert, ac_expert, gamma_expert = self.expert_dataset.ep_obs[rand_traj], \
self.expert_dataset.ep_acs[rand_traj], \
self.expert_dataset.ep_gammas[rand_traj]
ob_batch, ac_batch, gamma_batch = np.array(ob_batch), np.array(ac_batch), np.array(
gamma_batch)
while True:
ob_reg_expert, ac_reg_expert = self.expert_dataset.get_next_batch()
ob_reg_expert, ac_reg_expert = np.array(ob_reg_expert), np.array(ac_reg_expert)
if ob_reg_expert.shape[0] == ob_reg_batch.shape[0] and ac_reg_expert.shape[0] == \
ac_reg_batch.shape[0]:
break
alpha = np.random.uniform(0.0, 1.0, size=(ob_reg_batch.shape[0], 1))
ob_mix_batch = alpha * ob_reg_batch + (1 - alpha) * ob_reg_expert
ac_mix_batch = alpha * ac_reg_batch + (1 - alpha) * ac_reg_expert
with self.sess.as_default():
*newlosses, grad = self.reward_giver.lossandgrad(
ob_batch, ac_batch, np.expand_dims(gamma_batch, axis=1),
ob_expert, ac_expert, np.expand_dims(gamma_expert, axis=1),
ob_mix_batch, ac_mix_batch)
self.d_adam.update(self.allmean(grad), self.d_stepsize)
# self.reward_giver.train(ob_batch, ac_batch, np.expand_dims(gamma_batch, axis=1),
# ob_expert, ac_expert,
# np.expand_dims(gamma_expert, axis=1),
# ob_mix_batch, ac_mix_batch)
if self.using_gail or self.using_mdal:
# lr: lengths and rewards
lr_local = (seg["ep_lens"], seg["ep_rets"], seg["ep_true_rets"]) # local values
list_lr_pairs = MPI.COMM_WORLD.allgather(lr_local) # list of tuples
lens, rews, true_rets = map(flatten_lists, zip(*list_lr_pairs))
true_reward_buffer.extend(true_rets)
else:
lr_local = (seg["ep_lens"], seg["ep_rets"]) # local values
list_lr_pairs = MPI.COMM_WORLD.allgather(lr_local) # list of tuples
lens, rews = map(flatten_lists, zip(*list_lr_pairs))
len_buffer.extend(lens)
reward_buffer.extend(rews)
if len(len_buffer) > 0:
if self.using_gail or self.using_mdal:
logger.record_tabular("EpTrueRewMean", np.mean(true_reward_buffer))
logger.record_tabular("EpRewMean", np.mean(reward_buffer))
logger.record_tabular("EpLenMean", np.mean(len_buffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
current_it_timesteps = MPI.COMM_WORLD.allreduce(seg["total_timestep"])
timesteps_so_far += current_it_timesteps
self.num_timesteps += current_it_timesteps
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", self.num_timesteps)
logger.record_tabular("TimeElapsed", time.time() - t_start)
logger.record_tabular("Tsallis-q", self.tsallis_q)
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("seed", self.seed)
if self.verbose >= 1 and self.rank == 0:
logger.dump_tabular()
callback.on_training_end()
return self
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="TRPO",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn(seed)
with self.sess.as_default():
seg_gen = traj_segment_generator(
self.policy_pi,
self.env,
self.timesteps_per_batch,
reward_giver=self.reward_giver,
gail=self.using_gail)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
t_start = time.time()
len_buffer = deque(
maxlen=40) # rolling buffer for episode lengths
reward_buffer = deque(
maxlen=40) # rolling buffer for episode rewards
self.episode_reward = np.zeros((self.n_envs, ))
true_reward_buffer = None
if self.using_gail:
true_reward_buffer = deque(maxlen=40)
# Initialize dataloader
batchsize = self.timesteps_per_batch // self.d_step
self.expert_dataset.init_dataloader(batchsize)
# Stats not used for now
# TODO: replace with normal tb logging
# g_loss_stats = Stats(loss_names)
# d_loss_stats = Stats(reward_giver.loss_name)
# ep_stats = Stats(["True_rewards", "Rewards", "Episode_length"])
while True:
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
if total_timesteps and timesteps_so_far >= total_timesteps:
break
logger.log("********** Iteration %i ************" %
iters_so_far)
def fisher_vector_product(vec):
return self.allmean(
self.compute_fvp(
vec, *fvpargs,
sess=self.sess)) + self.cg_damping * vec
# ------------------ Update G ------------------
logger.log("Optimizing Policy...")
# g_step = 1 when not using GAIL
mean_losses = None
vpredbefore = None
tdlamret = None
observation = None
action = None
seg = None
for k in range(self.g_step):
with self.timed("sampling"):
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, self.gamma, self.lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
observation, action = seg["observations"], seg[
"actions"]
atarg, tdlamret = seg["adv"], seg["tdlamret"]
vpredbefore = seg[
"vpred"] # predicted value function before update
atarg = (atarg - atarg.mean()) / atarg.std(
) # standardized advantage function estimate
# true_rew is the reward without discount
if writer is not None:
self.episode_reward = total_episode_reward_logger(
self.episode_reward,
seg["true_rewards"].reshape(
(self.n_envs, -1)), seg["dones"].reshape(
(self.n_envs, -1)), writer,
self.num_timesteps)
args = seg["observations"], seg["observations"], seg[
"actions"], atarg
# Subsampling: see p40-42 of John Schulman thesis
# http://joschu.net/docs/thesis.pdf
fvpargs = [arr[::5] for arr in args]
self.assign_old_eq_new(sess=self.sess)
with self.timed("computegrad"):
steps = self.num_timesteps + (k + 1) * (
seg["total_timestep"] / self.g_step)
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata(
) if self.full_tensorboard_log else None
# run loss backprop with summary, and save the metadata (memory, compute time, ...)
if writer is not None:
summary, grad, *lossbefore = self.compute_lossandgrad(
*args,
tdlamret,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
if self.full_tensorboard_log:
writer.add_run_metadata(
run_metadata, 'step%d' % steps)
writer.add_summary(summary, steps)
else:
_, grad, *lossbefore = self.compute_lossandgrad(
*args,
tdlamret,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
lossbefore = self.allmean(np.array(lossbefore))
grad = self.allmean(grad)
if np.allclose(grad, 0):
logger.log("Got zero gradient. not updating")
else:
with self.timed("conjugate_gradient"):
stepdir = conjugate_gradient(
fisher_vector_product,
grad,
cg_iters=self.cg_iters,
verbose=self.rank == 0
and self.verbose >= 1)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(
fisher_vector_product(stepdir))
# abs(shs) to avoid taking square root of negative values
lagrange_multiplier = np.sqrt(
abs(shs) / self.max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lagrange_multiplier
expectedimprove = grad.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = self.get_flat()
thnew = None
for _ in range(10):
thnew = thbefore + fullstep * stepsize
self.set_from_flat(thnew)
mean_losses = surr, kl_loss, *_ = self.allmean(
np.array(
self.compute_losses(*args,
sess=self.sess)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" %
(expectedimprove, improve))
if not np.isfinite(mean_losses).all():
logger.log(
"Got non-finite value of losses -- bad!"
)
elif kl_loss > self.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")
self.set_from_flat(thbefore)
if self.nworkers > 1 and iters_so_far % 20 == 0:
# list of tuples
paramsums = MPI.COMM_WORLD.allgather(
(thnew.sum(), self.vfadam.getflat().sum()))
assert all(
np.allclose(ps, paramsums[0])
for ps in paramsums[1:])
with self.timed("vf"):
for _ in range(self.vf_iters):
# NOTE: for recurrent policies, use shuffle=False?
for (mbob, mbret) in dataset.iterbatches(
(seg["observations"], seg["tdlamret"]),
include_final_partial_batch=False,
batch_size=128,
shuffle=True):
grad = self.allmean(
self.compute_vflossandgrad(
mbob, mbob, mbret, sess=self.sess))
self.vfadam.update(grad, self.vf_stepsize)
for (loss_name, loss_val) in zip(self.loss_names,
mean_losses):
logger.record_tabular(loss_name, loss_val)
logger.record_tabular(
"explained_variance_tdlam_before",
explained_variance(vpredbefore, tdlamret))
if self.using_gail:
# ------------------ Update D ------------------
logger.log("Optimizing Discriminator...")
logger.log(fmt_row(13, self.reward_giver.loss_name))
assert len(observation) == self.timesteps_per_batch
batch_size = self.timesteps_per_batch // self.d_step
# NOTE: uses only the last g step for observation
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
# NOTE: for recurrent policies, use shuffle=False?
for ob_batch, ac_batch in dataset.iterbatches(
(observation, action),
include_final_partial_batch=False,
batch_size=batch_size,
shuffle=True):
ob_expert, ac_expert = self.expert_dataset.get_next_batch(
)
# update running mean/std for reward_giver
if self.reward_giver.normalize:
self.reward_giver.obs_rms.update(
np.concatenate((ob_batch, ob_expert), 0))
# Reshape actions if needed when using discrete actions
if isinstance(self.action_space,
gym.spaces.Discrete):
if len(ac_batch.shape) == 2:
ac_batch = ac_batch[:, 0]
if len(ac_expert.shape) == 2:
ac_expert = ac_expert[:, 0]
*newlosses, grad = self.reward_giver.lossandgrad(
ob_batch, ac_batch, ob_expert, ac_expert)
self.d_adam.update(self.allmean(grad),
self.d_stepsize)
d_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
# lr: lengths and rewards
lr_local = (seg["ep_lens"], seg["ep_rets"],
seg["ep_true_rets"]) # local values
list_lr_pairs = MPI.COMM_WORLD.allgather(
lr_local) # list of tuples
lens, rews, true_rets = map(flatten_lists,
zip(*list_lr_pairs))
true_reward_buffer.extend(true_rets)
else:
# lr: lengths and rewards
lr_local = (seg["ep_lens"], seg["ep_rets"]
) # local values
list_lr_pairs = MPI.COMM_WORLD.allgather(
lr_local) # list of tuples
lens, rews = map(flatten_lists, zip(*list_lr_pairs))
len_buffer.extend(lens)
reward_buffer.extend(rews)
if len(len_buffer) > 0:
logger.record_tabular("EpLenMean", np.mean(len_buffer))
logger.record_tabular("EpRewMean",
np.mean(reward_buffer))
if self.using_gail:
logger.record_tabular("EpTrueRewMean",
np.mean(true_reward_buffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
current_it_timesteps = MPI.COMM_WORLD.allreduce(
seg["total_timestep"])
timesteps_so_far += current_it_timesteps
self.num_timesteps += current_it_timesteps
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", self.num_timesteps)
logger.record_tabular("TimeElapsed", time.time() - t_start)
if self.verbose >= 1 and self.rank == 0:
logger.dump_tabular()
return self
def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="ACER"):
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name) as writer:
self._setup_learn(seed)
self.learning_rate_schedule = Scheduler(initial_value=self.learning_rate, n_values=total_timesteps,
schedule=self.lr_schedule)
episode_stats = EpisodeStats(self.n_steps, self.n_envs)
runner = _Runner(env=self.env, model=self, n_steps=self.n_steps)
self.episode_reward = np.zeros((self.n_envs,))
if self.replay_ratio > 0:
buffer = Buffer(env=self.env, n_steps=self.n_steps, size=self.buffer_size)
else:
buffer = None
t_start = time.time()
# n_batch samples, 1 on_policy call and multiple off-policy calls
for steps in range(0, total_timesteps, self.n_batch):
enc_obs, obs, actions, rewards, mus, dones, masks = runner.run()
episode_stats.feed(rewards, dones)
if buffer is not None:
buffer.put(enc_obs, actions, rewards, mus, dones, masks)
if writer is not None:
self.episode_reward = total_episode_reward_logger(self.episode_reward,
rewards.reshape((self.n_envs, self.n_steps)),
dones.reshape((self.n_envs, self.n_steps)),
writer, steps)
# reshape stuff correctly
obs = obs.reshape(runner.batch_ob_shape)
actions = actions.reshape([runner.n_batch])
rewards = rewards.reshape([runner.n_batch])
mus = mus.reshape([runner.n_batch, runner.n_act])
dones = dones.reshape([runner.n_batch])
masks = masks.reshape([runner.batch_ob_shape[0]])
names_ops, values_ops = self._train_step(obs, actions, rewards, dones, mus, self.initial_state, masks,
steps, writer)
if callback is not None:
callback(locals(), globals())
if self.verbose >= 1 and (int(steps / runner.n_batch) % log_interval == 0):
logger.record_tabular("total_timesteps", steps)
logger.record_tabular("fps", int(steps / (time.time() - t_start)))
# IMP: In EpisodicLife env, during training, we get done=True at each loss of life,
# not just at the terminal state. Thus, this is mean until end of life, not end of episode.
# For true episode rewards, see the monitor files in the log folder.
logger.record_tabular("mean_episode_length", episode_stats.mean_length())
logger.record_tabular("mean_episode_reward", episode_stats.mean_reward())
for name, val in zip(names_ops, values_ops):
logger.record_tabular(name, float(val))
logger.dump_tabular()
if self.replay_ratio > 0 and buffer.has_atleast(self.replay_start):
samples_number = np.random.poisson(self.replay_ratio)
for _ in range(samples_number):
# get obs, actions, rewards, mus, dones from buffer.
obs, actions, rewards, mus, dones, masks = buffer.get()
# reshape stuff correctly
obs = obs.reshape(runner.batch_ob_shape)
actions = actions.reshape([runner.n_batch])
rewards = rewards.reshape([runner.n_batch])
mus = mus.reshape([runner.n_batch, runner.n_act])
dones = dones.reshape([runner.n_batch])
masks = masks.reshape([runner.batch_ob_shape[0]])
self._train_step(obs, actions, rewards, dones, mus, self.initial_state, masks, steps)
return self
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="ACKTR"):
with SetVerbosity(self.verbose), TensorboardWriter(
self.graph, self.tensorboard_log, tb_log_name) as writer:
self._setup_learn(seed)
self.n_batch = self.n_envs * self.n_steps
self.learning_rate_schedule = Scheduler(
initial_value=self.learning_rate,
n_values=total_timesteps,
schedule=self.lr_schedule)
# FIFO queue of the q_runner thread is closed at the end of the learn function.
# As a result, it needs to be redefinied at every call
with self.graph.as_default():
with tf.variable_scope(
"kfac_apply",
reuse=self.trained,
custom_getter=tf_util.outer_scope_getter(
"kfac_apply")):
# Some of the variables are not in a scope when they are create
# so we make a note of any previously uninitialized variables
tf_vars = tf.global_variables()
is_uninitialized = self.sess.run(
[tf.is_variable_initialized(var) for var in tf_vars])
old_uninitialized_vars = [
v for (v, f) in zip(tf_vars, is_uninitialized) if not f
]
self.train_op, self.q_runner = self.optim.apply_gradients(
list(zip(self.grads_check, self.params)))
# then we check for new uninitialized variables and initialize them
tf_vars = tf.global_variables()
is_uninitialized = self.sess.run(
[tf.is_variable_initialized(var) for var in tf_vars])
new_uninitialized_vars = [
v for (v, f) in zip(tf_vars, is_uninitialized)
if not f and v not in old_uninitialized_vars
]
if len(new_uninitialized_vars) != 0:
self.sess.run(
tf.variables_initializer(new_uninitialized_vars))
self.trained = True
runner = A2CRunner(self.env,
self,
n_steps=self.n_steps,
gamma=self.gamma)
self.episode_reward = np.zeros((self.n_envs, ))
t_start = time.time()
coord = tf.train.Coordinator()
enqueue_threads = self.q_runner.create_threads(self.sess,
coord=coord,
start=True)
for update in range(1, total_timesteps // self.n_batch + 1):
# true_reward is the reward without discount
obs, states, rewards, masks, actions, values, true_reward = runner.run(
)
policy_loss, value_loss, policy_entropy = self._train_step(
obs, states, rewards, masks, actions, values, update,
writer)
n_seconds = time.time() - t_start
fps = int((update * self.n_batch) / n_seconds)
if writer is not None:
self.episode_reward = total_episode_reward_logger(
self.episode_reward,
true_reward.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)), writer,
update * (self.n_batch + 1))
if callback is not None:
callback(locals(), globals())
if self.verbose >= 1 and (update % log_interval == 0
or update == 1):
explained_var = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps",
update * self.n_batch)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy",
float(policy_entropy))
logger.record_tabular("policy_loss", float(policy_loss))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance",
float(explained_var))
logger.dump_tabular()
coord.request_stop()
coord.join(enqueue_threads)
return self
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="PPO1",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn(seed)
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the PPO1 model must be " \
"an instance of common.policies.ActorCriticPolicy."
with self.sess.as_default():
self.adam.sync()
# Prepare for rollouts
seg_gen = traj_segment_generator(self.policy_pi, self.env,
self.timesteps_per_actorbatch)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
t_start = time.time()
# rolling buffer for episode lengths
lenbuffer = deque(maxlen=100)
# rolling buffer for episode rewards
rewbuffer = deque(maxlen=100)
self.episode_reward = np.zeros((self.n_envs, ))
while True:
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
if total_timesteps and timesteps_so_far >= total_timesteps:
break
if self.schedule == 'constant':
cur_lrmult = 1.0
elif self.schedule == 'linear':
cur_lrmult = max(
1.0 - float(timesteps_so_far) / total_timesteps, 0)
else:
raise NotImplementedError
logger.log("********** Iteration %i ************" %
iters_so_far)
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, self.gamma, self.lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
obs_ph, action_ph, atarg, tdlamret = seg["ob"], seg[
"ac"], seg["adv"], seg["tdlamret"]
# true_rew is the reward without discount
if writer is not None:
self.episode_reward = total_episode_reward_logger(
self.episode_reward, seg["true_rew"].reshape(
(self.n_envs, -1)), seg["dones"].reshape(
(self.n_envs, -1)), writer,
self.num_timesteps)
# predicted value function before udpate
vpredbefore = seg["vpred"]
# standardized advantage function estimate
atarg = (atarg - atarg.mean()) / atarg.std()
dataset = Dataset(dict(ob=obs_ph,
ac=action_ph,
atarg=atarg,
vtarg=tdlamret),
shuffle=not self.policy.recurrent)
optim_batchsize = self.optim_batchsize or obs_ph.shape[0]
# set old parameter values to new parameter values
self.assign_old_eq_new(sess=self.sess)
logger.log("Optimizing...")
logger.log(fmt_row(13, self.loss_names))
# Here we do a bunch of optimization epochs over the data
for k in range(self.optim_epochs):
# list of tuples, each of which gives the loss for a minibatch
losses = []
for i, batch in enumerate(
dataset.iterate_once(optim_batchsize)):
steps = (
self.num_timesteps + k * optim_batchsize +
int(i *
(optim_batchsize / len(dataset.data_map))))
if writer is not None:
# run loss backprop with summary, but once every 10 runs save the metadata
# (memory, compute time, ...)
if self.full_tensorboard_log and (1 +
k) % 10 == 0:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, grad, *newlosses = self.lossandgrad(
batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(
run_metadata, 'step%d' % steps)
else:
summary, grad, *newlosses = self.lossandgrad(
batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess)
writer.add_summary(summary, steps)
else:
_, grad, *newlosses = self.lossandgrad(
batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess)
self.adam.update(grad,
self.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 dataset.iterate_once(optim_batchsize):
newlosses = self.compute_losses(batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess)
losses.append(newlosses)
mean_losses, _, _ = mpi_moments(losses, axis=0)
logger.log(fmt_row(13, mean_losses))
for (loss_val, name) in zipsame(mean_losses,
self.loss_names):
logger.record_tabular("loss_" + name, loss_val)
logger.record_tabular(
"ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
# local values
lrlocal = (seg["ep_lens"], seg["ep_rets"])
# list of tuples
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal)
lens, rews = map(flatten_lists, zip(*listoflrpairs))
lenbuffer.extend(lens)
rewbuffer.extend(rews)
if len(lenbuffer) > 0:
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)
current_it_timesteps = MPI.COMM_WORLD.allreduce(
seg["total_timestep"])
timesteps_so_far += current_it_timesteps
self.num_timesteps += current_it_timesteps
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", self.num_timesteps)
logger.record_tabular("TimeElapsed", time.time() - t_start)
if self.verbose >= 1 and MPI.COMM_WORLD.Get_rank() == 0:
logger.dump_tabular()
return self
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="DQN",
reset_num_timesteps=True,
replay_wrapper=None,
learning_curve=False,
test_t=None):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn(seed)
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
else:
prioritized_replay_beta_iters = self.prioritized_replay_beta_iters
self.beta_schedule = LinearSchedule(
prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0)
else:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
if replay_wrapper is not None:
assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER"
self.replay_buffer = replay_wrapper(self.replay_buffer)
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(
schedule_timesteps=int(self.exploration_fraction *
total_timesteps),
initial_p=1.0,
final_p=self.exploration_final_eps)
episode_rewards = [0.0]
self.cumul_reward = [0.0]
episode_successes = []
obs = self.env.reset()
reset = True
self.episode_reward = np.zeros((1, ))
# variables for test eval ##
test_step = test_t * 3
test_results = {'sum': []}
test_ts = []
for _ in range(total_timesteps):
## Test eval period ##
if learning_curve and _ % test_step == 0 and _ > 0:
print("--> Simulating test period")
self.env.reset()
test_r = 0.0
for i in range(test_t):
feasible_actions = AllocationEnv.get_feasible_actions(
obs["board_config"])
action_mask = AllocationEnv.get_action_mask(
feasible_actions, self.env.action_space.n)
action, _states = self.predict(obs, mask=action_mask)
action = AllocationEnv.check_action(
obs['board_config'], action)
obs, rewards, dones, info = self.env.step(action)
test_r += rewards
test_results["sum"].append(test_r)
test_ts.append(_)
self.env.reset()
# plot test eval progress
plt.plot(test_ts, test_results["sum"])
# plt.errorbar(iteration_cuts, results["mean"], yerr=results["std"], fmt='.k')
plt.xlabel("Iteration count")
plt.ylabel("Total (sum) test reward")
plt.savefig("figs/rl-learning-curve-{}.pdf".format(
cfg.vals['prj_name']))
plt.clf()
plt.close()
# write test eval progress
write_results = {}
for k, v in test_results.items():
write_results[k] = serialize_floats(v)
with open(
"output/rl-learning-curve-{}.json".format(
cfg.vals['prj_name']), 'w') as f:
json.dump(write_results, f)
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
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. - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
feasible_actions = AllocationEnv.get_feasible_actions(
obs["board_config"])
action_mask = AllocationEnv.get_action_mask(
feasible_actions, self.action_space.n)
with self.sess.as_default():
action = self.act(State.get_vec_observation(obs)[None],
update_eps=update_eps,
**kwargs,
mask=action_mask)[0]
reset = False
# CHECK IF ACTIONS IS FEASIBLE
action = AllocationEnv.check_action(obs['board_config'],
action)
env_action = action
new_obs, rew, done, info = self.env.step(env_action)
print("action: {} - reward: {} - eps: {:.4}".format(
action, rew, update_eps))
print(new_obs['day_vec'])
print(new_obs['board_config'])
# Store transition in the replay buffer.
self.replay_buffer.add(State.get_vec_observation(obs), action,
rew, State.get_vec_observation(new_obs),
float(done))
obs = new_obs
if writer is not None:
ep_rew = np.array([rew]).reshape((1, -1))
ep_done = np.array([done]).reshape((1, -1))
self.episode_reward = total_episode_reward_logger(
self.episode_reward, ep_rew, ep_done, writer,
self.num_timesteps)
episode_rewards[-1] += rew
self.cumul_reward.append(self.cumul_reward[-1] + rew)
if done:
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
# Do not train if the warmup phase is not over
# or if there are not enough samples in the replay buffer
can_sample = self.replay_buffer.can_sample(self.batch_size)
if can_sample and self.num_timesteps > self.learning_starts \
and self.num_timesteps % self.train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if self.prioritized_replay:
experience = self.replay_buffer.sample(
self.batch_size,
beta=self.beta_schedule.value(self.num_timesteps))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(
self.batch_size)
weights, batch_idxes = np.ones_like(rewards), None
if writer is not None:
# run loss backprop with summary, but once every 100 steps save the metadata
# (memory, compute time, ...)
if (1 + self.num_timesteps) % 100 == 0:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(
run_metadata, 'step%d' % self.num_timesteps)
else:
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
writer.add_summary(summary, self.num_timesteps)
else:
_, td_errors = self._train_step(obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
if self.prioritized_replay:
new_priorities = np.abs(
td_errors) + self.prioritized_replay_eps
self.replay_buffer.update_priorities(
batch_idxes, new_priorities)
if can_sample and self.num_timesteps > self.learning_starts and \
self.num_timesteps % self.target_network_update_freq == 0:
# Update target network periodically.
self.update_target(sess=self.sess)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
if self.verbose >= 1 and done and log_interval is not None and len(
episode_rewards) % log_interval == 0:
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate",
np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular(
"% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)))
logger.dump_tabular()
print('timestamp: {}'.format(self.num_timesteps, end='\r\n'))
self.num_timesteps += 1
return self
def learn(self, total_timesteps, callback=None, log_interval=100, tb_log_name="DQN",
reset_num_timesteps=True, replay_wrapper=None):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
else:
prioritized_replay_beta_iters = self.prioritized_replay_beta_iters
self.beta_schedule = LinearSchedule(prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0)
else:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
if replay_wrapper is not None:
assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER"
self.replay_buffer = replay_wrapper(self.replay_buffer)
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(schedule_timesteps=int(self.exploration_fraction * total_timesteps),
initial_p=self.exploration_initial_eps,
final_p=self.exploration_final_eps)
episode_rewards = [0.0]
episode_successes = []
callback.on_training_start(locals(), globals())
callback.on_rollout_start()
reset = True
obs = self.env.reset()
# Retrieve unnormalized observation for saving into the buffer
if self._vec_normalize_env is not None:
obs_ = self._vec_normalize_env.get_original_obs().squeeze()
for _ in range(total_timesteps):
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
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. - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs['update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
with self.sess.as_default():
action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0]
env_action = action
reset = False
new_obs, rew, done, info = self.env.step(env_action)
self.num_timesteps += 1
# Stop training if return value is False
callback.update_locals(locals())
if callback.on_step() is False:
break
# Store only the unnormalized version
if self._vec_normalize_env is not None:
new_obs_ = self._vec_normalize_env.get_original_obs().squeeze()
reward_ = self._vec_normalize_env.get_original_reward().squeeze()
else:
# Avoid changing the original ones
obs_, new_obs_, reward_ = obs, new_obs, rew
# Store transition in the replay buffer.
self.replay_buffer_add(obs_, action, reward_, new_obs_, done, info)
obs = new_obs
# Save the unnormalized observation
if self._vec_normalize_env is not None:
obs_ = new_obs_
if writer is not None:
ep_rew = np.array([reward_]).reshape((1, -1))
ep_done = np.array([done]).reshape((1, -1))
tf_util.total_episode_reward_logger(self.episode_reward, ep_rew, ep_done, writer,
self.num_timesteps)
episode_rewards[-1] += reward_
if done:
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
# Do not train if the warmup phase is not over
# or if there are not enough samples in the replay buffer
can_sample = self.replay_buffer.can_sample(self.batch_size)
if can_sample and self.num_timesteps > self.learning_starts \
and self.num_timesteps % self.train_freq == 0:
callback.on_rollout_end()
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
# pytype:disable=bad-unpacking
if self.prioritized_replay:
assert self.beta_schedule is not None, \
"BUG: should be LinearSchedule when self.prioritized_replay True"
experience = self.replay_buffer.sample(self.batch_size,
beta=self.beta_schedule.value(self.num_timesteps),
env=self._vec_normalize_env)
(obses_t, actions, rewards, obses_tp1, dones, weights, batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(self.batch_size,
env=self._vec_normalize_env)
weights, batch_idxes = np.ones_like(rewards), None
# pytype:enable=bad-unpacking
if writer is not None:
# run loss backprop with summary, but once every 100 steps save the metadata
# (memory, compute time, ...)
if (1 + self.num_timesteps) % 100 == 0:
run_options = tf.compat.v1.RunOptions(trace_level=tf.compat.v1.RunOptions.FULL_TRACE)
run_metadata = tf.compat.v1.RunMetadata()
summary, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1,
dones, weights, sess=self.sess, options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(run_metadata, 'step%d' % self.num_timesteps)
else:
summary, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1,
dones, weights, sess=self.sess)
writer.add_summary(summary, self.num_timesteps)
else:
_, td_errors = self._train_step(obses_t, actions, rewards, obses_tp1, obses_tp1, dones, weights,
sess=self.sess)
if self.prioritized_replay:
new_priorities = np.abs(td_errors) + self.prioritized_replay_eps
assert isinstance(self.replay_buffer, PrioritizedReplayBuffer)
self.replay_buffer.update_priorities(batch_idxes, new_priorities)
callback.on_rollout_start()
if can_sample and self.num_timesteps > self.learning_starts and \
self.num_timesteps % self.target_network_update_freq == 0:
# Update target network periodically.
self.update_target(sess=self.sess)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
if self.verbose >= 1 and done and log_interval is not None and len(episode_rewards) % log_interval == 0:
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate", np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward", mean_100ep_reward)
logger.record_tabular("% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)))
logger.dump_tabular()
callback.on_training_end()
return self
def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="DDPG"):
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name) as writer:
self._setup_learn(seed)
# a list for tensorboard logging, to prevent logging with the same step number, if it already occured
self.tb_seen_steps = []
rank = MPI.COMM_WORLD.Get_rank()
# we assume symmetric actions.
assert np.all(np.abs(self.env.action_space.low) == self.env.action_space.high)
if self.verbose >= 2:
logger.log('Using agent with the following configuration:')
logger.log(str(self.__dict__.items()))
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
self.episode_reward = np.zeros((1,))
with self.sess.as_default(), self.graph.as_default():
# Prepare everything.
self._reset()
obs = self.env.reset()
eval_obs = None
if self.eval_env is not None:
eval_obs = self.eval_env.reset()
episode_reward = 0.
episode_step = 0
episodes = 0
step = 0
total_steps = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
eval_episode_rewards = []
eval_qs = []
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
epoch = 0
while True:
for _ in range(log_interval):
# Perform rollouts.
for _ in range(self.nb_rollout_steps):
if total_steps >= total_timesteps:
return self
# Predict next action.
action, q_value = self._policy(obs, apply_noise=True, compute_q=True)
assert action.shape == self.env.action_space.shape
# Execute next action.
if rank == 0 and self.render:
self.env.render()
new_obs, reward, done, _ = self.env.step(action * np.abs(self.action_space.low))
if writer is not None:
ep_rew = np.array([reward]).reshape((1, -1))
ep_done = np.array([done]).reshape((1, -1))
self.episode_reward = total_episode_reward_logger(self.episode_reward, ep_rew, ep_done,
writer, total_steps)
step += 1
total_steps += 1
if rank == 0 and self.render:
self.env.render()
episode_reward += reward
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
epoch_qs.append(q_value)
self._store_transition(obs, action, reward, new_obs, done)
obs = new_obs
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) == False:
return self
if done:
# Episode done.
epoch_episode_rewards.append(episode_reward)
episode_rewards_history.append(episode_reward)
epoch_episode_steps.append(episode_step)
episode_reward = 0.
episode_step = 0
epoch_episodes += 1
episodes += 1
self._reset()
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(self.nb_train_steps):
# Adapt param noise, if necessary.
if self.memory.nb_entries >= self.batch_size and \
t_train % self.param_noise_adaption_interval == 0:
distance = self._adapt_param_noise()
epoch_adaptive_distances.append(distance)
# weird equation to deal with the fact the nb_train_steps will be different
# to nb_rollout_steps
step = (int(t_train * (self.nb_rollout_steps / self.nb_train_steps)) +
total_steps - self.nb_rollout_steps)
critic_loss, actor_loss = self._train_step(step, writer, log=t_train == 0)
epoch_critic_losses.append(critic_loss)
epoch_actor_losses.append(actor_loss)
self._update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if self.eval_env is not None:
eval_episode_reward = 0.
for _ in range(self.nb_eval_steps):
if total_steps >= total_timesteps:
return self
eval_action, eval_q = self._policy(eval_obs, apply_noise=False, compute_q=True)
eval_obs, eval_r, eval_done, _ = self.eval_env.step(eval_action *
np.abs(self.action_space.low))
if self.render_eval:
self.eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
if eval_done:
if not isinstance(self.env, VecEnv):
eval_obs = self.eval_env.reset()
eval_episode_rewards.append(eval_episode_reward)
eval_episode_rewards_history.append(eval_episode_reward)
eval_episode_reward = 0.
mpi_size = MPI.COMM_WORLD.Get_size()
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = self._get_stats()
combined_stats = stats.copy()
combined_stats['rollout/return'] = np.mean(epoch_episode_rewards)
combined_stats['rollout/return_history'] = np.mean(episode_rewards_history)
combined_stats['rollout/episode_steps'] = np.mean(epoch_episode_steps)
combined_stats['rollout/actions_mean'] = np.mean(epoch_actions)
combined_stats['rollout/Q_mean'] = np.mean(epoch_qs)
combined_stats['train/loss_actor'] = np.mean(epoch_actor_losses)
combined_stats['train/loss_critic'] = np.mean(epoch_critic_losses)
if len(epoch_adaptive_distances) != 0:
combined_stats['train/param_noise_distance'] = np.mean(epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(step) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats['rollout/episodes'] = epoch_episodes
combined_stats['rollout/actions_std'] = np.std(epoch_actions)
# Evaluation statistics.
if self.eval_env is not None:
combined_stats['eval/return'] = eval_episode_rewards
combined_stats['eval/return_history'] = np.mean(eval_episode_rewards_history)
combined_stats['eval/Q'] = eval_qs
combined_stats['eval/episodes'] = len(eval_episode_rewards)
def as_scalar(scalar):
"""
check and return the input if it is a scalar, otherwise raise ValueError
:param scalar: (Any) the object to check
:return: (Number) the scalar if x is a scalar
"""
if isinstance(scalar, np.ndarray):
assert scalar.size == 1
return scalar[0]
elif np.isscalar(scalar):
return scalar
else:
raise ValueError('expected scalar, got %s' % scalar)
combined_stats_sums = MPI.COMM_WORLD.allreduce(
np.array([as_scalar(x) for x in combined_stats.values()]))
combined_stats = {k: v / mpi_size for (k, v) in zip(combined_stats.keys(), combined_stats_sums)}
# Total statistics.
combined_stats['total/epochs'] = epoch + 1
combined_stats['total/steps'] = step
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(self.env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as file_handler:
pickle.dump(self.env.get_state(), file_handler)
if self.eval_env and hasattr(self.eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'), 'wb') as file_handler:
pickle.dump(self.eval_env.get_state(), file_handler)
def learn(self,
total_timesteps,
callback=None,
log_interval=100,
tb_log_name="DQN",
reset_num_timesteps=True,
replay_wrapper=None):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
else:
prioritized_replay_beta_iters = self.prioritized_replay_beta_iters
self.beta_schedule = LinearSchedule(
prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0)
else:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
if replay_wrapper is not None:
assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER"
self.replay_buffer = replay_wrapper(self.replay_buffer)
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(
schedule_timesteps=int(self.exploration_fraction *
total_timesteps),
initial_p=self.exploration_initial_eps,
final_p=self.exploration_final_eps)
episode_rewards = [0.0]
episode_successes = []
obs = self.env.reset()
obs_hdqn_old = None
action_hdqn = None
reset = True
F = 0
for _ in range(total_timesteps):
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
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).
update_param_noise_threshold = \
-np.log(1. - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
# Check if agent is busy or idle
OBS_IS_IDLE = True
if (OBS_IS_IDLE):
if not reset:
# Store HDQN transition
self.replay_buffer.add(obs_hdqn_old, action_hdqn, F,
obs, float(done))
# Select new goal for the agent using the current Q function
action = self.act(np.array(obs)[None],
update_eps=update_eps,
**kwargs)[0]
env_action = action
# Update bookkeepping for next HDQN buffer update
obs_hdqn_old = obs
action_hdqn = env_action
F = 0.
else:
# Agent is busy, so select a dummy action (it will be ignored anyway)
env_action = 0
reset = False
new_obs, rew, done, info = self.env.step(env_action)
F = F + rew
if writer is not None:
ep_rew = np.array([rew]).reshape((1, -1))
ep_done = np.array([done]).reshape((1, -1))
total_episode_reward_logger(self.episode_reward, ep_rew,
ep_done, writer,
self.num_timesteps)
episode_rewards[-1] += rew
if done:
# Store HDQN transition
self.replay_buffer.add(obs_hdqn_old, action_hdqn, F, obs,
float(done))
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
# Do not train if the warmup phase is not over
# or if there are not enough samples in the replay buffer
can_sample = self.replay_buffer.can_sample(self.batch_size)
if can_sample and self.num_timesteps > self.learning_starts \
and self.num_timesteps % self.train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
# pytype:disable=bad-unpacking
if self.prioritized_replay:
assert self.beta_schedule is not None, \
"BUG: should be LinearSchedule when self.prioritized_replay True"
experience = self.replay_buffer.sample(
self.batch_size,
beta=self.beta_schedule.value(self.num_timesteps))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(
self.batch_size)
weights, batch_idxes = np.ones_like(rewards), None
# pytype:enable=bad-unpacking
if writer is not None:
# run loss backprop with summary, but once every 100 steps save the metadata
# (memory, compute time, ...)
if (1 + self.num_timesteps) % 100 == 0:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(
run_metadata, 'step%d' % self.num_timesteps)
else:
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
writer.add_summary(summary, self.num_timesteps)
else:
_, td_errors = self._train_step(obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
if self.prioritized_replay:
new_priorities = np.abs(
td_errors) + self.prioritized_replay_eps
assert isinstance(self.replay_buffer,
PrioritizedReplayBuffer)
self.replay_buffer.update_priorities(
batch_idxes, new_priorities)
if can_sample and self.num_timesteps > self.learning_starts and \
self.num_timesteps % self.target_network_update_freq == 0:
# Update target network periodically.
self.update_target(sess=self.sess)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
if self.verbose >= 1 and done and log_interval is not None and len(
episode_rewards) % log_interval == 0:
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate",
np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular(
"% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)))
logger.dump_tabular()
self.num_timesteps += 1
return self
def learn(self,
total_timesteps,
callback=None,
log_interval=100,
tb_log_name="PPO1",
reset_num_timesteps=True,
save_path=None,
save_iters=20):
is_root = (MPI.COMM_WORLD.Get_rank() == 0)
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the PPO1 model must be " \
"an instance of common.policies.ActorCriticPolicy."
with self.sess.as_default():
self.adam.sync()
callback.on_training_start(locals(), globals())
# Prepare for rollouts
seg_gen = traj_segment_generator(self.policy_pi,
self.env,
self.timesteps_per_actorbatch,
callback=callback)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
t_start = time.time()
# rolling buffer for episode lengths
len_buffer = deque(maxlen=100)
# rolling buffer for episode rewards
reward_buffer = deque(maxlen=100)
while True:
t_episode = time.time()
if timesteps_so_far >= total_timesteps:
break
if self.schedule == 'constant':
cur_lrmult = 1.0
elif self.schedule == 'linear':
cur_lrmult = max(
1.0 - float(timesteps_so_far) / total_timesteps, 0)
else:
raise NotImplementedError
if is_root:
logger.log("********** Iteration %i ************" %
iters_so_far)
seg = seg_gen.__next__()
# Stop training early (triggered by the callback)
if not seg.get('continue_training', True): # pytype: disable=attribute-error
break
add_vtarg_and_adv(seg, self.num_robot, self.gamma,
self.lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
observations, actions = seg["observations"], seg["actions"]
atarg, tdlamret = seg["adv"], seg["tdlamret"]
# true_rew is the reward without discount
# if writer is not None:
# total_episode_reward_logger(self.episode_reward,
# seg["true_rewards"].reshape((self.n_envs, -1)),
# writer, self.num_timesteps, int(self.timesteps_per_actorbatch/100)) # step write reward sum
# predicted value function before udpate
vpredbefore = seg["vpred"]
# standardized advantage function estimate
# atarg = (atarg - atarg.mean()) / atarg.std()
temp_atarg = [[] for _ in range(self.num_robot)]
for i in range(
int(self.timesteps_per_actorbatch /
self.num_robot)):
for j in range(self.num_robot):
temp_atarg[j].append(atarg[i * self.num_robot + j])
for i in range(self.num_robot):
temp_atarg[i] = np.array(temp_atarg[i])
temp_atarg[i] = (temp_atarg[i] - temp_atarg[i].mean()
) / temp_atarg[i].std()
for i in range(
int(self.timesteps_per_actorbatch /
self.num_robot)):
for j in range(self.num_robot):
atarg[i * self.num_robot + j] = temp_atarg[j][i]
dataset = Dataset(dict(ob=observations,
ac=actions,
atarg=atarg,
vtarg=tdlamret),
shuffle=not self.policy.recurrent)
optim_batchsize = self.optim_batchsize or observations.shape[
0]
# set old parameter values to new parameter values
self.assign_old_eq_new(sess=self.sess)
if is_root:
logger.log("Optimizing...")
logger.log(fmt_row(13, self.loss_names))
# Here we do a bunch of optimization epochs over the data
for k in range(self.optim_epochs):
# list of tuples, each of which gives the loss for a minibatch
losses = []
for i, batch in enumerate(
dataset.iterate_once(optim_batchsize)):
steps = (
self.num_timesteps + k * optim_batchsize +
int(i *
(optim_batchsize / len(dataset.data_map))))
if writer is not None:
# run loss backprop with summary, but once every 10 runs save the metadata
# (memory, compute time, ...)
if self.full_tensorboard_log and (1 +
k) % 10 == 0:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, grad, *newlosses = self.lossandgrad(
batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(
run_metadata, 'step%d' % steps)
else:
summary, grad, *newlosses = self.lossandgrad(
batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess)
writer.add_summary(summary, steps)
else:
_, grad, *newlosses = self.lossandgrad(
batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess)
self.adam.update(grad,
self.optim_stepsize * cur_lrmult)
losses.append(newlosses)
if is_root:
logger.log(fmt_row(13, np.mean(losses, axis=0)))
if is_root:
logger.log("Evaluating losses...")
losses = []
for batch in dataset.iterate_once(optim_batchsize):
newlosses = self.compute_losses(batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess)
losses.append(newlosses)
mean_losses, _, _ = mpi_moments(losses, axis=0)
if is_root:
logger.log(fmt_row(13, mean_losses))
for (loss_val, name) in zipsame(mean_losses,
self.loss_names):
logger.record_tabular("loss_" + name, loss_val)
logger.record_tabular(
"ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
# local values
lrlocal = (seg["ep_lens"], seg["ep_rets"])
# list of tuples
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal)
lens, rews = map(flatten_lists, zip(*listoflrpairs))
if writer is not None:
for i in range(len(rews)):
summary = tf.Summary(value=[
tf.Summary.Value(tag="episode_reward",
simple_value=rews[i])
])
writer.add_summary(summary, self.num_timesteps + i)
len_buffer.extend(lens)
reward_buffer.extend(rews)
if len(len_buffer) > 0:
logger.record_tabular("EpLenMean", np.mean(len_buffer))
logger.record_tabular("EpRewMean",
np.mean(reward_buffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
current_it_timesteps = MPI.COMM_WORLD.allreduce(
seg["total_timestep"])
timesteps_so_far += current_it_timesteps
self.num_timesteps += current_it_timesteps
if is_root and (save_path
is not None) and (iters_so_far % save_iters
== 0):
self.save(save_path)
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", self.num_timesteps)
logger.record_tabular("TimeElapsed", time.time() - t_start)
logger.record_tabular("TimePerEpisode",
time.time() - t_episode)
if self.verbose >= 1 and is_root:
logger.dump_tabular()
callback.on_training_end()
if is_root:
self.save(save_path)
return self
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100):
with SetVerbosity(self.verbose):
self._setup_learn(seed)
with self.sess.as_default():
self.adam.sync()
# Prepare for rollouts
seg_gen = traj_segment_generator(self.policy_pi, self.env,
self.timesteps_per_actorbatch)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
t_start = time.time()
# rolling buffer for episode lengths
lenbuffer = deque(maxlen=100)
# rolling buffer for episode rewards
rewbuffer = deque(maxlen=100)
while True:
if callback:
callback(locals(), globals())
if total_timesteps and timesteps_so_far >= total_timesteps:
break
if self.schedule == 'constant':
cur_lrmult = 1.0
elif self.schedule == 'linear':
cur_lrmult = max(
1.0 - float(timesteps_so_far) / total_timesteps, 0)
else:
raise NotImplementedError
logger.log("********** Iteration %i ************" %
iters_so_far)
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, self.gamma, self.lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
obs_ph, action_ph, atarg, tdlamret = seg["ob"], seg[
"ac"], seg["adv"], seg["tdlamret"]
# predicted value function before udpate
vpredbefore = seg["vpred"]
# standardized advantage function estimate
atarg = (atarg - atarg.mean()) / atarg.std()
dataset = Dataset(
dict(ob=obs_ph,
ac=action_ph,
atarg=atarg,
vtarg=tdlamret),
shuffle=not issubclass(self.policy, LstmPolicy))
optim_batchsize = self.optim_batchsize or obs_ph.shape[0]
# set old parameter values to new parameter values
self.assign_old_eq_new(sess=self.sess)
logger.log("Optimizing...")
logger.log(fmt_row(13, self.loss_names))
# Here we do a bunch of optimization epochs over the data
for _ in range(self.optim_epochs):
# list of tuples, each of which gives the loss for a minibatch
losses = []
for batch in dataset.iterate_once(optim_batchsize):
*newlosses, grad = self.lossandgrad(batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess)
self.adam.update(grad,
self.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 dataset.iterate_once(optim_batchsize):
newlosses = self.compute_losses(batch["ob"],
batch["ob"],
batch["ac"],
batch["atarg"],
batch["vtarg"],
cur_lrmult,
sess=self.sess)
losses.append(newlosses)
mean_losses, _, _ = mpi_moments(losses, axis=0)
logger.log(fmt_row(13, mean_losses))
for (loss_val, name) in zipsame(mean_losses,
self.loss_names):
logger.record_tabular("loss_" + name, loss_val)
logger.record_tabular(
"ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
# local values
lrlocal = (seg["ep_lens"], seg["ep_rets"])
# list of tuples
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal)
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 += seg["total_timestep"]
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - t_start)
if self.verbose >= 1 and MPI.COMM_WORLD.Get_rank() == 0:
logger.dump_tabular()
return self
def learn(self,
total_timesteps,
callback=None,
log_interval=100,
tb_log_name="TRPO",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(
self.graph, self.tensorboard_log, tb_log_name,
new_tb_log) as writer:
self._setup_learn()
with self.sess.as_default():
callback.on_training_start(locals(), globals())
seg_gen = traj_segment_generator(self.policy_pi,
self.env,
self.timesteps_per_batch,
callback=callback)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
t_start = time.time()
len_buffer = deque(
maxlen=40) # rolling buffer for episode lengths
reward_buffer = deque(
maxlen=40) # rolling buffer for episode rewards
while True:
if timesteps_so_far >= total_timesteps:
break
logger.log("********** Iteration %i ************" %
iters_so_far)
def fisher_vector_product(vec):
return self.allmean(
self.compute_fvp(
vec, *fvpargs,
sess=self.sess)) + self.cg_damping * vec
# ------------------ Update G ------------------
logger.log("Optimizing Policy...")
# g_step = 1 when not using GAIL
mean_losses = None
vpredbefore = None
tdlamret = None
observation = None
action = None
seg = None
for k in range(self.g_step):
with self.timed("sampling"):
seg = seg_gen.__next__()
# Stop training early (triggered by the callback)
if not seg.get('continue_training', True): # pytype: disable=attribute-error
break
add_vtarg_and_adv(seg, self.gamma, self.lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
observation, action = seg["observations"], seg[
"actions"]
atarg, tdlamret = seg["adv"], seg["tdlamret"]
vpredbefore = seg[
"vpred"] # predicted value function before update
atarg = (atarg - atarg.mean()) / (
atarg.std() + 1e-8
) # standardized advantage function estimate
print('advantages: ', np.min(atarg), np.max(atarg),
np.mean(atarg))
# true_rew is the reward without discount
if writer is not None:
total_episode_reward_logger(
self.episode_reward,
seg["true_rewards"].reshape(
(self.n_envs, -1)), seg["dones"].reshape(
(self.n_envs, -1)), writer,
self.num_timesteps)
args = seg["observations"], seg["observations"], seg[
"actions"], atarg
# Subsampling: see p40-42 of John Schulman thesis
# http://joschu.net/docs/thesis.pdf
fvpargs = [arr[::5] for arr in args]
self.assign_old_eq_new(sess=self.sess)
with self.timed("computegrad"):
steps = self.num_timesteps + (k + 1) * (
seg["total_timestep"] / self.g_step)
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata(
) if self.full_tensorboard_log else None
_, grad, *lossbefore = self.compute_lossandgrad(
*args,
tdlamret,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
print(f'losses before', lossbefore)
lossbefore = self.allmean(np.array(lossbefore))
grad = self.allmean(grad)
if np.allclose(grad, 0):
logger.log("Got zero gradient. not updating")
else:
with self.timed("conjugate_gradient"):
stepdir = conjugate_gradient(
fisher_vector_product,
grad,
cg_iters=self.cg_iters,
verbose=self.rank == 0
and self.verbose >= 1)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(
fisher_vector_product(stepdir))
# abs(shs) to avoid taking square root of negative values
lagrange_multiplier = np.sqrt(
abs(shs) / self.max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lagrange_multiplier
expectedimprove = grad.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = self.get_flat()
for _ in range(10):
thnew = thbefore + fullstep * stepsize
self.set_from_flat(thnew)
mean_losses = surr, kl_loss, *_ = self.allmean(
np.array(
self.compute_losses(*args,
sess=self.sess)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" %
(expectedimprove, improve))
if not np.isfinite(mean_losses).all():
logger.log(
"Got non-finite value of losses -- bad!"
)
elif kl_loss > self.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")
self.set_from_flat(thbefore)
if self.nworkers > 1 and iters_so_far % 20 == 0:
# list of tuples
paramsums = MPI.COMM_WORLD.allgather(
(thnew.sum(), self.vfadam.getflat().sum()))
assert all(
np.allclose(ps, paramsums[0])
for ps in paramsums[1:])
for (loss_name,
loss_val) in zip(self.loss_names,
mean_losses):
logger.record_tabular(loss_name, loss_val)
with self.timed("vf"):
for _ in range(self.vf_iters):
# NOTE: for recurrent policies, use shuffle=False?
for (mbob, mbret) in dataset.iterbatches(
(seg["observations"], seg["tdlamret"]),
include_final_partial_batch=False,
batch_size=128,
shuffle=True):
grad = self.allmean(
self.compute_vflossandgrad(
mbob, mbob, mbret, sess=self.sess))
self.vfadam.update(grad, self.vf_stepsize)
# Stop training early (triggered by the callback)
if not seg.get('continue_training', True): # pytype: disable=attribute-error
break
logger.record_tabular(
"explained_variance_tdlam_before",
explained_variance(vpredbefore, tdlamret))
# lr: lengths and rewards
lr_local = (seg["ep_lens"], seg["ep_rets"]) # local values
list_lr_pairs = MPI.COMM_WORLD.allgather(
lr_local) # list of tuples
lens, rews = map(flatten_lists, zip(*list_lr_pairs))
len_buffer.extend(lens)
reward_buffer.extend(rews)
if len(len_buffer) > 0:
logger.record_tabular("EpLenMean", np.mean(len_buffer))
logger.record_tabular("EpRewMean",
np.mean(reward_buffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
current_it_timesteps = MPI.COMM_WORLD.allreduce(
seg["total_timestep"])
timesteps_so_far += current_it_timesteps
self.num_timesteps += current_it_timesteps
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", self.num_timesteps)
logger.record_tabular("TimeElapsed", time.time() - t_start)
if self.verbose >= 1 and self.rank == 0:
logger.dump_tabular()
callback.on_training_end()
return self
def train(policy, rollout_worker, evaluator, n_epochs, n_test_rollouts,
n_cycles, n_batches, policy_save_interval, save_policies):
"""
train the given policy
:param policy: (her.DDPG) the policy to train
:param rollout_worker: (RolloutWorker) Rollout worker generates experience for training.
:param evaluator: (RolloutWorker) Rollout worker for evalutation
:param n_epochs: (int) the number of epochs
:param n_test_rollouts: (int) the number of for the evalutation RolloutWorker
:param n_cycles: (int) the number of cycles for training per epoch
:param n_batches: (int) the batch size
:param policy_save_interval: (int) the interval with which policy pickles are saved.
If set to 0, only the best and latest policy will be pickled.
:param save_policies: (bool) whether or not to save the policies
"""
rank = MPI.COMM_WORLD.Get_rank()
latest_policy_path = os.path.join(logger.get_dir(), 'policy_latest.pkl')
best_policy_path = os.path.join(logger.get_dir(), 'policy_best.pkl')
periodic_policy_path = os.path.join(logger.get_dir(), 'policy_{}.pkl')
logger.info("Training...")
best_success_rate = -1
for epoch in range(n_epochs):
# train
rollout_worker.clear_history()
for _ in range(n_cycles):
episode = rollout_worker.generate_rollouts()
policy.store_episode(episode)
for _ in range(n_batches):
policy.train_step()
policy.update_target_net()
# test
evaluator.clear_history()
for _ in range(n_test_rollouts):
evaluator.generate_rollouts()
# record logs
logger.record_tabular('epoch', epoch)
for key, val in evaluator.logs('test'):
logger.record_tabular(key, mpi_average(val))
for key, val in rollout_worker.logs('train'):
logger.record_tabular(key, mpi_average(val))
for key, val in policy.logs():
logger.record_tabular(key, mpi_average(val))
if rank == 0:
logger.dump_tabular()
# save the policy if it's better than the previous ones
success_rate = mpi_average(evaluator.current_success_rate())
if rank == 0 and success_rate >= best_success_rate and save_policies:
best_success_rate = success_rate
logger.info(
'New best success rate: {}. Saving policy to {} ...'.format(
best_success_rate, best_policy_path))
evaluator.save_policy(best_policy_path)
evaluator.save_policy(latest_policy_path)
if rank == 0 and policy_save_interval > 0 and epoch % policy_save_interval == 0 and save_policies:
policy_path = periodic_policy_path.format(epoch)
logger.info('Saving periodic policy to {} ...'.format(policy_path))
evaluator.save_policy(policy_path)
# make sure that different threads have different seeds
local_uniform = np.random.uniform(size=(1, ))
root_uniform = local_uniform.copy()
MPI.COMM_WORLD.Bcast(root_uniform, root=0)
if rank != 0:
assert local_uniform[0] != root_uniform[0]
def learn(
self,
total_timesteps,
model_coworker,
role,
callback=None,
log_interval=100,
tb_log_name="DQN",
reset_num_timesteps=True,
replay_wrapper=None,
clipping_during_training=True,
):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
with SetVerbosity(self.verbose), TensorboardWriter(
self.graph, self.tensorboard_log, tb_log_name,
new_tb_log) as writer:
self._setup_learn()
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
else:
prioritized_replay_beta_iters = self.prioritized_replay_beta_iters
self.beta_schedule = LinearSchedule(
prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0,
)
else:
if self.replay_buffer is None:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
if replay_wrapper is not None:
assert (not self.prioritized_replay
), "Prioritized replay buffer is not supported by HER"
self.replay_buffer = replay_wrapper(self.replay_buffer)
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(
schedule_timesteps=int(self.exploration_fraction *
total_timesteps),
initial_p=self.exploration_initial_eps,
final_p=self.exploration_final_eps,
)
episode_rewards = [0.0]
episode_successes = []
callback.on_training_start(locals(), globals())
callback.on_rollout_start()
reset = True
obs = self.env.reset()
# Retrieve unnormalized observation for saving into the buffer
if self._vec_normalize_env is not None:
obs_ = self._vec_normalize_env.get_original_obs().squeeze()
for _ in range(total_timesteps):
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
update_param_noise_threshold = 0.0
else:
update_eps = 0.0
# 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(
1.0 - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) /
float(self.env.action_space.n))
kwargs["reset"] = reset
kwargs[
"update_param_noise_threshold"] = update_param_noise_threshold
kwargs["update_param_noise_scale"] = True
with self.sess.as_default():
action = self.act(np.array(obs)[None],
update_eps=update_eps,
**kwargs)[0]
turn, speed = None, None
if role == "turn":
turn = action
speed, nothing = model_coworker.predict(np.array(obs))
else:
turn, nothing = model_coworker.predict(np.array(obs))
speed = action
if clipping_during_training:
# check if next state (after action) would be outside of fish tank (CLIPPING)
env_state = self.env.get_state()
turn_speed = self.env.action([turn, speed])
global_turn = env_state[0][2] + turn_speed[0]
coords = np.array([
env_state[0][0] + turn_speed[1] * np.cos(global_turn),
env_state[0][1] + turn_speed[1] * np.sin(global_turn),
])
changed = False
if coords[0] < -0.49:
coords[0] = -0.47
changed = True
elif coords[0] > 0.49:
coords[0] = 0.47
changed = True
if coords[1] < -0.49:
coords[1] = -0.47
changed = True
elif coords[1] > 0.49:
coords[1] = 0.47
changed = True
if changed:
diff = coords - env_state[0, :2]
speed = np.linalg.norm(diff)
angles = np.arctan2(diff[1], diff[0])
turn = angles - env_state[0, 2]
turn = turn - 2 * np.pi if turn > np.pi else turn
turn = turn + 2 * np.pi if turn < -np.pi else turn
# convert to DQN output
dist_turn = np.abs(self.env.turn_rate_bins - turn)
dist_speed = np.abs(self.env.speed_bins - speed)
# convert to bins
turn = np.argmin(dist_turn, axis=0)
speed = np.argmin(dist_speed, axis=0)
if role == "turn":
action = turn
else:
action = speed
reset = False
new_obs, rew, done, info = self.env.step([turn, speed])
self.num_timesteps += 1
# Stop training if return value is False
if callback.on_step() is False:
break
# Store only the unnormalized version
if self._vec_normalize_env is not None:
new_obs_ = self._vec_normalize_env.get_original_obs(
).squeeze()
reward_ = self._vec_normalize_env.get_original_reward(
).squeeze()
else:
# Avoid changing the original ones
obs_, new_obs_, reward_ = obs, new_obs, rew
# Store transition in the replay buffer, but change reward to 0 (use it for plot later though)
self.replay_buffer.add(obs_, action, 0, new_obs_, float(done))
# Also give transition to model coworker
if model_coworker.replay_buffer is None:
model_coworker.replay_buffer = ReplayBuffer(
self.buffer_size)
if role == "turn":
model_coworker.replay_buffer.add(obs_, speed, 0, new_obs_,
float(done))
else:
model_coworker.replay_buffer.add(obs_, turn, 0, new_obs_,
float(done))
obs = new_obs
# Save the unnormalized observation
if self._vec_normalize_env is not None:
obs_ = new_obs_
if writer is not None:
ep_rew = np.array([reward_]).reshape((1, -1))
ep_done = np.array([done]).reshape((1, -1))
tf_util.total_episode_reward_logger(
self.episode_reward, ep_rew, ep_done, writer,
self.num_timesteps)
episode_rewards[-1] += reward_
if done:
maybe_is_success = info.get("is_success")
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
# Do not train if the warmup phase is not over
# or if there are not enough samples in the replay buffer
can_sample = self.replay_buffer.can_sample(self.batch_size)
if (can_sample and self.num_timesteps > self.learning_starts
and self.num_timesteps % self.train_freq == 0):
callback.on_rollout_end()
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
# pytype:disable=bad-unpacking
if self.prioritized_replay:
assert (
self.beta_schedule is not None
), "BUG: should be LinearSchedule when self.prioritized_replay True"
experience = self.replay_buffer.sample(
self.batch_size,
beta=self.beta_schedule.value(self.num_timesteps),
env=self._vec_normalize_env,
)
(
obses_t,
actions,
rewards,
obses_tp1,
dones,
weights,
batch_idxes,
) = experience
else:
(
obses_t,
actions,
rewards,
obses_tp1,
dones,
) = self.replay_buffer.sample(
self.batch_size, env=self._vec_normalize_env)
# also sample from expert buffer
(
obses_t_exp,
actions_exp,
rewards_exp,
obses_tp1_exp,
dones_exp,
) = self.expert_buffer.sample(
self.batch_size, env=self._vec_normalize_env)
weights, batch_idxes = np.ones_like(rewards), None
weights_exp, batch_idxes_exp = np.ones_like(
rewards_exp), None
# pytype:enable=bad-unpacking
if writer is not None:
# run loss backprop with summary, but once every 100 steps save the metadata
# (memory, compute time, ...)
if (1 + self.num_timesteps) % 100 == 0:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, td_errors = self._train_step(
np.append(obses_t, obses_t_exp, axis=0),
np.append(actions,
actions_exp.flatten(),
axis=0),
np.append(rewards,
rewards_exp.flatten(),
axis=0),
np.append(obses_tp1, obses_tp1_exp, axis=0),
np.append(obses_tp1, obses_tp1_exp, axis=0),
np.append(dones.flatten(),
dones_exp.flatten(),
axis=0),
np.append(weights, weights_exp),
sess=self.sess,
options=run_options,
run_metadata=run_metadata,
)
writer.add_run_metadata(
run_metadata, "step%d" % self.num_timesteps)
else:
summary, td_errors = self._train_step(
np.append(obses_t, obses_t_exp, axis=0),
np.append(actions,
actions_exp.flatten(),
axis=0),
np.append(rewards,
rewards_exp.flatten(),
axis=0),
np.append(obses_tp1, obses_tp1_exp, axis=0),
np.append(obses_tp1, obses_tp1_exp, axis=0),
np.append(dones.flatten(),
dones_exp.flatten(),
axis=0),
np.append(weights, weights_exp),
sess=self.sess,
options=run_options,
run_metadata=run_metadata,
)
writer.add_summary(summary, self.num_timesteps)
else:
_, td_errors = self._train_step(
np.append(obses_t, obses_t_exp, axis=0),
np.append(actions, actions_exp.flatten(), axis=0),
np.append(rewards, rewards_exp.flatten(), axis=0),
np.append(obses_tp1, obses_tp1_exp, axis=0),
np.append(obses_tp1, obses_tp1_exp, axis=0),
np.append(dones.flatten(),
dones_exp.flatten(),
axis=0),
np.append(weights, weights_exp),
sess=self.sess,
)
if self.prioritized_replay:
new_priorities = np.abs(
td_errors) + self.prioritized_replay_eps
assert isinstance(self.replay_buffer,
PrioritizedReplayBuffer)
self.replay_buffer.update_priorities(
batch_idxes, new_priorities)
callback.on_rollout_start()
if (can_sample and self.num_timesteps > self.learning_starts
and self.num_timesteps %
self.target_network_update_freq == 0):
# Update target network periodically.
self.update_target(sess=self.sess)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
if (self.verbose >= 1 and done and log_interval is not None
and len(episode_rewards) % log_interval == 0):
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate",
np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular(
"% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)),
)
logger.dump_tabular()
callback.on_training_end()
return self
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="SIL_A2C"):
with SetVerbosity(self.verbose), \
TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name) as writer: # type: tf.summary.FileWriter
self._setup_learn(seed)
self.learning_rate_schedule = Scheduler(
initial_value=self.learning_rate,
n_values=total_timesteps,
schedule=self.lr_schedule)
runner = SelfImitationA2CRunner(self.env,
self,
n_steps=self.n_steps,
gamma=self.gamma)
self.episode_reward = np.zeros((self.n_envs, ))
t_start = time.time()
for update in range(1, total_timesteps // self.n_batch + 1):
# true_reward is the reward without discount
obs, states, rewards, masks, actions, values, true_reward, raw_rewards = runner.run(
)
_, value_loss, policy_entropy = self._train_step(
obs, states, rewards, masks, actions, values, update,
writer)
sil_loss, sil_adv, sil_samples, sil_nlogp = self._train_sil()
n_seconds = time.time() - t_start
fps = int((update * self.n_batch) / n_seconds)
if writer is not None:
self.episode_reward = total_episode_reward_logger(
self.episode_reward,
raw_rewards.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)), writer,
update * (self.n_batch + 1))
summary = tf.Summary(value=[
tf.Summary.Value(
tag="episode_reward/best_reward",
simple_value=self.sil.get_best_reward())
])
writer.add_summary(summary, update * (self.n_batch + 1))
if callback is not None:
callback(locals(), globals())
if self.verbose >= 1 and (update % log_interval == 0
or update == 1):
explained_var = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps",
update * self.n_batch)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy",
float(policy_entropy))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance",
float(explained_var))
logger.record_tabular("best_episode_reward",
float(self.sil.get_best_reward()))
if self.sil_update > 0:
logger.record_tabular("sil_num_episodes",
float(self.sil.num_episodes()))
logger.record_tabular("sil_valid_samples",
float(sil_samples))
logger.record_tabular("sil_steps",
float(self.sil.num_steps()))
logger.dump_tabular()
if update % (log_interval * 20) == 0:
self.save(writer.get_logdir())
return self
def learn(self, total_timesteps, callback=None, vae=None, skip_episodes=5):
rank = MPI.COMM_WORLD.Get_rank()
# we assume symmetric actions.
assert np.all(
np.abs(self.env.action_space.low) == self.env.action_space.high)
self.episode_reward = np.zeros((1, ))
with self.sess.as_default(), self.graph.as_default():
# Prepare everything.
self._reset()
episode_reward = 0.
episode_step = 0
episodes = 0
step = 0
total_steps = 0
start_time = time.time()
actor_losses = []
critic_losses = []
while True:
obs = self.env.reset()
# Rollout one episode.
while True:
if total_steps >= total_timesteps:
return self
# Predict next action.
action, q_value = self._policy(obs,
apply_noise=True,
compute_q=True)
print(action)
assert action.shape == self.env.action_space.shape
# Execute next action.
if rank == 0 and self.render:
self.env.render()
new_obs, reward, done, _ = self.env.step(
action * np.abs(self.action_space.low))
step += 1
total_steps += 1
if rank == 0 and self.render:
self.env.render()
episode_reward += reward
episode_step += 1
# Book-keeping.
# Do not record observations, while we skip DDPG training.
if (episodes + 1) > skip_episodes:
self._store_transition(obs, action, reward, new_obs,
done)
obs = new_obs
if callback is not None:
callback(locals(), globals())
if done:
print("episode finished. Reward: ", episode_reward)
# Episode done.
episode_reward = 0.
episode_step = 0
episodes += 1
self._reset()
obs = self.env.reset()
# Finish rollout on episode finish.
break
print("rollout finished")
# Train VAE.
train_start = time.time()
vae.optimize()
print("VAE training duration:", time.time() - train_start)
# Train DDPG.
actor_losses = []
critic_losses = []
train_start = time.time()
if episodes > skip_episodes:
for t_train in range(self.nb_train_steps):
critic_loss, actor_loss = self._train_step(
0, None, log=t_train == 0)
critic_losses.append(critic_loss)
actor_losses.append(actor_loss)
self._update_target_net()
print("DDPG training duration:", time.time() - train_start)
mpi_size = MPI.COMM_WORLD.Get_size()
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = self._get_stats()
combined_stats = stats.copy()
combined_stats['train/loss_actor'] = np.mean(actor_losses)
combined_stats['train/loss_critic'] = np.mean(
critic_losses)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(
step) / float(duration)
combined_stats['total/episodes'] = episodes
def as_scalar(scalar):
"""
check and return the input if it is a scalar, otherwise raise ValueError
:param scalar: (Any) the object to check
:return: (Number) the scalar if x is a scalar
"""
if isinstance(scalar, np.ndarray):
assert scalar.size == 1
return scalar[0]
elif np.isscalar(scalar):
return scalar
else:
raise ValueError('expected scalar, got %s' %
scalar)
combined_stats_sums = MPI.COMM_WORLD.allreduce(
np.array(
[as_scalar(x) for x in combined_stats.values()]))
combined_stats = {
k: v / mpi_size
for (k, v) in zip(combined_stats.keys(),
combined_stats_sums)
}
# Total statistics.
combined_stats['total/steps'] = step
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
logger.dump_tabular()
logger.info('')
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="A2C",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn(seed)
self.learning_rate_schedule = Scheduler(
initial_value=self.learning_rate,
n_values=total_timesteps,
schedule=self.lr_schedule)
runner = A2CRunner(self.env,
self,
n_steps=self.n_steps,
gamma=self.gamma)
self.episode_reward = np.zeros((self.n_envs, ))
# Training stats (when using Monitor wrapper)
ep_info_buf = deque(maxlen=100)
t_start = time.time()
for update in range(1, total_timesteps // self.n_batch + 1):
# true_reward is the reward without discount
obs, states, rewards, masks, actions, values, ep_infos, true_reward = runner.run(
)
ep_info_buf.extend(ep_infos)
_, value_loss, policy_entropy = self._train_step(
obs, states, rewards, masks, actions, values,
self.num_timesteps // self.n_batch, writer)
n_seconds = time.time() - t_start
fps = int((update * self.n_batch) / n_seconds)
if writer is not None:
self.episode_reward = total_episode_reward_logger(
self.episode_reward,
true_reward.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)), writer,
self.num_timesteps)
self.num_timesteps += self.n_batch
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
if self.verbose >= 1 and (update % log_interval == 0
or update == 1):
explained_var = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps",
self.num_timesteps)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy",
float(policy_entropy))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance",
float(explained_var))
if len(ep_info_buf) > 0 and len(ep_info_buf[0]) > 0:
logger.logkv(
'ep_reward_mean',
safe_mean(
[ep_info['r'] for ep_info in ep_info_buf]))
logger.logkv(
'ep_len_mean',
safe_mean(
[ep_info['l'] for ep_info in ep_info_buf]))
logger.dump_tabular()
return self
def main(args):
"""
Train a DeepQ agent on cartpole env
:param args: (Parsed Arguments) the input arguments
"""
with tf_utils.make_session(8):
# Create the environment
env = gym.make("CartPole-v0")
# Create all the functions necessary to train the model
act, train, update_target, _ = deepq.build_train(
make_obs_ph=lambda name: ObservationInput(env.observation_space, 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.
tf_utils.initialize()
update_target()
episode_rewards = [0.0]
obs = env.reset()
for step in itertools.count():
# Take action and update exploration to the newest value
action = act(obs[None], update_eps=exploration.value(step))[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()
episode_rewards.append(0)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(float(np.mean(episode_rewards[-101:-1])), 1)
is_solved = step > 100 and mean_100ep_reward >= 200
if args.no_render and step > args.max_timesteps:
break
if is_solved:
if args.no_render:
break
# Show off the result
env.render()
else:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if step > 1000:
obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(32)
train(obses_t, actions, rewards, obses_tp1, dones, np.ones_like(rewards))
# Update target network periodically.
if step % 1000 == 0:
update_target()
if done and len(episode_rewards) % 10 == 0:
logger.record_tabular("steps", step)
logger.record_tabular("episodes", len(episode_rewards))
logger.record_tabular("mean episode reward", mean_100ep_reward)
logger.record_tabular("% time spent exploring", int(100 * exploration.value(step)))
logger.dump_tabular()
def learn(self, total_timesteps, callback=None, seed=None, log_interval=100, tb_log_name="DQN",
reset_num_timesteps=True, replay_wrapper=None):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
print("args are", self.kappa, self.phi_grad_update_freq, self.seed, np.random.randint(100))
with SetVerbosity(self.verbose):
# Create the replay buffer
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(schedule_timesteps=int(self.exploration_fraction * total_timesteps),
initial_p=1.0,
final_p=self.exploration_final_eps)
#self.exploration = PiecewiseSchedule([(0, 1.0), (int(1e6), 0.1), (int(1e7), 0.01)], outside_value=0.01)
episode_rewards = [0.0]
episode_successes = []
#td_errors_mean = []
#td_phi_errors_mean = []
obs = self.env.reset()
reset = True
self.episode_reward = np.zeros((1,))
for _ in range(total_timesteps):
#if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
# if callback(locals(), globals()) is False:
# break
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
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. - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs['update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
with self.sess.as_default():
action = self.act(np.array(obs)[None], update_eps=update_eps, **kwargs)[0]
env_action = action
reset = False
new_obs, rew, done, info = self.env.step(env_action)
# Store transition in the replay buffer.
self.replay_buffer.add(obs, action, rew, new_obs, float(done))
obs = new_obs
episode_rewards[-1] += rew
if done:
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
# Do not train if the warmup phase is not over
# or if there are not enough samples in the replay buffer
can_sample = self.replay_buffer.can_sample(self.batch_size)
if can_sample and self.num_timesteps > self.learning_starts \
and self.num_timesteps % self.train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(self.batch_size)
weights, batch_idxes = np.ones_like(rewards), None
# Use v_phi as zero until buffere is filled
if self.num_timesteps <= self.buffer_size:
weights = np.zeros_like(rewards)
with self.sess.as_default():
#actions_policy = self.act(obses_t)
actions_policy_phi = self.act(obses_tp1)
_, td_errors = self._train_step(obses_t, actions, actions_policy_phi, actions_policy_phi, rewards, obses_tp1, obses_tp1, obses_tp1, obses_tp1, dones, weights,
sess=self.sess)
#td_errors_mean.append(np.mean(td_errors))
if can_sample and self.num_timesteps > self.learning_starts and \
self.num_timesteps % self.target_network_update_freq == 0:
# Update target network periodically.
self.update_target(sess=self.sess)
if can_sample and self.kappa != 1.0 and self.num_timesteps >= self.buffer_size and \
self.num_timesteps % (self.phi_grad_update_freq * self.train_freq) == 0:
# Update target network periodically.
self.update_target_phi(sess=self.sess)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
if self.verbose >= 1 and done and log_interval is not None and len(episode_rewards) % log_interval == 0:
with self.timed("eval time"):
if self.test_env is not None and len(episode_rewards) % (10 * log_interval) == 0:
eval_return, actual_return = self.evaluate_agent(self.test_env)
else:
eval_return, actual_return = None, None
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate", np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward", mean_100ep_reward)
logger.record_tabular("eval return", eval_return)
logger.record_tabular("actual return", actual_return)
#logger.record_tabular("td errors", np.mean(td_errors_mean))
#logger.record_tabular("td errors phi", np.mean(td_phi_errors_mean))
logger.record_tabular("% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)))
logger.dump_tabular()
#td_errors_mean = []
#td_phi_errors_mean = []
if self.checkpoint_path is not None and self.num_timesteps % self.checkpoint_freq == 0:
self.save(self.checkpoint_path)
self.num_timesteps += 1
return self
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
reset_num_timesteps=True):
"""
Return a trained model.
:param total_timesteps: (int) The total number of samples to train on
:param seed: (int) The initial seed for training, if None: keep current seed
:param callback: (function (dict, dict)) -> boolean function called at every steps with state of the algorithm.
It takes the local and global variables. If it returns False, training is aborted.
:param log_interval: (int) The number of timesteps before logging.
:param reset_num_timesteps: (bool) whether or not to reset the current timestep number (used in logging)
:return: (BaseRLModel) the trained model
"""
self._setup_learn(seed)
# Initialize variables
episode_rewards = [0.0]
episode_successes = []
obs = self.env.reset()
episode_length = 0
for _ in range(total_timesteps):
num_episodes = len(episode_rewards)
if callback is not None:
# Stop training if return value is False, not when it is None.
if callback(locals(), globals()) is False:
break
# Act
action = self.act(np.array(obs))
new_obs, reward, done, info = self.env.step(action)
episode_rewards[-1] += reward
# Update data set
self._train_step(obs, action, reward, new_obs, done, lr=None)
obs = new_obs
# Restart if necesary
if done:
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
obs = self.env.reset()
episode_rewards.append(0.0)
episode_length = 0
# Performance in last 100 episodes
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 6)
# Logging
if self.verbose >= 1 and done and log_interval is not None and num_episodes % log_interval == 0:
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate",
np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.dump_tabular()
self.num_timesteps += 1
episode_length += 1
return self
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100):
with SetVerbosity(self.verbose):
self._setup_learn(seed)
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
self.beta_schedule = LinearSchedule(
prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0)
else:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(
schedule_timesteps=int(self.exploration_fraction *
total_timesteps),
initial_p=1.0,
final_p=self.exploration_final_eps)
episode_rewards = [0.0]
obs = self.env.reset()
reset = True
for step in range(total_timesteps):
if callback is not None:
callback(locals(), globals())
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(step)
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. - self.exploration.value(step) +
self.exploration.value(step) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
with self.sess.as_default():
action = self.act(np.array(obs)[None],
update_eps=update_eps,
**kwargs)[0]
env_action = action
reset = False
new_obs, rew, done, _ = self.env.step(env_action)
# Store transition in the replay buffer.
self.replay_buffer.add(obs, action, rew, new_obs, float(done))
obs = new_obs
episode_rewards[-1] += rew
if done:
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
if step > self.learning_starts and step % self.train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if self.prioritized_replay:
experience = self.replay_buffer.sample(
self.batch_size,
beta=self.beta_schedule.value(step))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(
self.batch_size)
weights, batch_idxes = np.ones_like(rewards), None
td_errors = self._train_step(obses_t,
actions,
rewards,
obses_tp1,
dones,
weights,
sess=self.sess)
if self.prioritized_replay:
new_priorities = np.abs(
td_errors) + self.prioritized_replay_eps
self.replay_buffer.update_priorities(
batch_idxes, new_priorities)
if step > self.learning_starts and step % self.target_network_update_freq == 0:
# Update target network periodically.
self.update_target(sess=self.sess)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
if self.verbose >= 1 and done and log_interval is not None and len(
episode_rewards) % log_interval == 0:
logger.record_tabular("steps", step)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular(
"% time spent exploring",
int(100 * self.exploration.value(step)))
logger.dump_tabular()
return self
def learn(self,
total_timesteps,
callback=None,
log_interval=100,
tb_log_name="DQN",
reset_num_timesteps=True,
replay_wrapper=None,
distinct_replay_buffer=False):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
for i, m in enumerate(self.sub_models):
m.learning_rate = get_schedule_fn(m.learning_rate)
if len(self.replay_wrappers) != 0:
m.replay_buffer = self.replay_wrappers[i](m.replay_buffer)
m._setup_learn()
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
else:
prioritized_replay_beta_iters = self.prioritized_replay_beta_iters
self.beta_schedule = LinearSchedule(
prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0)
else:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
if replay_wrapper is not None:
assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER"
self.replay_buffer = replay_wrapper(self.replay_buffer)
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(
schedule_timesteps=int(self.exploration_fraction *
total_timesteps),
initial_p=self.exploration_initial_eps,
final_p=self.exploration_final_eps)
episode_rewards = [0.0]
episode_successes = []
obs = self.env.reset()
reset = True
macro_count = 0
macro_len = self.macro_len
macro_choices = []
n_updates = 0
for step in range(total_timesteps):
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
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. - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
with self.sess.as_default():
if reset or macro_count % macro_len == 0:
macro_action = self.act(np.array(obs)[None],
update_eps=update_eps,
**kwargs)[0]
# macro_action = 1
macro_obs = obs
reward_in_one_macro = 0
macro_count += 1
macro_choices.append(macro_action)
# use sub_model to decide action
# env_action = self.sub_models[macro_action]
current_sub = self.sub_models[macro_action]
if self.num_timesteps < self.learning_starts or np.random.rand(
) < current_sub.random_exploration:
# actions sampled from action space are from range specific to the environment
# but algorithm operates on tanh-squashed actions therefore simple scaling is used
unscaled_action = self.env.action_space.sample()
action = scale_action(self.env.action_space,
unscaled_action)
else:
action = current_sub.policy_tf.step(
obs[None], deterministic=False).flatten()
# Add noise to the action (improve exploration,
# not needed in general)
if current_sub.action_noise is not None:
action = np.clip(action + current_sub.action_noise(),
-1, 1)
# inferred actions need to be transformed to environment action_space before stepping
unscaled_action = unscale_action(self.env.action_space,
action)
assert action.shape == self.env.action_space.shape
reset = False
new_obs, rew, done, info = self.env.step(unscaled_action)
episode_rewards[-1] += rew
# rew -= self.args.policy_cost_coef * self.args.sub_policy_costs[macro_action]
reward_in_one_macro += rew - self.args.policy_cost_coef * self.args.sub_policy_costs[
macro_action]
# Store transition in the replay buffer.
if macro_count % macro_len == 0 or done:
self.replay_buffer.add(macro_obs, macro_action,
reward_in_one_macro, new_obs,
float(done))
for i, m in enumerate(self.sub_models):
if distinct_replay_buffer:
if i == macro_action:
m.replay_buffer.add(obs, action, rew, new_obs,
float(done))
else:
m.replay_buffer.add(obs, action, rew, new_obs,
float(done))
obs = new_obs
if writer is not None:
ep_rew = np.array([rew]).reshape((1, -1))
ep_done = np.array([done]).reshape((1, -1))
total_episode_reward_logger(self.episode_reward, ep_rew,
ep_done, writer,
self.num_timesteps)
# print("step: %d, done: %d" % (self.num_timesteps, done))
if done:
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
macro_action = None
macro_count = 0
prev_macro_choices = macro_choices
macro_choices = []
# Do not train if the warmup phase is not over
# or if there are not enough samples in the replay buffer
can_sample = self.replay_buffer.can_sample(self.batch_size)
if can_sample and self.num_timesteps > self.learning_starts \
and self.num_timesteps % self.train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
# pytype:disable=bad-unpacking
if self.prioritized_replay:
assert self.beta_schedule is not None, \
"BUG: should be LinearSchedule when self.prioritized_replay True"
experience = self.replay_buffer.sample(
self.batch_size,
beta=self.beta_schedule.value(self.num_timesteps))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(
self.batch_size)
weights, batch_idxes = np.ones_like(rewards), None
# pytype:enable=bad-unpacking
if writer is not None:
# run loss backprop with summary, but once every 100 steps save the metadata
# (memory, compute time, ...)
if (1 + self.num_timesteps) % 100 == 0:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(
run_metadata, 'step%d' % self.num_timesteps)
else:
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
writer.add_summary(summary, self.num_timesteps)
else:
_, td_errors = self._train_step(obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
if self.prioritized_replay:
new_priorities = np.abs(
td_errors) + self.prioritized_replay_eps
assert isinstance(self.replay_buffer,
PrioritizedReplayBuffer)
self.replay_buffer.update_priorities(
batch_idxes, new_priorities)
if can_sample and self.num_timesteps > self.learning_starts and \
self.num_timesteps % self.target_network_update_freq == 0:
# Update target network periodically.
self.update_target(sess=self.sess)
if step % self.sub_models[0].train_freq == 0:
mb_infos_vals = []
for m in self.sub_models:
# Update policy, critics and target networks
for grad_step in range(m.gradient_steps):
# Break if the warmup phase is not over
# or if there are not enough samples in the replay buffer
if not m.replay_buffer.can_sample(m.batch_size) \
or self.num_timesteps < m.learning_starts:
break
n_updates += 1
# Compute current learning_rate
frac = 1.0 - step / total_timesteps
current_lr = m.learning_rate(frac)
# Update policy and critics (q functions)
mb_infos_vals.append(
m._train_step(step, writer, current_lr))
# Update target network
if (step +
grad_step) % m.target_update_interval == 0:
# Update target network
m.sess.run(m.target_update_op)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
# print(done, log_interval, len(episode_rewards), self.num_timesteps)
if self.verbose >= 1 and done and log_interval is not None and len(
episode_rewards) % log_interval == 0:
logger.record_tabular("steps", self.num_timesteps)
prev_macro_choices = np.array(prev_macro_choices)
macro_choices_ratio = [
'%.2f' %
((prev_macro_choices[prev_macro_choices == i]).size /
prev_macro_choices.size)
for i in range(self.n_actions)
]
logger.record_tabular("macro choices", macro_choices_ratio)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate",
np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular(
"% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)))
logger.logkv("n_updates_of_sub", n_updates)
logger.dump_tabular()
print("macro choices", prev_macro_choices)
self.num_timesteps += 1
return self
def learn(self,
total_timesteps,
callback=None,
log_interval=100,
tb_log_name="ACKTR",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
self.n_batch = self.n_envs * self.n_steps
self.learning_rate_schedule = Scheduler(
initial_value=self.learning_rate,
n_values=total_timesteps,
schedule=self.lr_schedule)
# FIFO queue of the q_runner thread is closed at the end of the learn function.
# As a result, it needs to be redefinied at every call
with self.graph.as_default():
with tf.variable_scope(
"kfac_apply",
reuse=self.trained,
custom_getter=tf_util.outer_scope_getter(
"kfac_apply")):
# Some of the variables are not in a scope when they are create
# so we make a note of any previously uninitialized variables
tf_vars = tf.global_variables()
is_uninitialized = self.sess.run(
[tf.is_variable_initialized(var) for var in tf_vars])
old_uninitialized_vars = [
v for (v, f) in zip(tf_vars, is_uninitialized) if not f
]
self.train_op, self.q_runner = self.optim.apply_gradients(
list(zip(self.grads_check, self.params)))
# then we check for new uninitialized variables and initialize them
tf_vars = tf.global_variables()
is_uninitialized = self.sess.run(
[tf.is_variable_initialized(var) for var in tf_vars])
new_uninitialized_vars = [
v for (v, f) in zip(tf_vars, is_uninitialized)
if not f and v not in old_uninitialized_vars
]
if len(new_uninitialized_vars) != 0:
self.sess.run(
tf.variables_initializer(new_uninitialized_vars))
self.trained = True
# Use GAE
if self.gae_lambda is not None:
runner = PPO2Runner(env=self.env,
model=self,
n_steps=self.n_steps,
gamma=self.gamma,
lam=self.gae_lambda)
else:
runner = A2CRunner(self.env,
self,
n_steps=self.n_steps,
gamma=self.gamma)
self.episode_reward = np.zeros((self.n_envs, ))
t_start = time.time()
coord = tf.train.Coordinator()
if self.q_runner is not None:
enqueue_threads = self.q_runner.create_threads(self.sess,
coord=coord,
start=True)
else:
enqueue_threads = []
# Training stats (when using Monitor wrapper)
ep_info_buf = deque(maxlen=100)
for update in range(1, total_timesteps // self.n_batch + 1):
# pytype:disable=bad-unpacking
# true_reward is the reward without discount
if isinstance(runner, PPO2Runner):
# We are using GAE
obs, returns, masks, actions, values, _, states, ep_infos, true_reward = runner.run(
)
else:
obs, states, returns, masks, actions, values, ep_infos, true_reward = runner.run(
)
# pytype:enable=bad-unpacking
ep_info_buf.extend(ep_infos)
policy_loss, value_loss, policy_entropy = self._train_step(
obs, states, returns, masks, actions, values,
self.num_timesteps // (self.n_batch + 1), writer)
n_seconds = time.time() - t_start
fps = int((update * self.n_batch) / n_seconds)
if writer is not None:
total_episode_reward_logger(
self.episode_reward,
true_reward.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)), writer,
self.num_timesteps)
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
if self.verbose >= 1 and (update % log_interval == 0
or update == 1):
explained_var = explained_variance(values, returns)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps",
self.num_timesteps)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy",
float(policy_entropy))
logger.record_tabular("policy_loss", float(policy_loss))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance",
float(explained_var))
if len(ep_info_buf) > 0 and len(ep_info_buf[0]) > 0:
logger.logkv(
'ep_reward_mean',
safe_mean(
[ep_info['r'] for ep_info in ep_info_buf]))
logger.logkv(
'ep_len_mean',
safe_mean(
[ep_info['l'] for ep_info in ep_info_buf]))
logger.dump_tabular()
self.num_timesteps += self.n_batch + 1
coord.request_stop()
coord.join(enqueue_threads)
return self
def eval(self,
total_episodes,
callback=None,
log_interval=100,
tb_log_name="DQN",
reset_num_timesteps=True,
replay_wrapper=None):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
episode_rewards = [0.0]
episode_successes = []
obs = self.env.reset()
reset = True
macro_count = 0
macro_len = self.macro_len
macro_choices = []
n_updates = 0
macro_action = None
# for step in range(total_timesteps):
while True:
with self.sess.as_default():
if reset or macro_count % macro_len == 0:
# macro_action = self.act(np.array(obs)[None], update_eps=0, **{})[0]
macro_actions, _, _ = self.step_model.step(
np.array(obs)[None], deterministic=False)
macro_action = macro_actions[0]
# macro_action = 1
macro_obs = obs
reward_in_one_macro = 0
macro_count += 1
macro_choices.append(macro_action)
current_sub = self.sub_models[macro_action]
action = current_sub.policy_tf.step(
obs[None], deterministic=True).flatten()
# Add noise to the action (improve exploration,
# not needed in general)
if current_sub.action_noise is not None:
action = np.clip(action + current_sub.action_noise(), -1,
1)
# inferred actions need to be transformed to environment action_space before stepping
unscaled_action = unscale_action(self.env.action_space, action)
assert action.shape == self.env.action_space.shape
reset = False
new_obs, rew, done, info = self.env.step(unscaled_action)
episode_rewards[-1] += rew
rew -= self.args.policy_cost_coef * self.args.sub_policy_costs[
macro_action]
reward_in_one_macro += rew
obs = new_obs
# print("step: %d, done: %d" % (self.num_timesteps, done))
if done:
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
macro_action = None
macro_count = 0
print("=" * 70)
print("macro_choices:", macro_choices)
print("return:", episode_rewards[-2])
print("=" * 70)
prev_macro_choices = macro_choices
macro_choices = []
if len(episode_rewards) - 1 == total_episodes:
break
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
# print(done, log_interval, len(episode_rewards), self.num_timesteps)
if self.verbose >= 1 and done and log_interval is not None and len(
episode_rewards) % log_interval == 0:
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("macro choices",
np.mean(prev_macro_choices))
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate",
np.mean(episode_successes[-100:]))
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.logkv("n_updates_of_sub", n_updates)
logger.dump_tabular()
print("macro choices", prev_macro_choices)
self.num_timesteps += 1
return self
def learn(self,
total_timesteps,
callback=None,
log_interval=1,
tb_log_name="DDPG",
print_freq=100):
with TensorboardWriter(self.graph, self.tensorboard_log,
tb_log_name) as writer:
rank = MPI.COMM_WORLD.Get_rank()
# we assume symmetric actions.
assert np.all(
np.abs(self.env.action_space.low) ==
self.env.action_space.high)
self.episode_reward = np.zeros((1, ))
with self.sess.as_default(), self.graph.as_default():
# Prepare everything.
self._reset()
episode_reward = 0.
episode_step = 0
episodes = 0
step = 0
total_steps = 0
start_time = time.time()
actor_losses = []
critic_losses = []
should_return = False
while True:
obs = self.env.reset()
# Rollout one episode.
while True:
if total_steps >= total_timesteps:
if should_return:
return self
should_return = True
break
# Predict next action.
action, q_value = self._policy(obs,
apply_noise=True,
compute_q=True)
if self.verbose >= 2:
print(action)
assert action.shape == self.env.action_space.shape
# Execute next action.
new_obs, reward, done, info = self.env.step(
action * np.abs(self.action_space.low))
step += 1
total_steps += 1
if rank == 0 and self.render:
self.env.render()
episode_reward += reward
episode_step += 1
if print_freq > 0 and episode_step % print_freq == 0 and episode_step > 0:
print("{} steps".format(episode_step))
# Book-keeping.
self._store_transition(obs, action, reward, new_obs,
done)
obs = new_obs
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
return self
if done:
print("Episode finished. Reward: {:.2f} {} Steps".
format(episode_reward, episode_step))
# Episode done.
episode_reward = 0.
episode_step = 0
episodes += 1
self._reset()
obs = self.env.reset()
# Finish rollout on episode finish.
break
# Train DDPG.
actor_losses = []
critic_losses = []
train_start = time.time()
for t_train in range(self.nb_train_steps):
critic_loss, actor_loss = self._train_step(
0, None, log=t_train == 0)
critic_losses.append(critic_loss)
actor_losses.append(actor_loss)
self._update_target_net()
print(
"DDPG training duration: {:.2f}s".format(time.time() -
train_start))
mpi_size = MPI.COMM_WORLD.Get_size()
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = self._get_stats()
combined_stats = stats.copy()
combined_stats['train/loss_actor'] = np.mean(actor_losses)
combined_stats['train/loss_critic'] = np.mean(
critic_losses)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(
step) / float(duration)
combined_stats['total/episodes'] = episodes
combined_stats_sums = MPI.COMM_WORLD.allreduce(
np.array(
[as_scalar(x) for x in combined_stats.values()]))
combined_stats = {
k: v / mpi_size
for (k, v) in zip(combined_stats.keys(),
combined_stats_sums)
}
# Total statistics.
combined_stats['total/steps'] = step
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
logger.dump_tabular()
logger.info('')
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100):
with SetVerbosity(self.verbose):
self._setup_learn(seed)
with self.sess.as_default():
seg_gen = traj_segment_generator(
self.policy_pi,
self.env,
self.timesteps_per_batch,
reward_giver=self.reward_giver,
gail=self.using_gail)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
t_start = time.time()
lenbuffer = deque(
maxlen=40) # rolling buffer for episode lengths
rewbuffer = deque(
maxlen=40) # rolling buffer for episode rewards
true_rewbuffer = None
if self.using_gail:
true_rewbuffer = deque(maxlen=40)
# Stats not used for now
# 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 self.pretrained_weight is not None:
tf_util.load_state(
self.pretrained_weight,
var_list=tf_util.get_globals_vars("pi"),
sess=self.sess)
while True:
if callback:
callback(locals(), globals())
if total_timesteps and timesteps_so_far >= total_timesteps:
break
logger.log("********** Iteration %i ************" %
iters_so_far)
def fisher_vector_product(vec):
return self.allmean(
self.compute_fvp(
vec, *fvpargs,
sess=self.sess)) + self.cg_damping * vec
# ------------------ Update G ------------------
logger.log("Optimizing Policy...")
# g_step = 1 when not using GAIL
mean_losses = None
vpredbefore = None
tdlamret = None
observation = None
action = None
seg = None
for _ in range(self.g_step):
with self.timed("sampling"):
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, self.gamma, self.lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
observation, action, 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
args = seg["ob"], seg["ob"], seg["ac"], atarg
fvpargs = [arr[::5] for arr in args]
self.assign_old_eq_new(sess=self.sess)
with self.timed("computegrad"):
*lossbefore, grad = self.compute_lossandgrad(
*args, sess=self.sess)
lossbefore = self.allmean(np.array(lossbefore))
grad = self.allmean(grad)
if np.allclose(grad, 0):
logger.log("Got zero gradient. not updating")
else:
with self.timed("cg"):
stepdir = conjugate_gradient(
fisher_vector_product,
grad,
cg_iters=self.cg_iters,
verbose=self.rank == 0
and self.verbose >= 1)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(
fisher_vector_product(stepdir))
# abs(shs) to avoid taking square root of negative values
lagrange_multiplier = np.sqrt(
abs(shs) / self.max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lagrange_multiplier
expectedimprove = grad.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = self.get_flat()
thnew = None
for _ in range(10):
thnew = thbefore + fullstep * stepsize
self.set_from_flat(thnew)
mean_losses = surr, kl_loss, *_ = self.allmean(
np.array(
self.compute_losses(*args,
sess=self.sess)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" %
(expectedimprove, improve))
if not np.isfinite(mean_losses).all():
logger.log(
"Got non-finite value of losses -- bad!"
)
elif kl_loss > self.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")
self.set_from_flat(thbefore)
if self.nworkers > 1 and iters_so_far % 20 == 0:
# list of tuples
paramsums = MPI.COMM_WORLD.allgather(
(thnew.sum(), self.vfadam.getflat().sum()))
assert all(
np.allclose(ps, paramsums[0])
for ps in paramsums[1:])
with self.timed("vf"):
for _ in range(self.vf_iters):
for (mbob, mbret) in dataset.iterbatches(
(seg["ob"], seg["tdlamret"]),
include_final_partial_batch=False,
batch_size=128):
grad = self.allmean(
self.compute_vflossandgrad(
mbob, mbob, mbret, sess=self.sess))
self.vfadam.update(grad, self.vf_stepsize)
for (loss_name, loss_val) in zip(self.loss_names,
mean_losses):
logger.record_tabular(loss_name, loss_val)
logger.record_tabular(
"ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
if self.using_gail:
# ------------------ Update D ------------------
logger.log("Optimizing Discriminator...")
logger.log(fmt_row(13, self.reward_giver.loss_name))
ob_expert, ac_expert = self.expert_dataset.get_next_batch(
len(observation))
batch_size = len(observation) // self.d_step
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
for ob_batch, ac_batch in dataset.iterbatches(
(observation, action),
include_final_partial_batch=False,
batch_size=batch_size):
ob_expert, ac_expert = self.expert_dataset.get_next_batch(
len(ob_batch))
# update running mean/std for reward_giver
if hasattr(self.reward_giver, "obs_rms"):
self.reward_giver.obs_rms.update(
np.concatenate((ob_batch, ob_expert), 0))
*newlosses, grad = self.reward_giver.lossandgrad(
ob_batch, ac_batch, ob_expert, ac_expert)
self.d_adam.update(self.allmean(grad),
self.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)
else:
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))
if self.using_gail:
logger.record_tabular("EpTrueRewMean",
np.mean(true_rewbuffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
timesteps_so_far += seg["total_timestep"]
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - t_start)
if self.verbose >= 1 and self.rank == 0:
logger.dump_tabular()
return self
def learn(self,
total_timesteps,
callback=None,
seed=None,
log_interval=100,
tb_log_name="DQN",
reset_num_timesteps=True):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn(seed)
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
else:
prioritized_replay_beta_iters = self.prioritized_replay_beta_iters
self.beta_schedule = LinearSchedule(
prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0)
else:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(
schedule_timesteps=int(self.exploration_fraction *
total_timesteps),
initial_p=1.0,
final_p=self.exploration_final_eps)
episode_rewards = [0.0]
obs = self.env.reset()
reset = True
self.episode_reward = np.zeros((1, ))
for _ in range(total_timesteps):
if callback is not None:
# Only stop training if return value is False, not when it is None. This is for backwards
# compatibility with callbacks that have no return statement.
if callback(locals(), globals()) is False:
break
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
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. - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
with self.sess.as_default():
action = self.act(np.array(obs)[None],
update_eps=update_eps,
**kwargs)[0]
env_action = action
reset = False
new_obs, rew, done, _ = self.env.step(env_action)
# Store transition in the replay buffer.
self.replay_buffer.add(obs, action, rew, new_obs, float(done))
obs = new_obs
if writer is not None:
ep_rew = np.array([rew]).reshape((1, -1))
ep_done = np.array([done]).reshape((1, -1))
self.episode_reward = total_episode_reward_logger(
self.episode_reward, ep_rew, ep_done, writer,
self.num_timesteps)
episode_rewards[-1] += rew
if done:
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append(0.0)
reset = True
if self.num_timesteps > self.learning_starts and self.num_timesteps % self.train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if self.prioritized_replay:
experience = self.replay_buffer.sample(
self.batch_size,
beta=self.beta_schedule.value(self.num_timesteps))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(
self.batch_size)
weights, batch_idxes = np.ones_like(rewards), None
if writer is not None:
# run loss backprop with summary, but once every 100 steps save the metadata
# (memory, compute time, ...)
if (1 + self.num_timesteps) % 100 == 0:
run_options = tf.RunOptions(
trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess,
options=run_options,
run_metadata=run_metadata)
writer.add_run_metadata(
run_metadata, 'step%d' % self.num_timesteps)
else:
summary, td_errors = self._train_step(
obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
writer.add_summary(summary, self.num_timesteps)
else:
_, td_errors = self._train_step(obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
if self.prioritized_replay:
new_priorities = np.abs(
td_errors) + self.prioritized_replay_eps
self.replay_buffer.update_priorities(
batch_idxes, new_priorities)
if self.num_timesteps > self.learning_starts and \
self.num_timesteps % self.target_network_update_freq == 0:
# Update target network periodically.
self.update_target(sess=self.sess)
if len(episode_rewards[-101:-1]) == 0:
mean_100ep_reward = -np.inf
else:
mean_100ep_reward = round(
float(np.mean(episode_rewards[-101:-1])), 1)
num_episodes = len(episode_rewards)
if self.verbose >= 1 and done and log_interval is not None and len(
episode_rewards) % log_interval == 0:
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular(
"% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)))
logger.dump_tabular()
self.num_timesteps += 1
return self
def learn(self,
total_timesteps,
callback=None,
log_interval=100,
tb_log_name="DQN",
reset_num_timesteps=True,
replay_wrapper=None):
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
# callback = self._init_callback(callback)
# with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
# as writer:
self._setup_learn()
# Create the replay buffer
if self.prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
self.buffer_size, alpha=self.prioritized_replay_alpha)
if self.prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
else:
prioritized_replay_beta_iters = self.prioritized_replay_beta_iters
self.beta_schedule = LinearSchedule(
prioritized_replay_beta_iters,
initial_p=self.prioritized_replay_beta0,
final_p=1.0)
else:
self.replay_buffer = ReplayBuffer(self.buffer_size)
self.beta_schedule = None
if replay_wrapper is not None:
assert not self.prioritized_replay, "Prioritized replay buffer is not supported by HER"
self.replay_buffer = replay_wrapper(self.replay_buffer)
# Create the schedule for exploration starting from 1.
self.exploration = LinearSchedule(
schedule_timesteps=int(self.exploration_fraction *
total_timesteps),
initial_p=self.exploration_initial_eps,
final_p=self.exploration_final_eps)
episode_rewards = [[0.0] * self.num_agents] #MA-MOD
episode_successes = []
#callback.on_training_start(locals(), globals())
#callback.on_rollout_start()
reset = True
obs = self.env.reset()
for _ in range(total_timesteps):
# Take action and update exploration to the newest value
kwargs = {}
if not self.param_noise:
update_eps = self.exploration.value(self.num_timesteps)
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. - self.exploration.value(self.num_timesteps) +
self.exploration.value(self.num_timesteps) / float(self.env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
with self.sess.as_default():
env_action = [] # MA-MOD
for i in range(self.num_agents
): # MA-MOD. This is fine for one policy.
action = self.act[i](
np.array(obs[i])[None],
update_eps=update_eps,
**kwargs
)[0] # TODO: Is this the correct way to get the correct agent obs?
env_action.append(action)
reset = False
new_obs, rew, done, info = self.env.step(
env_action
) # NOUPDATE - env.step should take a vector of actions
# print("Agent 1", type(new_obs[0]))
# for row in new_obs[0]:
# print(' '.join(str(int(item)) for item in row))
# print("Agent 2", type(new_obs[1]))
# for row in new_obs[1]:
# print(' '.join(str(int(item)) for item in row))
'''
Obs: x_me, x_opp --- agent 1. In env: x_1, x_2
Obs: x_me, x_opp -- agent 2. In env: x_2, x_1
Env: (n_agents, state_dim)
'''
self.num_timesteps += 1
# Store transition in the replay buffer.
# Loop for replay buffer -- either separate or joined. obs[agent_index], action[agent_index], reward[agent_index]
# Joey: Does this look right to you?
for num_agent in range(self.num_agents):
self.replay_buffer.add(obs[num_agent], env_action[num_agent],
rew[num_agent], new_obs[num_agent],
float(done[num_agent]))
obs = new_obs
# if writer is not None:
# ep_rew = np.array([rew]).reshape((1, -1))
# ep_done = np.array([done]).reshape((1, -1))
# tf_util.total_episode_reward_logger(self.episode_reward, ep_rew, ep_done, writer,
# self.num_timesteps)
# TODO: current episode_rewards is a list, make it a list of lists where each list is the reward for each agent in all timesteps
# append the newest reward to the end of each list for each agent
if isinstance(done, list):
done = np.array(done)
if done.any():
for num_agent in range(self.num_agents): #MA-MOD
episode_rewards[-1][num_agent] += rew[num_agent]
maybe_is_success = info.get('is_success')
if maybe_is_success is not None:
episode_successes.append(float(maybe_is_success))
if not isinstance(self.env, VecEnv):
obs = self.env.reset()
episode_rewards.append([0.0] * self.num_agents)
reset = True
# Do not train if the warmup phase is not over
# or if there are not enough samples in the replay buffer
can_sample = self.replay_buffer.can_sample(self.batch_size)
if can_sample and self.num_timesteps > self.learning_starts \
and self.num_timesteps % self.train_freq == 0:
# callback.on_rollout_end()
for i in range(self.num_agents): # MA-MOD
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
# pytype:disable=bad-unpacking
if self.prioritized_replay:
assert self.beta_schedule is not None, \
"BUG: should be LinearSchedule when self.prioritized_replay True"
experience = self.replay_buffer.sample(
self.batch_size,
beta=self.beta_schedule.value(self.num_timesteps))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = self.replay_buffer.sample(
self.batch_size)
weights, batch_idxes = np.ones_like(rewards), None
# pytype:enable=bad-unpacking
# if writer is not None:
# # run loss backprop with summary, but once every 100 steps save the metadata
# # (memory, compute time, ...)
# if (1 + self.num_timesteps) % 100 == 0:
# run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
# run_metadata = tf.RunMetadata()
# summary, td_errors = self._train_step[i](obses_t, actions, rewards, obses_tp1, obses_tp1,
# dones, weights, sess=self.sess, options=run_options,
# run_metadata=run_metadata)
# writer.add_run_metadata(run_metadata, 'step%d_agent%d' % (self.num_timesteps, i))
# else:
# summary, td_errors = self._train_step[i](obses_t, actions, rewards, obses_tp1, obses_tp1,
# dones, weights, sess=self.sess)
# writer.add_summary(summary, self.num_timesteps)
# else:
td_errors = self._train_step[i](obses_t,
actions,
rewards,
obses_tp1,
obses_tp1,
dones,
weights,
sess=self.sess)
if self.prioritized_replay: # NOUPDATE - not inside main agent for loop
new_priorities = np.abs(
td_errors) + self.prioritized_replay_eps # NOUPDATE
assert isinstance(self.replay_buffer,
PrioritizedReplayBuffer)
self.replay_buffer.update_priorities(
batch_idxes, new_priorities)
# callback.on_rollout_start()
if can_sample and self.num_timesteps > self.learning_starts and \
self.num_timesteps % self.target_network_update_freq == 0:
# Update target network periodically.
for i in range(self.num_agents):
self.update_target[i](sess=self.sess) # MA-MOD
if len(episode_rewards[-101:-1]) == 0: # MA-MOD
mean_100ep_reward = -np.inf * np.ones((self.num_agents, ))
else:
mean_100ep_reward = np.mean(episode_rewards[-101:-1], axis=0)
# below is what's logged in terminal.
num_episodes = len(episode_rewards) #MA-MOD
if self.verbose >= 1 and done.any(
) and log_interval is not None and len(
episode_rewards) % log_interval == 0: #MA-MOD
logger.record_tabular("steps", self.num_timesteps)
logger.record_tabular("episodes", num_episodes)
if len(episode_successes) > 0:
logger.logkv("success rate",
np.mean(episode_successes[-100:], axis=0))
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular(
"% time spent exploring",
int(100 * self.exploration.value(self.num_timesteps)))
logger.dump_tabular()
return self