def __init__(
self,
config_name,
record_samples,
actor_critic=MLPActorCritic,
ac_kwargs=dict(),
seed=0,
penalty_init=5e-3,
):
self.config_name = config_name
self.record_samples = record_samples
self.seed = seed
self.replay_buffer = None
self.actor_critic = actor_critic
self.ac_kwargs = ac_kwargs
self.penalty_init = penalty_init
self.penalty_init_param = np.log(
max(np.exp(self.penalty_init) - 1, 1e-8))
self.max_steps = 1000
# Paths
self._project_dir = '/home/tyna/Documents/openai/research-project/'
self._root_data_path = self._project_dir + 'data/'
self._expert_path = self._project_dir + 'expert_data/'
self._clone_path = self._project_dir + 'clone_data/'
self._demo_dir = os.path.join(self._expert_path,
self.config_name + '_episodes/')
self.file_name = 'ppo_penalized_' + self.config_name + '_128x4'
# self.benchmark_project_name = 'clone_benchmarking_' + self.config_name
# Special function to avoid certain slowdowns from PyTorch + MPI combo.
setup_pytorch_for_mpi()
# Random seed # seed = 0
self.seed += 10000 * proc_id()
torch.manual_seed(self.seed)
np.random.seed(self.seed)
def ppo_penalized(
env_fn,
actor_critic=ActorCritic,
ac_kwargs=dict(),
seed=0,
episodes_per_epoch=40,
epochs=500,
gamma=0.99,
lam=0.98,
pi_lr=3e-4,
vf_lr=1e-3,
train_v_iters=80,
train_pi_iters=1, ## NOTE: Incredibly Important That This Be Low For Penalized Learning
max_ep_len=1000,
logger_kwargs=dict(),
clip_ratio=0.2, # tuned????
# Cost constraints / penalties:
cost_lim=25,
penalty_init=1.,
penalty_lr=5e-3,
config_name='standard',
save_freq=10):
# W&B Logging
wandb.login()
composite_name = 'new_ppo_penalized_' + config_name
wandb.init(project="LearningCurves",
group="PPO Expert",
name=composite_name)
# 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())
seed += 10000 * proc_id()
torch.manual_seed(seed)
np.random.seed(seed)
# Instantiate environment
env = env_fn()
obs_dim = env.observation_space.shape
act_dim = env.action_space.shape
ac_kwargs['action_space'] = env.action_space
# Models
# Create actor-critic module and monitor it
ac = actor_critic(input_dim=obs_dim[0], **ac_kwargs)
# Set up model saving
logger.setup_pytorch_saver(ac)
# Sync params across processes
sync_params(ac)
# Buffers
local_episodes_per_epoch = int(episodes_per_epoch / num_procs())
buf = BufferActor(obs_dim[0], act_dim[0], local_episodes_per_epoch,
max_ep_len)
# Count variables
var_counts = tuple(count_vars(module) for module in [ac.pi, ac.v])
logger.log('\nNumber of parameters: \t pi: %d, \t v: %d\n' % var_counts)
# Optimizers
pi_optimizer = torch.optim.Adam(ac.pi.parameters(), lr=pi_lr)
vf_optimizer = torch.optim.Adam(ac.v.parameters(), lr=vf_lr)
# pi_optimizer = AdaBelief(ac.pi.parameters(), betas=(0.9, 0.999), eps=1e-8)
# vf_optimizer = AdaBelief(ac.v.parameters(), betas=(0.9, 0.999), eps=1e-8)
# # Parameters Sync
# sync_all_params(ac.parameters())
# Set up function for computing PPO policy loss
def compute_loss_pi(obs, act, adv, logp_old):
# Policy loss # policy gradient term + entropy term
# Policy loss with clipping (without clipping, loss_pi = -(logp*adv).mean()).
# TODO: Think about removing clipping
_, 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()
return loss_pi
def penalty_update(cur_penalty):
cur_cost = logger.get_stats('EpCost')[0]
cur_rew = logger.get_stats('EpRet')[0]
# Penalty update
cur_penalty = max(0, cur_penalty + penalty_lr * (cur_cost - cost_lim))
return cur_penalty
def update(e):
obs, act, adv, ret, logp_old = [
torch.Tensor(x) for x in buf.retrieve_all()
]
# Policy
_, logp, _ = ac.pi(obs, act)
entropy = (-logp).mean()
# Train policy with multiple steps of gradient descent
for _ in range(train_pi_iters):
pi_optimizer.zero_grad()
loss_pi = compute_loss_pi(obs, act, adv, ret)
loss_pi.backward()
# average_gradients(train_pi.param_groups)
# mpi_avg_grads(pi_optimizer.param_groups)
mpi_avg_grads(ac.pi)
pi_optimizer.step()
# Value function training
v = ac.v(obs)
v_l_old = F.mse_loss(v, ret) # old loss
for _ in range(train_v_iters):
v = ac.v(obs)
v_loss = F.mse_loss(
v, ret) # how well did our value function predict loss?
# Value function train
vf_optimizer.zero_grad()
v_loss.backward()
# average_gradients(vf_optimizer.param_groups)
mpi_avg_grads(ac.v) # average gradients across MPI processes
vf_optimizer.step()
# Log the changes
_, logp, _, v = ac(obs, act)
# entropy_new = (-logp).mean()
pi_loss_new = -(logp * adv).mean()
v_loss_new = F.mse_loss(v, ret)
kl = (logp_old - logp).mean()
logger.store(LossPi=loss_pi,
LossV=v_l_old,
DeltaLossPi=(pi_loss_new - loss_pi),
DeltaLossV=(v_loss_new - v_l_old),
Entropy=entropy,
KL=kl)
# Prepare for interaction with the environment
start_time = time.time()
o, r, d, ep_ret, ep_cost, ep_len = env.reset(), 0, False, 0, 0, 0
total_t = 0
# Initialize penalty
cur_penalty = np.log(max(np.exp(penalty_init) - 1, 1e-8))
for epoch in range(epochs):
ac.eval() # eval mode
# Policy rollout
for _ in range(local_episodes_per_epoch):
for _ in range(max_ep_len):
# obs =
a, _, lopg_t, v_t = ac(torch.Tensor(o.reshape(1, -1)))
logger.store(VVals=v_t)
o, r, d, info = env.step(a.detach().numpy()[0])
c = info.get('cost', 0) # Include penalty on cost
r_total = r - cur_penalty * c
r_total /= (1 + cur_penalty)
# store
buf.store(o,
a.detach().numpy(), r_total, v_t.item(),
lopg_t.detach().numpy())
ep_ret += r
ep_cost += c
ep_len += 1
total_t += 1
terminal = d or (ep_len == max_ep_len)
if terminal:
# buf.end_episode()
buf.finish_path()
logger.store(EpRet=ep_ret, EpCost=ep_cost, EpLen=ep_len)
print("end of episode return: ", ep_ret)
episode_metrics = {
'average ep ret': ep_ret,
'average ep cost': ep_cost
}
wandb.log(episode_metrics)
o, r, d, ep_ret, ep_cost, ep_len = env.reset(
), 0, False, 0, 0, 0
if (epoch % save_freq == 0) or (epoch == epochs - 1):
# logger._torch_save(ac, fname="expert_torch_save.pt")
# logger._torch_save(ac, fname="model.pt")
logger.save_state({'env': env}, None, None)
# Update
ac.train()
# update penalty
cur_penalty = penalty_update(cur_penalty)
# update networks
update(epoch)
# Log
logger.log_tabular('Epoch', epoch)
# logger.log_tabular('EpRet', with_min_and_max=True)
logger.log_tabular('EpRet', average_only=True)
logger.log_tabular('EpCost', average_only=True)
logger.log_tabular('EpLen', average_only=True)
logger.log_tabular('VVals', average_only=True)
# logger.log_tabular('VVals', with_min_and_max=True)
logger.log_tabular('TotalEnvInteracts', total_t)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('DeltaLossPi', average_only=True)
logger.log_tabular('LossV', 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('Time', time.time() - start_time)
logger.dump_tabular()
wandb.finish()
parser.add_argument('--steps', type=int, default=4000)
parser.add_argument('--epochs', type=int, default=500)
parser.add_argument('--clip_ratio', type=float, default=0.2)
parser.add_argument('--actor_lr', type=float, default=3e-4)
parser.add_argument('--critic_lr', type=float, default=1e-3)
parser.add_argument('--train_pi_iters', type=int, default=80)
parser.add_argument('--train_v_iters', type=int, default=80)
parser.add_argument('--lam', type=float, default=0.97)
parser.add_argument('--max_ep_len', type=int, default=1000)
parser.add_argument('--target_kl', type=float, default=0.01)
parser.add_argument('--save_freq', type=int, default=10)
args = parser.parse_args()
mpi_fork(args.cpu)
seed = 10000 * proc_id()
tf.set_random_seed(seed)
np.random.seed(seed)
env = gym.make(args.env)
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.shape[0]
# Inputs to computation graph
x_ph = tf.placeholder(dtype=tf.float32, shape=(None, obs_dim))
a_ph = tf.placeholder(dtype=tf.float32, shape=(None, act_dim))
adv_ph = tf.placeholder(dtype=tf.float32, shape=(None, ))
ret_ph = tf.placeholder(dtype=tf.float32, shape=(None, ))
logp_old_ph = tf.placeholder(dtype=tf.float32, shape=(None, ))
# Main outputs from computation graph
def valor(env_fn,
actor_critic=ActorCritic,
ac_kwargs=dict(),
disc=Discriminator,
dc_kwargs=dict(),
seed=0,
episodes_per_epoch=40,
epochs=50,
gamma=0.99,
pi_lr=3e-4,
vf_lr=1e-3,
dc_lr=5e-4,
train_v_iters=80,
train_dc_iters=10,
train_dc_interv=10,
lam=0.97,
max_ep_len=1000,
logger_kwargs=dict(),
con_dim=5,
save_freq=10,
k=1):
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
seed += 10000 * proc_id()
torch.manual_seed(seed)
np.random.seed(seed)
env = env_fn()
obs_dim = env.observation_space.shape
act_dim = env.action_space.shape
ac_kwargs['action_space'] = env.action_space
# Model
ac = actor_critic(input_dim=obs_dim[0] + con_dim, **ac_kwargs)
disc = disc(input_dim=obs_dim[0], context_dim=con_dim, **dc_kwargs)
# Set up model saving
logger.setup_pytorch_saver([ac, disc])
# Buffer
local_episodes_per_epoch = int(episodes_per_epoch / num_procs())
buffer = VALORBuffer(con_dim, obs_dim[0], act_dim[0],
local_episodes_per_epoch, max_ep_len, train_dc_interv)
# Count variables
var_counts = tuple(
count_vars(module) for module in [ac.policy, ac.value_f, disc.policy])
logger.log('\nNumber of parameters: \t pi: %d, \t v: %d, \t d: %d\n' %
var_counts)
# Optimizers
train_pi = torch.optim.Adam(ac.policy.parameters(), lr=pi_lr)
train_v = torch.optim.Adam(ac.value_f.parameters(), lr=vf_lr)
train_dc = torch.optim.Adam(disc.policy.parameters(), lr=dc_lr)
# Parameters Sync
sync_all_params(ac.parameters())
sync_all_params(disc.parameters())
def update(e):
obs, act, adv, pos, ret, logp_old = [
torch.Tensor(x) for x in buffer.retrieve_all()
]
# Policy
_, logp, _ = ac.policy(obs, act)
entropy = (-logp).mean()
# Policy loss
pi_loss = -(logp * (k * adv + pos)).mean()
# Train policy
train_pi.zero_grad()
pi_loss.backward()
average_gradients(train_pi.param_groups)
train_pi.step()
# Value function
v = ac.value_f(obs)
v_l_old = F.mse_loss(v, ret)
for _ in range(train_v_iters):
v = ac.value_f(obs)
v_loss = F.mse_loss(v, ret)
# Value function train
train_v.zero_grad()
v_loss.backward()
average_gradients(train_v.param_groups)
train_v.step()
# Discriminator
if (e + 1) % train_dc_interv == 0:
print('Discriminator Update!')
con, s_diff = [torch.Tensor(x) for x in buffer.retrieve_dc_buff()]
_, logp_dc, _ = disc(s_diff, con)
d_l_old = -logp_dc.mean()
# Discriminator train
for _ in range(train_dc_iters):
_, logp_dc, _ = disc(s_diff, con)
d_loss = -logp_dc.mean()
train_dc.zero_grad()
d_loss.backward()
average_gradients(train_dc.param_groups)
train_dc.step()
_, logp_dc, _ = disc(s_diff, con)
dc_l_new = -logp_dc.mean()
else:
d_l_old = 0
dc_l_new = 0
# Log the changes
_, logp, _, v = ac(obs, act)
pi_l_new = -(logp * (k * adv + pos)).mean()
v_l_new = F.mse_loss(v, ret)
kl = (logp_old - logp).mean()
logger.store(LossPi=pi_loss,
LossV=v_l_old,
KL=kl,
Entropy=entropy,
DeltaLossPi=(pi_l_new - pi_loss),
DeltaLossV=(v_l_new - v_l_old),
LossDC=d_l_old,
DeltaLossDC=(dc_l_new - d_l_old))
# logger.store(Adv=adv.reshape(-1).numpy().tolist(), Pos=pos.reshape(-1).numpy().tolist())
start_time = time.time()
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
context_dist = Categorical(logits=torch.Tensor(np.ones(con_dim)))
total_t = 0
for epoch in range(epochs):
ac.eval()
disc.eval()
for _ in range(local_episodes_per_epoch):
c = context_dist.sample()
c_onehot = F.one_hot(c, con_dim).squeeze().float()
for _ in range(max_ep_len):
concat_obs = torch.cat(
[torch.Tensor(o.reshape(1, -1)),
c_onehot.reshape(1, -1)], 1)
a, _, logp_t, v_t = ac(concat_obs)
buffer.store(c,
concat_obs.squeeze().detach().numpy(),
a.detach().numpy(), r, v_t.item(),
logp_t.detach().numpy())
logger.store(VVals=v_t)
o, r, d, _ = env.step(a.detach().numpy()[0])
ep_ret += r
ep_len += 1
total_t += 1
terminal = d or (ep_len == max_ep_len)
if terminal:
dc_diff = torch.Tensor(buffer.calc_diff()).unsqueeze(0)
con = torch.Tensor([float(c)]).unsqueeze(0)
_, _, log_p = disc(dc_diff, con)
buffer.end_episode(log_p.detach().numpy())
logger.store(EpRet=ep_ret, EpLen=ep_len)
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
if (epoch % save_freq == 0) or (epoch == epochs - 1):
logger.save_state({'env': env}, [ac, disc], None)
# Update
ac.train()
disc.train()
update(epoch)
# Log
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', total_t)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('DeltaLossPi', average_only=True)
logger.log_tabular('LossV', average_only=True)
logger.log_tabular('DeltaLossV', average_only=True)
logger.log_tabular('LossDC', average_only=True)
logger.log_tabular('DeltaLossDC', average_only=True)
logger.log_tabular('Entropy', average_only=True)
logger.log_tabular('KL', average_only=True)
logger.log_tabular('Time', time.time() - start_time)
logger.dump_tabular()
def policyg(env_fn,
actor_critic=ActorCritic,
ac_kwargs=dict(),
seed=0,
episodes_per_epoch=40,
epochs=500,
gamma=0.99,
lam=0.97,
pi_lr=3e-4,
vf_lr=1e-3,
train_v_iters=80,
max_ep_len=1000,
logger_kwargs=dict(),
save_freq=10):
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
seed += 10000 * proc_id()
torch.manual_seed(seed)
np.random.seed(seed)
env = env_fn()
obs_dim = env.observation_space.shape
act_dim = env.action_space.shape
ac_kwargs['action_space'] = env.action_space
# Models
ac = actor_critic(input_dim=obs_dim[0], **ac_kwargs)
# Set up model saving
logger.setup_pytorch_saver(ac)
# Buffers
local_episodes_per_epoch = int(episodes_per_epoch / num_procs())
buf = BufferActor(obs_dim[0], act_dim[0], local_episodes_per_epoch,
max_ep_len)
# Count variables
var_counts = tuple(
count_vars(module) for module in [ac.policy, ac.value_f])
print("POLICY GRADIENT")
logger.log('\nNumber of parameters: \t pi: %d, \t v: %d\n' % var_counts)
# Optimizers
train_pi = torch.optim.Adam(ac.policy.parameters(), lr=pi_lr)
train_v = torch.optim.Adam(ac.value_f.parameters(), lr=vf_lr)
# Parameters Sync
sync_all_params(ac.parameters())
def update(e):
obs, act, adv, ret, lgp_old = [
torch.Tensor(x) for x in buf.retrieve_all()
]
# Policy
_, lgp, _ = ac.policy(obs, act)
entropy = (-lgp).mean()
# Policy loss # policy gradient term + entropy term
pi_loss = -(lgp * adv).mean()
# Train policy
train_pi.zero_grad()
pi_loss.backward()
average_gradients(train_pi.param_groups)
train_pi.step()
# Value function
v = ac.value_f(obs)
v_l_old = F.mse_loss(v, ret)
for _ in range(train_v_iters):
v = ac.value_f(obs)
v_loss = F.mse_loss(v, ret)
# Value function train
train_v.zero_grad()
v_loss.backward()
average_gradients(train_v.param_groups)
train_v.step()
# Log the changes
_, lgp, _, v = ac(obs, act)
entropy_new = (-lgp).mean()
pi_loss_new = -(lgp * adv).mean()
v_loss_new = F.mse_loss(v, ret)
kl = (lgp_old - lgp).mean()
logger.store(LossPi=pi_loss,
LossV=v_l_old,
DeltaLossPi=(pi_loss_new - pi_loss),
DeltaLossV=(v_loss_new - v_l_old),
Entropy=entropy,
KL=kl)
start_time = time.time()
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
total_t = 0
for epoch in range(epochs):
ac.eval()
# Policy rollout
for _ in range(local_episodes_per_epoch):
for _ in range(max_ep_len):
obs = torch.Tensor(o.reshape(1, -1))
a, _, lopg_t, v_t = ac(obs)
buf.store(o,
a.detach().numpy(), r, v_t.item(),
lopg_t.detach().numpy())
logger.store(VVals=v_t)
o, r, d, _ = env.step(a.detach().numpy()[0])
ep_ret += r
ep_len += 1
total_t += 1
terminal = d or (ep_len == max_ep_len)
if terminal:
buf.end_episode()
logger.store(EpRet=ep_ret, EpLen=ep_len)
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
if (epoch % save_freq == 0) or (epoch == epochs - 1):
# logger._torch_save(ac, fname="expert_torch_save.pt")
# logger._torch_save(ac, fname="model.pt")
logger.save_state({'env': env}, None, None)
# Update
ac.train()
update(epoch)
# Log
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', total_t)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('DeltaLossPi', average_only=True)
logger.log_tabular('LossV', 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('Time', time.time() - start_time)
logger.dump_tabular()
def gail(env_fn,
actor_critic=ActorCritic,
ac_kwargs=dict(),
disc=Discriminator,
dc_kwargs=dict(),
seed=0,
episodes_per_epoch=40,
epochs=500,
gamma=0.99,
lam=0.97,
pi_lr=3e-3,
vf_lr=3e-3,
dc_lr=5e-4,
train_v_iters=80,
train_dc_iters=80,
max_ep_len=1000,
logger_kwargs=dict(),
save_freq=10):
l_lam = 0 # balance two loss term
print("starting now")
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
seed += 10000 * proc_id()
torch.manual_seed(seed)
np.random.seed(seed)
env = env_fn()
obs_dim = env.observation_space.shape
act_dim = env.action_space.shape
ac_kwargs['action_space'] = env.action_space
# Models
ac = actor_critic(input_dim=obs_dim[0], **ac_kwargs)
disc = disc(input_dim=obs_dim[0], **dc_kwargs)
# Set up model saving
logger.setup_pytorch_saver([ac, disc])
# TODO: Load expert policy here
expert = actor_critic(input_dim=obs_dim[0], **ac_kwargs)
# expert_name = "expert_torch_save.pt"
expert_name = "model.pt"
# expert = torch.load(osp.join(logger_kwargs['output_dir'],'pyt_save' , expert_name))
expert = torch.load(
'/home/tyna/Documents/openai/research-project/data/anonymous-expert/anonymous-expert_s0/pyt_save/model.pt'
)
print('RUNNING GAIL')
# Buffers
local_episodes_per_epoch = int(episodes_per_epoch / num_procs())
buff_s = BufferS(obs_dim[0], act_dim[0], local_episodes_per_epoch,
max_ep_len)
buff_t = BufferT(obs_dim[0], act_dim[0], local_episodes_per_epoch,
max_ep_len)
# Count variables
var_counts = tuple(
count_vars(module) for module in [ac.policy, ac.value_f, disc.policy])
print("GAIL")
logger.log('\nNumber of parameters: \t pi: %d, \t v: %d, \t d: %d\n' %
var_counts)
# Optimizers
train_pi = torch.optim.Adam(ac.policy.parameters(), lr=pi_lr)
train_v = torch.optim.Adam(ac.value_f.parameters(), lr=vf_lr)
train_dc = torch.optim.Adam(disc.policy.parameters(), lr=dc_lr)
# Parameters Sync
sync_all_params(ac.parameters())
sync_all_params(disc.parameters())
def update(e):
obs_s, act, adv, ret, lgp_old = [
torch.Tensor(x) for x in buff_s.retrieve_all()
]
obs_t, _ = [torch.Tensor(x) for x in buff_t.retrieve_all()]
# Policy
_, lgp, _ = ac.policy(obs_s, act)
entropy = (-lgp).mean()
# Policy loss
# policy gradient term + entropy term
pi_loss = -(lgp * adv).mean() - l_lam * entropy
# Train policy
if e > 10:
train_pi.zero_grad()
pi_loss.backward()
average_gradients(train_pi.param_groups)
train_pi.step()
# Value function
v = ac.value_f(obs_s)
v_l_old = F.mse_loss(v, ret)
for _ in range(train_v_iters):
v = ac.value_f(obs_s)
v_loss = F.mse_loss(v, ret)
# Value function train
train_v.zero_grad()
v_loss.backward()
average_gradients(train_v.param_groups)
train_v.step()
# Discriminator
gt1 = torch.ones(obs_s.size()[0], dtype=torch.int)
gt2 = torch.zeros(obs_t.size()[0], dtype=torch.int)
_, lgp_s, _ = disc(obs_s, gt=gt1)
_, lgp_t, _ = disc(obs_t, gt=gt2)
dc_loss_old = -lgp_s.mean() - lgp_t.mean()
for _ in range(train_dc_iters):
_, lgp_s, _ = disc(obs_s, gt=gt1)
_, lgp_t, _ = disc(obs_t, gt=gt2)
dc_loss = -lgp_s.mean() - lgp_t.mean()
# Discriminator train
train_dc.zero_grad()
dc_loss.backward()
average_gradients(train_dc.param_groups)
train_dc.step()
_, lgp_s, _ = disc(obs_s, gt=gt1)
_, lgp_t, _ = disc(obs_t, gt=gt2)
dc_loss_new = -lgp_s.mean() - lgp_t.mean()
# Log the changes
_, lgp, _, v = ac(obs, act)
entropy_new = (-lgp).mean()
pi_loss_new = -(lgp * adv).mean() - l_lam * entropy
v_loss_new = F.mse_loss(v, ret)
kl = (lgp_old - lgp).mean()
logger.store(LossPi=pi_loss,
LossV=v_l_old,
LossDC=dc_loss_old,
DeltaLossPi=(pi_loss_new - pi_loss),
DeltaLossV=(v_loss_new - v_l_old),
DeltaLossDC=(dc_loss_new - dc_loss_old),
DeltaEnt=(entropy_new - entropy),
Entropy=entropy,
KL=kl)
start_time = time.time()
o, r, sdr, d, ep_ret, ep_sdr, ep_len = env.reset(), 0, 0, False, 0, 0, 0
total_t = 0
ep_len_t = 0
for epoch in range(epochs):
ac.eval()
disc.eval()
# We recognize the probability term of index [0] correspond to the teacher's policy
# Student's policy rollout
for _ in range(local_episodes_per_epoch):
for _ in range(max_ep_len):
obs = torch.Tensor(o.reshape(1, -1))
a, _, lopg_t, v_t = ac(obs)
buff_s.store(o,
a.detach().numpy(), r, sdr, v_t.item(),
lopg_t.detach().numpy())
logger.store(VVals=v_t)
o, r, d, _ = env.step(a.detach().numpy()[0])
_, sdr, _ = disc(torch.Tensor(o.reshape(1, -1)),
gt=torch.Tensor([0]))
if sdr < -4: # Truncate rewards
sdr = -4
ep_ret += r
ep_sdr += sdr
ep_len += 1
total_t += 1
terminal = d or (ep_len == max_ep_len)
if terminal:
buff_s.end_episode()
logger.store(EpRetS=ep_ret, EpLenS=ep_len, EpSdrS=ep_sdr)
print("Student Episode Return: \t", ep_ret)
o, r, sdr, d, ep_ret, ep_sdr, ep_len = env.reset(
), 0, 0, False, 0, 0, 0
# Teacher's policy rollout
for _ in range(local_episodes_per_epoch):
for _ in range(max_ep_len):
obs = torch.Tensor(o.reshape(1, -1))
a, _, _, _ = expert(obs)
buff_t.store(o, a.detach().numpy(), r)
o, r, d, _ = env.step(a.detach().numpy()[0])
ep_ret += r
ep_len += 1
total_t += 1
terminal = d or (ep_len == max_ep_len)
if terminal:
buff_t.end_episode()
logger.store(EpRetT=ep_ret, EpLenT=ep_len)
print("Teacher Episode Return: \t", ep_ret)
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
if (epoch % save_freq == 0) or (epoch == epochs - 1):
logger.save_state({'env': env}, [ac, disc], None)
# Update
ac.train()
disc.train()
update(epoch)
# Log
logger.log_tabular('Epoch', epoch)
logger.log_tabular('EpRetS', with_min_and_max=True)
logger.log_tabular('EpSdrS', with_min_and_max=True)
logger.log_tabular('EpLenS', average_only=True)
logger.log_tabular('EpRetT', with_min_and_max=True)
logger.log_tabular('EpLenT', average_only=True)
logger.log_tabular('VVals', with_min_and_max=True)
logger.log_tabular('TotalEnvInteracts', total_t)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('DeltaLossPi', average_only=True)
logger.log_tabular('LossV', average_only=True)
logger.log_tabular('DeltaLossV', average_only=True)
logger.log_tabular('LossDC', average_only=True)
logger.log_tabular('DeltaLossDC', average_only=True)
logger.log_tabular('Entropy', average_only=True)
logger.log_tabular('DeltaEnt', average_only=True)
logger.log_tabular('KL', average_only=True)
logger.log_tabular('Time', time.time() - start_time)
logger.dump_tabular()
def gail_penalized(
env_fn,
actor_critic=ActorCritic,
ac_kwargs=dict(),
disc=Discriminator,
dc_kwargs=dict(),
seed=0,
episodes_per_epoch=40,
epochs=500,
gamma=0.99,
lam=0.97,
# Cost constraints / penalties:
cost_lim=25,
penalty_init=1.,
penalty_lr=5e-3,
clip_ratio=0.2,
pi_lr=3e-3,
vf_lr=3e-3,
dc_lr=5e-4,
train_v_iters=80,
train_pi_iters=80,
train_dc_iters=80,
max_ep_len=1000,
logger_kwargs=dict(),
config_name='standard',
save_freq=10):
# W&B Logging
wandb.login()
composite_name = 'new_gail_penalized_' + config_name
wandb.init(project="LearningCurves",
group="GAIL Clone",
name=composite_name)
# Special function to avoid certain slowdowns from PyTorch + MPI combo.
setup_pytorch_for_mpi()
l_lam = 0 # balance two loss terms
# Set up logger and save configuration
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
seed += 10000 * proc_id()
torch.manual_seed(seed)
np.random.seed(seed)
# Instantiate environment
env = env_fn()
obs_dim = env.observation_space.shape
act_dim = env.action_space.shape
ac_kwargs['action_space'] = env.action_space
# Models # Create actor-critic and discriminator modules
ac = actor_critic(input_dim=obs_dim[0], **ac_kwargs)
discrim = disc(input_dim=obs_dim[0], **dc_kwargs)
# Set up model saving
logger.setup_pytorch_saver([ac, discrim])
# Sync params across processes
sync_params(ac)
sync_params(discrim)
# Load expert policy here
expert = actor_critic(input_dim=obs_dim[0], **ac_kwargs)
# expert_name = "expert_torch_save.pt"
expert_name = "model.pt"
# expert = torch.load(osp.join(logger_kwargs['output_dir'],'pyt_save' , expert_name))
# expert = torch.load('/home/tyna/Documents/openai/research-project/data/anonymous-expert/anonymous-expert_s0/pyt_save/model.pt')
expert = torch.load(
'/home/tyna/Documents/openai/research-project/data/test-pen-ppo/test-pen-ppo_s0/pyt_save/model.pt'
)
print('RUNNING GAIL')
# Buffers
local_episodes_per_epoch = int(episodes_per_epoch / num_procs())
buff_s = BufferStudent(obs_dim[0], act_dim[0], local_episodes_per_epoch,
max_ep_len)
buff_t = BufferTeacher(obs_dim[0], act_dim[0], local_episodes_per_epoch,
max_ep_len)
# Count variables
var_counts = tuple(
count_vars(module) for module in [ac.pi, ac.v, discrim.pi])
logger.log('\nNumber of parameters: \t pi: %d, \t v: %d, \t d: %d\n' %
var_counts)
# Optimizers
pi_optimizer = torch.optim.Adam(ac.pi.parameters(), lr=pi_lr)
vf_optimizer = torch.optim.Adam(ac.v.parameters(), lr=vf_lr)
discrim_optimizer = torch.optim.Adam(discrim.pi.parameters(), lr=dc_lr)
# # Parameters Sync
# sync_all_params(ac.parameters())
# sync_all_params(disc.parameters())
# Set up function for computing PPO policy loss
def compute_loss_pi(obs, act, adv, logp_old):
# Policy loss # policy gradient term + entropy term
# Policy loss with clipping (without clipping, loss_pi = -(logp*adv).mean()).
# TODO: Think about removing clipping
_, 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()
return loss_pi
def penalty_update(cur_penalty):
cur_cost = logger.get_stats('EpCostS')[0]
cur_rew = logger.get_stats('EpRetS')[0]
# Penalty update
cur_penalty = max(0, cur_penalty + penalty_lr * (cur_cost - cost_lim))
return cur_penalty
def update(e):
obs_s, act, adv, ret, log_pi_old = [
torch.Tensor(x) for x in buff_s.retrieve_all()
]
obs_t, _ = [torch.Tensor(x) for x in buff_t.retrieve_all()]
# Policy
_, logp, _ = ac.pi(obs_s, act)
entropy = (-logp).mean()
# Policy loss # policy gradient term + entropy term
# loss_pi = -(logp * adv).mean() - l_lam * entropy
# Train policy
if e > 10:
# Train policy with multiple steps of gradient descent
for _ in range(train_pi_iters):
pi_optimizer.zero_grad()
loss_pi = compute_loss_pi(obs, act, adv, ret)
loss_pi.backward()
mpi_avg_grads(ac.pi)
pi_optimizer.step()
# Value function
v = ac.v(obs_s)
v_l_old = F.mse_loss(v, ret)
for _ in range(train_v_iters):
v = ac.v(obs_s)
v_loss = F.mse_loss(v, ret)
# Value function train
vf_optimizer.zero_grad()
v_loss.backward()
mpi_avg_grads(ac.v) # average gradients across MPI processes
vf_optimizer.step()
# Discriminator
gt1 = torch.ones(obs_s.size()[0], dtype=torch.int)
gt2 = torch.zeros(obs_t.size()[0], dtype=torch.int)
_, logp_student, _ = discrim(obs_s, gt=gt1)
_, logp_teacher, _ = discrim(obs_t, gt=gt2)
discrim_l_old = -logp_student.mean() - logp_teacher.mean()
for _ in range(train_dc_iters):
_, logp_student, _ = discrim(obs_s, gt=gt1)
_, logp_teacher, _ = discrim(obs_t, gt=gt2)
dc_loss = -logp_student.mean() - logp_teacher.mean()
# Discriminator train
discrim_optimizer.zero_grad()
dc_loss.backward()
# average_gradients(discrim_optimizer.param_groups)
mpi_avg_grads(discrim.pi)
discrim_optimizer.step()
_, logp_student, _ = discrim(obs_s, gt=gt1)
_, logp_teacher, _ = discrim(obs_t, gt=gt2)
dc_loss_new = -logp_student.mean() - logp_teacher.mean()
# Log the changes
_, logp, _, v = ac(obs, act)
entropy_new = (-logp).mean()
pi_loss_new = -(logp * adv).mean() - l_lam * entropy
v_loss_new = F.mse_loss(v, ret)
kl = (log_pi_old - logp).mean()
logger.store(
# LossPi=loss_pi,
LossV=v_l_old,
LossDC=discrim_l_old,
# DeltaLossPi=(pi_loss_new - loss_pi),
DeltaLossV=(v_loss_new - v_l_old),
DeltaLossDC=(dc_loss_new - discrim_l_old),
DeltaEnt=(entropy_new - entropy),
Entropy=entropy,
KL=kl)
start_time = time.time()
o, r, sdr, d, ep_ret, ep_cost, ep_sdr, ep_len = env.reset(
), 0, 0, False, 0, 0, 0, 0
total_t = 0
ep_len_t = 0
# Initialize penalty
cur_penalty = np.log(max(np.exp(penalty_init) - 1, 1e-8))
for epoch in range(epochs):
ac.eval()
discrim.eval()
# We recognize the probability term of index [0] correspond to the teacher's policy
# Student's policy rollout
for _ in range(local_episodes_per_epoch):
for _ in range(max_ep_len):
obs = torch.Tensor(o.reshape(1, -1))
a, _, lopg_t, v_t = ac(obs)
buff_s.store(o,
a.detach().numpy(), r, sdr, v_t.item(),
lopg_t.detach().numpy())
logger.store(VVals=v_t)
o, r, d, info = env.step(a.detach().numpy()[0])
# print("INFO: ", info)
c = info.get("cost")
_, sdr, _ = discrim(torch.Tensor(o.reshape(1, -1)),
gt=torch.Tensor([0]))
if sdr < -4: # Truncate rewards
sdr = -4
ep_ret += r
ep_cost += c
ep_sdr += sdr
ep_len += 1
total_t += 1
terminal = d or (ep_len == max_ep_len)
if terminal:
buff_s.end_episode()
logger.store(EpRetS=ep_ret,
EpCostS=ep_cost,
EpLenS=ep_len,
EpSdrS=ep_sdr)
print("Student Episode Return: \t", ep_ret)
o, r, sdr, d, ep_ret, ep_cost, ep_sdr, ep_len = env.reset(
), 0, 0, False, 0, 0, 0, 0
# Teacher's policy rollout
for _ in range(local_episodes_per_epoch):
for _ in range(max_ep_len):
# obs =
a, _, _, _ = expert(torch.Tensor(o.reshape(1, -1)))
buff_t.store(o, a.detach().numpy(), r)
o, r, d, info = env.step(a.detach().numpy()[0])
c = info.get("cost")
ep_ret += r
ep_cost += c
ep_len += 1
total_t += 1
terminal = d or (ep_len == max_ep_len)
if terminal:
buff_t.end_episode()
logger.store(EpRetT=ep_ret, EpCostT=ep_cost, EpLenT=ep_len)
print("Teacher Episode Return: \t", ep_ret)
o, r, d, ep_ret, ep_cost, ep_len = env.reset(
), 0, False, 0, 0, 0
if (epoch % save_freq == 0) or (epoch == epochs - 1):
logger.save_state({'env': env}, [ac, discrim], None)
# Update
ac.train()
discrim.train()
# update penalty
cur_penalty = penalty_update(cur_penalty)
# update networks
update(epoch)
# Log
logger.log_tabular('Epoch', epoch)
logger.log_tabular('EpRetS', average_only=True)
logger.log_tabular('EpSdrS', average_only=True)
logger.log_tabular('EpLenS', average_only=True)
logger.log_tabular('EpRetT', average_only=True)
logger.log_tabular('EpLenT', average_only=True)
logger.log_tabular('VVals', with_min_and_max=True)
logger.log_tabular('TotalEnvInteracts', total_t)
# logger.log_tabular('LossPi', average_only=True)
# logger.log_tabular('DeltaLossPi', average_only=True)
logger.log_tabular('LossV', average_only=True)
logger.log_tabular('DeltaLossV', average_only=True)
logger.log_tabular('LossDC', average_only=True)
logger.log_tabular('DeltaLossDC', average_only=True)
# logger.log_tabular('Entropy', average_only=True)
# logger.log_tabular('DeltaEnt', average_only=True)
# logger.log_tabular('KL', average_only=True)
logger.log_tabular('Time', time.time() - start_time)
logger.dump_tabular()
def valor_penalized(
env_fn,
actor_critic=ActorCritic,
ac_kwargs=dict(),
disc=ValorDiscriminator,
dc_kwargs=dict(),
seed=0,
episodes_per_epoch=40,
epochs=50,
gamma=0.99,
pi_lr=3e-4,
vf_lr=1e-3,
dc_lr=5e-4,
train_pi_iters=1,
train_v_iters=80,
train_dc_iters=10,
train_dc_interv=10,
lam=0.97,
# Cost constraints / penalties:
cost_lim=25,
penalty_init=1.,
penalty_lr=5e-3,
clip_ratio=0.2,
max_ep_len=1000,
logger_kwargs=dict(),
con_dim=5,
config_name='standard',
save_freq=10,
k=1):
# W&B Logging
wandb.login()
composite_name = 'new_valor_penalized_' + config_name
wandb.init(project="LearningCurves",
group="VALOR Expert",
name=composite_name)
# 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())
seed += 10000 * proc_id()
torch.manual_seed(seed)
np.random.seed(seed)
# Instantiate environment
env = env_fn()
obs_dim = env.observation_space.shape
act_dim = env.action_space.shape
ac_kwargs['action_space'] = env.action_space
# Model # Create actor-critic modules and discriminator and monitor them
ac = actor_critic(input_dim=obs_dim[0] + con_dim, **ac_kwargs)
discrim = disc(input_dim=obs_dim[0], context_dim=con_dim, **dc_kwargs)
# Set up model saving
logger.setup_pytorch_saver([ac, discrim])
# Sync params across processes
sync_params(ac)
sync_params(discrim)
# Buffer
local_episodes_per_epoch = int(episodes_per_epoch / num_procs())
buffer = VALORBuffer(con_dim, obs_dim[0], act_dim[0],
local_episodes_per_epoch, max_ep_len, train_dc_interv)
# Count variables
var_counts = tuple(
count_vars(module) for module in [ac.pi, ac.v, discrim.pi])
logger.log('\nNumber of parameters: \t pi: %d, \t v: %d, \t d: %d\n' %
var_counts)
# Optimizers
pi_optimizer = torch.optim.Adam(ac.pi.parameters(), lr=pi_lr)
vf_optimizer = torch.optim.Adam(ac.v.parameters(), lr=vf_lr)
discrim_optimizer = torch.optim.Adam(discrim.pi.parameters(), lr=dc_lr)
def compute_loss_pi(obs, act, adv, logp_old):
# Policy loss # policy gradient term + entropy term
# Policy loss with clipping (without clipping, loss_pi = -(logp*adv).mean()).
# TODO: Think about removing clipping
_, 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()
return loss_pi
# # Parameters Sync
# sync_all_params(ac.parameters())
# sync_all_params(disc.parameters())
def penalty_update(cur_penalty):
cur_cost = logger.get_stats('EpCost')[0]
cur_rew = logger.get_stats('EpRet')[0]
# Penalty update
cur_penalty = max(0, cur_penalty + penalty_lr * (cur_cost - cost_lim))
return cur_penalty
def update(e):
obs, act, adv, pos, ret, logp_old = [
torch.Tensor(x) for x in buffer.retrieve_all()
]
# Policy
_, logp, _ = ac.pi(obs, act)
entropy = (-logp).mean()
# Train policy with multiple steps of gradient descent
for _ in range(train_pi_iters):
pi_optimizer.zero_grad()
loss_pi = compute_loss_pi(obs, act, adv, ret)
loss_pi.backward()
mpi_avg_grads(ac.pi)
pi_optimizer.step()
# Value function
v = ac.v(obs)
v_l_old = F.mse_loss(v, ret)
for _ in range(train_v_iters):
v = ac.v(obs)
v_loss = F.mse_loss(v, ret)
# Value function train
vf_optimizer.zero_grad()
v_loss.backward()
mpi_avg_grads(ac.v)
vf_optimizer.step()
# Discriminator
if (e + 1) % train_dc_interv == 0:
print('Discriminator Update!')
# Remove BiLSTM, take FFNN, take state_diff and predict what the context was
# Predict what was the context based on the tuple (or just context from just the current state)
con, s_diff = [torch.Tensor(x) for x in buffer.retrieve_dc_buff()]
print("s diff: ", s_diff)
_, logp_dc, _ = discrim(s_diff, con)
d_l_old = -logp_dc.mean()
# Discriminator train
for _ in range(train_dc_iters):
_, logp_dc, _ = discrim(s_diff, con)
d_loss = -logp_dc.mean()
discrim_optimizer.zero_grad()
d_loss.backward()
mpi_avg_grads(discrim.pi)
discrim_optimizer.step()
_, logp_dc, _ = discrim(s_diff, con)
dc_l_new = -logp_dc.mean()
else:
d_l_old = 0
dc_l_new = 0
# Log the changes
_, logp, _, v = ac(obs, act)
pi_l_new = -(logp * (k * adv + pos)).mean()
v_l_new = F.mse_loss(v, ret)
kl = (logp_old - logp).mean()
logger.store(LossPi=loss_pi,
LossV=v_l_old,
KL=kl,
Entropy=entropy,
DeltaLossPi=(pi_l_new - loss_pi),
DeltaLossV=(v_l_new - v_l_old),
LossDC=d_l_old,
DeltaLossDC=(dc_l_new - d_l_old))
# logger.store(Adv=adv.reshape(-1).numpy().tolist(), Pos=pos.reshape(-1).numpy().tolist())
start_time = time.time()
o, r, d, ep_ret, ep_cost, ep_len = env.reset(), 0, False, 0, 0, 0
context_dist = Categorical(logits=torch.Tensor(np.ones(con_dim)))
print("context distribution:", context_dist)
total_t = 0
# Initialize penalty
cur_penalty = np.log(max(np.exp(penalty_init) - 1, 1e-8))
for epoch in range(epochs):
ac.eval()
discrim.eval()
for _ in range(local_episodes_per_epoch):
c = context_dist.sample()
print("context sample: ", c)
c_onehot = F.one_hot(c, con_dim).squeeze().float()
# print("one hot sample: ", c_onehot)
for _ in range(max_ep_len):
concat_obs = torch.cat(
[torch.Tensor(o.reshape(1, -1)),
c_onehot.reshape(1, -1)], 1)
a, _, logp_t, v_t = ac(concat_obs)
o, r, d, info = env.step(a.detach().numpy()[0])
# print("info", info)
# time.sleep(0.002)
# cost = info.get("cost")
# ep_cost += cost
ep_ret += r
ep_len += 1
total_t += 1
# r_total = r - cur_penalty * cost
# r_total /= (1 + cur_penalty)
# buffer.store(c, concat_obs.squeeze().detach().numpy(), a.detach().numpy(), r_total, v_t.item(),
# logp_t.detach().numpy())
buffer.store(c,
concat_obs.squeeze().detach().numpy(),
a.detach().numpy(), r, v_t.item(),
logp_t.detach().numpy())
logger.store(VVals=v_t)
terminal = d or (ep_len == max_ep_len)
if terminal:
dc_diff = torch.Tensor(buffer.calc_diff()).unsqueeze(0)
con = torch.Tensor([float(c)]).unsqueeze(0)
_, _, log_p = discrim(dc_diff, con)
buffer.finish_path(log_p.detach().numpy())
logger.store(EpRet=ep_ret, EpCost=ep_cost, EpLen=ep_len)
o, r, d, ep_ret, ep_cost, ep_len = env.reset(
), 0, False, 0, 0, 0
if (epoch % save_freq == 0) or (epoch == epochs - 1):
logger.save_state({'env': env}, [ac, discrim], None)
# Update
ac.train()
discrim.train()
# update penalty
cur_penalty = penalty_update(cur_penalty)
# update models
update(epoch)
# Log
logger.log_tabular('Epoch', epoch)
logger.log_tabular('EpRet', with_min_and_max=True)
logger.log_tabular('EpCost', 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', total_t)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('DeltaLossPi', average_only=True)
logger.log_tabular('LossV', average_only=True)
logger.log_tabular('DeltaLossV', average_only=True)
logger.log_tabular('LossDC', average_only=True)
logger.log_tabular('DeltaLossDC', average_only=True)
logger.log_tabular('Entropy', average_only=True)
logger.log_tabular('KL', average_only=True)
logger.log_tabular('Time', time.time() - start_time)
logger.dump_tabular()