def train(
env,
cost_fn,
logdir=None,
render=False,
learning_rate=1e-3,
onpol_iters=10,
dynamics_iters=60,
batch_size=512,
num_paths_random=10,
num_paths_onpol=10,
num_simulated_paths=10000,
env_horizon=1000,
mpc_horizon=15,
n_layers=2,
size=500,
activation=tf.nn.relu,
output_activation=None,
clip_param=0.2,
entcoeff=0.0,
gamma=0.99,
lam=0.95,
optim_epochs=10,
optim_batchsize=64,
schedule='linear',
bc_lr=1e-3,
ppo_lr=3e-4,
timesteps_per_actorbatch=1000,
MPC=True,
BEHAVIORAL_CLONING=True,
PPO=True,
):
start = time.time()
logz.configure_output_dir(logdir)
merged_summary, summary_writer, ppo_return_op, mpc_return_op, model_loss_op, reward_loss_op, ppo_std_op, mpc_std_op = build_summary_ops(
logdir, env)
print("-------- env info --------")
print("Environment: ", FLAGS.env_name)
print("observation_space: ", env.observation_space.shape)
print("action_space: ", env.action_space.shape)
print("action_space low: ", env.action_space.low)
print("action_space high: ", env.action_space.high)
print("BEHAVIORAL_CLONING: ", BEHAVIORAL_CLONING)
print("PPO: ", PPO)
print("MPC-AUG: ", MPC)
print(" ")
random_controller = RandomController(env)
# Creat buffers
model_data_buffer = DataBufferGeneral(FLAGS.MODELBUFFER_SIZE, 5)
ppo_data_buffer = DataBufferGeneral(10000, 4)
bc_data_buffer = DataBufferGeneral(2000, 2)
# Random sample path
print("collecting random data ..... ")
paths = sample(env,
random_controller,
num_paths=num_paths_random,
horizon=env_horizon,
render=False,
verbose=False)
# add into buffer
for path in paths:
for n in range(len(path['observations'])):
model_data_buffer.add([
path['observations'][n], path['actions'][n],
path['rewards'][n], path['next_observations'][n],
path['next_observations'][n] - path['observations'][n]
])
print("model data buffer size: ", model_data_buffer.size)
normalization = compute_normalization(model_data_buffer)
#========================================================
#
# Build dynamics model and MPC controllers and Behavioral cloning network.
#
# tf_config = tf.ConfigProto(inter_op_parallelism_threads=1, intra_op_parallelism_threads=1)
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
sess = tf.Session(config=tf_config)
policy_nn = MlpPolicy(sess=sess,
env=env,
hid_size=128,
num_hid_layers=2,
clip_param=clip_param,
entcoeff=entcoeff)
if FLAGS.LEARN_REWARD:
print("Learn reward function")
dyn_model = NNDynamicsRewardModel(env=env,
normalization=normalization,
batch_size=batch_size,
iterations=dynamics_iters,
learning_rate=learning_rate,
sess=sess)
mpc_ppo_controller = MPCcontrollerPolicyNetReward(
env=env,
dyn_model=dyn_model,
explore=FLAGS.MPC_EXP,
policy_net=policy_nn,
self_exp=FLAGS.SELFEXP,
horizon=mpc_horizon,
num_simulated_paths=num_simulated_paths)
else:
print("Use predefined cost function")
dyn_model = NNDynamicsModel(env=env,
n_layers=n_layers,
size=size,
activation=activation,
output_activation=output_activation,
normalization=normalization,
batch_size=batch_size,
iterations=dynamics_iters,
learning_rate=learning_rate,
sess=sess)
mpc_ppo_controller = MPCcontrollerPolicyNet(
env=env,
dyn_model=dyn_model,
explore=FLAGS.MPC_EXP,
policy_net=policy_nn,
self_exp=FLAGS.SELFEXP,
horizon=mpc_horizon,
cost_fn=cost_fn,
num_simulated_paths=num_simulated_paths)
mpc_controller = MPCcontroller(env=env,
dyn_model=dyn_model,
horizon=mpc_horizon,
cost_fn=cost_fn,
num_simulated_paths=num_simulated_paths)
# if not PPO:
# mpc_ppo_controller = mpc_controller
#========================================================
#
# Tensorflow session building.
#
sess.__enter__()
tf.global_variables_initializer().run()
# init or load checkpoint with saver
saver = tf.train.Saver()
checkpoint = tf.train.get_checkpoint_state(logdir)
if checkpoint and checkpoint.model_checkpoint_path and FLAGS.LOAD_MODEL:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("checkpoint loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old checkpoint")
if not os.path.exists(logdir):
os.mkdir(logdir)
#========================================================
#
# Prepare for rollouts
#
episodes_so_far = 0
timesteps_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=100) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=100) # rolling buffer for episode rewards
max_timesteps = num_paths_onpol * env_horizon
bc = False
ppo_mpc = False
mpc_returns = 0
model_loss = 0
for itr in range(onpol_iters):
print(" ")
print("onpol_iters: ", itr)
if schedule == 'constant':
cur_lrmult = 1.0
elif schedule == 'linear':
cur_lrmult = max(1.0 - float(timesteps_so_far) / max_timesteps, 0)
print("bc learning_rate: ", bc_lr)
print("ppo learning_rate: ", ppo_lr)
################## fit mpc model
if MPC:
model_loss, reward_loss = dyn_model.fit(model_data_buffer)
################## ppo seg data
ppo_data_buffer.clear()
# ppo_seg = traj_segment_generator_ppo(policy_nn, env, env_horizon)
ppo_mpc = False
mpc = False
ppo_seg = traj_segment_generator(policy_nn, mpc_controller,
mpc_ppo_controller, bc_data_buffer,
env, mpc, ppo_mpc, env_horizon)
add_vtarg_and_adv(ppo_seg, gamma, lam)
ob, ac, rew, nxt_ob, atarg, tdlamret = \
ppo_seg["ob"], ppo_seg["ac"], ppo_seg["rew"], ppo_seg["nxt_ob"], ppo_seg["adv"], ppo_seg["tdlamret"]
atarg = (atarg - atarg.mean()
) / atarg.std() # standardized advantage function estimate
# add into buffer
for n in range(len(ob)):
ppo_data_buffer.add([ob[n], ac[n], atarg[n], tdlamret[n]])
model_data_buffer.add(
[ob[n], ac[n], rew[n], nxt_ob[n], nxt_ob[n] - ob[n]])
ppo_std = np.std(ac, axis=0)
print("ppo_std: ", ppo_std)
################## mpc augmented seg data
if MPC:
print("MPC AUG PPO")
ppo_mpc = True
mpc = True
mpc_seg = traj_segment_generator(policy_nn, mpc_controller,
mpc_ppo_controller,
bc_data_buffer, env, mpc, ppo_mpc,
env_horizon)
add_vtarg_and_adv(mpc_seg, gamma, lam)
ob, ac, mpcac, rew, nxt_ob, atarg, tdlamret = mpc_seg[
"ob"], mpc_seg["ac"], mpc_seg["mpcac"], mpc_seg[
"rew"], mpc_seg["nxt_ob"], mpc_seg["adv"], mpc_seg[
"tdlamret"]
atarg = (atarg - atarg.mean()) / atarg.std(
) # standardized advantage function estimate
mpc_returns = mpc_seg["ep_rets"]
mpc_std = np.std(mpcac)
if not MPC:
mpc_std = 0
################## mpc random seg data
if FLAGS.mpc_rand:
print("MPC Random base policy")
ppo_mpc = False
mpc = True
mpc_random_seg = traj_segment_generator(policy_nn, mpc_controller,
mpc_ppo_controller,
bc_data_buffer, env, mpc,
ppo_mpc, env_horizon)
add_vtarg_and_adv(mpc_random_seg, gamma, lam)
ob, ac, mpcac, rew, nxt_ob, atarg, tdlamret = mpc_random_seg[
"ob"], mpc_random_seg["ac"], mpc_random_seg[
"mpcac"], mpc_random_seg["rew"], mpc_random_seg[
"nxt_ob"], mpc_random_seg["adv"], mpc_random_seg[
"tdlamret"]
atarg = (atarg - atarg.mean()) / atarg.std(
) # standardized advantage function estimate
mpc_rand_returns = mpc_random_seg["ep_rets"]
################# PPO deterministic evaluation
ppo_determinisitc_return = policy_net_eval(sess,
env,
policy_nn,
env_horizon,
stochastic=False)
################## optimization
print("ppo_data_buffer size", ppo_data_buffer.size)
print("bc_data_buffer size", bc_data_buffer.size)
print("model data buffer size: ", model_data_buffer.size)
# optim_batchsize = optim_batchsize or ob.shape[0]
if hasattr(policy_nn, "ob_rms"):
policy_nn.ob_rms.update(ob) # update running mean/std for policy
policy_nn.assign_old_eq_new(
) # set old parameter values to new parameter values
for op_ep in range(optim_epochs):
# losses = [] # list of tuples, each of which gives the loss for a minibatch
# for i in range(int(timesteps_per_actorbatch/optim_batchsize)):
if PPO:
sample_ob_no, sample_ac_na, sample_adv_n, sample_b_n_target = ppo_data_buffer.sample(
optim_batchsize)
newlosses = policy_nn.lossandupdate_ppo(
sample_ob_no, sample_ac_na, sample_adv_n,
sample_b_n_target, cur_lrmult, ppo_lr * cur_lrmult)
# losses.append(newlosses)
if BEHAVIORAL_CLONING and bc:
sample_ob_no, sample_ac_na = bc_data_buffer.sample(
optim_batchsize)
# print("sample_ob_no", sample_ob_no.shape)
# print("sample_ac_na", sample_ac_na.shape)
policy_nn.update_bc(sample_ob_no, sample_ac_na,
bc_lr * cur_lrmult)
if op_ep % (100) == 0 and BEHAVIORAL_CLONING and bc:
print('epcho: ', op_ep)
policy_net_eval(sess, env, policy_nn, env_horizon)
################## print and save data
seg = ppo_seg
ep_lengths = seg["ep_lens"]
returns = seg["ep_rets"]
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)
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
# log ppo
logz.log_tabular("TimeSoFar", time.time() - start)
logz.log_tabular("TimeEp", time.time() - tstart)
logz.log_tabular("Iteration", itr)
logz.log_tabular("AverageReturn", np.mean(returns))
logz.log_tabular("StdReturn", np.std(returns))
logz.log_tabular("MaxReturn", np.max(returns))
logz.log_tabular("MinReturn", np.min(returns))
logz.log_tabular("EpLenMean", np.mean(ep_lengths))
logz.log_tabular("EpLenStd", np.std(ep_lengths))
logz.log_tabular("TimestepsSoFar", timesteps_so_far)
logz.log_tabular("Condition", "PPO")
logz.dump_tabular()
# log ppo deterministic
logz.log_tabular("Iteration", itr)
logz.log_tabular("AverageReturn", ppo_determinisitc_return)
logz.log_tabular("Condition", "PPO_DETERMINISTIC")
logz.dump_tabular()
# log mpc
if MPC:
logz.log_tabular("TimeSoFar", time.time() - start)
logz.log_tabular("TimeEp", time.time() - tstart)
logz.log_tabular("Iteration", itr)
logz.log_tabular("AverageReturn", np.mean(mpc_returns))
logz.log_tabular("StdReturn", np.std(mpc_returns))
logz.log_tabular("MaxReturn", np.max(mpc_returns))
logz.log_tabular("MinReturn", np.min(mpc_returns))
logz.log_tabular("EpLenMean", np.mean(ep_lengths))
logz.log_tabular("EpLenStd", np.std(ep_lengths))
logz.log_tabular("TimestepsSoFar", timesteps_so_far)
logz.log_tabular("Condition", "MPC_PPO")
logz.dump_tabular()
if FLAGS.mpc_rand:
logz.log_tabular("TimeSoFar", time.time() - start)
logz.log_tabular("TimeEp", time.time() - tstart)
logz.log_tabular("Iteration", itr)
logz.log_tabular("AverageReturn", np.mean(mpc_rand_returns))
logz.log_tabular("StdReturn", np.std(mpc_rand_returns))
logz.log_tabular("MaxReturn", np.max(mpc_rand_returns))
logz.log_tabular("MinReturn", np.min(mpc_rand_returns))
logz.log_tabular("EpLenMean", np.mean(ep_lengths))
logz.log_tabular("EpLenStd", np.std(ep_lengths))
logz.log_tabular("TimestepsSoFar", timesteps_so_far)
logz.log_tabular("Condition", "MPC_RAND")
logz.dump_tabular()
# logz.pickle_tf_vars()
tstart = time.time()
################### TF Summaries
summary_str = sess.run(merged_summary,
feed_dict={
ppo_return_op: np.mean(returns),
mpc_return_op: np.mean(mpc_returns),
model_loss_op: model_loss,
ppo_std_op: ppo_std,
reward_loss_op: reward_loss,
mpc_std_op: mpc_std,
})
summary_writer.add_summary(summary_str, itr)
summary_writer.flush()
################ TF SAVE
if itr % FLAGS.SAVE_ITER == 0 and itr != 0:
save_path = saver.save(sess, logdir + "/model.ckpt")
print("Model saved in path: %s" % save_path)
def train(env,
cost_fn,
logdir=None,
render=False,
learning_rate=1e-3,
onpol_iters=10,
dynamics_iters=60,
batch_size=512,
num_paths_random=10,
num_paths_onpol=10,
num_simulated_paths=10000,
env_horizon=1000,
mpc_horizon=15,
n_layers=2,
size=500,
activation=tf.nn.relu,
output_activation=None,
clip_param=0.2 ,
entcoeff=0.0,
gamma=0.99,
lam=0.95,
optim_epochs=10,
optim_batchsize=64,
schedule='linear',
bc_lr=1e-3,
ppo_lr=3e-4,
timesteps_per_actorbatch=1000,
MPC = True,
BEHAVIORAL_CLONING = True,
PPO = True,
):
start = time.time()
logz.configure_output_dir(logdir)
print("-------- env info --------")
print("observation_space: ", env.observation_space.shape)
print("action_space: ", env.action_space.shape)
print("BEHAVIORAL_CLONING: ", BEHAVIORAL_CLONING)
print("PPO: ", PPO)
print("MPC-AUG: ", MPC)
print(" ")
# initialize buffers
model_data_buffer = DataBufferGeneral(1000000, 5)
ppo_data_buffer = DataBufferGeneral(10000, 4)
bc_data_buffer = DataBufferGeneral(BC_BUFFER_SIZE, 2)
# random sample path
print("collecting random data ..... ")
random_controller = RandomController(env)
paths = sample(env,
random_controller,
num_paths=num_paths_random,
horizon=env_horizon,
render=False,
verbose=False)
# add into buffer
for path in paths:
for n in range(len(path['observations'])):
model_data_buffer.add([path['observations'][n],
path['actions'][n],
path['rewards'][n],
path['next_observations'][n],
path['next_observations'][n] - path['observations'][n]])
print("model data buffer size: ", model_data_buffer.size)
normalization = compute_normalization(model_data_buffer)
#========================================================
#
# Build dynamics model and MPC controllers and Behavioral cloning network.
#
# tf_config = tf.ConfigProto(inter_op_parallelism_threads=1, intra_op_parallelism_threads=1)
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
sess = tf.Session(config=tf_config)
dyn_model = NNDynamicsRewardModel(env=env,
normalization=normalization,
batch_size=batch_size,
iterations=dynamics_iters,
learning_rate=learning_rate,
sess=sess)
mpc_controller = MPCcontroller(env=env,
dyn_model=dyn_model,
horizon=mpc_horizon,
cost_fn=cost_fn,
num_simulated_paths=num_simulated_paths)
policy_nn = MlpPolicy(sess=sess, env=env, hid_size=256, num_hid_layers=2, clip_param=clip_param , entcoeff=entcoeff)
mpc_ppo_controller = MPCcontrollerPolicyNetReward(env=env,
dyn_model=dyn_model,
policy_net=policy_nn,
self_exp=False,
horizon=mpc_horizon,
num_simulated_paths=num_simulated_paths)
#========================================================
#
# Tensorflow session building.
#
sess.__enter__()
tf.global_variables_initializer().run()
# init or load checkpoint with saver
saver = tf.train.Saver()
checkpoint = tf.train.get_checkpoint_state(CHECKPOINT_DIR)
if checkpoint and checkpoint.model_checkpoint_path and LOAD_MODEL:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("checkpoint loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old checkpoint")
if not os.path.exists(CHECKPOINT_DIR):
os.mkdir(CHECKPOINT_DIR)
#========================================================
#
# Prepare for rollouts
#
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
max_timesteps = num_paths_onpol * env_horizon
bc = False
ppo_mpc = False
mpc_returns = 0
for itr in range(onpol_iters):
print(" ")
print("onpol_iters: ", itr)
if schedule == 'constant':
cur_lrmult = 1.0
elif schedule == 'linear':
cur_lrmult = max(1.0 - float(timesteps_so_far) / max_timesteps, 0)
print("bc learning_rate: ", bc_lr)
print("ppo learning_rate: ", ppo_lr)
################## fit mpc model
if MPC:
dyn_model.fit(model_data_buffer)
################## ppo seg data
if PPO:
ppo_data_buffer.clear()
# ppo_seg = traj_segment_generator_ppo(policy_nn, env, env_horizon)
mpc = False
ppo_seg = traj_segment_generator(policy_nn, mpc_controller, mpc_ppo_controller, bc_data_buffer, env, mpc, ppo_mpc, env_horizon)
add_vtarg_and_adv(ppo_seg, gamma, lam)
ob, ac, rew, nxt_ob, atarg, tdlamret = \
ppo_seg["ob"], ppo_seg["ac"], ppo_seg["rew"], ppo_seg["nxt_ob"], ppo_seg["adv"], ppo_seg["tdlamret"]
atarg = (atarg - atarg.mean()) / atarg.std() # standardized advantage function estimate
# add into buffer
for n in range(len(ob)):
ppo_data_buffer.add([ob[n], ac[n], atarg[n], tdlamret[n]])
if MPC:
model_data_buffer.add([ob[n], ac[n], rew[n], nxt_ob[n], nxt_ob[n]-ob[n]])
################## mpc augmented seg data
if itr % MPC_AUG_GAP == 0 and MPC:
print("MPC AUG PPO")
ppo_mpc = True
mpc = True
mpc_seg = traj_segment_generator(policy_nn, mpc_controller, mpc_ppo_controller, bc_data_buffer, env, mpc, ppo_mpc, env_horizon)
add_vtarg_and_adv(mpc_seg, gamma, lam)
ob, ac, mpcac, rew, nxt_ob, atarg, tdlamret = mpc_seg["ob"], mpc_seg["ac"], mpc_seg["mpcac"], mpc_seg["rew"], mpc_seg["nxt_ob"], mpc_seg["adv"], mpc_seg["tdlamret"]
atarg = (atarg - atarg.mean()) / atarg.std() # standardized advantage function estimate
# add into buffer
for n in range(len(ob)):
# if PPO:
# ppo_data_buffer.add([ob[n], ac[n], atarg[n], tdlamret[n]])
if BEHAVIORAL_CLONING and bc:
bc_data_buffer.add([ob[n], mpcac[n]])
if MPC:
model_data_buffer.add([ob[n], mpcac[n], rew[n], nxt_ob[n], nxt_ob[n]-ob[n]])
mpc_returns = mpc_seg["ep_rets"]
seg = ppo_seg
# check if seg is good
ep_lengths = seg["ep_lens"]
returns = seg["ep_rets"]
# saver.save(sess, CHECKPOINT_DIR)
if BEHAVIORAL_CLONING:
if np.mean(returns) > 100:
bc = True
else:
bc = False
print("BEHAVIORAL_CLONING: ", bc)
bc_return = behavioral_cloning_eval(sess, env, policy_nn, env_horizon)
if bc_return > 100:
ppo_mpc = True
else:
ppo_mpc = False
################## optimization
print("ppo_data_buffer size", ppo_data_buffer.size)
print("bc_data_buffer size", bc_data_buffer.size)
print("model data buffer size: ", model_data_buffer.size)
# optim_batchsize = optim_batchsize or ob.shape[0]
if hasattr(policy_nn, "ob_rms"): policy_nn.ob_rms.update(ob) # update running mean/std for policy
policy_nn.assign_old_eq_new() # set old parameter values to new parameter values
for op_ep in range(optim_epochs):
# losses = [] # list of tuples, each of which gives the loss for a minibatch
# for i in range(int(timesteps_per_actorbatch/optim_batchsize)):
if PPO:
sample_ob_no, sample_ac_na, sample_adv_n, sample_b_n_target = ppo_data_buffer.sample(optim_batchsize)
newlosses = policy_nn.lossandupdate_ppo(sample_ob_no, sample_ac_na, sample_adv_n, sample_b_n_target, cur_lrmult, ppo_lr*cur_lrmult)
# losses.append(newlosses)
if BEHAVIORAL_CLONING and bc:
sample_ob_no, sample_ac_na = bc_data_buffer.sample(optim_batchsize)
# print("sample_ob_no", sample_ob_no.shape)
# print("sample_ac_na", sample_ac_na.shape)
policy_nn.update_bc(sample_ob_no, sample_ac_na, bc_lr*cur_lrmult)
if op_ep % (100) == 0 and BEHAVIORAL_CLONING and bc:
print('epcho: ', op_ep)
behavioral_cloning_eval(sess, env, policy_nn, env_horizon)
################## print and save data
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)
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
iters_so_far += 1
# if np.mean(returns) > 1000:
# filename = "seg_data.pkl"
# pickle.dump(seg, open(filename, 'wb'))
# print("saved", filename)
logz.log_tabular("TimeSoFar", time.time() - start)
logz.log_tabular("TimeEp", time.time() - tstart)
logz.log_tabular("Iteration", iters_so_far)
logz.log_tabular("AverageReturn", np.mean(returns))
logz.log_tabular("MpcReturn", np.mean(mpc_returns))
logz.log_tabular("StdReturn", np.std(returns))
logz.log_tabular("MaxReturn", np.max(returns))
logz.log_tabular("MinReturn", np.min(returns))
logz.log_tabular("EpLenMean", np.mean(ep_lengths))
logz.log_tabular("EpLenStd", np.std(ep_lengths))
# logz.log_tabular("TimestepsThisBatch", timesteps_this_batch)
logz.log_tabular("TimestepsSoFar", timesteps_so_far)
logz.dump_tabular()
logz.pickle_tf_vars()
tstart = time.time()
def train(env,
cost_fn,
logdir=None,
render=False,
learning_rate=1e-3,
onpol_iters=10,
dynamics_iters=60,
batch_size=512,
num_paths_random=10,
num_paths_onpol=10,
num_simulated_paths=10000,
env_horizon=1000,
mpc_horizon=15,
n_layers=2,
size=500,
activation=tf.nn.relu,
output_activation=None,
clip_param=0.2 ,
entcoeff=0.0,
gamma=0.99,
lam=0.95,
optim_epochs=10,
optim_batchsize=64,
schedule='linear',
bc_lr=1e-3,
ppo_lr=3e-4,
timesteps_per_actorbatch=1000,
MPC = True,
BEHAVIORAL_CLONING = True,
PPO = True,
):
start = time.time()
print("-------- env info --------")
print("Environment: ", FLAGS.env_name)
print("observation_space: ", env.observation_space.shape)
print("action_space: ", env.action_space.shape)
print("action_space low: ", env.action_space.low)
print("action_space high: ", env.action_space.high)
print("BEHAVIORAL_CLONING: ", BEHAVIORAL_CLONING)
print("PPO: ", PPO)
print("MPC-AUG: ", MPC)
print(" ")
random_controller = RandomController(env)
# Creat buffers
model_data_buffer = DataBufferGeneral(FLAGS.MODELBUFFER_SIZE, 5)
ppo_data_buffer = DataBufferGeneral(10000, 4)
bc_data_buffer = DataBufferGeneral(2000, 2)
# Random sample path
print("collecting random data ..... ")
paths = sample(env,
random_controller,
num_paths=num_paths_random,
horizon=env_horizon,
render=False,
verbose=False)
# add into buffer
for path in paths:
for n in range(len(path['observations'])):
model_data_buffer.add([path['observations'][n],
path['actions'][n],
path['rewards'][n],
path['next_observations'][n],
path['next_observations'][n] - path['observations'][n]])
print("model data buffer size: ", model_data_buffer.size)
normalization = compute_normalization(model_data_buffer)
#========================================================
#
# Build dynamics model and MPC controllers and Behavioral cloning network.
#
# tf_config = tf.ConfigProto(inter_op_parallelism_threads=1, intra_op_parallelism_threads=1)
tf_config = tf.ConfigProto()
tf_config.gpu_options.allow_growth = True
sess = tf.Session(config=tf_config)
policy_nn = MlpPolicy(sess=sess, env=env, hid_size=128, num_hid_layers=2, clip_param=clip_param , entcoeff=entcoeff)
if FLAGS.LEARN_REWARD:
print("Learn reward function")
dyn_model = NNDynamicsRewardModel(env=env,
normalization=normalization,
batch_size=batch_size,
iterations=dynamics_iters,
learning_rate=learning_rate,
sess=sess)
mpc_ppo_controller = MPCcontrollerPolicyNetReward(env=env,
dyn_model=dyn_model,
explore=FLAGS.MPC_EXP,
policy_net=policy_nn,
self_exp=FLAGS.SELFEXP,
horizon=mpc_horizon,
num_simulated_paths=num_simulated_paths)
else:
print("Use predefined cost function")
dyn_model = NNDynamicsModel(env=env,
n_layers=n_layers,
size=size,
activation=activation,
output_activation=output_activation,
normalization=normalization,
batch_size=batch_size,
iterations=dynamics_iters,
learning_rate=learning_rate,
sess=sess)
mpc_ppo_controller = MPCcontrollerPolicyNet(env=env,
dyn_model=dyn_model,
explore=FLAGS.MPC_EXP,
policy_net=policy_nn,
self_exp=FLAGS.SELFEXP,
horizon=mpc_horizon,
cost_fn=cost_fn,
num_simulated_paths=num_simulated_paths)
mpc_controller = MPCcontroller(env=env,
dyn_model=dyn_model,
horizon=mpc_horizon,
cost_fn=cost_fn,
num_simulated_paths=num_simulated_paths)
# if not PPO:
# mpc_ppo_controller = mpc_controller
#========================================================
#
# Tensorflow session building.
#
sess.__enter__()
tf.global_variables_initializer().run()
# init or load checkpoint with saver
saver = tf.train.Saver()
checkpoint = tf.train.get_checkpoint_state(FLAGS.model_path)
print("checkpoint", checkpoint)
if checkpoint and checkpoint.model_checkpoint_path and FLAGS.LOAD_MODEL:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("checkpoint loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old checkpoint")
if not os.path.exists(FLAGS.model_path):
os.mkdir(FLAGS.model_path)
#========================================================
#
# Prepare for rollouts
#
tstart = time.time()
states_true = []
states_predict = []
rewards_true = []
rewards_predict = []
ob = env.reset()
ob_pre = np.expand_dims(ob, axis=0)
states_true.append(ob)
states_predict.append(ob_pre)
for step in range(100):
# ac = env.action_space.sample() # not used, just so we have the datatype
ac, _ = policy_nn.act(ob, stochastic=True)
ob, rew, done, _ = env.step(ac)
ob_pre, r_pre = dyn_model.predict(ob_pre, ac)
states_true.append(ob)
rewards_true.append(rew)
states_predict.append(ob_pre)
rewards_predict.append(r_pre[0][0])
states_true = np.asarray(states_true)
states_predict = np.asarray(states_predict)
states_predict = np.squeeze(states_predict, axis=1)
rewards_true = np.asarray(rewards_true)
rewards_predict = np.asarray(rewards_predict)
print("states_true", states_true.shape)
print("states_predict", states_predict.shape)
print("rewards_true", rewards_true.shape)
print("rewards_predict", rewards_predict.shape)
np.savetxt('./data/eval_model/states_true.out', states_true, delimiter=',')
np.savetxt('./data/eval_model/states_predict.out', states_predict, delimiter=',')
np.savetxt('./data/eval_model/rewards_true.out', rewards_true, delimiter=',')
np.savetxt('./data/eval_model/rewards_predict.out', rewards_predict, delimiter=',')
def train_PG(
exp_name='',
env_name='',
n_iter=100,
gamma=1.0,
min_timesteps_per_batch=1000,
max_path_length=None,
learning_rate=5e-3,
reward_to_go=False,
animate=True,
logdir=None,
normalize_advantages=False,
nn_baseline=False,
seed=0,
# network arguments
n_layers=1,
size=32,
# mb mpc arguments
model_learning_rate=1e-3,
onpol_iters=10,
dynamics_iters=260,
batch_size=512,
num_paths_random=10,
num_paths_onpol=10,
num_simulated_paths=1000,
env_horizon=1000,
mpc_horizon=10,
m_n_layers=2,
m_size=500,
):
start = time.time()
# Configure output directory for logging
logz.configure_output_dir(logdir)
# Log experimental parameters
args = inspect.getargspec(train_PG)[0]
locals_ = locals()
params = {k: locals_[k] if k in locals_ else None for k in args}
logz.save_params(params)
# Set random seeds
tf.set_random_seed(seed)
np.random.seed(seed)
# Make the gym environment
# env = gym.make(env_name)
env = HalfCheetahEnvNew()
cost_fn = cheetah_cost_fn
activation=tf.nn.relu
output_activation=None
# Is this env continuous, or discrete?
discrete = isinstance(env.action_space, gym.spaces.Discrete)
# Maximum length for episodes
# max_path_length = max_path_length or env.spec.max_episode_steps
max_path_length = max_path_length
# Observation and action sizes
ob_dim = env.observation_space.shape[0]
ac_dim = env.action_space.n if discrete else env.action_space.shape[0]
# Print environment infomation
print("-------- env info --------")
print("Environment name: ", env_name)
print("Action space is discrete: ", discrete)
print("Action space dim: ", ac_dim)
print("Observation space dim: ", ob_dim)
print("Max_path_length ", max_path_length)
#========================================================================================#
# Random data collection
#========================================================================================#
random_controller = RandomController(env)
data_buffer_model = DataBuffer()
data_buffer_ppo = DataBuffer_general(10000, 4)
# sample path
print("collecting random data ..... ")
paths = sample(env,
random_controller,
num_paths=num_paths_random,
horizon=env_horizon,
render=False,
verbose=False)
# add into buffer
for path in paths:
for n in range(len(path['observations'])):
data_buffer_model.add(path['observations'][n], path['actions'][n], path['next_observations'][n])
print("data buffer size: ", data_buffer_model.size)
normalization = compute_normalization(data_buffer_model)
#========================================================================================#
# Tensorflow Engineering: Config, Session, Variable initialization
#========================================================================================#
tf_config = tf.ConfigProto()
tf_config.allow_soft_placement = True
tf_config.intra_op_parallelism_threads =4
tf_config.inter_op_parallelism_threads = 1
sess = tf.Session(config=tf_config)
dyn_model = NNDynamicsModel(env=env,
n_layers=n_layers,
size=size,
activation=activation,
output_activation=output_activation,
normalization=normalization,
batch_size=batch_size,
iterations=dynamics_iters,
learning_rate=learning_rate,
sess=sess)
mpc_controller = MPCcontroller(env=env,
dyn_model=dyn_model,
horizon=mpc_horizon,
cost_fn=cost_fn,
num_simulated_paths=num_simulated_paths)
policy_nn = policy_network_ppo(sess, ob_dim, ac_dim, discrete, n_layers, size, learning_rate)
if nn_baseline:
value_nn = value_network(sess, ob_dim, n_layers, size, learning_rate)
sess.__enter__() # equivalent to `with sess:`
tf.global_variables_initializer().run()
#========================================================================================#
# Training Loop
#========================================================================================#
total_timesteps = 0
for itr in range(n_iter):
print("********** Iteration %i ************"%itr)
if MPC:
dyn_model.fit(data_buffer_model)
returns = []
costs = []
# Collect paths until we have enough timesteps
timesteps_this_batch = 0
paths = []
while True:
# print("data buffer size: ", data_buffer_model.size)
current_path = {'observations': [], 'actions': [], 'reward': [], 'next_observations':[]}
ob = env.reset()
obs, acs, mpc_acs, rewards = [], [], [], []
animate_this_episode=(len(paths)==0 and (itr % 10 == 0) and animate)
steps = 0
return_ = 0
while True:
# print("steps ", steps)
if animate_this_episode:
env.render()
time.sleep(0.05)
obs.append(ob)
if MPC:
mpc_ac = mpc_controller.get_action(ob)
else:
mpc_ac = random_controller.get_action(ob)
ac = policy_nn.predict(ob, mpc_ac)
ac = ac[0]
if not PG:
ac = mpc_ac
acs.append(ac)
mpc_acs.append(mpc_ac)
current_path['observations'].append(ob)
ob, rew, done, _ = env.step(ac)
current_path['reward'].append(rew)
current_path['actions'].append(ac)
current_path['next_observations'].append(ob)
return_ += rew
rewards.append(rew)
steps += 1
if done or steps > max_path_length:
break
if MPC:
# cost & return
cost = path_cost(cost_fn, current_path)
costs.append(cost)
returns.append(return_)
print("total return: ", return_)
print("costs: ", cost)
# add into buffers
for n in range(len(current_path['observations'])):
data_buffer_model.add(current_path['observations'][n], current_path['actions'][n], current_path['next_observations'][n])
for n in range(len(current_path['observations'])):
data_buffer_ppo.add(current_path['observations'][n], current_path['actions'][n], current_path['reward'][n], current_path['next_observations'][n])
path = {"observation" : np.array(obs),
"reward" : np.array(rewards),
"action" : np.array(acs),
"mpc_action" : np.array(mpc_acs)}
paths.append(path)
timesteps_this_batch += pathlength(path)
# print("timesteps_this_batch", timesteps_this_batch)
if timesteps_this_batch > min_timesteps_per_batch:
break
total_timesteps += timesteps_this_batch
print("data_buffer_ppo.size:", data_buffer_ppo.size)
# Build arrays for observation, action for the policy gradient update by concatenating
# across paths
ob_no = np.concatenate([path["observation"] for path in paths])
ac_na = np.concatenate([path["action"] for path in paths])
mpc_ac_na = np.concatenate([path["mpc_action"] for path in paths])
# Computing Q-values
if reward_to_go:
q_n = []
for path in paths:
for t in range(len(path["reward"])):
t_ = 0
q = 0
while t_ < len(path["reward"]):
if t_ >= t:
q += gamma**(t_-t) * path["reward"][t_]
t_ += 1
q_n.append(q)
q_n = np.asarray(q_n)
else:
q_n = []
for path in paths:
for t in range(len(path["reward"])):
t_ = 0
q = 0
while t_ < len(path["reward"]):
q += gamma**t_ * path["reward"][t_]
t_ += 1
q_n.append(q)
q_n = np.asarray(q_n)
# Computing Baselines
if nn_baseline:
# b_n = sess.run(baseline_prediction, feed_dict={sy_ob_no :ob_no})
b_n = value_nn.predict(ob_no)
b_n = normalize(b_n)
b_n = denormalize(b_n, np.std(q_n), np.mean(q_n))
adv_n = q_n - b_n
else:
adv_n = q_n.copy()
# Advantage Normalization
if normalize_advantages:
adv_n = normalize(adv_n)
# Optimizing Neural Network Baseline
if nn_baseline:
b_n_target = normalize(q_n)
value_nn.fit(ob_no, b_n_target)
# sess.run(baseline_update_op, feed_dict={sy_ob_no :ob_no, sy_baseline_target_n:b_n_target})
# Performing the Policy Update
# policy_nn.fit(ob_no, ac_na, adv_n)
policy_nn.fit(ob_no, ac_na, adv_n, mpc_ac_na)
# sess.run(update_op, feed_dict={sy_ob_no :ob_no, sy_ac_na:ac_na, sy_adv_n:adv_n})
# Log diagnostics
returns = [path["reward"].sum() for path in paths]
ep_lengths = [pathlength(path) for path in paths]
logz.log_tabular("Time", time.time() - start)
logz.log_tabular("Iteration", itr)
logz.log_tabular("AverageReturn", np.mean(returns))
logz.log_tabular("StdReturn", np.std(returns))
logz.log_tabular("MaxReturn", np.max(returns))
logz.log_tabular("MinReturn", np.min(returns))
logz.log_tabular("EpLenMean", np.mean(ep_lengths))
logz.log_tabular("EpLenStd", np.std(ep_lengths))
logz.log_tabular("TimestepsThisBatch", timesteps_this_batch)
logz.log_tabular("TimestepsSoFar", total_timesteps)
logz.dump_tabular()
logz.pickle_tf_vars()
def train(
env,
cost_fn,
logdir=None,
render=False,
learning_rate=1e-3,
onpol_iters=10,
dynamics_iters=60,
batch_size=512,
num_paths_random=10,
num_paths_onpol=10,
num_simulated_paths=10000,
env_horizon=1000,
mpc_horizon=15,
n_layers=2,
size=500,
activation=tf.nn.relu,
output_activation=None,
clip_param=0.2,
entcoeff=0.0,
gamma=0.99,
lam=0.95,
optim_epochs=10,
optim_batchsize=64,
schedule='linear',
optim_stepsize=3e-4,
timesteps_per_actorbatch=1000,
BEHAVIORAL_CLONING=True,
PPO=True,
):
start = time.time()
logz.configure_output_dir(logdir)
print("-------- env info --------")
print("observation_space: ", env.observation_space.shape)
print("action_space: ", env.action_space.shape)
print("BEHAVIORAL_CLONING: ", BEHAVIORAL_CLONING)
print("PPO: ", PPO)
print(" ")
random_controller = RandomController(env)
model_data_buffer = DataBuffer()
ppo_data_buffer = DataBuffer_general(BC_BUFFER_SIZE, 6)
bc_data_buffer = DataBuffer_general(BC_BUFFER_SIZE, 2)
# sample path
print("collecting random data ..... ")
paths = sample(env,
random_controller,
num_paths=num_paths_random,
horizon=env_horizon,
render=False,
verbose=False)
# add into buffer
for path in paths:
for n in range(len(path['observations'])):
model_data_buffer.add(path['observations'][n], path['actions'][n],
path['next_observations'][n])
print("model data buffer size: ", model_data_buffer.size)
normalization = compute_normalization(model_data_buffer)
#========================================================
#
# Build dynamics model and MPC controllers and Behavioral cloning network.
#
sess = tf.Session()
dyn_model = NNDynamicsModel(env=env,
n_layers=n_layers,
size=size,
activation=activation,
output_activation=output_activation,
normalization=normalization,
batch_size=batch_size,
iterations=dynamics_iters,
learning_rate=learning_rate,
sess=sess)
mpc_controller = MPCcontroller(env=env,
dyn_model=dyn_model,
horizon=mpc_horizon,
cost_fn=cost_fn,
num_simulated_paths=num_simulated_paths)
policy_nn = MlpPolicy_bc(sess=sess,
env=env,
hid_size=64,
num_hid_layers=2,
clip_param=clip_param,
entcoeff=entcoeff)
bc_net = BCnetwork(sess, env, BATCH_SIZE_BC, learning_rate)
mpc_controller_bc_ppo = MPCcontroller_BC_PPO(
env=env,
dyn_model=dyn_model,
bc_ppo_network=policy_nn,
horizon=mpc_horizon,
cost_fn=cost_fn,
num_simulated_paths=num_simulated_paths)
#========================================================
#
# Tensorflow session building.
#
sess.__enter__()
tf.global_variables_initializer().run()
# init or load checkpoint with saver
saver = tf.train.Saver()
checkpoint = tf.train.get_checkpoint_state(CHECKPOINT_DIR)
if checkpoint and checkpoint.model_checkpoint_path and LOAD_MODEL:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("checkpoint loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old checkpoint")
if not os.path.exists(CHECKPOINT_DIR):
os.mkdir(CHECKPOINT_DIR)
#========================================================
#
# Prepare for rollouts
#
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
max_timesteps = num_paths_onpol * env_horizon
for itr in range(onpol_iters):
print("onpol_iters: ", itr)
dyn_model.fit(model_data_buffer)
if schedule == 'constant':
cur_lrmult = 1.0
elif schedule == 'linear':
cur_lrmult = max(1.0 - float(timesteps_so_far) / max_timesteps, 0)
# saver.save(sess, CHECKPOINT_DIR)
behavioral_cloning_eval(sess, env, policy_nn, env_horizon)
ppo_data_buffer.clear()
seg = traj_segment_generator(policy_nn, mpc_controller,
mpc_controller_bc_ppo, bc_data_buffer,
env, env_horizon)
add_vtarg_and_adv(seg, gamma, lam)
ob, ac, rew, nxt_ob, atarg, tdlamret = seg["ob"], seg["ac"], seg[
"rew"], seg["nxt_ob"], seg["adv"], seg["tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()
) / atarg.std() # standardized advantage function estimate
for n in range(len(ob)):
ppo_data_buffer.add(
(ob[n], ac[n], rew[n], nxt_ob[n], atarg[n], tdlamret[n]))
bc_data_buffer.add((ob[n], ac[n]))
model_data_buffer.add(ob[n], ac[n], nxt_ob[n])
print("ppo_data_buffer size", ppo_data_buffer.size)
print("bc_data_buffer size", bc_data_buffer.size)
print("model data buffer size: ", model_data_buffer.size)
# optim_batchsize = optim_batchsize or ob.shape[0]
# behavioral_cloning(sess, env, bc_net, mpc_controller, env_horizon, bc_data_buffer, Training_epoch=1000)
if hasattr(policy_nn, "ob_rms"):
policy_nn.ob_rms.update(ob) # update running mean/std for policy
policy_nn.assign_old_eq_new(
) # set old parameter values to new parameter values
for op_ep in range(optim_epochs):
# losses = [] # list of tuples, each of which gives the loss for a minibatch
# for i in range(int(timesteps_per_actorbatch/optim_batchsize)):
if PPO:
sample_ob_no, sample_ac_na, sample_rew, sample_nxt_ob_no, sample_adv_n, sample_b_n_target = ppo_data_buffer.sample(
optim_batchsize)
newlosses = policy_nn.lossandupdate_ppo(
sample_ob_no, sample_ac_na, sample_adv_n,
sample_b_n_target, cur_lrmult, optim_stepsize * cur_lrmult)
# losses.append(newlosses)
if BEHAVIORAL_CLONING:
sample_ob_no, sample_ac_na = bc_data_buffer.sample(
optim_batchsize)
policy_nn.update_bc(sample_ob_no, sample_ac_na,
optim_stepsize * cur_lrmult)
if op_ep % 100 == 0:
print('epcho: ', op_ep)
behavioral_cloning_eval(sess, env, policy_nn, env_horizon)
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)
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
iters_so_far += 1
ep_lengths = seg["ep_lens"]
returns = seg["ep_rets"]
logz.log_tabular("Time", time.time() - start)
logz.log_tabular("Iteration", iters_so_far)
logz.log_tabular("AverageReturn", np.mean(returns))
logz.log_tabular("StdReturn", np.std(returns))
logz.log_tabular("MaxReturn", np.max(returns))
logz.log_tabular("MinReturn", np.min(returns))
logz.log_tabular("EpLenMean", np.mean(ep_lengths))
logz.log_tabular("EpLenStd", np.std(ep_lengths))
# logz.log_tabular("TimestepsThisBatch", timesteps_this_batch)
logz.log_tabular("TimestepsSoFar", timesteps_so_far)
logz.dump_tabular()
logz.pickle_tf_vars()