def call(self, on_policy):
runner, model, buffer, steps = self.runner, self.model, self.buffer, self.steps
if on_policy:
enc_obs, obs, actions, rewards, mus, dones, masks = runner.run()
self.episode_stats.feed(rewards, dones)
if buffer is not None:
buffer.put(enc_obs, actions, rewards, mus, dones, masks)
else:
# 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.nbatch])
rewards = rewards.reshape([runner.nbatch])
mus = mus.reshape([runner.nbatch, runner.nact])
dones = dones.reshape([runner.nbatch])
masks = masks.reshape([runner.batch_ob_shape[0]])
names_ops, values_ops = model.train(obs, actions, rewards, dones, mus, model.initial_state, masks, steps)
if on_policy and (int(steps/runner.nbatch) % self.log_interval == 0):
logger.record_tabular("total_timesteps", steps)
logger.record_tabular("fps", int(steps/(time.time() - self.tstart)))
# 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", self.episode_stats.mean_length())
logger.record_tabular("mean_episode_reward", self.episode_stats.mean_reward())
for name, val in zip(names_ops, values_ops):
logger.record_tabular(name, float(val))
logger.dump_tabular()
def main(policy_file, seed, n_test_rollouts, render):
set_global_seeds(seed)
# Load policy.
with open(policy_file, 'rb') as f:
policy = pickle.load(f)
env_name = policy.info['env_name']
# Prepare params.
params = config.DEFAULT_PARAMS
if env_name in config.DEFAULT_ENV_PARAMS:
# merge env-specific parameters in
params.update(config.DEFAULT_ENV_PARAMS[env_name])
params['env_name'] = env_name
params = config.prepare_params(params)
config.log_params(params, logger=logger)
dims = config.configure_dims(params)
eval_params = {
'exploit': True,
'use_target_net': params['test_with_polyak'],
'compute_Q': True,
'rollout_batch_size': 1,
'render': bool(render),
}
for name in ['T', 'gamma', 'noise_eps', 'random_eps']:
eval_params[name] = params[name]
evaluator = RolloutWorker(params['make_env'], policy, dims, logger,
**eval_params)
evaluator.seed(seed)
# Run evaluation.
evaluator.clear_history()
for _ in range(n_test_rollouts):
evaluator.generate_rollouts()
# record logs
for key, val in evaluator.logs('test'):
logger.record_tabular(key, np.mean(val))
logger.dump_tabular()
def learn(env,
policy_func,
reward_giver,
expert_dataset,
rank,
pretrained,
pretrained_weight,
*,
g_step,
d_step,
entcoeff,
save_per_iter,
ckpt_dir,
log_dir,
timesteps_per_batch,
task_name,
gamma,
lam,
max_kl,
cg_iters,
cg_damping=1e-2,
vf_stepsize=3e-4,
d_stepsize=3e-4,
vf_iters=3,
max_timesteps=0,
max_episodes=0,
max_iters=0,
callback=None):
nworkers = MPI.COMM_WORLD.Get_size()
rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_func("pi",
ob_space,
ac_space,
reuse=(pretrained_weight != None))
oldpi = policy_func("oldpi", ob_space, ac_space)
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
ob = U.get_placeholder_cached(name="ob")
ac = pi.pdtype.sample_placeholder([None])
kloldnew = oldpi.pd.kl(pi.pd)
ent = pi.pd.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = entcoeff * meanent
vferr = tf.reduce_mean(tf.square(pi.vpred - ret))
ratio = tf.exp(pi.pd.logp(ac) -
oldpi.pd.logp(ac)) # advantage * pnew / pold
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy"]
dist = meankl
all_var_list = pi.get_trainable_variables()
var_list = [
v for v in all_var_list
if v.name.startswith("pi/pol") or v.name.startswith("pi/logstd")
]
vf_var_list = [v for v in all_var_list if v.name.startswith("pi/vff")]
assert len(var_list) == len(vf_var_list) + 1
d_adam = MpiAdam(reward_giver.get_trainable_variables())
vfadam = MpiAdam(vf_var_list)
get_flat = U.GetFlat(var_list)
set_from_flat = U.SetFromFlat(var_list)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32,
shape=[None],
name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
sz = U.intprod(shape)
tangents.append(tf.reshape(flat_tangent[start:start + sz], shape))
start += sz
gvp = tf.add_n([
tf.reduce_sum(g * tangent)
for (g, tangent) in zipsame(klgrads, tangents)
]) # pylint: disable=E1111
fvp = U.flatgrad(gvp, var_list)
assign_old_eq_new = U.function(
[], [],
updates=[
tf.assign(oldv, newv)
for (oldv,
newv) in zipsame(oldpi.get_variables(), pi.get_variables())
])
compute_losses = U.function([ob, ac, atarg], losses)
compute_lossandgrad = U.function([ob, ac, atarg], losses +
[U.flatgrad(optimgain, var_list)])
compute_fvp = U.function([flat_tangent, ob, ac, atarg], fvp)
compute_vflossandgrad = U.function([ob, ret],
U.flatgrad(vferr, vf_var_list))
@contextmanager
def timed(msg):
if rank == 0:
print(colorize(msg, color='magenta'))
tstart = time.time()
yield
print(
colorize("done in %.3f seconds" % (time.time() - tstart),
color='magenta'))
else:
yield
def allmean(x):
assert isinstance(x, np.ndarray)
out = np.empty_like(x)
MPI.COMM_WORLD.Allreduce(x, out, op=MPI.SUM)
out /= nworkers
return out
U.initialize()
th_init = get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
set_from_flat(th_init)
d_adam.sync()
vfadam.sync()
if rank == 0:
print("Init param sum", th_init.sum(), flush=True)
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi,
env,
reward_giver,
timesteps_per_batch,
stochastic=True)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=40) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=40) # rolling buffer for episode rewards
true_rewbuffer = deque(maxlen=40)
assert sum([max_iters > 0, max_timesteps > 0, max_episodes > 0]) == 1
g_loss_stats = stats(loss_names)
d_loss_stats = stats(reward_giver.loss_name)
ep_stats = stats(["True_rewards", "Rewards", "Episode_length"])
# if provide pretrained weight
if pretrained_weight is not None:
U.load_state(pretrained_weight, var_list=pi.get_variables())
while True:
if callback: callback(locals(), globals())
if max_timesteps and timesteps_so_far >= max_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
# Save model
if rank == 0 and iters_so_far % save_per_iter == 0 and ckpt_dir is not None:
fname = os.path.join(ckpt_dir, task_name)
os.makedirs(os.path.dirname(fname), exist_ok=True)
saver = tf.train.Saver()
saver.save(tf.get_default_session(), fname)
logger.log("********** Iteration %i ************" % iters_so_far)
def fisher_vector_product(p):
return allmean(compute_fvp(p, *fvpargs)) + cg_damping * p
# ------------------ Update G ------------------
logger.log("Optimizing Policy...")
for _ in range(g_step):
with timed("sampling"):
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, gamma, lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
ob, ac, atarg, tdlamret = seg["ob"], seg["ac"], seg["adv"], seg[
"tdlamret"]
vpredbefore = seg[
"vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()) / atarg.std(
) # standardized advantage function estimate
if hasattr(pi, "ob_rms"):
pi.ob_rms.update(ob) # update running mean/std for policy
args = seg["ob"], seg["ac"], atarg
fvpargs = [arr[::5] for arr in args]
assign_old_eq_new(
) # set old parameter values to new parameter values
with timed("computegrad"):
*lossbefore, g = compute_lossandgrad(*args)
lossbefore = allmean(np.array(lossbefore))
g = allmean(g)
if np.allclose(g, 0):
logger.log("Got zero gradient. not updating")
else:
with timed("cg"):
stepdir = cg(fisher_vector_product,
g,
cg_iters=cg_iters,
verbose=rank == 0)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(fisher_vector_product(stepdir))
lm = np.sqrt(shs / max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lm
expectedimprove = g.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = get_flat()
for _ in range(10):
thnew = thbefore + fullstep * stepsize
set_from_flat(thnew)
meanlosses = surr, kl, *_ = allmean(
np.array(compute_losses(*args)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" %
(expectedimprove, improve))
if not np.isfinite(meanlosses).all():
logger.log("Got non-finite value of losses -- bad!")
elif kl > max_kl * 1.5:
logger.log("violated KL constraint. shrinking step.")
elif improve < 0:
logger.log("surrogate didn't improve. shrinking step.")
else:
logger.log("Stepsize OK!")
break
stepsize *= .5
else:
logger.log("couldn't compute a good step")
set_from_flat(thbefore)
if nworkers > 1 and iters_so_far % 20 == 0:
paramsums = MPI.COMM_WORLD.allgather(
(thnew.sum(),
vfadam.getflat().sum())) # list of tuples
assert all(
np.allclose(ps, paramsums[0]) for ps in paramsums[1:])
with timed("vf"):
for _ in range(vf_iters):
for (mbob, mbret) in dataset.iterbatches(
(seg["ob"], seg["tdlamret"]),
include_final_partial_batch=False,
batch_size=128):
if hasattr(pi, "ob_rms"):
pi.ob_rms.update(
mbob) # update running mean/std for policy
g = allmean(compute_vflossandgrad(mbob, mbret))
vfadam.update(g, vf_stepsize)
g_losses = meanlosses
for (lossname, lossval) in zip(loss_names, meanlosses):
logger.record_tabular(lossname, lossval)
logger.record_tabular("ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
# ------------------ Update D ------------------
logger.log("Optimizing Discriminator...")
logger.log(fmt_row(13, reward_giver.loss_name))
ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob))
batch_size = len(ob) // d_step
d_losses = [
] # list of tuples, each of which gives the loss for a minibatch
for ob_batch, ac_batch in dataset.iterbatches(
(ob, ac), include_final_partial_batch=False,
batch_size=batch_size):
ob_expert, ac_expert = expert_dataset.get_next_batch(len(ob_batch))
# update running mean/std for reward_giver
if hasattr(reward_giver, "obs_rms"):
reward_giver.obs_rms.update(
np.concatenate((ob_batch, ob_expert), 0))
*newlosses, g = reward_giver.lossandgrad(ob_batch, ac_batch,
ob_expert, ac_expert)
d_adam.update(allmean(g), d_stepsize)
d_losses.append(newlosses)
logger.log(fmt_row(13, np.mean(d_losses, axis=0)))
lrlocal = (seg["ep_lens"], seg["ep_rets"], seg["ep_true_rets"]
) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
lens, rews, true_rets = map(flatten_lists, zip(*listoflrpairs))
true_rewbuffer.extend(true_rets)
lenbuffer.extend(lens)
rewbuffer.extend(rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpTrueRewMean", np.mean(true_rewbuffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - tstart)
if rank == 0:
logger.dump_tabular()
def learn(env, policy_fn, *,
timesteps_per_actorbatch, # timesteps per actor per update
clip_param, entcoeff, # clipping parameter epsilon, entropy coeff
optim_epochs, optim_stepsize, optim_batchsize,# optimization hypers
gamma, lam, # advantage estimation
max_timesteps=0, max_episodes=0, max_iters=0, max_seconds=0, # time constraint
callback=None, # you can do anything in the callback, since it takes locals(), globals()
adam_epsilon=1e-5,
schedule='constant' # annealing for stepsize parameters (epsilon and adam)
):
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_fn("pi", ob_space, ac_space) # Construct network for new policy
oldpi = policy_fn("oldpi", ob_space, ac_space) # Network for old policy
atarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
lrmult = tf.placeholder(name='lrmult', dtype=tf.float32, shape=[]) # learning rate multiplier, updated with schedule
ob = U.get_placeholder_cached(name="ob")
ac = pi.pdtype.sample_placeholder([None])
kloldnew = oldpi.pd.kl(pi.pd)
ent = pi.pd.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
pol_entpen = (-entcoeff) * meanent
ratio = tf.exp(pi.pd.logp(ac) - oldpi.pd.logp(ac)) # pnew / pold
surr1 = ratio * atarg # surrogate from conservative policy iteration
surr2 = tf.clip_by_value(ratio, 1.0 - clip_param, 1.0 + clip_param) * atarg #
pol_surr = - tf.reduce_mean(tf.minimum(surr1, surr2)) # PPO's pessimistic surrogate (L^CLIP)
vf_loss = tf.reduce_mean(tf.square(pi.vpred - ret))
total_loss = pol_surr + pol_entpen + vf_loss
losses = [pol_surr, pol_entpen, vf_loss, meankl, meanent]
loss_names = ["pol_surr", "pol_entpen", "vf_loss", "kl", "ent"]
var_list = pi.get_trainable_variables()
lossandgrad = U.function([ob, ac, atarg, ret, lrmult], losses + [U.flatgrad(total_loss, var_list)])
adam = MpiAdam(var_list, epsilon=adam_epsilon)
assign_old_eq_new = U.function([],[], updates=[tf.assign(oldv, newv)
for (oldv, newv) in zipsame(oldpi.get_variables(), pi.get_variables())])
compute_losses = U.function([ob, ac, atarg, ret, lrmult], losses)
U.initialize()
adam.sync()
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi, env, timesteps_per_actorbatch, stochastic=True)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=100) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=100) # rolling buffer for episode rewards
assert sum([max_iters>0, max_timesteps>0, max_episodes>0, max_seconds>0])==1, "Only one time constraint permitted"
while True:
if callback: callback(locals(), globals())
if max_timesteps and timesteps_so_far >= max_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
elif max_seconds and time.time() - tstart >= max_seconds:
break
if schedule == 'constant':
cur_lrmult = 1.0
elif schedule == 'linear':
cur_lrmult = max(1.0 - float(timesteps_so_far) / max_timesteps, 0)
else:
raise NotImplementedError
logger.log("********** Iteration %i ************"%iters_so_far)
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, gamma, lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
ob, ac, atarg, tdlamret = seg["ob"], seg["ac"], seg["adv"], seg["tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()) / atarg.std() # standardized advantage function estimate
d = Dataset(dict(ob=ob, ac=ac, atarg=atarg, vtarg=tdlamret), deterministic=pi.recurrent)
optim_batchsize = optim_batchsize or ob.shape[0]
if hasattr(pi, "ob_rms"): pi.ob_rms.update(ob) # update running mean/std for policy
assign_old_eq_new() # set old parameter values to new parameter values
logger.log("Optimizing...")
logger.log(fmt_row(13, loss_names))
# Here we do a bunch of optimization epochs over the data
for _ in range(optim_epochs):
losses = [] # list of tuples, each of which gives the loss for a minibatch
for batch in d.iterate_once(optim_batchsize):
*newlosses, g = lossandgrad(batch["ob"], batch["ac"], batch["atarg"], batch["vtarg"], cur_lrmult)
adam.update(g, optim_stepsize * cur_lrmult)
losses.append(newlosses)
logger.log(fmt_row(13, np.mean(losses, axis=0)))
logger.log("Evaluating losses...")
losses = []
for batch in d.iterate_once(optim_batchsize):
newlosses = compute_losses(batch["ob"], batch["ac"], batch["atarg"], batch["vtarg"], cur_lrmult)
losses.append(newlosses)
meanlosses,_,_ = mpi_moments(losses, axis=0)
logger.log(fmt_row(13, meanlosses))
for (lossval, name) in zipsame(meanlosses, loss_names):
logger.record_tabular("loss_"+name, lossval)
logger.record_tabular("ev_tdlam_before", explained_variance(vpredbefore, tdlamret))
lrlocal = (seg["ep_lens"], seg["ep_rets"]) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
lens, rews = map(flatten_lists, zip(*listoflrpairs))
lenbuffer.extend(lens)
rewbuffer.extend(rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - tstart)
if MPI.COMM_WORLD.Get_rank()==0:
logger.dump_tabular()
return pi
def learn(network, env,
seed=None,
total_timesteps=None,
nb_epochs=None, # with default settings, perform 1M steps total
nb_epoch_cycles=20,
nb_rollout_steps=100,
reward_scale=1.0,
render=False,
render_eval=False,
noise_type='adaptive-param_0.2',
normalize_returns=False,
normalize_observations=True,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50, # per epoch cycle and MPI worker,
nb_eval_steps=100,
batch_size=64, # per MPI worker
tau=0.01,
eval_env=None,
param_noise_adaption_interval=50,
**network_kwargs):
set_global_seeds(seed)
if total_timesteps is not None:
assert nb_epochs is None
nb_epochs = int(total_timesteps) // (nb_epoch_cycles * nb_rollout_steps)
else:
nb_epochs = 500
if MPI is not None:
rank = MPI.COMM_WORLD.Get_rank()
else:
rank = 0
nb_actions = env.action_space.shape[-1]
assert (np.abs(env.action_space.low) == env.action_space.high).all() # we assume symmetric actions.
memory = Memory(limit=int(1e6), action_shape=env.action_space.shape, observation_shape=env.observation_space.shape)
critic = Critic(network=network, **network_kwargs)
actor = Actor(nb_actions, network=network, **network_kwargs)
action_noise = None
param_noise = None
if noise_type is not None:
for current_noise_type in noise_type.split(','):
current_noise_type = current_noise_type.strip()
if current_noise_type == 'none':
pass
elif 'adaptive-param' in current_noise_type:
_, stddev = current_noise_type.split('_')
param_noise = AdaptiveParamNoiseSpec(initial_stddev=float(stddev), desired_action_stddev=float(stddev))
elif 'normal' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = NormalActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
elif 'ou' in current_noise_type:
_, stddev = current_noise_type.split('_')
action_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(nb_actions), sigma=float(stddev) * np.ones(nb_actions))
else:
raise RuntimeError('unknown noise type "{}"'.format(current_noise_type))
max_action = env.action_space.high
logger.info('scaling actions by {} before executing in env'.format(max_action))
agent = DDPG(actor, critic, memory, env.observation_space.shape, env.action_space.shape,
gamma=gamma, tau=tau, normalize_returns=normalize_returns, normalize_observations=normalize_observations,
batch_size=batch_size, action_noise=action_noise, param_noise=param_noise, critic_l2_reg=critic_l2_reg,
actor_lr=actor_lr, critic_lr=critic_lr, enable_popart=popart, clip_norm=clip_norm,
reward_scale=reward_scale)
logger.info('Using agent with the following configuration:')
logger.info(str(agent.__dict__.items()))
eval_episode_rewards_history = deque(maxlen=100)
episode_rewards_history = deque(maxlen=100)
sess = U.get_session()
# Prepare everything.
agent.initialize(sess)
sess.graph.finalize()
agent.reset()
obs = env.reset()
if eval_env is not None:
eval_obs = eval_env.reset()
nenvs = obs.shape[0]
episode_reward = np.zeros(nenvs, dtype = np.float32) #vector
episode_step = np.zeros(nenvs, dtype = int) # vector
episodes = 0 #scalar
t = 0 # scalar
epoch = 0
start_time = time.time()
epoch_episode_rewards = []
epoch_episode_steps = []
epoch_actions = []
epoch_qs = []
epoch_episodes = 0
for epoch in range(nb_epochs):
for cycle in range(nb_epoch_cycles):
# Perform rollouts.
if nenvs > 1:
# if simulating multiple envs in parallel, impossible to reset agent at the end of the episode in each
# of the environments, so resetting here instead
agent.reset()
for t_rollout in range(nb_rollout_steps):
# Predict next action.
action, q, _, _ = agent.step(obs, apply_noise=True, compute_Q=True)
# Execute next action.
if rank == 0 and render:
env.render()
# max_action is of dimension A, whereas action is dimension (nenvs, A) - the multiplication gets broadcasted to the batch
new_obs, r, done, info = env.step(max_action * action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
# note these outputs are batched from vecenv
t += 1
if rank == 0 and render:
env.render()
episode_reward += r
episode_step += 1
# Book-keeping.
epoch_actions.append(action)
epoch_qs.append(q)
agent.store_transition(obs, action, r, new_obs, done) #the batched data will be unrolled in memory.py's append.
obs = new_obs
for d in range(len(done)):
if done[d]:
# Episode done.
epoch_episode_rewards.append(episode_reward[d])
episode_rewards_history.append(episode_reward[d])
epoch_episode_steps.append(episode_step[d])
episode_reward[d] = 0.
episode_step[d] = 0
epoch_episodes += 1
episodes += 1
if nenvs == 1:
agent.reset()
# Train.
epoch_actor_losses = []
epoch_critic_losses = []
epoch_adaptive_distances = []
for t_train in range(nb_train_steps):
# Adapt param noise, if necessary.
if memory.nb_entries >= batch_size and t_train % param_noise_adaption_interval == 0:
distance = agent.adapt_param_noise()
epoch_adaptive_distances.append(distance)
cl, al = agent.train()
epoch_critic_losses.append(cl)
epoch_actor_losses.append(al)
agent.update_target_net()
# Evaluate.
eval_episode_rewards = []
eval_qs = []
if eval_env is not None:
nenvs_eval = eval_obs.shape[0]
eval_episode_reward = np.zeros(nenvs_eval, dtype = np.float32)
for t_rollout in range(nb_eval_steps):
eval_action, eval_q, _, _ = agent.step(eval_obs, apply_noise=False, compute_Q=True)
eval_obs, eval_r, eval_done, eval_info = eval_env.step(max_action * eval_action) # scale for execution in env (as far as DDPG is concerned, every action is in [-1, 1])
if render_eval:
eval_env.render()
eval_episode_reward += eval_r
eval_qs.append(eval_q)
for d in range(len(eval_done)):
if eval_done[d]:
eval_episode_rewards.append(eval_episode_reward[d])
eval_episode_rewards_history.append(eval_episode_reward[d])
eval_episode_reward[d] = 0.0
if MPI is not None:
mpi_size = MPI.COMM_WORLD.Get_size()
else:
mpi_size = 1
# Log stats.
# XXX shouldn't call np.mean on variable length lists
duration = time.time() - start_time
stats = agent.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)
combined_stats['train/param_noise_distance'] = np.mean(epoch_adaptive_distances)
combined_stats['total/duration'] = duration
combined_stats['total/steps_per_second'] = float(t) / 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 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(x):
if isinstance(x, np.ndarray):
assert x.size == 1
return x[0]
elif np.isscalar(x):
return x
else:
raise ValueError('expected scalar, got %s'%x)
combined_stats_sums = np.array([ np.array(x).flatten()[0] for x in combined_stats.values()])
if MPI is not None:
combined_stats_sums = MPI.COMM_WORLD.allreduce(combined_stats_sums)
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'] = t
for key in sorted(combined_stats.keys()):
logger.record_tabular(key, combined_stats[key])
if rank == 0:
logger.dump_tabular()
logger.info('')
logdir = logger.get_dir()
if rank == 0 and logdir:
if hasattr(env, 'get_state'):
with open(os.path.join(logdir, 'env_state.pkl'), 'wb') as f:
pickle.dump(env.get_state(), f)
if eval_env and hasattr(eval_env, 'get_state'):
with open(os.path.join(logdir, 'eval_env_state.pkl'), 'wb') as f:
pickle.dump(eval_env.get_state(), f)
return agent
def learn(network,
env,
seed=None,
nsteps=5,
total_timesteps=int(80e6),
vf_coef=0.5,
ent_coef=0.01,
max_grad_norm=0.5,
lr=7e-4,
lrschedule='linear',
epsilon=1e-5,
alpha=0.99,
gamma=0.99,
log_interval=100,
load_path=None,
**network_kwargs):
'''
Main entrypoint for A2C algorithm. Train a policy with given network search_space on a given environment using a2c algorithm.
Parameters:
-----------
network: policy network search_space. Either string (mlp, lstm, lnlstm, cnn_lstm, cnn, cnn_small, conv_only - see baselines.common/models.py for full list)
specifying the standard network search_space, or a function that takes tensorflow tensor as input and returns
tuple (output_tensor, extra_feed) where output tensor is the last network layer output, extra_feed is None for feed-forward
neural nets, and extra_feed is a dictionary describing how to feed state into the network for recurrent neural nets.
See baselines.common/policies.py/lstm for more details on using recurrent nets in policies
env: RL environment. Should implement interface similar to VecEnv (baselines.common/vec_env) or be wrapped with DummyVecEnv (baselines.common/vec_env/dummy_vec_env.py)
seed: seed to make random number sequence in the alorightm reproducible. By default is None which means seed from system noise generator (not reproducible)
nsteps: int, number of steps of the vectorized environment per update (i.e. batch size is nsteps * nenv where
nenv is number of environment copies simulated in parallel)
total_timesteps: int, total number of timesteps to train on (default: 80M)
vf_coef: float, coefficient in front of value function loss in the total loss function (default: 0.5)
ent_coef: float, coeffictiant in front of the policy entropy in the total loss function (default: 0.01)
max_gradient_norm: float, gradient is clipped to have global L2 norm no more than this value (default: 0.5)
lr: float, learning rate for RMSProp (current implementation has RMSProp hardcoded in) (default: 7e-4)
lrschedule: schedule of learning rate. Can be 'linear', 'constant', or a function [0..1] -> [0..1] that takes fraction of the training progress as input and
returns fraction of the learning rate (specified as lr) as output
epsilon: float, RMSProp epsilon (stabilizes square root computation in denominator of RMSProp update) (default: 1e-5)
alpha: float, RMSProp decay parameter (default: 0.99)
gamma: float, reward discounting parameter (default: 0.99)
log_interval: int, specifies how frequently the logs are printed out (default: 100)
**network_kwargs: keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
For instance, 'mlp' network search_space has arguments num_hidden and num_layers.
'''
set_global_seeds(seed)
# Get the nb of env
nenvs = env.num_envs
policy = build_policy(env, network, **network_kwargs)
# Instantiate the model object (that creates step_model and train_model)
model = Model(policy=policy,
env=env,
nsteps=nsteps,
ent_coef=ent_coef,
vf_coef=vf_coef,
max_grad_norm=max_grad_norm,
lr=lr,
alpha=alpha,
epsilon=epsilon,
total_timesteps=total_timesteps,
lrschedule=lrschedule)
if load_path is not None:
model.load(load_path)
# Instantiate the runner object
runner = Runner(env, model, nsteps=nsteps, gamma=gamma)
epinfobuf = deque(maxlen=100)
# Calculate the batch_size
nbatch = nenvs * nsteps
# Start total timer
tstart = time.time()
for update in range(1, total_timesteps // nbatch + 1):
# Get mini batch of experiences
obs, states, rewards, masks, actions, values, epinfos = runner.run()
epinfobuf.extend(epinfos)
policy_loss, value_loss, policy_entropy = model.train(
obs, states, rewards, masks, actions, values)
nseconds = time.time() - tstart
# Calculate the fps (frame per second)
fps = int((update * nbatch) / nseconds)
if update % log_interval == 0 or update == 1:
# Calculates if value function is a good predicator of the returns (ev > 1)
# or if it's just worse than predicting nothing (ev =< 0)
ev = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps", update * nbatch)
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(ev))
logger.record_tabular(
"eprewmean", safemean([epinfo['r'] for epinfo in epinfobuf]))
logger.record_tabular(
"eplenmean", safemean([epinfo['l'] for epinfo in epinfobuf]))
logger.dump_tabular()
return model
def learn(
*,
network,
env,
total_timesteps,
timesteps_per_batch=1024, # what to train on
max_kl=0.001,
cg_iters=10,
gamma=0.99,
lam=1.0, # advantage estimation
seed=None,
ent_coef=0.0,
cg_damping=1e-2,
vf_stepsize=3e-4,
vf_iters=3,
max_episodes=0,
max_iters=0, # time constraint
callback=None,
load_path=None,
**network_kwargs):
'''
learn a policy function with TRPO algorithm
Parameters:
----------
network neural network to learn. Can be either string ('mlp', 'cnn', 'lstm', 'lnlstm' for basic types)
or function that takes input placeholder and returns tuple (output, None) for feedforward nets
or (output, (state_placeholder, state_output, mask_placeholder)) for recurrent nets
env environment (one of the gym environments or wrapped via baselines.common.vec_env.VecEnv-type class
timesteps_per_batch timesteps per gradient estimation batch
max_kl max KL divergence between old policy and new policy ( KL(pi_old || pi) )
ent_coef coefficient of policy entropy term in the optimization objective
cg_iters number of iterations of conjugate gradient algorithm
cg_damping conjugate gradient damping
vf_stepsize learning rate for adam optimizer used to optimie value function loss
vf_iters number of iterations of value function optimization iterations per each policy optimization step
total_timesteps max number of timesteps
max_episodes max number of episodes
max_iters maximum number of policy optimization iterations
callback function to be called with (locals(), globals()) each policy optimization step
load_path str, path to load the model from (default: None, i.e. no model is loaded)
**network_kwargs keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
Returns:
-------
learnt model
'''
if MPI is not None:
nworkers = MPI.COMM_WORLD.Get_size()
rank = MPI.COMM_WORLD.Get_rank()
else:
nworkers = 1
rank = 0
cpus_per_worker = 1
U.get_session(
config=tf.ConfigProto(allow_soft_placement=True,
inter_op_parallelism_threads=cpus_per_worker,
intra_op_parallelism_threads=cpus_per_worker))
policy = build_policy(env, network, value_network='copy', **network_kwargs)
set_global_seeds(seed)
np.set_printoptions(precision=3)
# Setup losses and stuff
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
ob = observation_placeholder(ob_space)
with tf.variable_scope("pi"):
pi = policy(observ_placeholder=ob)
with tf.variable_scope("oldpi"):
oldpi = policy(observ_placeholder=ob)
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
ac = pi.pdtype.sample_placeholder([None])
kloldnew = oldpi.pd.kl(pi.pd)
ent = pi.pd.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = ent_coef * meanent
vferr = tf.reduce_mean(tf.square(pi.vf - ret))
ratio = tf.exp(pi.pd.logp(ac) -
oldpi.pd.logp(ac)) # advantage * pnew / pold
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy"]
dist = meankl
all_var_list = get_trainable_variables("pi")
# var_list = [v for v in all_var_list if v.name.split("/")[1].startswith("pol")]
# vf_var_list = [v for v in all_var_list if v.name.split("/")[1].startswith("vf")]
var_list = get_pi_trainable_variables("pi")
vf_var_list = get_vf_trainable_variables("pi")
vfadam = MpiAdam(vf_var_list)
get_flat = U.GetFlat(var_list)
set_from_flat = U.SetFromFlat(var_list)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32,
shape=[None],
name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
sz = U.intprod(shape)
tangents.append(tf.reshape(flat_tangent[start:start + sz], shape))
start += sz
gvp = tf.add_n([
tf.reduce_sum(g * tangent)
for (g, tangent) in zipsame(klgrads, tangents)
]) #pylint: disable=E1111
fvp = U.flatgrad(gvp, var_list)
assign_old_eq_new = U.function(
[], [],
updates=[
tf.assign(oldv, newv)
for (oldv,
newv) in zipsame(get_variables("oldpi"), get_variables("pi"))
])
compute_losses = U.function([ob, ac, atarg], losses)
compute_lossandgrad = U.function([ob, ac, atarg], losses +
[U.flatgrad(optimgain, var_list)])
compute_fvp = U.function([flat_tangent, ob, ac, atarg], fvp)
compute_vflossandgrad = U.function([ob, ret],
U.flatgrad(vferr, vf_var_list))
@contextmanager
def timed(msg):
if rank == 0:
print(colorize(msg, color='magenta'))
tstart = time.time()
yield
print(
colorize("done in %.3f seconds" % (time.time() - tstart),
color='magenta'))
else:
yield
def allmean(x):
assert isinstance(x, np.ndarray)
if MPI is not None:
out = np.empty_like(x)
MPI.COMM_WORLD.Allreduce(x, out, op=MPI.SUM)
out /= nworkers
else:
out = np.copy(x)
return out
U.initialize()
if load_path is not None:
pi.load(load_path)
th_init = get_flat()
if MPI is not None:
MPI.COMM_WORLD.Bcast(th_init, root=0)
set_from_flat(th_init)
vfadam.sync()
print("Init param sum", th_init.sum(), flush=True)
# Prepare for rollouts
# ----------------------------------------
seg_gen = traj_segment_generator(pi,
env,
timesteps_per_batch,
stochastic=True)
episodes_so_far = 0
timesteps_so_far = 0
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=40) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=40) # rolling buffer for episode rewards
if sum([max_iters > 0, total_timesteps > 0, max_episodes > 0]) == 0:
# noththing to be done
return pi
assert sum([max_iters>0, total_timesteps>0, max_episodes>0]) < 2, \
'out of max_iters, total_timesteps, and max_episodes only one should be specified'
while True:
if callback: callback(locals(), globals())
if total_timesteps and timesteps_so_far >= total_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
logger.log("********** Iteration %i ************" % iters_so_far)
with timed("sampling"):
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, gamma, lam)
# ob, ac, atarg, ret, td1ret = map(np.concatenate, (obs, acs, atargs, rets, td1rets))
ob, ac, atarg, tdlamret = seg["ob"], seg["ac"], seg["adv"], seg[
"tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()
) / atarg.std() # standardized advantage function estimate
if hasattr(pi, "ret_rms"): pi.ret_rms.update(tdlamret)
if hasattr(pi, "ob_rms"):
pi.ob_rms.update(ob) # update running mean/std for policy
args = seg["ob"], seg["ac"], atarg
fvpargs = [arr[::5] for arr in args]
def fisher_vector_product(p):
return allmean(compute_fvp(p, *fvpargs)) + cg_damping * p
assign_old_eq_new() # set old parameter values to new parameter values
with timed("computegrad"):
*lossbefore, g = compute_lossandgrad(*args)
lossbefore = allmean(np.array(lossbefore))
g = allmean(g)
if np.allclose(g, 0):
logger.log("Got zero gradient. not updating")
else:
with timed("cg"):
stepdir = cg(fisher_vector_product,
g,
cg_iters=cg_iters,
verbose=rank == 0)
assert np.isfinite(stepdir).all()
shs = .5 * stepdir.dot(fisher_vector_product(stepdir))
lm = np.sqrt(shs / max_kl)
# logger.log("lagrange multiplier:", lm, "gnorm:", np.linalg.norm(g))
fullstep = stepdir / lm
expectedimprove = g.dot(fullstep)
surrbefore = lossbefore[0]
stepsize = 1.0
thbefore = get_flat()
for _ in range(10):
thnew = thbefore + fullstep * stepsize
set_from_flat(thnew)
meanlosses = surr, kl, *_ = allmean(
np.array(compute_losses(*args)))
improve = surr - surrbefore
logger.log("Expected: %.3f Actual: %.3f" %
(expectedimprove, improve))
if not np.isfinite(meanlosses).all():
logger.log("Got non-finite value of losses -- bad!")
elif kl > max_kl * 1.5:
logger.log("violated KL constraint. shrinking step.")
elif improve < 0:
logger.log("surrogate didn't improve. shrinking step.")
else:
logger.log("Stepsize OK!")
break
stepsize *= .5
else:
logger.log("couldn't compute a good step")
set_from_flat(thbefore)
if nworkers > 1 and iters_so_far % 20 == 0:
paramsums = MPI.COMM_WORLD.allgather(
(thnew.sum(), vfadam.getflat().sum())) # list of tuples
assert all(
np.allclose(ps, paramsums[0]) for ps in paramsums[1:])
for (lossname, lossval) in zip(loss_names, meanlosses):
logger.record_tabular(lossname, lossval)
with timed("vf"):
for _ in range(vf_iters):
for (mbob, mbret) in dataset.iterbatches(
(seg["ob"], seg["tdlamret"]),
include_final_partial_batch=False,
batch_size=64):
g = allmean(compute_vflossandgrad(mbob, mbret))
vfadam.update(g, vf_stepsize)
logger.record_tabular("ev_tdlam_before",
explained_variance(vpredbefore, tdlamret))
lrlocal = (seg["ep_lens"], seg["ep_rets"]) # local values
if MPI is not None:
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
else:
listoflrpairs = [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 += sum(lens)
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - tstart)
if rank == 0:
logger.dump_tabular()
return pi
action = act(obs[None], update_eps=exploration.value(t))[0]
new_obs, rew, done, _ = env.step(action)
# Store transition in the replay buffer.
replay_buffer.add(obs, action, rew, new_obs, float(done))
obs = new_obs
episode_rewards[-1] += rew
if done:
obs = env.reset()
episode_rewards.append(0)
is_solved = t > 100 and np.mean(episode_rewards[-101:-1]) >= 200
if is_solved:
# Show off the result
env.render()
else:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if t > 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 t % 1000 == 0:
update_target()
if done and len(episode_rewards) % 10 == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", len(episode_rewards))
logger.record_tabular("mean episode reward", round(np.mean(episode_rewards[-101:-1]), 1))
logger.record_tabular("% time spent exploring", int(100 * exploration.value(t)))
logger.dump_tabular()
def learn(network,
env,
seed,
total_timesteps=int(40e6),
gamma=0.99,
log_interval=1,
nprocs=32,
nsteps=20,
ent_coef=0.01,
vf_coef=0.5,
vf_fisher_coef=1.0,
lr=0.25,
max_grad_norm=0.5,
kfac_clip=0.001,
save_interval=None,
lrschedule='linear',
load_path=None,
is_async=True,
**network_kwargs):
set_global_seeds(seed)
if network == 'cnn':
network_kwargs['one_dim_bias'] = True
policy = build_policy(env, network, **network_kwargs)
nenvs = env.num_envs
ob_space = env.observation_space
ac_space = env.action_space
make_model = lambda: Model(policy,
ob_space,
ac_space,
nenvs,
total_timesteps,
nprocs=nprocs,
nsteps=nsteps,
ent_coef=ent_coef,
vf_coef=vf_coef,
vf_fisher_coef=vf_fisher_coef,
lr=lr,
max_grad_norm=max_grad_norm,
kfac_clip=kfac_clip,
lrschedule=lrschedule,
is_async=is_async)
if save_interval and logger.get_dir():
import cloudpickle
with open(osp.join(logger.get_dir(), 'make_model.pkl'), 'wb') as fh:
fh.write(cloudpickle.dumps(make_model))
model = make_model()
if load_path is not None:
model.load(load_path)
runner = Runner(env, model, nsteps=nsteps, gamma=gamma)
epinfobuf = deque(maxlen=100)
nbatch = nenvs * nsteps
tstart = time.time()
coord = tf.train.Coordinator()
if is_async:
enqueue_threads = model.q_runner.create_threads(model.sess,
coord=coord,
start=True)
else:
enqueue_threads = []
for update in range(1, total_timesteps // nbatch + 1):
obs, states, rewards, masks, actions, values, epinfos = runner.run()
epinfobuf.extend(epinfos)
policy_loss, value_loss, policy_entropy = model.train(
obs, states, rewards, masks, actions, values)
model.old_obs = obs
nseconds = time.time() - tstart
fps = int((update * nbatch) / nseconds)
if update % log_interval == 0 or update == 1:
ev = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps", update * nbatch)
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(ev))
logger.record_tabular(
"eprewmean", safemean([epinfo['r'] for epinfo in epinfobuf]))
logger.record_tabular(
"eplenmean", safemean([epinfo['l'] for epinfo in epinfobuf]))
logger.dump_tabular()
if save_interval and (update % save_interval == 0
or update == 1) and logger.get_dir():
savepath = osp.join(logger.get_dir(), 'checkpoint%.5i' % update)
print('Saving to', savepath)
model.save(savepath)
coord.request_stop()
coord.join(enqueue_threads)
return model
def learn(env,
network,
seed=None,
lr=5e-4,
total_timesteps=100000,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02,
train_freq=1,
batch_size=32,
print_freq=100,
checkpoint_freq=10000,
checkpoint_path=None,
learning_starts=1000,
gamma=1.0,
target_network_update_freq=500,
prioritized_replay=False,
prioritized_replay_alpha=0.6,
prioritized_replay_beta0=0.4,
prioritized_replay_beta_iters=None,
prioritized_replay_eps=1e-6,
param_noise=False,
callback=None,
load_path=None,
**network_kwargs):
"""Train a deepq model.
Parameters
-------
env: gym.Env
environment to train on
network: string or a function
neural network to use as a q function approximator. If string, has to be one of the names of registered models in baselines.common.models
(mlp, cnn, conv_only). If a function, should take an observation tensor and return a latent variable tensor, which
will be mapped to the Q function heads (see build_q_func in baselines.deepq.models for details on that)
seed: int or None
prng seed. The runs with the same seed "should" give the same results. If None, no seeding is used.
lr: float
learning rate for adam optimizer
total_timesteps: int
number of env steps to optimizer for
buffer_size: int
size of the replay buffer
exploration_fraction: float
fraction of entire training period over which the exploration rate is annealed
exploration_final_eps: float
final value of random action probability
train_freq: int
update the model every `train_freq` steps.
set to None to disable printing
batch_size: int
size of a batched sampled from replay buffer for training
print_freq: int
how often to print out training progress
set to None to disable printing
checkpoint_freq: int
how often to save the model. This is so that the best version is restored
at the end of the training. If you do not wish to restore the best version at
the end of the training set this variable to None.
learning_starts: int
how many steps of the model to collect transitions for before learning starts
gamma: float
discount factor
target_network_update_freq: int
update the target network every `target_network_update_freq` steps.
prioritized_replay: True
if True prioritized replay buffer will be used.
prioritized_replay_alpha: float
alpha parameter for prioritized replay buffer
prioritized_replay_beta0: float
initial value of beta for prioritized replay buffer
prioritized_replay_beta_iters: int
number of iterations over which beta will be annealed from initial value
to 1.0. If set to None equals to total_timesteps.
prioritized_replay_eps: float
epsilon to add to the TD errors when updating priorities.
param_noise: bool
whether or not to use parameter space noise (https://arxiv.org/abs/1706.01905)
callback: (locals, globals) -> None
function called at every steps with state of the algorithm.
If callback returns true training stops.
load_path: str
path to load the model from. (default: None)
**network_kwargs
additional keyword arguments to pass to the network builder.
Returns
-------
act: ActWrapper
Wrapper over act function. Adds ability to save it and load it.
See header of baselines/deepq/categorical.py for details on the act function.
"""
# Create all the functions necessary to train the model
sess = get_session()
set_global_seeds(seed)
q_func = build_q_func(network, **network_kwargs)
# capture the shape outside the closure so that the env object is not serialized
# by cloudpickle when serializing make_obs_ph
observation_space = env.observation_space
def make_obs_ph(name):
return ObservationInput(observation_space, name=name)
act, train, update_target, debug = deepq.build_train(
make_obs_ph=make_obs_ph,
q_func=q_func,
num_actions=env.action_space.n,
optimizer=tf.train.AdamOptimizer(learning_rate=lr),
gamma=gamma,
grad_norm_clipping=10,
param_noise=param_noise)
act_params = {
'make_obs_ph': make_obs_ph,
'q_func': q_func,
'num_actions': env.action_space.n,
}
act = ActWrapper(act, act_params)
# Create the replay buffer
if prioritized_replay:
replay_buffer = PrioritizedReplayBuffer(buffer_size,
alpha=prioritized_replay_alpha)
if prioritized_replay_beta_iters is None:
prioritized_replay_beta_iters = total_timesteps
beta_schedule = LinearSchedule(prioritized_replay_beta_iters,
initial_p=prioritized_replay_beta0,
final_p=1.0)
else:
replay_buffer = ReplayBuffer(buffer_size)
beta_schedule = None
# Create the schedule for exploration starting from 1.
exploration = LinearSchedule(schedule_timesteps=int(exploration_fraction *
total_timesteps),
initial_p=1.0,
final_p=exploration_final_eps)
# Initialize the parameters and copy them to the target network.
U.initialize()
update_target()
episode_rewards = [0.0]
saved_mean_reward = None
obs = env.reset()
reset = True
with tempfile.TemporaryDirectory() as td:
td = checkpoint_path or td
model_file = os.path.join(td, "model")
model_saved = False
if tf.train.latest_checkpoint(td) is not None:
load_variables(model_file)
logger.log('Loaded model from {}'.format(model_file))
model_saved = True
elif load_path is not None:
load_variables(load_path)
logger.log('Loaded model from {}'.format(load_path))
for t in range(total_timesteps):
if callback is not None:
if callback(locals(), globals()):
break
# Take action and update exploration to the newest value
kwargs = {}
if not param_noise:
update_eps = exploration.value(t)
update_param_noise_threshold = 0.
else:
update_eps = 0.
# Compute the threshold such that the KL divergence between perturbed and non-perturbed
# policy is comparable to eps-greedy exploration with eps = exploration.value(t).
# See Appendix C.1 in Parameter Space Noise for Exploration, Plappert et al., 2017
# for detailed explanation.
update_param_noise_threshold = -np.log(1. - exploration.value(
t) + exploration.value(t) / float(env.action_space.n))
kwargs['reset'] = reset
kwargs[
'update_param_noise_threshold'] = update_param_noise_threshold
kwargs['update_param_noise_scale'] = True
action = act(np.array(obs)[None], update_eps=update_eps,
**kwargs)[0]
env_action = action
reset = False
new_obs, rew, done, _ = env.step(env_action)
# Store transition in the replay buffer.
replay_buffer.add(obs, action, rew, new_obs, float(done))
obs = new_obs
episode_rewards[-1] += rew
if done:
obs = env.reset()
episode_rewards.append(0.0)
reset = True
if t > learning_starts and t % train_freq == 0:
# Minimize the error in Bellman's equation on a batch sampled from replay buffer.
if prioritized_replay:
experience = replay_buffer.sample(
batch_size, beta=beta_schedule.value(t))
(obses_t, actions, rewards, obses_tp1, dones, weights,
batch_idxes) = experience
else:
obses_t, actions, rewards, obses_tp1, dones = replay_buffer.sample(
batch_size)
weights, batch_idxes = np.ones_like(rewards), None
td_errors = train(obses_t, actions, rewards, obses_tp1, dones,
weights)
if prioritized_replay:
new_priorities = np.abs(td_errors) + prioritized_replay_eps
replay_buffer.update_priorities(batch_idxes,
new_priorities)
if t > learning_starts and t % target_network_update_freq == 0:
# Update target network periodically.
update_target()
mean_100ep_reward = round(np.mean(episode_rewards[-101:-1]), 1)
num_episodes = len(episode_rewards)
if done and print_freq is not None and len(
episode_rewards) % print_freq == 0:
logger.record_tabular("steps", t)
logger.record_tabular("episodes", num_episodes)
logger.record_tabular("mean 100 episode reward",
mean_100ep_reward)
logger.record_tabular("% time spent exploring",
int(100 * exploration.value(t)))
logger.dump_tabular()
if (checkpoint_freq is not None and t > learning_starts
and num_episodes > 100 and t % checkpoint_freq == 0):
if saved_mean_reward is None or mean_100ep_reward > saved_mean_reward:
if print_freq is not None:
logger.log(
"Saving model due to mean reward increase: {} -> {}"
.format(saved_mean_reward, mean_100ep_reward))
save_variables(model_file)
model_saved = True
saved_mean_reward = mean_100ep_reward
if model_saved:
if print_freq is not None:
logger.log("Restored model with mean reward: {}".format(
saved_mean_reward))
load_variables(model_file)
return act
def train(*, policy, rollout_worker, evaluator,
n_epochs, n_test_rollouts, n_cycles, n_batches, policy_save_interval,
save_path, demo_file, **kwargs):
rank = MPI.COMM_WORLD.Get_rank()
if save_path:
latest_policy_path = os.path.join(save_path, 'policy_latest.pkl')
best_policy_path = os.path.join(save_path, 'policy_best.pkl')
periodic_policy_path = os.path.join(save_path, 'policy_{}.pkl')
logger.info("Training...")
best_success_rate = -1
if policy.bc_loss == 1: policy.init_demo_buffer(demo_file) #initialize demo buffer if training with demonstrations
# num_timesteps = n_epochs * n_cycles * rollout_length * number of rollout workers
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()
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_path:
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_path:
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]
return policy
