def mpi_avg_grads(module):
if num_procs() == 1:
return
for p in module.parameters():
p_grad_np = p.grad.numpy()
avg_p_grad = mpi_avg(p_grad_np)
p_grad_np[:] = avg_p_grad[:]
def sync_params(module):
if num_procs() == 1:
return
for p in module.parameters():
p_np = p.data.numpy()
broadcast(p_np)
def sync_params(module):
""" Sync all parameters of module across all MPI processes. """
if num_procs() == 1:
return
for p in module.parameters():
p_numpy = p.data.numpy()
broadcast(p_numpy)
def setup_pytorch_for_mpi():
"""
Avoid slowdowns caused by each separate process's Pytorch using more than its fair share of CPU resources.
"""
if torch.get_num_threads() == 1:
return
fair_num_threads = max(torch.get_num_threads() // num_procs(), 1)
torch.set_num_threads(fair_num_threads)
def mpi_avg_grads(module):
""" Average contents of gradient buffers across MPI processes. """
if num_procs() == 1:
return
for p in module.parameters():
p_grad_numpy = p.grad.numpy() # numpy view of tensor data
avg_p_grad = mpi_avg(p.grad)
p_grad_numpy[:] = avg_p_grad[:]
def setup_pytorch_for_mpi():
"""
Avoid slowdowns caused by each separate process's PyTorch using
more than its fair share of CPU resources.
"""
#print('Proc %d: Reporting original number of Torch threads as %d.'%(proc_id(), torch.get_num_threads()), flush=True)
if torch.get_num_threads() == 1:
return
fair_num_threads = max(int(torch.get_num_threads() / num_procs()), 1)
torch.set_num_threads(fair_num_threads)
def ppo(env_fn,
actor_critic=core.MLPActorCritic,
ac_kwargs=dict(),
seed=0,
steps_per_epoch=4000,
epochs=50,
gamma=0.99,
clip_ratio=0.2,
pi_lr=3e-4,
vf_lr=1e-3,
train_pi_iters=80,
train_v_iters=80,
lam=0.97,
max_ep_len=None,
target_kl=0.01,
logger_kwargs=dict(),
save_freq=10,
TensorBoard=True,
save_nn=True,
save_every=1000,
load_latest=False,
load_custom=False,
LoadPath=None,
RTA_type=None):
"""
Proximal Policy Optimization (by clipping),
with early stopping based on approximate KL
Args:
env_fn : A function which creates a copy of the environment.
The environment must satisfy the OpenAI Gym API.
actor_critic: The constructor method for a PyTorch Module with a
``step`` method, an ``act`` method, a ``pi`` module, and a ``v``
module. The ``step`` method should accept a batch of observations
and return:
=========== ================ ======================================
Symbol Shape Description
=========== ================ ======================================
``a`` (batch, act_dim) | Numpy array of actions for each
| observation.
``v`` (batch,) | Numpy array of value estimates
| for the provided observations.
``logp_a`` (batch,) | Numpy array of log probs for the
| actions in ``a``.
=========== ================ ======================================
The ``act`` method behaves the same as ``step`` but only returns ``a``.
The ``pi`` module's forward call should accept a batch of
observations and optionally a batch of actions, and return:
=========== ================ ======================================
Symbol Shape Description
=========== ================ ======================================
``pi`` N/A | Torch Distribution object, containing
| a batch of distributions describing
| the policy for the provided observations.
``logp_a`` (batch,) | Optional (only returned if batch of
| actions is given). Tensor containing
| the log probability, according to
| the policy, of the provided actions.
| If actions not given, will contain
| ``None``.
=========== ================ ======================================
The ``v`` module's forward call should accept a batch of observations
and return:
=========== ================ ======================================
Symbol Shape Description
=========== ================ ======================================
``v`` (batch,) | Tensor containing the value estimates
| for the provided observations. (Critical:
| make sure to flatten this!)
=========== ================ ======================================
ac_kwargs (dict): Any kwargs appropriate for the ActorCritic object
you provided to PPO.
seed (int): Seed for random number generators.
steps_per_epoch (int): Number of steps of interaction (state-action pairs)
for the agent and the environment in each epoch.
epochs (int): Number of epochs of interaction (equivalent to
number of policy updates) to perform.
gamma (float): Discount factor. (Always between 0 and 1.)
clip_ratio (float): Hyperparameter for clipping in the policy objective.
Roughly: how far can the new policy go from the old policy while
still profiting (improving the objective function)? The new policy
can still go farther than the clip_ratio says, but it doesn't help
on the objective anymore. (Usually small, 0.1 to 0.3.) Typically
denoted by :math:`\epsilon`.
pi_lr (float): Learning rate for policy optimizer.
vf_lr (float): Learning rate for value function optimizer.
train_pi_iters (int): Maximum number of gradient descent steps to take
on policy loss per epoch. (Early stopping may cause optimizer
to take fewer than this.)
train_v_iters (int): Number of gradient descent steps to take on
value function per epoch.
lam (float): Lambda for GAE-Lambda. (Always between 0 and 1,
close to 1.)
max_ep_len (int): Maximum length of trajectory / episode / rollout.
target_kl (float): Roughly what KL divergence we think is appropriate
between new and old policies after an update. This will get used
for early stopping. (Usually small, 0.01 or 0.05.)
logger_kwargs (dict): Keyword args for EpochLogger.
save_freq (int): How often (in terms of gap between epochs) to save
the current policy and value function.
TensorBoard (bool): True plots to TensorBoard, False does not
save_nn (bool): True saves neural network data, False does not
save_every (int): How often to save neural network
load_latest (bool): Load last saved neural network data before training
load_custom (bool): Load custom neural network data file before training
LoadPath (str): Path for custom neural network data file
RTA_type (str): RTA framework, either 'CBF', 'SVL', 'ASIF', or
'SBSF'
"""
# Special function to avoid certain slowdowns from PyTorch + MPI combo.
setup_pytorch_for_mpi()
# Set up logger and save configuration
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
# Instantiate environment
env = env_fn()
obs_dim = env.observation_space.shape
act_dim = env.action_space.shape
# Random seed for each cpu
seed += 1 * proc_id()
env.seed(seed)
# Create actor-critic module
ac = actor_critic(env.observation_space, env.action_space, **ac_kwargs)
# Load model if True
if load_latest:
models = glob.glob(f"{PATH}/models/PPO/*")
LoadPath = max(models, key=os.path.getctime)
ac.load_state_dict(torch.load(LoadPath))
elif load_custom:
ac.load_state_dict(torch.load(LoadPath))
# Sync params across processes
sync_params(ac)
# Count variables
var_counts = tuple(core.count_vars(module) for module in [ac.pi, ac.v])
logger.log('\nNumber of parameters: \t pi: %d, \t v: %d\n' % var_counts)
# Set up experience buffer
local_steps_per_epoch = int(steps_per_epoch / num_procs())
buf = PPOBuffer(obs_dim, act_dim, local_steps_per_epoch, gamma, lam)
# Set up function for computing PPO policy loss
def compute_loss_pi(data):
obs, act, adv, logp_old = data['obs'], data['act'], data['adv'], data[
'logp']
# Policy loss
pi, logp = ac.pi(obs, act)
ratio = torch.exp(logp - logp_old)
clip_adv = torch.clamp(ratio, 1 - clip_ratio, 1 + clip_ratio) * adv
loss_pi = -(torch.min(ratio * adv, clip_adv)).mean()
# Useful extra info
approx_kl = (logp_old - logp).mean().item()
ent = pi.entropy().mean().item()
clipped = ratio.gt(1 + clip_ratio) | ratio.lt(1 - clip_ratio)
clipfrac = torch.as_tensor(clipped, dtype=torch.float32).mean().item()
pi_info = dict(kl=approx_kl, ent=ent, cf=clipfrac)
return loss_pi, pi_info
# Set up function for computing value loss
def compute_loss_v(data):
obs, ret = data['obs'], data['ret']
return ((ac.v(obs) - ret)**2).mean()
# Set up optimizers for policy and value function
pi_optimizer = Adam(ac.pi.parameters(), lr=pi_lr)
vf_optimizer = Adam(ac.v.parameters(), lr=vf_lr)
# Set up model saving
logger.setup_pytorch_saver(ac)
def update():
data = buf.get()
pi_l_old, pi_info_old = compute_loss_pi(data)
pi_l_old = pi_l_old.item()
v_l_old = compute_loss_v(data).item()
# Train policy with multiple steps of gradient descent
for i in range(train_pi_iters):
pi_optimizer.zero_grad()
loss_pi, pi_info = compute_loss_pi(data)
kl = mpi_avg(pi_info['kl'])
if kl > 1.5 * target_kl:
logger.log(
'Early stopping at step %d due to reaching max kl.' % i)
break
loss_pi.backward()
mpi_avg_grads(ac.pi) # average grads across MPI processes
pi_optimizer.step()
logger.store(StopIter=i)
# Value function learning
for i in range(train_v_iters):
vf_optimizer.zero_grad()
loss_v = compute_loss_v(data)
loss_v.backward()
mpi_avg_grads(ac.v) # average grads across MPI processes
vf_optimizer.step()
# Log changes from update
kl, ent, cf = pi_info['kl'], pi_info_old['ent'], pi_info['cf']
logger.store(LossPi=pi_l_old,
LossV=v_l_old,
KL=kl,
Entropy=ent,
ClipFrac=cf,
DeltaLossPi=(loss_pi.item() - pi_l_old),
DeltaLossV=(loss_v.item() - v_l_old))
# Import RTA
if RTA_type == 'CBF':
from CBF_for_speed_limit import RTA
elif RTA_type == 'SVL':
from Simple_velocity_limit import RTA
elif RTA_type == 'ASIF':
from IASIF import RTA
elif RTA_type == 'SBSF':
from ISimplex import RTA
# Call RTA, define action conversion
if RTA_type != 'off':
env.RTA_reward = RTA_type
rta = RTA(env)
def RTA_act(obs, act):
act = np.clip(act, -env.force_magnitude, env.force_magnitude)
x0 = [obs[0], obs[1], 0, obs[2], obs[3], 0]
u_des = np.array([[act[0]], [act[1]], [0]])
u = rta.main(x0, u_des)
new_act = [u[0, 0], u[1, 0]]
if np.sqrt((act[0] - new_act[0])**2 +
(act[1] - new_act[1])**2) < 0.0001:
env.RTA_on = False
else:
env.RTA_on = True
return new_act
# Prepare for interaction with environment
start_time = time.time()
o, ep_ret, ep_len = env.reset(), 0, 0
total_episodes = 0
RTA_percent = 0
# Create TensorBoard file if True
if TensorBoard and proc_id() == 0:
if env_name == 'spacecraft-docking-continuous-v0' or env_name == 'spacecraft-docking-v0':
Name = f"{PATH}/runs/Spacecraft-docking-" + current_time
elif env_name == 'dubins-aircraft-v0' or env_name == 'dubins-aircraft-continuous-v0':
Name = f"{PATH}/runs/Dubins-aircraft-" + current_time
writer = SummaryWriter(Name)
# Main loop: collect experience in env and update/log each epoch
for epoch in range(epochs):
batch_ret = [] # Track episode returns
batch_len = [] # Track episode lengths
batch_RTA_percent = [] # Track precentage of time RTA is on
env.success = 0 # Track episode success rate
env.failure = 0 # Track episode failure rate
env.crash = 0 # Track episode crash rate
env.overtime = 0 # Track episode over max time/control rate
episodes = 0 # Track episodes
delta_v = [] # Track episode total delta v
for t in range(local_steps_per_epoch):
a, v, logp = ac.step(torch.as_tensor(o, dtype=torch.float32))
if RTA_type != 'off': # If RTA is on, get RTA action
RTA_a = RTA_act(o, a)
if env.RTA_on:
RTA_percent += 1
next_o, r, d, _ = env.step(RTA_a)
else: # If RTA is off, pass through desired action
next_o, r, d, _ = env.step(a)
if env_name == 'spacecraft-docking-continuous-v0' or env_name == 'spacecraft-docking-v0':
over_max_vel, _, _ = env.check_velocity(a[0], a[1])
if over_max_vel:
RTA_percent += 1
ep_ret += r
ep_len += 1
# save and log
buf.store(o, a, r, v, logp)
logger.store(VVals=v)
# Update obs (critical!)
o = next_o
timeout = ep_len == max_ep_len
terminal = d or timeout
epoch_ended = t == local_steps_per_epoch - 1
if terminal or epoch_ended:
if epoch_ended and not (terminal):
print('Warning: trajectory cut off by epoch at %d steps.' %
ep_len,
flush=True)
# if trajectory didn't reach terminal state, bootstrap value target
if timeout or epoch_ended:
_, v, _ = ac.step(torch.as_tensor(o, dtype=torch.float32))
else:
v = 0
buf.finish_path(v)
if terminal:
# only save EpRet / EpLen if trajectory finished
logger.store(EpRet=ep_ret, EpLen=ep_len)
batch_ret.append(ep_ret)
batch_len.append(ep_len)
episodes += 1
if env_name == 'spacecraft-docking-continuous-v0' or env_name == 'spacecraft-docking-v0':
delta_v.append(env.control_input / env.mass_deputy)
batch_RTA_percent.append(RTA_percent / ep_len * 100)
RTA_percent = 0
o, ep_ret, ep_len = env.reset(), 0, 0
total_episodes += episodes
# Track success, failure, crash, overtime rates
if episodes != 0:
success_rate = env.success / episodes
failure_rate = env.failure / episodes
crash_rate = env.crash / episodes
overtime_rate = env.overtime / episodes
else:
success_rate = 0
failure_rate = 0
crash_rate = 0
overtime_rate = 0
raise (
"No completed episodes logging will break [increase steps per epoch]"
)
# Save model
if (epoch % save_freq == 0) or (epoch == epochs - 1):
logger.save_state({'env': env}, None)
# Perform PPO update!
update()
# Log info about epoch
logger.log_tabular('Epoch', epoch)
logger.log_tabular('EpRet', with_min_and_max=True)
logger.log_tabular('EpLen', average_only=True)
logger.log_tabular('VVals', with_min_and_max=True)
logger.log_tabular('TotalEnvInteracts', (epoch + 1) * steps_per_epoch)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('LossV', average_only=True)
logger.log_tabular('DeltaLossPi', average_only=True)
logger.log_tabular('DeltaLossV', average_only=True)
logger.log_tabular('Entropy', average_only=True)
logger.log_tabular('KL', average_only=True)
logger.log_tabular('ClipFrac', average_only=True)
logger.log_tabular('StopIter', average_only=True)
logger.log_tabular('Time', time.time() - start_time)
logger.dump_tabular()
# Average data over all cpus
avg_batch_ret = mpi_avg(np.mean(batch_ret))
avg_batch_len = mpi_avg(np.mean(batch_len))
avg_success_rate = mpi_avg(success_rate)
avg_failure_rate = mpi_avg(failure_rate)
avg_crash_rate = mpi_avg(crash_rate)
avg_overtime_rate = mpi_avg(overtime_rate)
if env_name == 'spacecraft-docking-continuous-v0' or env_name == 'spacecraft-docking-v0':
avg_delta_v = mpi_avg(np.mean(delta_v))
avg_RTA_percent = mpi_avg(np.mean(batch_RTA_percent))
if proc_id() == 0: # Only on one cpu
# Plot to TensorBoard if True, only on one cpu
if TensorBoard:
writer.add_scalar('Return', avg_batch_ret, epoch)
writer.add_scalar('Episode-Length', avg_batch_len * env.tau,
epoch)
writer.add_scalar('Success-Rate', avg_success_rate * 100,
epoch)
writer.add_scalar('Failure-Rate', avg_failure_rate * 100,
epoch)
writer.add_scalar('Crash-Rate', avg_crash_rate * 100, epoch)
writer.add_scalar('Overtime-Rate', avg_overtime_rate * 100,
epoch)
if env_name == 'spacecraft-docking-continuous-v0' or env_name == 'spacecraft-docking-v0':
writer.add_scalar('Delta-V', avg_delta_v, epoch)
writer.add_scalar('RTA-on-percent', avg_RTA_percent, epoch)
# Save neural network if true, can change to desired location
if save_nn and epoch % save_every == 0 and epoch != 0:
if not os.path.isdir(f"{PATH}/models"):
os.mkdir(f"{PATH}/models")
if not os.path.isdir(f"{PATH}/models/PPO"):
os.mkdir(f"{PATH}/models/PPO")
if env_name == 'spacecraft-docking-continuous-v0' or env_name == 'spacecraft-docking-v0':
Name2 = f"{PATH}/models/PPO/Spacecraft-docking-" + current_time + f"-epoch{epoch}.dat"
elif env_name == 'dubins-aircraft-v0' or env_name == 'dubins-aircraft-continuous-v0':
Name2 = f"{PATH}/models/PPO/Dubins-aircraft-" + current_time + f"-epoch{epoch}.dat"
torch.save(ac.state_dict(), Name2)
# Average episodes per hour, episode per epoch
ep_hr = mpi_avg(total_episodes) * args.cpu / (time.time() -
start_time) * 3600
ep_Ep = mpi_avg(total_episodes) * args.cpu / (epoch + 1)
# Plot on one cpu
if proc_id() == 0:
# Save neural network
if save_nn:
if not os.path.isdir(f"{PATH}/models"):
os.mkdir(f"{PATH}/models")
if not os.path.isdir(f"{PATH}/models/PPO"):
os.mkdir(f"{PATH}/models/PPO")
if env_name == 'spacecraft-docking-continuous-v0' or env_name == 'spacecraft-docking-v0':
Name2 = f"{PATH}/models/PPO/Spacecraft-docking-" + current_time + "-final.dat"
elif env_name == 'dubins-aircraft-v0' or env_name == 'dubins-aircraft-continuous-v0':
Name2 = f"{PATH}/models/PPO/Dubins-aircraft-" + current_time + "-final.dat"
torch.save(ac.state_dict(), Name2)
# Print statistics on episodes
print(
f"Episodes per hour: {ep_hr:.0f}, Episodes per epoch: {ep_Ep:.0f}, Epochs per hour: {(epoch+1)/(time.time()-start_time)*3600:.0f}"
)
def ppo(env_fn,
actor_critic=core.mlp_actor_critic,
ac_kwargs=dict(),
seed=0,
trials_per_epoch=2500,
steps_per_trial=100,
epochs=50,
gamma=0.99,
clip_ratio=0.2,
pi_lr=3e-4,
vf_lr=1e-3,
train_pi_iters=1000,
train_v_iters=80,
lam=0.97,
max_ep_len=1000,
target_kl=0.01,
logger_kwargs=dict(),
save_freq=10):
"""
Args:
env_fn : A function which creates a copy of the environment.
The environment must satisfy the OpenAI Gym API.
actor_critic: A function which takes in placeholder symbols
for state, ``x_ph``, and action, ``a_ph``, and returns the main
outputs from the agent's Tensorflow computation graph:
=========== ================ ======================================
Symbol Shape Description
=========== ================ ======================================
``pi`` (batch, act_dim) | Samples actions from policy given
| states.
``logp`` (batch,) | Gives log probability, according to
| the policy, of taking actions ``a_ph``
| in states ``x_ph``.
``logp_pi`` (batch,) | Gives log probability, according to
| the policy, of the action sampled by
| ``pi``.
``v`` (batch,) | Gives the value estimate for states
| in ``x_ph``. (Critical: make sure
| to flatten this!)
=========== ================ ======================================
ac_kwargs (dict): Any kwargs appropriate for the actor_critic
function you provided to PPO.
seed (int): Seed for random number generators.
steps_per_epoch (int): Number of steps of interaction (state-action pairs)
for the agent and the environment in each epoch.
epochs (int): Number of epochs of interaction (equivalent to
number of policy updates) to perform.
gamma (float): Discount factor. (Always between 0 and 1.)
clip_ratio (float): Hyperparameter for clipping in the policy objective.
Roughly: how far can the new policy go from the old policy while
still profiting (improving the objective function)? The new policy
can still go farther than the clip_ratio says, but it doesn't help
on the objective anymore. (Usually small, 0.1 to 0.3.)
pi_lr (float): Learning rate for policy optimizer.
vf_lr (float): Learning rate for value function optimizer.
train_pi_iters (int): Maximum number of gradient descent steps to take
on policy loss per epoch. (Early stopping may cause optimizer
to take fewer than this.)
train_v_iters (int): Number of gradient descent steps to take on
value function per epoch.
lam (float): Lambda for GAE-Lambda. (Always between 0 and 1,
close to 1.)
max_ep_len (int): Maximum length of trajectory / episode / rollout.
target_kl (float): Roughly what KL divergence we think is appropriate
between new and old policies after an update. This will get used
for early stopping. (Usually small, 0.01 or 0.05.)
logger_kwargs (dict): Keyword args for EpochLogger.
save_freq (int): How often (in terms of gap between epochs) to save
the current policy and value function.
"""
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
seed += 10000 * proc_id()
tf.set_random_seed(seed)
np.random.seed(seed)
env = env_fn()
obs_dim = env.observation_space.shape
act_dim = env.action_space.shape
# Share information about action space with policy architecture
ac_kwargs['action_space'] = env.action_space
# Inputs to computation graph
# x_ph, a_ph = core.placeholders_from_spaces(env.observation_space, env.action_space)
x_ph = tf.placeholder(dtype=tf.float32, shape=(None, None, 1), name='x_ph')
a_ph = tf.placeholder(dtype=tf.int32, shape=(None, None), name='a_ph')
# adv_ph, ret_ph, logp_old_ph, rew_ph = core.placeholders(None, None, None, 1)
adv_ph = tf.placeholder(dtype=tf.float32,
shape=(None, None),
name='adv_ph')
ret_ph = tf.placeholder(dtype=tf.float32,
shape=(None, None),
name='ret_ph')
logp_old_ph = tf.placeholder(dtype=tf.float32,
shape=(None, None),
name='logp_old_ph')
rew_ph = tf.placeholder(dtype=tf.float32,
shape=(None, None, 1),
name='rew_ph')
pi_state_ph = tf.placeholder(dtype=tf.float32,
shape=(None, NUM_GRU_UNITS),
name='pi_state_ph')
v_state_ph = tf.placeholder(dtype=tf.float32,
shape=(None, NUM_GRU_UNITS),
name='v_state_ph')
# Initialize rnn states for pi and v
# Main outputs from computation graph
pi, logp, logp_pi, v, new_pi_state, new_v_state = actor_critic(
x_ph,
a_ph,
rew_ph,
pi_state_ph,
v_state_ph,
NUM_GRU_UNITS,
action_space=env.action_space)
# Need all placeholders in *this* order later (to zip with data from buffer)
all_phs = [x_ph, a_ph, adv_ph, ret_ph, logp_old_ph, rew_ph]
# Every step, get: action, value, and logprob and reward
get_action_ops = [pi, v, logp_pi, new_pi_state, new_v_state]
# Experience buffer
steps_per_epoch = trials_per_epoch * steps_per_trial
local_steps_per_epoch = int(steps_per_epoch / num_procs())
buf = PPOBuffer(obs_dim, act_dim, local_steps_per_epoch, gamma, lam)
# Count variables
var_counts = tuple(core.count_vars(scope) for scope in ['pi', 'v'])
logger.log('\nNumber of parameters: \t pi: %d, \t v: %d\n' % var_counts)
# PPO objectives
ratio = tf.exp(logp - logp_old_ph) # pi(a|s) / pi_old(a|s)
min_adv = tf.where(adv_ph > 0, (1 + clip_ratio) * adv_ph,
(1 - clip_ratio) * adv_ph)
pi_loss = -tf.reduce_mean(tf.minimum(ratio * adv_ph, min_adv))
v_loss = tf.reduce_mean((ret_ph - v)**2)
# Info (useful to watch during learning)
approx_kl = tf.reduce_mean(
logp_old_ph -
logp) # a sample estimate for KL-divergence, easy to compute
approx_ent = tf.reduce_mean(
-logp) # a sample estimate for entropy, also easy to compute
clipped = tf.logical_or(ratio > (1 + clip_ratio), ratio < (1 - clip_ratio))
clipfrac = tf.reduce_mean(tf.cast(clipped, tf.float32))
# Optimizers
train_pi = MpiAdamOptimizer(
learning_rate=pi_lr).minimize(pi_loss - 0.01 * approx_ent)
train_v = MpiAdamOptimizer(learning_rate=vf_lr).minimize(v_loss)
sess = tf.Session()
sess.run(tf.global_variables_initializer())
# Sync params across processes
sess.run(sync_all_params())
# Setup model saving
logger.setup_tf_saver(sess, inputs={'x': x_ph}, outputs={'pi': pi, 'v': v})
# tf.reset_default_graph()
# restore_tf_graph(sess, '..//data//ppo//ppo_s0//simple_save')
def update():
inputs = {k: v for k, v in zip(all_phs, buf.get())}
inputs[pi_state_ph] = np.zeros((trials_per_epoch, NUM_GRU_UNITS))
inputs[v_state_ph] = np.zeros((trials_per_epoch, NUM_GRU_UNITS))
pi_l_old, v_l_old, ent = sess.run([pi_loss, v_loss, approx_ent],
feed_dict=inputs)
print(pi_l_old, v_l_old)
# Training
for i in range(train_pi_iters):
# print(f'pi:{i}')
_, kl = sess.run([train_pi, approx_kl], feed_dict=inputs)
# print(sess.run(pi_loss, feed_dict=inputs))
kl = mpi_avg(kl)
if kl > 1.5 * target_kl:
logger.log(
'Early stopping at step %d due to reaching max kl.' % i)
break
logger.store(StopIter=i)
for _ in range(train_v_iters):
# print(f'v:{_}')
sess.run(train_v, feed_dict=inputs)
# Log changes from update
import datetime
print(f'finish one batch training at {datetime.datetime.now()}')
pi_l_new, v_l_new, kl, cf = sess.run(
[pi_loss, v_loss, approx_kl, clipfrac], feed_dict=inputs)
logger.store(LossPi=pi_l_old,
LossV=v_l_old,
KL=kl,
Entropy=ent,
ClipFrac=cf,
DeltaLossPi=(pi_l_new - pi_l_old),
DeltaLossV=(v_l_new - v_l_old))
start_time = time.time()
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
# Main loop: collect experience in env and update/log each epoch
for epoch in range(epochs):
for trial in range(trials_per_epoch):
print(f'trial: {trial}')
old_a = np.array([0]).reshape(1, 1)
old_r = np.array([0]).reshape((1, 1, 1))
means = env.sample_tasks(1)[0]
action_dict = defaultdict(int)
for i in range(env.action_space.n):
action_dict[i] = 0
env.reset_task_simple(means)
task_avg = 0.0
pi_state_t = np.zeros((1, NUM_GRU_UNITS))
v_state_t = np.zeros((1, NUM_GRU_UNITS))
for step in range(steps_per_trial):
a, v_t, logp_t, pi_state_t, v_state_t = sess.run(
get_action_ops,
feed_dict={
x_ph: o.reshape(1, 1, -1),
a_ph: old_a,
rew_ph: old_r,
pi_state_ph: pi_state_t,
v_state_ph: v_state_t
})
# save and log
buf.store(o, a, r, v_t, logp_t)
logger.store(VVals=v_t)
try:
o, r, d, _ = env.step(a[0][0])
except:
print(a)
raise AssertionError
action_dict[a[0][0]] += 1
old_a = np.array(a).reshape(1, 1)
old_r = np.array([r]).reshape(1, 1, 1)
ep_ret += r
task_avg += r
ep_len += 1
terminal = d or (ep_len == max_ep_len)
if terminal or (step == local_steps_per_epoch - 1):
if not (terminal):
print(
'Warning: trajectory cut off by epoch at %d steps.'
% ep_len)
# if trajectory didn't reach terminal state, bootstrap value target
last_val = r if d else sess.run(
v, feed_dict={x_ph: o.reshape(1, -1)})
buf.finish_path(last_val)
if terminal:
# only save EpRet / EpLen if trajectory finished
logger.store(EpRet=ep_ret, EpLen=ep_len)
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
# logger.log_tabular('Epoch', epoch)
# logger.log_tabular('EpRet', with_min_and_max=True)
# logger.log_tabular('Means', means)
# logger.dump_tabular()
print(f'avg in trial {trial}: {task_avg / steps_per_trial}')
print(f'Means in trial {trial}: {means}')
print(action_dict)
# Save model
if (epoch % save_freq == 0) or (epoch == epochs - 1):
logger.save_state({'env': env}, None)
# saved_path = saver.save(sess, f"/tmp/model_epoch{epoch}.ckpt")
# print(f'Model saved in {saved_path}')
# Perform PPO update!
update()
logger.log_tabular('Epoch', epoch)
logger.log_tabular('EpRet', with_min_and_max=True)
logger.log_tabular('EpLen', average_only=True)
logger.log_tabular('VVals', with_min_and_max=True)
logger.log_tabular('TotalEnvInteracts', (epoch + 1) * steps_per_epoch)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('LossV', average_only=True)
logger.log_tabular('DeltaLossPi', average_only=True)
logger.log_tabular('DeltaLossV', average_only=True)
logger.log_tabular('Entropy', average_only=True)
logger.log_tabular('KL', average_only=True)
logger.log_tabular('ClipFrac', average_only=True)
logger.log_tabular('StopIter', average_only=True)
logger.log_tabular('Time', time.time() - start_time)
logger.dump_tabular()
def ppo(env_fn,
actor_critic=core.MLPActorCritic,
ac_kwargs=dict(),
seed=0,
steps_per_epoch=4000,
epochs=50,
gamma=0.99,
clip_ratio=0.2,
pi_lr=3e-4,
vf_lr=1e-3,
train_pi_iters=80,
train_v_iters=80,
lam=0.97,
max_ep_len=2000,
target_kl=0.01,
logger_kwargs=dict(),
save_freq=10):
global RENDER, BONUS
"""
Proximal Policy Optimization (by clipping),
with early stopping based on approximate KL
Args:
env_fn : A function which creates a copy of the environment.
The environment must satisfy the OpenAI Gym API.
actor_critic: The constructor method for a PyTorch Module with a
``step`` method, an ``act`` method, a ``pi`` module, and a ``v``
module. The ``step`` method should accept a batch of observations
and return:
=========== ================ ======================================
Symbol Shape Description
=========== ================ ======================================
``a`` (batch, act_dim) | Numpy array of actions for each
| observation.
``v`` (batch,) | Numpy array of value estimates
| for the provided observations.
``logp_a`` (batch,) | Numpy array of log probs for the
| actions in ``a``.
=========== ================ ======================================
The ``act`` method behaves the same as ``step`` but only returns ``a``.
The ``pi`` module's forward call should accept a batch of
observations and optionally a batch of actions, and return:
=========== ================ ======================================
Symbol Shape Description
=========== ================ ======================================
``pi`` N/A | Torch Distribution object, containing
| a batch of distributions describing
| the policy for the provided observations.
``logp_a`` (batch,) | Optional (only returned if batch of
| actions is given). Tensor containing
| the log probability, according to
| the policy, of the provided actions.
| If actions not given, will contain
| ``None``.
=========== ================ ======================================
The ``v`` module's forward call should accept a batch of observations
and return:
=========== ================ ======================================
Symbol Shape Description
=========== ================ ======================================
``v`` (batch,) | Tensor containing the value estimates
| for the provided observations. (Critical:
| make sure to flatten this!)
=========== ================ ======================================
ac_kwargs (dict): Any kwargs appropriate for the ActorCritic object
you provided to PPO.
seed (int): Seed for random number generators.
steps_per_epoch (int): Number of steps of interaction (state-action pairs)
for the agent and the environment in each epoch.
epochs (int): Number of epochs of interaction (equivalent to
number of policy updates) to perform.
gamma (float): Discount factor. (Always between 0 and 1.)
clip_ratio (float): Hyperparameter for clipping in the policy objective.
Roughly: how far can the new policy go from the old policy while
still profiting (improving the objective function)? The new policy
can still go farther than the clip_ratio says, but it doesn't help
on the objective anymore. (Usually small, 0.1 to 0.3.) Typically
denoted by :math:`\epsilon`.
pi_lr (float): Learning rate for policy optimizer.
vf_lr (float): Learning rate for value function optimizer.
train_pi_iters (int): Maximum number of gradient descent steps to take
on policy loss per epoch. (Early stopping may cause optimizer
to take fewer than this.)
train_v_iters (int): Number of gradient descent steps to take on
value function per epoch.
lam (float): Lambda for GAE-Lambda. (Always between 0 and 1,
close to 1.)
max_ep_len (int): Maximum length of trajectory / episode / rollout.
target_kl (float): Roughly what KL divergence we think is appropriate
between new and old policies after an update. This will get used
for early stopping. (Usually small, 0.01 or 0.05.)
logger_kwargs (dict): Keyword args for EpochLogger.
save_freq (int): How often (in terms of gap between epochs) to save
the current policy and value function.
"""
# Reachability Trainer
r_network = R_Network().to(device)
trainer = R_Network_Trainer(r_network=r_network, exp_name="random1")
episodic_memory = EpisodicMemory(embedding_shape=[EMBEDDING_DIM])
# Special function to avoid certain slowdowns from PyTorch + MPI combo.
setup_pytorch_for_mpi()
# Set up logger and save configuration
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
# Random seed
seed += 10000 * proc_id()
torch.manual_seed(seed)
np.random.seed(seed)
# Instantiate environment
env = env_fn()
observation_space = gym.spaces.Box(low=0.0, high=1.0, shape=(3, 64, 64))
action_space = gym.spaces.Discrete(3)
obs_dim = observation_space.shape
act_dim = action_space.shape
# Create actor-critic module
ac = actor_critic(observation_space, action_space, **ac_kwargs)
# Sync params across processes
sync_params(ac)
# Count variables
var_counts = tuple(core.count_vars(module) for module in [ac.pi, ac.v])
logger.log('\nNumber of parameters: \t pi: %d, \t v: %d\n' % var_counts)
# Set up experience buffer
local_steps_per_epoch = int(steps_per_epoch / num_procs())
buf = PPOBuffer(obs_dim, act_dim, local_steps_per_epoch, gamma, lam)
# Set up function for computing PPO policy loss
def compute_loss_pi(data):
obs, act, adv, logp_old = data['obs'], data['act'], data['adv'], data[
'logp']
# Policy loss
pi, logp = ac.pi(obs, act)
ratio = torch.exp(logp - logp_old)
clip_adv = torch.clamp(ratio, 1 - clip_ratio, 1 + clip_ratio) * adv
loss_pi = -(torch.min(ratio * adv, clip_adv)).mean()
# Useful extra info
approx_kl = (logp_old - logp).mean().item()
ent = pi.entropy().mean().item()
clipped = ratio.gt(1 + clip_ratio) | ratio.lt(1 - clip_ratio)
clipfrac = torch.as_tensor(clipped, dtype=torch.float32).mean().item()
pi_info = dict(kl=approx_kl, ent=ent, cf=clipfrac)
return loss_pi, pi_info
# Set up function for computing value loss
def compute_loss_v(data):
obs, ret = data['obs'], data['ret']
return ((ac.v(obs) - ret)**2).mean()
# Set up optimizers for policy and value function
pi_optimizer = Adam(ac.pi.parameters(), lr=pi_lr)
vf_optimizer = Adam(ac.v.parameters(), lr=vf_lr)
# Set up model saving
logger.setup_pytorch_saver(ac)
def update():
data = buf.get()
pi_l_old, pi_info_old = compute_loss_pi(data)
pi_l_old = pi_l_old.item()
v_l_old = compute_loss_v(data).item()
# Train policy with multiple steps of gradient descent
for i in range(train_pi_iters):
pi_optimizer.zero_grad()
loss_pi, pi_info = compute_loss_pi(data)
# Entropy bonus
loss_pi += pi_info['ent'] * 0.0021
kl = mpi_avg(pi_info['kl'])
if kl > 1.5 * target_kl:
logger.log(
'Early stopping at step %d due to reaching max kl.' % i)
break
loss_pi.backward()
mpi_avg_grads(ac.pi) # average grads across MPI processes
pi_optimizer.step()
logger.store(StopIter=i)
# Value function learning
for i in range(train_v_iters):
vf_optimizer.zero_grad()
loss_v = compute_loss_v(data)
loss_v.backward()
mpi_avg_grads(ac.v) # average grads across MPI processes
vf_optimizer.step()
# Log changes from update
kl, ent, cf = pi_info['kl'], pi_info_old['ent'], pi_info['cf']
logger.store(LossPi=pi_l_old,
LossV=v_l_old,
KL=kl,
Entropy=ent,
ClipFrac=cf,
DeltaLossPi=(loss_pi.item() - pi_l_old),
DeltaLossV=(loss_v.item() - v_l_old))
# Prepare for interaction with environment
start_time = time.time()
o, _ = env.reset()
env.render()
o = o.astype(np.float32) / 255.
o = o.transpose(2, 0, 1)
ep_ret, ep_len = 0, 0
indices = []
# Main loop: collect experience in env and update/log each epoch
for epoch in range(epochs):
for t in range(local_steps_per_epoch):
state = torch.as_tensor(o[np.newaxis, ...], dtype=torch.float32)
a, v, logp = ac.step(state)
next_o, r, d, info = env.step(a)
next_o = next_o.astype(np.float32) / 255.
d = ep_len == max_ep_len
trainer.store_new_state([next_o], [r], [d], [None])
r_network.eval()
with torch.no_grad():
state_embedding = r_network.embed_observation(
torch.FloatTensor([o]).to(device)).cpu().numpy()[0]
aggregated, _, _ = similarity_to_memory(
state_embedding, episodic_memory, r_network)
curiosity_bonus = 0.03 * (0.5 - aggregated)
if BONUS:
print(f'{curiosity_bonus:.3f}')
if curiosity_bonus > 0 or len(episodic_memory) == 0:
idx = episodic_memory.store_new_state(state_embedding)
x = int(env.map_scale * info['pose']['x'])
y = int(env.map_scale * info['pose']['y'])
if idx == len(indices):
indices.append((x, y))
else:
indices[idx] = (x, y)
r_network.train()
next_o = next_o.transpose(2, 0, 1)
ep_ret += r + curiosity_bonus
ep_len += 1
# save and log
buf.store(o, a, r, v, logp)
logger.store(VVals=v)
k = cv2.waitKey(1)
if k == ord('s'):
RENDER = 1 - RENDER
elif k == ord('b'):
BONUS = 1 - BONUS
if RENDER:
env.info['map'] = cv2.flip(env.info['map'], 0)
for index in indices:
cv2.circle(env.info['map'], index, 3, (0, 0, 255), -1)
env.info['map'] = cv2.flip(env.info['map'], 0)
env.render()
# Update obs (critical!)
o = next_o
timeout = ep_len == max_ep_len
terminal = d or timeout
epoch_ended = t == local_steps_per_epoch - 1
if terminal or epoch_ended:
if epoch_ended and not (terminal):
print('Warning: trajectory cut off by epoch at %d steps.' %
ep_len,
flush=True)
# if trajectory didn't reach terminal state, bootstrap value target
if timeout or epoch_ended:
state = torch.as_tensor(o[np.newaxis, ...],
dtype=torch.float32)
_, v, _ = ac.step(state)
else:
v = 0
buf.finish_path(v)
if terminal:
# only save EpRet / EpLen if trajectory finished
logger.store(EpRet=ep_ret, EpLen=ep_len)
print(ep_ret, ep_len, len(episodic_memory))
ep_ret, ep_len = 0, 0
o, _ = env.reset()
o = o.astype(np.float32) / 255.
o = o.transpose(2, 0, 1)
episodic_memory.reset()
indices = []
# Save model
if (epoch % save_freq == 0) or (epoch == epochs - 1):
logger.save_state({'env': env}, None)
# Perform PPO update!
if epoch > 4:
update()
# Log info about epoch
logger.log_tabular('Epoch', epoch)
logger.log_tabular('EpRet', with_min_and_max=True)
logger.log_tabular('EpLen', average_only=True)
logger.log_tabular('VVals', with_min_and_max=True)
logger.log_tabular('TotalEnvInteracts',
(epoch + 1) * steps_per_epoch)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('LossV', average_only=True)
logger.log_tabular('DeltaLossPi', average_only=True)
logger.log_tabular('DeltaLossV', average_only=True)
logger.log_tabular('Entropy', average_only=True)
logger.log_tabular('KL', average_only=True)
logger.log_tabular('ClipFrac', average_only=True)
logger.log_tabular('StopIter', average_only=True)
logger.log_tabular('Time', time.time() - start_time)
logger.dump_tabular()
else:
buf.get()
env.eval(path=args.eval_path, play_speed=args.play_speed)
sys.exit()
mpi_tools.mpi_fork(args.num_procs)
proc_id = mpi_tools.proc_id()
mpi_tools.setup_pytorch_for_mpi()
# Random seed
seed = 0
seed += 10000 * proc_id
torch.manual_seed(seed)
np.random.seed(seed)
env = envs.env_by_name(args.env)
sample_size = int(args.sample_size / mpi_tools.num_procs())
print('sample_size ', sample_size, flush=True)
buffer = PPOBuffer(obs_dim=env.obs_dim,
act_dim=env.act_dim,
buffer_size=sample_size)
actor_std = 0.2 * np.ones(env.act_dim, dtype=np.float32)
actor = nns.MLPGaussianActor(env.obs_dim, env.act_dim, args.hid,
torch.nn.ReLU, actor_std)
critic = nns.MLPCritic(env.obs_dim, args.hid, torch.nn.ReLU)
mpi_tools.sync_params(actor)
mpi_tools.sync_params(critic)
agent = rlPPO(actor,
critic,
lr_a=args.lr_a,
lr_c=args.lr_c,