def test_mpi_adam():
"""
tests the MpiAdam object's functionality
"""
np.random.seed(0)
tf.set_random_seed(0)
a_var = tf.Variable(np.random.randn(3).astype('float32'))
b_var = tf.Variable(np.random.randn(2, 5).astype('float32'))
loss = tf.reduce_sum(tf.square(a_var)) + tf.reduce_sum(tf.sin(b_var))
learning_rate = 1e-2
update_op = tf.train.AdamOptimizer(learning_rate).minimize(loss)
do_update = tf_utils.function([], loss, updates=[update_op])
tf.get_default_session().run(tf.global_variables_initializer())
for step in range(10):
print(step, do_update())
tf.set_random_seed(0)
tf.get_default_session().run(tf.global_variables_initializer())
var_list = [a_var, b_var]
lossandgrad = tf_utils.function(
[], [loss, tf_utils.flatgrad(loss, var_list)], updates=[update_op])
adam = MpiAdam(var_list)
for step in range(10):
loss, grad = lossandgrad()
adam.update(grad, learning_rate)
print(step, loss)
def __init__(self, env, hidden_size, entcoeff=0.001, scope="adversary"):
"""
reward regression from observations and transitions
:param env: (Gym Environment)
:param hidden_size: ([int]) the hidden dimension for the MLP
:param entcoeff: (float) the entropy loss weight
:param scope: (str) tensorflow variable scope
"""
self.scope = scope
self.observation_shape = env.observation_space.shape
self.actions_shape = env.action_space.shape
self.input_shape = tuple([
o + a for o, a in zip(self.observation_shape, self.actions_shape)
])
self.num_actions = env.action_space.shape[0]
self.hidden_size = hidden_size
self.build_ph()
# Build grpah
generator_logits = self.build_graph(self.generator_obs_ph,
self.generator_acs_ph,
reuse=False)
expert_logits = self.build_graph(self.expert_obs_ph,
self.expert_acs_ph,
reuse=True)
# Build accuracy
generator_acc = tf.reduce_mean(
tf.to_float(tf.nn.sigmoid(generator_logits) < 0.5))
expert_acc = tf.reduce_mean(
tf.to_float(tf.nn.sigmoid(expert_logits) > 0.5))
# Build regression loss
# let x = logits, z = targets.
# z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))
generator_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=generator_logits, labels=tf.zeros_like(generator_logits))
generator_loss = tf.reduce_mean(generator_loss)
expert_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=expert_logits, labels=tf.ones_like(expert_logits))
expert_loss = tf.reduce_mean(expert_loss)
# Build entropy loss
logits = tf.concat([generator_logits, expert_logits], 0)
entropy = tf.reduce_mean(logit_bernoulli_entropy(logits))
entropy_loss = -entcoeff * entropy
# Loss + Accuracy terms
self.losses = [
generator_loss, expert_loss, entropy, entropy_loss, generator_acc,
expert_acc
]
self.loss_name = [
"generator_loss", "expert_loss", "entropy", "entropy_loss",
"generator_acc", "expert_acc"
]
self.total_loss = generator_loss + expert_loss + entropy_loss
# Build Reward for policy
self.reward_op = -tf.log(1 - tf.nn.sigmoid(generator_logits) + 1e-8)
var_list = self.get_trainable_variables()
self.lossandgrad = tf_util.function([
self.generator_obs_ph, self.generator_acs_ph, self.expert_obs_ph,
self.expert_acs_ph
], self.losses + [tf_util.flatgrad(self.total_loss, var_list)])
def _setup_critic_optimizer(self):
"""
setup the optimizer for the critic
"""
if self.verbose >= 2:
logger.info('setting up critic optimizer')
normalized_critic_target_tf = tf.clip_by_value(normalize(self.critic_target, self.ret_rms),
self.return_range[0], self.return_range[1])
self.critic_loss = tf.reduce_mean(tf.square(self.normalized_critic_tf - normalized_critic_target_tf))
if self.critic_l2_reg > 0.:
critic_reg_vars = [var for var in tf_util.get_trainable_vars('model/qf/')
if 'bias' not in var.name and 'output' not in var.name and 'b' not in var.name]
if self.verbose >= 2:
for var in critic_reg_vars:
logger.info(' regularizing: {}'.format(var.name))
logger.info(' applying l2 regularization with {}'.format(self.critic_l2_reg))
critic_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.critic_l2_reg),
weights_list=critic_reg_vars
)
self.critic_loss += critic_reg
critic_shapes = [var.get_shape().as_list() for var in tf_util.get_trainable_vars('model/qf/')]
critic_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
if self.verbose >= 2:
logger.info(' critic shapes: {}'.format(critic_shapes))
logger.info(' critic params: {}'.format(critic_nb_params))
self.critic_grads = tf_util.flatgrad(self.critic_loss, tf_util.get_trainable_vars('model/qf/'),
clip_norm=self.clip_norm)
self.critic_optimizer = MpiAdam(var_list=tf_util.get_trainable_vars('model/qf/'), beta1=0.9, beta2=0.999,
epsilon=1e-08)
def _setup_critic_optimizer(self):
"""
setup the optimizer for the critic
"""
if self.verbose >= 2:
logger.info('setting up critic optimizer')
### BSS LOSS ###
all_vars = [v for v in tf.global_variables()]
self.l2_loss = 0.0
for var in all_vars:
if 'qf' in var.name:
self.l2_loss += tf.losses.mean_squared_error(
tf.zeros(var.shape), var)
_, qf_features = self.policy_tf.feature_matrices()
singular_qf = tf.linalg.svd(qf_features, compute_uv=False)
self.bss_loss = tf.reduce_sum(tf.square(singular_qf[-1]))
### BSS LOSS ###
normalized_critic_target_tf = tf.clip_by_value(
normalize(self.critic_target, self.ret_rms), self.return_range[0],
self.return_range[1])
self.critic_loss = tf.reduce_mean(tf.square(self.normalized_critic_tf - normalized_critic_target_tf)) + \
self.bss_coef * self.bss_loss + self.l2_coef * self.l2_loss
if self.critic_l2_reg > 0.:
critic_reg_vars = [
var for var in tf_util.get_trainable_vars('model/qf/')
if 'bias' not in var.name and 'qf_output' not in var.name
and 'b' not in var.name
]
if self.verbose >= 2:
for var in critic_reg_vars:
logger.info(' regularizing: {}'.format(var.name))
logger.info(' applying l2 regularization with {}'.format(
self.critic_l2_reg))
critic_reg = tc.layers.apply_regularization(
tc.layers.l2_regularizer(self.critic_l2_reg),
weights_list=critic_reg_vars)
self.critic_loss += critic_reg
critic_shapes = [
var.get_shape().as_list()
for var in tf_util.get_trainable_vars('model/qf/')
]
critic_nb_params = sum(
[reduce(lambda x, y: x * y, shape) for shape in critic_shapes])
if self.verbose >= 2:
logger.info(' critic shapes: {}'.format(critic_shapes))
logger.info(' critic params: {}'.format(critic_nb_params))
self.critic_grads = tf_util.flatgrad(
self.critic_loss,
tf_util.get_trainable_vars('model/qf/'),
clip_norm=self.clip_norm)
self.critic_optimizer = MpiAdam(
var_list=tf_util.get_trainable_vars('model/qf/'),
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
def _setup_actor_optimizer(self):
"""
setup the optimizer for the actor
"""
if self.verbose >= 2:
logger.info('setting up actor optimizer')
self.actor_loss = -tf.reduce_mean(self.critic_with_actor_tf)
actor_shapes = [var.get_shape().as_list() for var in tf_util.get_trainable_vars('model/pi/')]
actor_nb_params = sum([reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
if self.verbose >= 2:
logger.info(' actor shapes: {}'.format(actor_shapes))
logger.info(' actor params: {}'.format(actor_nb_params))
self.actor_grads = tf_util.flatgrad(self.actor_loss, tf_util.get_trainable_vars('model/pi/'),
clip_norm=self.clip_norm)
self.actor_optimizer = MpiAdam(var_list=tf_util.get_trainable_vars('model/pi/'), beta1=0.9, beta2=0.999,
epsilon=1e-08)
def _setup_actor_optimizer(self):
"""
setup the optimizer for the actor
"""
if self.verbose >= 2:
logger.info('setting up actor optimizer')
### BSS LOSS ###
all_vars = [v for v in tf.global_variables()]
self.l2_loss = 0.0
for var in all_vars:
if 'pi' in var.name:
self.l2_loss += tf.losses.mean_squared_error(
tf.zeros(var.shape), var)
pi_features, _ = self.policy_tf.feature_matrices()
singular_pi = tf.linalg.svd(pi_features, compute_uv=False)
self.bss_loss = tf.reduce_sum(tf.square(singular_pi[-1]))
### BSS LOSS ###
self.actor_loss = -tf.reduce_mean(self.critic_with_actor_tf) + \
self.bss_coef * self.bss_loss + self.l2_coef * self.l2_loss
actor_shapes = [
var.get_shape().as_list()
for var in tf_util.get_trainable_vars('model/pi/')
]
actor_nb_params = sum(
[reduce(lambda x, y: x * y, shape) for shape in actor_shapes])
if self.verbose >= 2:
logger.info(' actor shapes: {}'.format(actor_shapes))
logger.info(' actor params: {}'.format(actor_nb_params))
self.actor_grads = tf_util.flatgrad(
self.actor_loss,
tf_util.get_trainable_vars('model/pi/'),
clip_norm=self.clip_norm)
self.actor_optimizer = MpiAdam(
var_list=tf_util.get_trainable_vars('model/pi/'),
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
def learn(
env,
policy_func,
*,
timesteps_per_batch, # timesteps per actor per update
clip_param,
entcoeff, # clipping parameter epsilon, entropy coeff
optim_epochs,
optim_stepsize,
optim_batchsize, # optimization hypers
gamma,
lam, # advantage estimation
max_timesteps=0,
max_episodes=0,
max_iters=0,
max_seconds=0, # time constraint
callback=None, # you can do anything in the callback, since it takes locals(), globals()
adam_epsilon=1e-5,
schedule='constant', # annealing for stepsize parameters (epsilon and adam)
num_options=2,
app='',
saves=False,
wsaves=False,
epoch=-1,
seed=1,
dc=0):
optim_batchsize_ideal = optim_batchsize
np.random.seed(seed)
tf.set_random_seed(seed)
# env._seed(seed)
gamename = env.spec.id[:-3].lower()
gamename += 'seed' + str(seed)
gamename += app
dirname = '{}_{}opts_saves/'.format(gamename, num_options)
if wsaves:
first = True
if not os.path.exists(dirname):
os.makedirs(dirname)
first = False
# while os.path.exists(dirname) and first:
# dirname += '0'
files = ['pposgd_simple.py', 'mlp_policy.py', 'run_main.py']
for i in range(len(files)):
src = os.path.expanduser('~/baselines/baselines/ppo1/') + files[i]
dest = os.path.expanduser('~/baselines/baselines/ppo1/') + dirname
shutil.copy2(src, dest)
# ----------------------------------------
ob_space = env.observation_space
ac_space = env.action_space
pi = policy_func("pi", ob_space,
ac_space) # Construct network for new policy
oldpi = policy_func("oldpi", ob_space, ac_space) # Network for old policy
atarg = tf.placeholder(
dtype=tf.float32,
shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
# option = tf.placeholder(dtype=tf.int32, shape=[None])
lrmult = tf.placeholder(
name='lrmult', dtype=tf.float32,
shape=[]) # learning rate multiplier, updated with schedule
clip_param = clip_param * lrmult # Annealed cliping parameter epislon
# pdb.set_trace()
ob = U.get_placeholder_cached(name="ob")
option = U.get_placeholder_cached(name="option")
term_adv = U.get_placeholder(name='term_adv',
dtype=tf.float32,
shape=[None])
ac = pi.pdtype.sample_placeholder([None])
kloldnew = oldpi.pd.kl(pi.pd)
ent = pi.pd.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
pol_entpen = (-entcoeff) * meanent
ratio = tf.exp(pi.pd.logp(ac) - oldpi.pd.logp(ac)) # pnew / pold
surr1 = ratio * atarg # surrogate from conservative policy iteration
surr2 = tf.clip_by_value(ratio, 1.0 - clip_param,
1.0 + clip_param) * atarg #
pol_surr = -tf.reduce_mean(tf.minimum(
surr1, surr2)) # PPO's pessimistic surrogate (L^CLIP)
vf_loss = tf.reduce_mean(tf.square(pi.vpred - ret))
total_loss = pol_surr + pol_entpen + vf_loss
losses = [pol_surr, pol_entpen, vf_loss, meankl, meanent]
loss_names = ["pol_surr", "pol_entpen", "vf_loss", "kl", "ent"]
term_loss = pi.tpred * term_adv
log_pi = tf.log(tf.clip_by_value(pi.op_pi, 1e-20, 1.0))
entropy = -tf.reduce_sum(pi.op_pi * log_pi, reduction_indices=1)
op_loss = -tf.reduce_sum(log_pi[0][option[0]] * atarg + entropy * 0.1)
total_loss += op_loss
var_list = pi.get_trainable_variables()
term_list = var_list[6:8]
lossandgrad = U.function([ob, ac, atarg, ret, lrmult, option, term_adv],
losses + [U.flatgrad(total_loss, var_list)])
termloss = U.function([ob, option, term_adv],
[U.flatgrad(term_loss, var_list)
]) # Since we will use a different step size.
adam = MpiAdam(var_list, epsilon=adam_epsilon)
assign_old_eq_new = U.function(
[], [],
updates=[
tf.assign(oldv, newv)
for (oldv,
newv) in zipsame(oldpi.get_variables(), pi.get_variables())
])
compute_losses = U.function([ob, ac, atarg, ret, lrmult, option], losses)
U.initialize()
adam.sync()
saver = tf.train.Saver(max_to_keep=10000)
results = []
if saves:
results = open(
gamename + '_' + str(num_options) + 'opts_' + '_results.csv', 'w')
out = 'epoch,avg_reward'
for opt in range(num_options):
out += ',option {} dur'.format(opt)
for opt in range(num_options):
out += ',option {} std'.format(opt)
for opt in range(num_options):
out += ',option {} term'.format(opt)
for opt in range(num_options):
out += ',option {} adv'.format(opt)
out += '\n'
results.write(out)
# results.write('epoch,avg_reward,option 1 dur, option 2 dur, option 1 term, option 2 term\n')
results.flush()
if epoch >= 0:
dirname = '{}_{}opts_saves/'.format(gamename, num_options)
print("Loading weights from iteration: " + str(epoch))
filename = dirname + '{}_epoch_{}.ckpt'.format(gamename, epoch)
saver.restore(U.get_session(), filename)
episodes_so_far = 0
timesteps_so_far = 0
global iters_so_far
iters_so_far = 0
tstart = time.time()
lenbuffer = deque(maxlen=100) # rolling buffer for episode lengths
rewbuffer = deque(maxlen=100) # rolling buffer for episode rewards
assert sum(
[max_iters > 0, max_timesteps > 0, max_episodes > 0,
max_seconds > 0]) == 1, "Only one time constraint permitted"
seg_gen = traj_segment_generator(pi,
env,
timesteps_per_batch,
stochastic=True,
num_options=num_options,
saves=saves,
results=results,
rewbuffer=rewbuffer,
dc=dc)
datas = [0 for _ in range(num_options)]
while True:
if callback: callback(locals(), globals())
if max_timesteps and timesteps_so_far >= max_timesteps:
break
elif max_episodes and episodes_so_far >= max_episodes:
break
elif max_iters and iters_so_far >= max_iters:
break
elif max_seconds and time.time() - tstart >= max_seconds:
break
if schedule == 'constant':
cur_lrmult = 1.0
elif schedule == 'linear':
cur_lrmult = max(1.0 - float(timesteps_so_far) / max_timesteps, 0)
else:
raise NotImplementedError
logger.log("********** Iteration %i ************" % iters_so_far)
seg = seg_gen.__next__()
add_vtarg_and_adv(seg, gamma, lam)
opt_d = []
for i in range(num_options):
dur = np.mean(
seg['opt_dur'][i]) if len(seg['opt_dur'][i]) > 0 else 0.
opt_d.append(dur)
std = []
for i in range(num_options):
logstd = np.mean(
seg['logstds'][i]) if len(seg['logstds'][i]) > 0 else 0.
std.append(np.exp(logstd))
print("mean opt dur:", opt_d)
print("mean op pol:", np.mean(np.array(seg['optpol_p']), axis=0))
print("mean term p:", np.mean(np.array(seg['term_p']), axis=0))
print("mean value val:", np.mean(np.array(seg['value_val']), axis=0))
ob, ac, opts, atarg, tdlamret = seg["ob"], seg["ac"], seg["opts"], seg[
"adv"], seg["tdlamret"]
vpredbefore = seg["vpred"] # predicted value function before udpate
atarg = (atarg - atarg.mean()
) / atarg.std() # standardized advantage function estimate
if hasattr(pi, "ob_rms"):
pi.ob_rms.update(ob) # update running mean/std for policy
assign_old_eq_new() # set old parameter values to new parameter values
if iters_so_far % 5 == 0 and wsaves:
print("weights are saved...")
filename = dirname + '{}_epoch_{}.ckpt'.format(
gamename, iters_so_far)
save_path = saver.save(U.get_session(), filename)
min_batch = 160
t_advs = [[] for _ in range(num_options)]
for opt in range(num_options):
indices = np.where(opts == opt)[0]
print("batch size:", indices.size)
opt_d[opt] = indices.size
if not indices.size:
t_advs[opt].append(0.)
continue
# This part is only necessasry when we use options.
# We proceed to these verifications in order not to discard any collected trajectories.
if datas[opt] != 0:
if (indices.size < min_batch and datas[opt].n > min_batch):
datas[opt] = Dataset(dict(ob=ob[indices],
ac=ac[indices],
atarg=atarg[indices],
vtarg=tdlamret[indices]),
shuffle=not pi.recurrent)
t_advs[opt].append(0.)
continue
elif indices.size + datas[opt].n < min_batch:
# pdb.set_trace()
oldmap = datas[opt].data_map
cat_ob = np.concatenate((oldmap['ob'], ob[indices]))
cat_ac = np.concatenate((oldmap['ac'], ac[indices]))
cat_atarg = np.concatenate(
(oldmap['atarg'], atarg[indices]))
cat_vtarg = np.concatenate(
(oldmap['vtarg'], tdlamret[indices]))
datas[opt] = Dataset(dict(ob=cat_ob,
ac=cat_ac,
atarg=cat_atarg,
vtarg=cat_vtarg),
shuffle=not pi.recurrent)
t_advs[opt].append(0.)
continue
elif (indices.size + datas[opt].n > min_batch and datas[opt].n
< min_batch) or (indices.size > min_batch
and datas[opt].n < min_batch):
oldmap = datas[opt].data_map
cat_ob = np.concatenate((oldmap['ob'], ob[indices]))
cat_ac = np.concatenate((oldmap['ac'], ac[indices]))
cat_atarg = np.concatenate(
(oldmap['atarg'], atarg[indices]))
cat_vtarg = np.concatenate(
(oldmap['vtarg'], tdlamret[indices]))
datas[opt] = d = Dataset(dict(ob=cat_ob,
ac=cat_ac,
atarg=cat_atarg,
vtarg=cat_vtarg),
shuffle=not pi.recurrent)
if (indices.size > min_batch and datas[opt].n > min_batch):
datas[opt] = d = Dataset(dict(ob=ob[indices],
ac=ac[indices],
atarg=atarg[indices],
vtarg=tdlamret[indices]),
shuffle=not pi.recurrent)
elif datas[opt] == 0:
datas[opt] = d = Dataset(dict(ob=ob[indices],
ac=ac[indices],
atarg=atarg[indices],
vtarg=tdlamret[indices]),
shuffle=not pi.recurrent)
optim_batchsize = optim_batchsize or ob.shape[0]
optim_epochs = np.clip(
np.int(10 * (indices.size /
(timesteps_per_batch / num_options))), 10,
10) if num_options > 1 else optim_epochs
print("optim epochs:", optim_epochs)
logger.log("Optimizing...")
# Here we do a bunch of optimization epochs over the data
for _ in range(optim_epochs):
losses = [
] # list of tuples, each of which gives the loss for a minibatch
for batch in d.iterate_once(optim_batchsize):
tadv, nodc_adv = pi.get_term_adv(batch["ob"], [opt])
tadv = tadv if num_options > 1 else np.zeros_like(tadv)
t_advs[opt].append(nodc_adv)
*newlosses, grads = lossandgrad(batch["ob"], batch["ac"],
batch["atarg"],
batch["vtarg"], cur_lrmult,
[opt], tadv)
termg = termloss(batch["ob"], [opt], tadv)
adam.update(termg[0], 5e-7 * cur_lrmult)
adam.update(grads, optim_stepsize * cur_lrmult)
losses.append(newlosses)
lrlocal = (seg["ep_lens"], seg["ep_rets"]) # local values
listoflrpairs = MPI.COMM_WORLD.allgather(lrlocal) # list of tuples
lens, rews = map(flatten_lists, zip(*listoflrpairs))
lenbuffer.extend(lens)
rewbuffer.extend(rews)
logger.record_tabular("EpLenMean", np.mean(lenbuffer))
logger.record_tabular("EpRewMean", np.mean(rewbuffer))
logger.record_tabular("EpThisIter", len(lens))
episodes_so_far += len(lens)
timesteps_so_far += sum(lens)
iters_so_far += 1
logger.record_tabular("EpisodesSoFar", episodes_so_far)
logger.record_tabular("TimestepsSoFar", timesteps_so_far)
logger.record_tabular("TimeElapsed", time.time() - tstart)
if MPI.COMM_WORLD.Get_rank() == 0:
logger.dump_tabular()
if saves:
out = "{},{}"
for _ in range(num_options):
out += ",{},{},{},{}"
out += "\n"
info = [iters_so_far, np.mean(rewbuffer)]
for i in range(num_options):
info.append(opt_d[i])
for i in range(num_options):
info.append(std[i])
for i in range(num_options):
info.append(np.mean(np.array(seg['term_p']), axis=0)[i])
for i in range(num_options):
info.append(np.mean(t_advs[i]))
results.write(out.format(*info))
results.flush()
def __init__(self,
observation_space,
action_space,
hidden_size,
entcoeff=0.001,
scope="adversary",
normalize=True):
"""
Reward regression from observations and transitions
:param observation_space: (gym.spaces)
:param action_space: (gym.spaces)
:param hidden_size: ([int]) the hidden dimension for the MLP
:param entcoeff: (float) the entropy loss weight
:param scope: (str) tensorflow variable scope
:param normalize: (bool) Whether to normalize the reward or not
"""
# TODO: support images properly (using a CNN)
self.scope = scope
self.observation_shape = observation_space.shape
self.actions_shape = action_space.shape
if isinstance(action_space, gym.spaces.Box):
# Continuous action space
self.discrete_actions = False
self.n_actions = action_space.shape[0]
elif isinstance(action_space, gym.spaces.Discrete):
self.n_actions = action_space.n
self.discrete_actions = True
else:
raise ValueError(
'Action space not supported: {}'.format(action_space))
self.hidden_size = hidden_size
self.normalize = normalize
self.obs_rms = None
# Placeholders
self.generator_obs_ph = tf.compat.v1.placeholder(
observation_space.dtype, (None, ) + self.observation_shape,
name="observations_ph")
self.generator_acs_ph = tf.compat.v1.placeholder(
action_space.dtype, (None, ) + self.actions_shape,
name="actions_ph")
self.expert_obs_ph = tf.compat.v1.placeholder(
observation_space.dtype, (None, ) + self.observation_shape,
name="expert_observations_ph")
self.expert_acs_ph = tf.compat.v1.placeholder(
action_space.dtype, (None, ) + self.actions_shape,
name="expert_actions_ph")
# Build graph
generator_logits = self.build_graph(self.generator_obs_ph,
self.generator_acs_ph,
reuse=False)
expert_logits = self.build_graph(self.expert_obs_ph,
self.expert_acs_ph,
reuse=True)
# Build accuracy
generator_acc = tf.reduce_mean(input_tensor=tf.cast(
tf.nn.sigmoid(generator_logits) < 0.5, tf.float32))
expert_acc = tf.reduce_mean(input_tensor=tf.cast(
tf.nn.sigmoid(expert_logits) > 0.5, tf.float32))
# Build regression loss
# let x = logits, z = targets.
# z * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))
generator_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=generator_logits, labels=tf.zeros_like(generator_logits))
generator_loss = tf.reduce_mean(input_tensor=generator_loss)
expert_loss = tf.nn.sigmoid_cross_entropy_with_logits(
logits=expert_logits, labels=tf.ones_like(expert_logits))
expert_loss = tf.reduce_mean(input_tensor=expert_loss)
# Build entropy loss
logits = tf.concat([generator_logits, expert_logits], 0)
entropy = tf.reduce_mean(input_tensor=logit_bernoulli_entropy(logits))
entropy_loss = -entcoeff * entropy
# Loss + Accuracy terms
self.losses = [
generator_loss, expert_loss, entropy, entropy_loss, generator_acc,
expert_acc
]
self.loss_name = [
"generator_loss", "expert_loss", "entropy", "entropy_loss",
"generator_acc", "expert_acc"
]
self.total_loss = generator_loss + expert_loss + entropy_loss
# Build Reward for policy
self.reward_op = -tf.math.log(1 - tf.nn.sigmoid(generator_logits) +
1e-8)
var_list = self.get_trainable_variables()
self.lossandgrad = tf_util.function([
self.generator_obs_ph, self.generator_acs_ph, self.expert_obs_ph,
self.expert_acs_ph
], self.losses + [tf_util.flatgrad(self.total_loss, var_list)])
def setup_model(self):
# prevent import loops
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the TRPO model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.nworkers = MPI.COMM_WORLD.Get_size()
print("number of workers are", self.nworkers)
self.rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.single_threaded_session(graph=self.graph)
self._setup_learn(self.seed)
# Construct network for new policy
self.policy_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
old_policy = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
# Network for phi
with tf.variable_scope("phi", reuse=False):
self.policy_phi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
# Network for phi old
with tf.variable_scope("oldphi", reuse=False):
self.policy_phi_old = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
atarg = tf.placeholder(dtype=tf.float32, shape=[
None
]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32,
shape=[None]) # Empirical return
observation = self.policy_pi.obs_ph
action = self.policy_pi.pdtype.sample_placeholder([None])
kloldnew = old_policy.proba_distribution.kl(
self.policy_pi.proba_distribution)
#kloldnew = self.policy_pi.proba_distribution.kl(old_policy.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = self.entcoeff * meanent
vferr = tf.reduce_mean(
tf.square(self.policy_pi.value_flat - ret))
vf_phi_err = tf.reduce_mean(
tf.square(self.policy_phi.value_flat - ret))
vf_phi_old_err = tf.reduce_mean(
tf.square(self.policy_phi_old.value_flat))
# advantage * pnew / pold
ratio = tf.exp(
self.policy_pi.proba_distribution.logp(action) -
old_policy.proba_distribution.logp(action))
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
self.loss_names = [
"optimgain", "meankl", "entloss", "surrgain", "entropy"
]
dist = meankl
all_var_list = tf_util.get_trainable_vars("model")
var_list = [
v for v in all_var_list
if "/vf" not in v.name and "/q/" not in v.name
]
vf_var_list = [
v for v in all_var_list
if "/pi" not in v.name and "/logstd" not in v.name
]
all_var_oldpi_list = tf_util.get_trainable_vars("oldpi")
var_oldpi_list = [
v for v in all_var_oldpi_list
if "/vf" not in v.name and "/q/" not in v.name
]
all_var_phi_list = tf_util.get_trainable_vars("phi")
vf_phi_var_list = [
v for v in all_var_phi_list if "/pi" not in v.name
and "/logstd" not in v.name and "/q" not in v.name
]
all_var_phi_old_list = tf_util.get_trainable_vars("oldphi")
vf_phi_old_var_list = [
v for v in all_var_phi_old_list if "/pi" not in v.name
and "/logstd" not in v.name and "/q" not in v.name
]
#print("vars", vf_var_list)
self.policy_vars = all_var_list
self.oldpolicy_vars = all_var_oldpi_list
print("all var list", all_var_list)
print("phi vars", vf_phi_var_list)
print("phi old vars", vf_phi_old_var_list)
self.get_flat = tf_util.GetFlat(var_list, sess=self.sess)
self.set_from_flat = tf_util.SetFromFlat(var_list,
sess=self.sess)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32,
shape=[None],
name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
var_size = tf_util.intprod(shape)
tangents.append(
tf.reshape(flat_tangent[start:start + var_size],
shape))
start += var_size
gvp = tf.add_n([
tf.reduce_sum(grad * tangent)
for (grad, tangent) in zipsame(klgrads, tangents)
]) # pylint: disable=E1111
fvp = tf_util.flatgrad(gvp, var_list)
tf.summary.scalar('entropy_loss', meanent)
tf.summary.scalar('policy_gradient_loss', optimgain)
tf.summary.scalar('value_function_loss', surrgain)
tf.summary.scalar('approximate_kullback-leibler', meankl)
tf.summary.scalar(
'loss',
optimgain + meankl + entbonus + surrgain + meanent)
self.assign_old_eq_new = \
tf_util.function([], [], updates=[tf.assign(oldv, newv) for (oldv, newv) in
zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))])
self.compute_losses = tf_util.function(
[observation, old_policy.obs_ph, action, atarg],
losses)
self.compute_fvp = tf_util.function([
flat_tangent, observation, old_policy.obs_ph, action,
atarg
], fvp)
self.compute_vflossandgrad = tf_util.function(
[observation, old_policy.obs_ph, ret],
tf_util.flatgrad(vferr, vf_var_list))
self.compute_vf_phi_lossandgrad = tf_util.function(
[observation, self.policy_phi.obs_ph, ret],
tf_util.flatgrad(vf_phi_err, vf_phi_var_list))
self.compute_vf_loss = tf_util.function(
[observation, old_policy.obs_ph, ret], vferr)
self.compute_vf_phi_loss = tf_util.function(
[observation, self.policy_phi.obs_ph, ret], vf_phi_err)
#self.compute_vf_phi_old_loss = tf_util.function([self.policy_phi_old.obs_ph], vf_phi_old_err)
#self.phi_old_obs = np.array([-0.012815 , -0.02076313, 0.07524705, 0.09407324, 0.0901745 , -0.09339058, 0.03544853, -0.03297224])
#self.phi_old_obs = self.phi_old_obs.reshape((1, 8))
update_phi_old_expr = []
for var, var_target in zip(
sorted(vf_phi_var_list, key=lambda v: v.name),
sorted(vf_phi_old_var_list, key=lambda v: v.name)):
update_phi_old_expr.append(var_target.assign(var))
update_phi_old_expr = tf.group(*update_phi_old_expr)
self.update_phi_old = tf_util.function(
[], [], updates=[update_phi_old_expr])
@contextmanager
def timed(msg):
if self.rank == 0 and self.verbose >= 1:
print(colorize(msg, color='magenta'))
start_time = time.time()
yield
print(
colorize("done in {:.3f} seconds".format(
(time.time() - start_time)),
color='magenta'))
else:
yield
@contextmanager
def temp_seed(seed):
state = np.random.get_state()
np.random.seed(seed)
try:
yield
finally:
np.random.set_state(state)
def allmean(arr):
assert isinstance(arr, np.ndarray)
out = np.empty_like(arr)
MPI.COMM_WORLD.Allreduce(arr, out, op=MPI.SUM)
out /= self.nworkers
return out
tf_util.initialize(sess=self.sess)
th_init = self.get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
self.set_from_flat(th_init)
with tf.variable_scope("Adam_mpi", reuse=False):
self.vfadam = MpiAdam(vf_var_list, sess=self.sess)
self.vf_phi_adam = MpiAdam(vf_phi_var_list, sess=self.sess)
self.vfadam.sync()
self.vf_phi_adam.sync()
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards',
tf.reduce_mean(ret))
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.vf_stepsize))
tf.summary.scalar('advantage', tf.reduce_mean(atarg))
tf.summary.scalar('kl_clip_range',
tf.reduce_mean(self.max_kl))
self.timed = timed
self.allmean = allmean
self.temp_seed = temp_seed
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
self.params = tf_util.get_trainable_vars(
"model") + tf_util.get_trainable_vars("oldpi")
self.summary = tf.summary.merge_all()
self.compute_lossandgrad = \
tf_util.function([observation, old_policy.obs_ph, action, atarg, ret],
[self.summary, tf_util.flatgrad(optimgain, var_list)] + losses)
def setup_model(self):
# prevent import loops
from stable_baselines.gail.adversary import TransitionClassifier
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the TRPO model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.nworkers = MPI.COMM_WORLD.Get_size()
self.rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess, graph=self.graph)
if self.using_gail:
self.reward_giver = TransitionClassifier(self.observation_space, self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
# Penalty related variable
with tf.variable_scope('penalty'):
cur_cost_ph = tf.placeholder(dtype=tf.float32, shape=[None]) # episodic cost
param_init = np.log(max(np.exp(self.penalty_init) - 1, 1e-8))
penalty_param = tf.get_variable('penalty_param',
initializer=float(param_init),
trainable=True,
dtype=tf.float32)
penalty = tf.nn.softplus(penalty_param)
penalty_loss = tf.reduce_mean(-penalty_param * (cur_cost_ph - self.cost_lim))
# Construct network for new policy
self.policy_pi = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
old_policy = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
# # Network for safety value function
# with tf.variable_Scope("vc",reuse=False):
# self.cost_value = MLPValue(self.sess, self.observation_spacem, self.n_envs, 1, None)
with tf.variable_scope("loss", reuse=False):
atarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
catarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target cost advantage function
cret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical cost
observation = self.policy_pi.obs_ph
action = self.policy_pi.pdtype.sample_placeholder([None])
kloldnew = old_policy.proba_distribution.kl(self.policy_pi.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = self.entcoeff * meanent
vferr = tf.reduce_mean(tf.square(self.policy_pi.value_flat - ret))
vcerr = tf.reduce_mean(tf.square(self.policy_pi.vcf_flat - cret))
# advantage * pnew / pold
ratio = tf.exp(self.policy_pi.proba_distribution.logp(action) -
old_policy.proba_distribution.logp(action))
surrgain = tf.reduce_mean(ratio * atarg)
# Surrogate for cost function
surrcost = tf.reduce_mean(ratio * catarg)
optimgain = surrgain + entbonus
# Include surr_cost in pi_objective
optimgain -= penalty * surrcost
optimgain /= (1 + penalty)
# # Loss function for pi is negative of pi_objective
# optimgain = -optimgain # Should we??
losses = [optimgain, meankl, entbonus, surrgain, meanent, surrcost]
self.loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy", "surrcost"]
dist = meankl
all_var_list = tf_util.get_trainable_vars("model")
var_list = [v for v in all_var_list if "/vf" not in v.name and "/q/" not in v.name and "/vcf" not in v.name] # policy parameters
vf_var_list = [v for v in all_var_list if "/pi" not in v.name and "/logstd" not in v.name and "/vcf" not in v.name] # value parameters
vcf_var_list = [v for v in all_var_list if "/pi" not in v.name and "/logstd" not in v.name and "/vf" not in v.name] # cost value parameters
self.get_flat = tf_util.GetFlat(var_list, sess=self.sess)
self.set_from_flat = tf_util.SetFromFlat(var_list, sess=self.sess)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32, shape=[None], name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
var_size = tf_util.intprod(shape)
tangents.append(tf.reshape(flat_tangent[start: start + var_size], shape))
start += var_size
gvp = tf.add_n([tf.reduce_sum(grad * tangent)
for (grad, tangent) in zipsame(klgrads, tangents)]) # pylint: disable=E1111
# Fisher vector products
fvp = tf_util.flatgrad(gvp, var_list)
tf.summary.scalar('penalty_loss', penalty_loss)
tf.summary.scalar('entropy_loss', meanent)
tf.summary.scalar('policy_gradient_loss', optimgain)
tf.summary.scalar('value_function_loss', surrgain)
tf.summary.scalar('constraint_cost_function_loss', surrcost)
tf.summary.scalar('approximate_kullback-leibler', meankl)
tf.summary.scalar('loss', optimgain + meankl + entbonus + surrgain + meanent + surrcost + penalty_loss)
self.assign_old_eq_new = \
tf_util.function([], [], updates=[tf.assign(oldv, newv) for (oldv, newv) in
zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))])
self.compute_losses = tf_util.function([observation, old_policy.obs_ph, action, atarg, catarg], losses)
self.compute_fvp = tf_util.function([flat_tangent, observation, old_policy.obs_ph, action, atarg, catarg],
fvp) # Why need all inputs? Might for implementation easiness
# self.compute_vflossandgrad = tf_util.function([observation, old_policy.obs_ph, ret],
# tf_util.flatgrad(vferr, vf_var_list)) # Why need old_policy.obs_ph? Doesn't seem to be used
# self.compute_vcflossandgrad = tf_util.function([observation, old_policy.obs_ph, cret],
# tf_util.flatgrad(vcerr, vcf_var_list))
self.compute_vflossandgrad = tf_util.function([observation, old_policy.obs_ph, ret, cret],
[tf_util.flatgrad(vferr, vf_var_list), tf_util.flatgrad(vcerr, vcf_var_list)])
self.compute_lagrangiangrad = tf_util.function([cur_cost_ph],
tf_util.flatgrad(penalty_loss, [penalty_param]))
@contextmanager
def timed(msg):
if self.rank == 0 and self.verbose >= 1:
print(colorize(msg, color='magenta'))
start_time = time.time()
yield
print(colorize("done in {:.3f} seconds".format((time.time() - start_time)),
color='magenta'))
else:
yield
def allmean(arr):
assert isinstance(arr, np.ndarray)
out = np.empty_like(arr)
MPI.COMM_WORLD.Allreduce(arr, out, op=MPI.SUM)
out /= self.nworkers
return out
tf_util.initialize(sess=self.sess)
th_init = self.get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
self.set_from_flat(th_init)
with tf.variable_scope("Adam_mpi", reuse=False):
self.vfadam = MpiAdam(vf_var_list, sess=self.sess)
if self.using_gail:
self.d_adam = MpiAdam(self.reward_giver.get_trainable_variables(), sess=self.sess)
self.d_adam.sync()
self.vfadam.sync()
# optimizer for constraint costs value function
self.vcadam = MpiAdam(vcf_var_list, sess=self.sess)
self.vcadam.sync()
# optimizer for lagragian value of safe RL
self.penaltyadam = MpiAdam([penalty_param], sess=self.sess)
self.penaltyadam.sync()
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards', tf.reduce_mean(ret))
tf.summary.scalar('discounted_costs', tf.reduce_mean(cret))
tf.summary.scalar('learning_rate', tf.reduce_mean(self.vf_stepsize))
tf.summary.scalar('advantage', tf.reduce_mean(atarg))
tf.summary.scalar('cost_advantage', tf.reduce_mean(catarg))
tf.summary.scalar('kl_clip_range', tf.reduce_mean(self.max_kl))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', ret)
tf.summary.histogram('discounted_rewards', cret)
tf.summary.histogram('learning_rate', self.vf_stepsize)
tf.summary.histogram('penalty_learning_rate', self.penalty_lr)
tf.summary.histogram('advantage', atarg)
tf.summary.histogram('cost_advantage', catarg)
tf.summary.histogram('kl_clip_range', self.max_kl)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', observation)
else:
tf.summary.histogram('observation', observation)
self.timed = timed
self.allmean = allmean
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
self.params = tf_util.get_trainable_vars("model") + tf_util.get_trainable_vars("oldpi")
if self.using_gail:
self.params.extend(self.reward_giver.get_trainable_variables())
self.summary = tf.summary.merge_all()
self.compute_lossandgrad = \
tf_util.function([observation, old_policy.obs_ph, action, atarg, catarg, ret, cret, cur_cost_ph],
[self.summary, tf_util.flatgrad(optimgain, var_list)] + losses)
def setup_model(self):
# prevent import loops
from stable_baselines.gail.adversary import TransitionClassifier
from stable_baselines.mdal.adversary import TabularAdversaryTF, NeuralAdversaryTRPO
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the MDPO model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.nworkers = MPI.COMM_WORLD.Get_size()
self.rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.single_threaded_session(graph=self.graph)
# self._setup_learn(self.seed)
self._setup_learn()
if self.using_gail:
self.reward_giver = TransitionClassifier(self.observation_space, self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
elif self.using_mdal:
if self.neural:
self.reward_giver = NeuralAdversaryTRPO(self.sess, self.observation_space, self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
else:
self.reward_giver = TabularAdversaryTF(self.sess, self.observation_space, self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff,
expert_features=self.expert_dataset.successor_features,
exploration_bonus=self.exploration_bonus,
bonus_coef=self.bonus_coef,
t_c=self.t_c,
is_action_features=self.is_action_features)
# Construct network for new policy
self.policy_pi = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
self.old_policy = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
# Network for fitting closed form
with tf.variable_scope("closedpi", reuse=False):
self.closed_policy = self.policy(self.sess, self.observation_space, self.action_space, self.n_envs, 1,
None, reuse=False, **self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
self.atarg = tf.placeholder(dtype=tf.float32, shape=[None]) # Target advantage function (if applicable)
self.vtarg = tf.placeholder(dtype=tf.float32, shape=[None])
self.ret = tf.placeholder(dtype=tf.float32, shape=[None]) # Empirical return
self.learning_rate_ph = tf.placeholder(dtype=tf.float32, shape=[], name="learning_rate_ph")
self.outer_learning_rate_ph = tf.placeholder(dtype=tf.float32, shape=[], name="outer_learning_rate_ph")
self.old_vpred_ph = tf.placeholder(dtype=tf.float32, shape=[None], name="old_vpred_ph")
self.clip_range_vf_ph = tf.placeholder(dtype=tf.float32, shape=[], name="clip_range_ph")
observation = self.policy_pi.obs_ph
self.action = self.policy_pi.pdtype.sample_placeholder([None])
if self.tsallis_q == 1.0:
kloldnew = self.policy_pi.proba_distribution.kl(self.old_policy.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
else:
logp_pi = self.policy_pi.proba_distribution.logp(self.action)
logp_pi_old = self.old_policy.proba_distribution.logp(self.action)
ent = self.policy_pi.proba_distribution.entropy()
#kloldnew = self.policy_pi.proba_distribution.kl_tsallis(self.old_policy.proba_distribution, self.tsallis_q)
tsallis_q = 2.0 - self.tsallis_q
meankl = tf.reduce_mean(tf_log_q(tf.exp(logp_pi), tsallis_q) - tf_log_q(tf.exp(logp_pi_old), tsallis_q)) #tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = self.entcoeff * meanent
if self.cliprange_vf is None:
vpred_clipped = self.policy_pi.value_flat
else:
vpred_clipped = self.old_vpred_ph + \
tf.clip_by_value(self.policy_pi.value_flat - self.old_vpred_ph,
- self.clip_range_vf_ph, self.clip_range_vf_ph)
vf_losses1 = tf.square(self.policy_pi.value_flat - self.ret)
vf_losses2 = tf.square(vpred_clipped - self.ret)
vferr = tf.reduce_mean(tf.maximum(vf_losses1, vf_losses2))
# advantage * pnew / pold
ratio = tf.exp(self.policy_pi.proba_distribution.logp(self.action) -
self.old_policy.proba_distribution.logp(self.action))
if self.method == "multistep-SGD":
surrgain = tf.reduce_mean(ratio * self.atarg) - meankl / self.learning_rate_ph
elif self.method == "closedreverse-KL":
surrgain = tf.reduce_mean(tf.exp(self.atarg) * self.policy_pi.proba_distribution.logp(self.action))
else:
policygain = tf.reduce_mean(tf.exp(self.atarg) * tf.log(self.closed_policy.proba_distribution.mean))
surrgain = tf.reduce_mean(ratio * self.atarg) - tf.reduce_mean(self.learning_rate_ph * ratio * self.policy_pi.proba_distribution.logp(self.action))
optimgain = surrgain #+ entbonus - self.learning_rate_ph * meankl
losses = [optimgain, meankl, entbonus, surrgain, meanent]
self.loss_names = ["optimgain", "meankl", "entloss", "surrgain", "entropy"]
dist = meankl
all_var_list = tf_util.get_trainable_vars("model")
var_list = [v for v in all_var_list if "/vf" not in v.name and "/q/" not in v.name]
vf_var_list = [v for v in all_var_list if "/pi" not in v.name and "/logstd" not in v.name]
print("policy vars", var_list)
all_closed_var_list = tf_util.get_trainable_vars("closedpi")
closed_var_list = [v for v in all_closed_var_list if "/vf" not in v.name and "/q" not in v.name]
self.get_flat = tf_util.GetFlat(var_list, sess=self.sess)
self.set_from_flat = tf_util.SetFromFlat(var_list, sess=self.sess)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32, shape=[None], name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
var_size = tf_util.intprod(shape)
tangents.append(tf.reshape(flat_tangent[start: start + var_size], shape))
start += var_size
gvp = tf.add_n([tf.reduce_sum(grad * tangent)
for (grad, tangent) in zipsame(klgrads, tangents)]) # pylint: disable=E1111
fvp = tf_util.flatgrad(gvp, var_list)
# tf.summary.scalar('entropy_loss', meanent)
# tf.summary.scalar('policy_gradient_loss', optimgain)
# tf.summary.scalar('value_function_loss', surrgain)
# tf.summary.scalar('approximate_kullback-leibler', meankl)
# tf.summary.scalar('loss', optimgain + meankl + entbonus + surrgain + meanent)
self.assign_old_eq_new = \
tf_util.function([], [], updates=[tf.assign(oldv, newv) for (oldv, newv) in
zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))])
self.compute_losses = tf_util.function([observation, self.old_policy.obs_ph, self.action, self.atarg, self.learning_rate_ph, self.vtarg], losses)
self.compute_fvp = tf_util.function([flat_tangent, observation, self.old_policy.obs_ph, self.action, self.atarg],
fvp)
self.compute_vflossandgrad = tf_util.function([observation, self.old_policy.obs_ph, self.ret, self.old_vpred_ph, self.clip_range_vf_ph],
tf_util.flatgrad(vferr, vf_var_list))
grads = tf.gradients(-optimgain, var_list)
grads, _grad_norm = tf.clip_by_global_norm(grads, 0.5)
trainer = tf.train.AdamOptimizer(learning_rate=self.outer_learning_rate_ph, epsilon=1e-5)
# trainer = tf.train.AdamOptimizer(learning_rate=3e-4, epsilon=1e-5)
grads = list(zip(grads, var_list))
self._train = trainer.apply_gradients(grads)
@contextmanager
def timed(msg):
if self.rank == 0 and self.verbose >= 1:
# print(colorize(msg, color='magenta'))
# start_time = time.time()
yield
# print(colorize("done in {:.3f} seconds".format((time.time() - start_time)),
# color='magenta'))
else:
yield
def allmean(arr):
assert isinstance(arr, np.ndarray)
out = np.empty_like(arr)
MPI.COMM_WORLD.Allreduce(arr, out, op=MPI.SUM)
out /= self.nworkers
return out
tf_util.initialize(sess=self.sess)
th_init = self.get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
self.set_from_flat(th_init)
with tf.variable_scope("Adam_mpi", reuse=False):
self.vfadam = MpiAdam(vf_var_list, sess=self.sess)
if self.using_gail or self.using_mdal:
self.d_adam = MpiAdam(self.reward_giver.get_trainable_variables(), sess=self.sess)
self.d_adam.sync()
self.vfadam.sync()
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards', tf.reduce_mean(self.ret))
tf.summary.scalar('learning_rate', tf.reduce_mean(self.vf_stepsize))
tf.summary.scalar('advantage', tf.reduce_mean(self.atarg))
tf.summary.scalar('kl_clip_range', tf.reduce_mean(self.max_kl))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', self.ret)
tf.summary.histogram('learning_rate', self.vf_stepsize)
tf.summary.histogram('advantage', self.atarg)
tf.summary.histogram('kl_clip_range', self.max_kl)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', observation)
else:
tf.summary.histogram('observation', observation)
self.timed = timed
self.allmean = allmean
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
self.params = tf_util.get_trainable_vars("model") + tf_util.get_trainable_vars("oldpi")
if self.using_gail:
self.params.extend(self.reward_giver.get_trainable_variables())
self.summary = tf.summary.merge_all()
self.compute_lossandgrad = \
tf_util.function([observation, self.old_policy.obs_ph, self.action, self.atarg, self.ret, self.learning_rate_ph, self.vtarg, self.closed_policy.obs_ph],
[self.summary, tf_util.flatgrad(optimgain, var_list)] + losses)
def _setup_model(self, rank, memory_size, alpha, obs_space, a_space,
noise_target_action, **kwargs):
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.single_threaded_session(graph=self.graph)
if self.use_prioritiy:
from .priority_memory import PrioritizedMemory
self.memory = PrioritizedMemory(capacity=memory_size,
alpha=alpha)
else:
from .memory import Memory
self.memory = Memory(limit=memory_size,
action_shape=a_space.shape,
observation_shape=obs_space.shape)
# 定义 placeholders
self.observe_Input = tf.placeholder(tf.float32,
[None] + list(obs_space.shape),
name='observe_Input')
self.observe_Input_ = tf.placeholder(tf.float32, [None] +
list(obs_space.shape),
name='observe_Input_')
self.R = tf.placeholder(tf.float32, [None, 1], 'r')
self.terminals1 = tf.placeholder(tf.float32,
shape=(None, 1),
name='terminals1')
self.ISWeights = tf.placeholder(tf.float32, [None, 1],
name='IS_weights')
self.n_step_steps = tf.placeholder(tf.float32,
shape=(None, 1),
name='n_step_reached')
self.q_demo = tf.placeholder(tf.float32, [None, 1],
name='Q_of_actions_from_memory')
self.come_from_demo = tf.placeholder(tf.float32, [None, 1],
name='Demo_index')
self.action_memory = tf.placeholder(tf.float32,
[None] + list(a_space.shape),
name='actions_from_memory')
with tf.variable_scope('obs_rms'):
self.obs_rms = RunningMeanStd(shape=obs_space.shape)
with tf.name_scope('obs_preprocess'):
self.normalized_observe_Input = tf.clip_by_value(
normalize(self.observe_Input, self.obs_rms), -5., 5.)
self.normalized_observe_Input_ = tf.clip_by_value(
normalize(self.observe_Input_, self.obs_rms), -5., 5.)
with tf.variable_scope('Actor'):
self.action = self.build_actor(self.normalized_observe_Input,
scope='eval',
trainable=True,
a_space=a_space)
self.action_ = self.build_actor(self.normalized_observe_Input_,
scope='target',
trainable=False,
a_space=a_space)
# Target policy smoothing, by adding clipped noise to target actions
if noise_target_action:
epsilon = tf.random_normal(tf.shape(self.action_),
stddev=0.007)
epsilon = tf.clip_by_value(epsilon, -0.01, 0.01)
a2 = self.action_ + epsilon
noised_action_ = tf.clip_by_value(a2, -1, 1)
else:
noised_action_ = self.action_
with tf.variable_scope('Critic'):
# Q值都要被clip 防止过估计.
self.q_1 = tf.clip_by_value(
self.build_critic(self.normalized_observe_Input,
self.action,
scope='eval_1',
trainable=True), self.Q_value_range[0],
self.Q_value_range[1])
q_1_ = self.build_critic(self.normalized_observe_Input_,
noised_action_,
scope='target_1',
trainable=False)
if self.use_TD3:
q_2 = tf.clip_by_value(
self.build_critic(self.normalized_observe_Input,
self.action,
scope='eval_2',
trainable=True),
self.Q_value_range[0], self.Q_value_range[1])
q_2_ = self.build_critic(self.normalized_observe_Input_,
noised_action_,
scope='target_2',
trainable=False)
# Collect networks parameters. It would make it more easily to manage them.
self.ae_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='Actor/eval')
self.at_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='Actor/target')
self.ce1_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='Critic/eval_1')
self.ct1_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='Critic/target_1')
if self.use_TD3:
self.ce2_params = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, scope='Critic/eval_2')
self.ct2_params = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, scope='Critic/target_2')
with tf.variable_scope('Soft_Update'):
self.soft_replace_a = [
tf.assign(t, (1 - TAU) * t + TAU * e)
for t, e in zip(self.at_params, self.ae_params)
]
self.soft_replace_c = [
tf.assign(t, (1 - TAU) * t + TAU * e)
for t, e in zip(self.ct1_params, self.ce1_params)
]
if self.use_TD3:
self.soft_replace_c += [
tf.assign(t, (1 - TAU) * t + TAU * e)
for t, e in zip(self.ct2_params, self.ce2_params)
]
# critic 的误差 为 (one-step-td 误差 + n-step-td 误差 + critic_online 的L2惩罚)
# TD3: critic一共有4个, 算两套 critic的误差, 秀儿.
with tf.variable_scope('Critic_Lose'):
if self.use_TD3:
min_q_ = tf.minimum(q_1_, q_2_)
else:
min_q_ = q_1_
self.q_target = self.R + (1. -
self.terminals1) * GAMMA * min_q_
if self.use_n_step:
self.n_step_target_q = self.R + (
1. - self.terminals1) * tf.pow(
GAMMA, self.n_step_steps) * min_q_
cliped_n_step_target_q = tf.clip_by_value(
self.n_step_target_q, self.Q_value_range[0],
self.Q_value_range[1])
cliped_q_target = tf.clip_by_value(self.q_target,
self.Q_value_range[0],
self.Q_value_range[1])
self.td_error_1 = tf.abs(cliped_q_target - self.q_1)
if self.use_TD3:
self.td_error_2 = tf.abs(cliped_q_target - q_2)
if self.use_n_step:
self.nstep_td_error_1 = tf.abs(cliped_n_step_target_q -
self.q_1)
if self.use_TD3:
self.nstep_td_error_2 = tf.abs(cliped_n_step_target_q -
q_2)
L2_regular_1 = tf.contrib.layers.apply_regularization(
tf.contrib.layers.l2_regularizer(0.001),
weights_list=self.ce1_params)
if self.use_TD3:
L2_regular_2 = tf.contrib.layers.apply_regularization(
tf.contrib.layers.l2_regularizer(0.001),
weights_list=self.ce2_params)
one_step_losse_1 = tf.reduce_mean(
tf.multiply(self.ISWeights, tf.square(
self.td_error_1))) * self.lambda_1_step
if self.use_TD3:
one_step_losse_2 = tf.reduce_mean(
tf.multiply(self.ISWeights, tf.square(
self.td_error_2))) * self.lambda_1_step
if self.use_n_step:
n_step_td_losses_1 = tf.reduce_mean(
tf.multiply(
self.ISWeights, tf.square(
self.nstep_td_error_1))) * self.lambda_n_step
c_loss_1 = one_step_losse_1 + n_step_td_losses_1 + L2_regular_1
if self.use_TD3:
n_step_td_losses_2 = tf.reduce_mean(
tf.multiply(self.ISWeights,
tf.square(self.nstep_td_error_2))
) * self.lambda_n_step
c_loss_2 = one_step_losse_2 + n_step_td_losses_2 + L2_regular_2
else:
c_loss_1 = one_step_losse_1 + L2_regular_1
if self.use_TD3:
c_loss_2 = one_step_losse_2 + L2_regular_2
# actor 的 loss 为 最大化q(s,a) 最小化行为克隆误差.
# (只有demo的transition 且 demo的action 比 actor生成的action q_1(s,a)高的时候 才会有克隆误差)
with tf.variable_scope('Actor_lose'):
Is_worse_than_demo = self.q_1 < self.q_demo
Is_worse_than_demo = tf.cast(Is_worse_than_demo, tf.float32)
worse_than_demo = tf.cast(tf.reduce_sum(Is_worse_than_demo),
tf.int8)
# 算action误差 我用的是平方和, 也有人用均方误差 reduce_mean. 其实都可以.
# 我的action本来都是很小的数.
action_diffs = Is_worse_than_demo * tf.reduce_sum(
self.come_from_demo *
tf.square(self.action - self.action_memory),
1,
keepdims=True)
L_BC = self.LAMBDA_BC * tf.reduce_sum(action_diffs)
a_loss = -tf.reduce_mean(self.q_1) + L_BC
# Setting optimizer for Actor and Critic
with tf.variable_scope('Critic_Optimizer'):
if self.use_TD3:
self.critic_grads_1 = tf_util.flatgrad(
loss=c_loss_1, var_list=self.ce1_params)
self.critic_grads_2 = tf_util.flatgrad(
loss=c_loss_2, var_list=self.ce2_params)
self.critic_optimizer_1 = MpiAdam(var_list=self.ce1_params,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
self.critic_optimizer_2 = MpiAdam(var_list=self.ce2_params,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
else:
self.critic_grads = tf_util.flatgrad(
loss=c_loss_1, var_list=self.ce1_params)
self.critic_optimizer = MpiAdam(var_list=self.ce1_params,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
with tf.variable_scope('Actor_Optimizer'):
self.actor_grads = tf_util.flatgrad(a_loss, self.ae_params)
self.actor_optimizer = MpiAdam(var_list=self.ae_params,
beta1=0.9,
beta2=0.999,
epsilon=1e-08)
with self.sess.as_default():
self._initialize(self.sess)
# 保存模型
var_list = tf.global_variables()
print(
"var_list!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n"
)
for v in var_list:
print(v)
self.saver = tf.train.Saver(var_list=var_list, max_to_keep=1)
self.writer = tf.summary.FileWriter(
"logs/" + self.experiment_name + "/DDPG_" + str(rank),
self.graph)
# TensorBoard summary
self.a_summary = tf.summary.merge([
tf.summary.scalar('a_loss', a_loss, family='actor'),
tf.summary.scalar('L_BC', L_BC, family='actor'),
tf.summary.scalar('worse_than_demo',
worse_than_demo,
family='actor')
])
if self.use_TD3:
self.c_summary = tf.summary.merge([
tf.summary.scalar('c_loss_1', c_loss_1, family='critic'),
tf.summary.scalar('c_loss_2', c_loss_2, family='critic')
])
else:
self.c_summary = tf.summary.merge(
[tf.summary.scalar('c_loss_1', c_loss_1, family='critic')])
# episode summary
self.episode_cumulate_reward = tf.placeholder(
tf.float32, name='episode_cumulate_reward')
self.episoed_length = tf.placeholder(
tf.int16, name='episode_cumulate_reward')
self.success_or_not = tf.placeholder(
tf.int8, name='episode_cumulate_reward')
self.eval_episode_cumulate_reward = tf.placeholder(
tf.float32, name='episode_cumulate_reward')
self.eval_episoed_length = tf.placeholder(
tf.int16, name='episode_cumulate_reward')
self.eval_success_or_not = tf.placeholder(
tf.int8, name='episode_cumulate_reward')
self.episode_summary = tf.summary.merge([
tf.summary.scalar('episode_cumulate_reward',
self.episode_cumulate_reward,
family='episoed_result'),
tf.summary.scalar('episoed_length',
self.episoed_length,
family='episoed_result'),
tf.summary.scalar('success_or_not',
self.success_or_not,
family='episoed_result'),
])
self.eval_episode_summary = tf.summary.merge([
tf.summary.scalar('eval_episode_cumulate_reward',
self.eval_episode_cumulate_reward,
family='Eval_episoed_result'),
tf.summary.scalar('eval_episoed_length',
self.eval_episoed_length,
family='Eval_episoed_result'),
tf.summary.scalar('eval_success_or_not',
self.eval_success_or_not,
family='Eval_episoed_result'),
])
def learn(env,
policy_func,
dataset,
optim_batch_size=128,
max_iters=1e4,
adam_epsilon=1e-5,
optim_stepsize=3e-4,
ckpt_dir=None,
task_name=None,
verbose=False):
"""
Learn a behavior clone policy, and return the save location
:param env: (Gym Environment) the environment
:param policy_func: (function (str, Gym Space, Gym Space): TensorFlow Tensor) creates the policy
:param dataset: (Dset or MujocoDset) the dataset manager
:param optim_batch_size: (int) the batch size
:param max_iters: (int) the maximum number of iterations
:param adam_epsilon: (float) the epsilon value for the adam optimizer
:param optim_stepsize: (float) the optimizer stepsize
:param ckpt_dir: (str) the save directory, can be None for temporary directory
:param task_name: (str) the save name, can be None for saving directly to the directory name
:param verbose: (bool)
:return: (str) the save location for the TensorFlow model
"""
val_per_iter = int(max_iters / 10)
ob_space = env.observation_space
ac_space = env.action_space
policy = policy_func("pi", ob_space,
ac_space) # Construct network for new policy
# placeholder
obs_ph = policy.obs_ph
action_ph = policy.pdtype.sample_placeholder([None])
stochastic_ph = policy.stochastic_ph
loss = tf.reduce_mean(tf.square(action_ph - policy.ac))
var_list = policy.get_trainable_variables()
adam = MpiAdam(var_list, epsilon=adam_epsilon)
lossandgrad = tf_util.function([obs_ph, action_ph, stochastic_ph],
[loss] + [tf_util.flatgrad(loss, var_list)])
tf_util.initialize()
adam.sync()
logger.log("Pretraining with Behavior Cloning...")
for iter_so_far in tqdm(range(int(max_iters))):
ob_expert, ac_expert = dataset.get_next_batch(optim_batch_size,
'train')
train_loss, grad = lossandgrad(ob_expert, ac_expert, True)
adam.update(grad, optim_stepsize)
if verbose and iter_so_far % val_per_iter == 0:
ob_expert, ac_expert = dataset.get_next_batch(-1, 'val')
val_loss, _ = lossandgrad(ob_expert, ac_expert, True)
logger.log("Training loss: {}, Validation loss: {}".format(
train_loss, val_loss))
if ckpt_dir is None:
savedir_fname = tempfile.TemporaryDirectory().name
else:
savedir_fname = os.path.join(ckpt_dir, task_name)
tf_util.save_state(savedir_fname, var_list=policy.get_variables())
return savedir_fname
def setup_model(self):
with SetVerbosity(self.verbose):
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess,
graph=self.graph)
# Construct network for new policy
self.policy_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
old_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
self.grad_inverter = grad_inverter(
[self.action_space.high, self.action_space.low])
# Target advantage function (if applicable)
atarg = tf.placeholder(dtype=tf.float32, shape=[None])
# Empirical return
ret = tf.placeholder(dtype=tf.float32, shape=[None])
# learning rate multiplier, updated with schedule
lrmult = tf.placeholder(name='lrmult',
dtype=tf.float32,
shape=[])
# Annealed cliping parameter epislon
clip_param = self.clip_param * lrmult
obs_ph = self.policy_pi.obs_ph
action_ph = self.policy_pi.pdtype.sample_placeholder(
[None])
if debug:
action_ph_val = tf.Print(
action_ph, [
action_ph,
],
'\n\n ====================Unclipped action in: \n',
summarize=-1)
action_ph_val = tf.Print(
action_ph_val, [],
'\n ======================================== \n',
summarize=-1)
kloldnew = old_pi.proba_distribution.kl(
self.policy_pi.proba_distribution)
# old_logstd = old_pi.proba_distribution.logstd
# new_logstd = self.policy_pi.proba_distribution.logstd
# old_std = old_pi.proba_distribution.std
# new_std = self.policy_pi.proba_distribution.std
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
# meankl = tf.Print(meankl, [meankl,], "kl value: ")
# meankl_log = tf.Print(meankl, [old_logstd,], "high kl, old logstd value: ", summarize=-1)
# meankl_log = tf.Print(meankl_log, [new_logstd,], "high kl, new logstd value: ", summarize=-1)
# meankl_log = tf.Print(meankl_log, [old_std,], "high kl, old std value: ", summarize=-1)
# meankl_log = tf.Print(meankl_log, [new_std,], "high kl, new std value: ", summarize=-1)
# meanklvalue_ = tf.where(
# tf.greater(meankl, tf.constant(1, dtype = tf.float32)),
# meankl_log,
# meankl)
meanent = tf.reduce_mean(ent)
pol_entpen = (-self.entcoeff) * meanent
# pnew / pold
if debug:
old_logp = old_pi.proba_distribution.logp(
action_ph_val)
old_mean = old_pi.proba_distribution.mode()
old_std = old_pi.proba_distribution.std
old_logp = tf.Print(old_logp, [
old_logp,
],
'====== OLD logp: \n',
summarize=-1)
old_logp = tf.Print(old_logp, [], '\n', summarize=-1)
old_logp = tf.Print(old_logp, [
old_mean,
],
'====== OLD mean: \n',
summarize=-1)
old_logp = tf.Print(old_logp, [], '\n', summarize=-1)
old_logp = tf.Print(old_logp, [
old_std,
],
'====== OLD std: \n',
summarize=-1)
old_logp = tf.Print(old_logp, [], '\n', summarize=-1)
now_logp = self.policy_pi.proba_distribution.logp(
action_ph_val)
now_mean = self.policy_pi.proba_distribution.mode()
now_std = self.policy_pi.proba_distribution.std
now_logp = tf.Print(now_logp, [
now_logp,
],
'====== NOW logp: \n',
summarize=-1)
now_logp = tf.Print(now_logp, [], '\n', summarize=-1)
now_logp = tf.Print(now_logp, [
now_mean,
],
'====== NOW mean: \n',
summarize=-1)
now_logp = tf.Print(now_logp, [], '\n', summarize=-1)
now_logp = tf.Print(now_logp, [
now_std,
],
'====== NOW std: \n',
summarize=-1)
now_logp = tf.Print(now_logp, [], '\n', summarize=-1)
else:
now_logp = self.policy_pi.proba_distribution.logp(
action_ph)
old_logp = old_pi.proba_distribution.logp(action_ph)
ratio = tf.exp(now_logp - old_logp)
if debug:
ratio = tf.Print(ratio, [
ratio,
],
'ratio: \n',
summarize=-1)
ratio = tf.Print(ratio, [], '\n', summarize=-1)
# surrogate from conservative policy iteration
surr1 = ratio * atarg
surr2 = tf.clip_by_value(ratio, 1.0 - clip_param,
1.0 + clip_param) * atarg
# PPO's pessimistic surrogate (L^CLIP)
pol_surr = -tf.reduce_mean(tf.minimum(surr1, surr2))
vf_loss = tf.reduce_mean(
tf.square(self.policy_pi.value_flat - ret))
total_loss = pol_surr + pol_entpen + vf_loss
losses = [pol_surr, pol_entpen, vf_loss, meankl, meanent]
# losses = [pol_surr, pol_entpen, vf_loss, meanklvalue_, meanent]
self.loss_names = [
"pol_surr", "pol_entpen", "vf_loss", "kl", "ent"
]
tf.summary.scalar('entropy_loss', pol_entpen)
tf.summary.scalar('policy_gradient_loss', pol_surr)
tf.summary.scalar('value_function_loss', vf_loss)
tf.summary.scalar('approximate_kullback-leibler', meankl)
tf.summary.scalar('clip_factor', clip_param)
tf.summary.scalar('loss', total_loss)
self.params = tf_util.get_trainable_vars("model")
self.assign_old_eq_new = tf_util.function(
[], [],
updates=[
tf.assign(oldv, newv) for (oldv, newv) in zipsame(
tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))
])
with tf.variable_scope("Adam_mpi", reuse=False):
self.adam = MpiAdam(self.params,
epsilon=self.adam_epsilon,
sess=self.sess)
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards',
tf.reduce_mean(ret))
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.optim_stepsize))
tf.summary.scalar('advantage', tf.reduce_mean(atarg))
tf.summary.scalar('clip_range',
tf.reduce_mean(self.clip_param))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', ret)
tf.summary.histogram('learning_rate',
self.optim_stepsize)
tf.summary.histogram('advantage', atarg)
tf.summary.histogram('clip_range', self.clip_param)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', obs_ph)
else:
tf.summary.histogram('observation', obs_ph)
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
tf_util.initialize(sess=self.sess)
self.summary = tf.summary.merge_all()
self.lossandgrad = tf_util.function(
[obs_ph, old_pi.obs_ph, action_ph, atarg, ret, lrmult],
[self.summary,
tf_util.flatgrad(total_loss, self.params)] + losses)
self.compute_losses = tf_util.function(
[obs_ph, old_pi.obs_ph, action_ph, atarg, ret, lrmult],
losses)
def setup_model(self):
with SetVerbosity(self.verbose):
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.single_threaded_session(graph=self.graph)
# Construct network for new policy
with tf.variable_scope("pi", reuse=False):
self.policy_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
old_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False)
# Target advantage function (if applicable)
atarg = tf.placeholder(dtype=tf.float32, shape=[None])
# Empirical return
ret = tf.placeholder(dtype=tf.float32, shape=[None])
# learning rate multiplier, updated with schedule
lrmult = tf.placeholder(name='lrmult',
dtype=tf.float32,
shape=[])
# Annealed cliping parameter epislon
clip_param = self.clip_param * lrmult
obs_ph = self.policy_pi.obs_ph
action_ph = self.policy_pi.pdtype.sample_placeholder([None])
kloldnew = old_pi.proba_distribution.kl(
self.policy_pi.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
pol_entpen = (-self.entcoeff) * meanent
# pnew / pold
ratio = tf.exp(
self.policy_pi.proba_distribution.logp(action_ph) -
old_pi.proba_distribution.logp(action_ph))
# surrogate from conservative policy iteration
surr1 = ratio * atarg
surr2 = tf.clip_by_value(ratio, 1.0 - clip_param,
1.0 + clip_param) * atarg
# PPO's pessimistic surrogate (L^CLIP)
pol_surr = -tf.reduce_mean(tf.minimum(surr1, surr2))
vf_loss = tf.reduce_mean(
tf.square(self.policy_pi.value_fn[:, 0] - ret))
total_loss = pol_surr + pol_entpen + vf_loss
losses = [pol_surr, pol_entpen, vf_loss, meankl, meanent]
self.loss_names = [
"pol_surr", "pol_entpen", "vf_loss", "kl", "ent"
]
self.params = tf_util.get_trainable_vars("pi")
self.lossandgrad = tf_util.function(
[obs_ph, old_pi.obs_ph, action_ph, atarg, ret, lrmult],
losses + [tf_util.flatgrad(total_loss, self.params)])
self.adam = MpiAdam(self.params,
epsilon=self.adam_epsilon,
sess=self.sess)
self.assign_old_eq_new = tf_util.function(
[], [],
updates=[
tf.assign(oldv, newv) for (
oldv,
newv) in zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("pi"))
])
self.compute_losses = tf_util.function(
[obs_ph, old_pi.obs_ph, action_ph, atarg, ret, lrmult],
losses)
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
tf_util.initialize(sess=self.sess)
def setup_model(self):
with SetVerbosity(self.verbose):
self.graph = tf.Graph()
with self.graph.as_default():
self.set_random_seed(self.seed)
self.sess = tf_util.make_session(num_cpu=self.n_cpu_tf_sess,
graph=self.graph)
# Construct network for new policy
self.policy_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
# Network for old policy
with tf.compat.v1.variable_scope("oldpi", reuse=False):
old_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
with tf.compat.v1.variable_scope("loss", reuse=False):
# Target advantage function (if applicable)
atarg = tf.compat.v1.placeholder(dtype=tf.float32,
shape=[None])
# Empirical return
ret = tf.compat.v1.placeholder(dtype=tf.float32,
shape=[None])
# learning rate multiplier, updated with schedule
lrmult = tf.compat.v1.placeholder(name='lrmult',
dtype=tf.float32,
shape=[])
# Annealed cliping parameter epislon
clip_param = self.clip_param * lrmult
obs_ph = self.policy_pi.obs_ph
action_ph = self.policy_pi.pdtype.sample_placeholder(
[None])
kloldnew = old_pi.proba_distribution.kl(
self.policy_pi.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(input_tensor=kloldnew)
meanent = tf.reduce_mean(input_tensor=ent)
pol_entpen = (-self.entcoeff) * meanent
# pnew / pold
ratio = tf.exp(
self.policy_pi.proba_distribution.logp(action_ph) -
old_pi.proba_distribution.logp(action_ph))
# surrogate from conservative policy iteration
surr1 = ratio * atarg
surr2 = tf.clip_by_value(ratio, 1.0 - clip_param,
1.0 + clip_param) * atarg
# PPO's pessimistic surrogate (L^CLIP)
pol_surr = -tf.reduce_mean(
input_tensor=tf.minimum(surr1, surr2))
vf_loss = tf.reduce_mean(
input_tensor=tf.square(self.policy_pi.value_flat -
ret))
total_loss = pol_surr + pol_entpen + vf_loss
losses = [pol_surr, pol_entpen, vf_loss, meankl, meanent]
self.loss_names = [
"pol_surr", "pol_entpen", "vf_loss", "kl", "ent"
]
tf.compat.v1.summary.scalar('entropy_loss', pol_entpen)
tf.compat.v1.summary.scalar('policy_gradient_loss',
pol_surr)
tf.compat.v1.summary.scalar('value_function_loss', vf_loss)
tf.compat.v1.summary.scalar('approximate_kullback-leibler',
meankl)
tf.compat.v1.summary.scalar('clip_factor', clip_param)
tf.compat.v1.summary.scalar('loss', total_loss)
self.params = tf_util.get_trainable_vars("model")
self.assign_old_eq_new = tf_util.function(
[], [],
updates=[
tf.compat.v1.assign(oldv, newv)
for (oldv, newv) in zipsame(
tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))
])
with tf.compat.v1.variable_scope("Adam_mpi", reuse=False):
self.adam = MpiAdam(self.params,
epsilon=self.adam_epsilon,
sess=self.sess)
with tf.compat.v1.variable_scope("input_info", reuse=False):
tf.compat.v1.summary.scalar(
'discounted_rewards', tf.reduce_mean(input_tensor=ret))
tf.compat.v1.summary.scalar(
'learning_rate',
tf.reduce_mean(input_tensor=self.optim_stepsize))
tf.compat.v1.summary.scalar(
'advantage', tf.reduce_mean(input_tensor=atarg))
tf.compat.v1.summary.scalar(
'clip_range',
tf.reduce_mean(input_tensor=self.clip_param))
if self.full_tensorboard_log:
tf.compat.v1.summary.histogram('discounted_rewards',
ret)
tf.compat.v1.summary.histogram('learning_rate',
self.optim_stepsize)
tf.compat.v1.summary.histogram('advantage', atarg)
tf.compat.v1.summary.histogram('clip_range',
self.clip_param)
if tf_util.is_image(self.observation_space):
tf.compat.v1.summary.image('observation', obs_ph)
else:
tf.compat.v1.summary.histogram(
'observation', obs_ph)
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
tf_util.initialize(sess=self.sess)
self.summary = tf.compat.v1.summary.merge_all()
self.lossandgrad = tf_util.function(
[obs_ph, old_pi.obs_ph, action_ph, atarg, ret, lrmult],
[self.summary,
tf_util.flatgrad(total_loss, self.params)] + losses)
self.compute_losses = tf_util.function(
[obs_ph, old_pi.obs_ph, action_ph, atarg, ret, lrmult],
losses)
def setup_model(self):
# prevent import loops
from stable_baselines.gail.adversary import TransitionClassifier
with SetVerbosity(self.verbose):
assert issubclass(self.policy, ActorCriticPolicy), "Error: the input policy for the TRPO model must be " \
"an instance of common.policies.ActorCriticPolicy."
self.nworkers = MPI.COMM_WORLD.Get_size()
self.rank = MPI.COMM_WORLD.Get_rank()
np.set_printoptions(precision=3)
self.graph = tf.Graph()
with self.graph.as_default():
self.sess = tf_util.single_threaded_session(graph=self.graph)
if self.using_gail:
self.reward_giver = TransitionClassifier(
self.observation_space,
self.action_space,
self.hidden_size_adversary,
entcoeff=self.adversary_entcoeff)
# Construct network for new policy
self.policy_pi = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
# Network for old policy
with tf.variable_scope("oldpi", reuse=False):
old_policy = self.policy(self.sess,
self.observation_space,
self.action_space,
self.n_envs,
1,
None,
reuse=False,
**self.policy_kwargs)
with tf.variable_scope("loss", reuse=False):
atarg = tf.placeholder(dtype=tf.float32, shape=[
None
]) # Target advantage function (if applicable)
ret = tf.placeholder(dtype=tf.float32,
shape=[None]) # Empirical return
observation = self.policy_pi.obs_ph
action = self.policy_pi.pdtype.sample_placeholder([None])
kloldnew = old_policy.proba_distribution.kl(
self.policy_pi.proba_distribution)
ent = self.policy_pi.proba_distribution.entropy()
meankl = tf.reduce_mean(kloldnew)
meanent = tf.reduce_mean(ent)
entbonus = self.entcoeff * meanent
vferr = tf.reduce_mean(
tf.square(self.policy_pi.value_fn[:, 0] - ret))
# advantage * pnew / pold
ratio = tf.exp(
self.policy_pi.proba_distribution.logp(action) -
old_policy.proba_distribution.logp(action))
surrgain = tf.reduce_mean(ratio * atarg)
optimgain = surrgain + entbonus
losses = [optimgain, meankl, entbonus, surrgain, meanent]
self.loss_names = [
"optimgain", "meankl", "entloss", "surrgain", "entropy"
]
dist = meankl
all_var_list = tf_util.get_trainable_vars("model")
var_list = [
v for v in all_var_list
if "/vf" not in v.name and "/q/" not in v.name
]
vf_var_list = [
v for v in all_var_list
if "/pi" not in v.name and "/logstd" not in v.name
]
self.get_flat = tf_util.GetFlat(var_list, sess=self.sess)
self.set_from_flat = tf_util.SetFromFlat(var_list,
sess=self.sess)
klgrads = tf.gradients(dist, var_list)
flat_tangent = tf.placeholder(dtype=tf.float32,
shape=[None],
name="flat_tan")
shapes = [var.get_shape().as_list() for var in var_list]
start = 0
tangents = []
for shape in shapes:
var_size = tf_util.intprod(shape)
tangents.append(
tf.reshape(flat_tangent[start:start + var_size],
shape))
start += var_size
gvp = tf.add_n([
tf.reduce_sum(grad * tangent)
for (grad, tangent) in zipsame(klgrads, tangents)
]) # pylint: disable=E1111
fvp = tf_util.flatgrad(gvp, var_list)
tf.summary.scalar('entropy_loss', meanent)
tf.summary.scalar('policy_gradient_loss', optimgain)
tf.summary.scalar('value_function_loss', surrgain)
tf.summary.scalar('approximate_kullback-leiber', meankl)
tf.summary.scalar(
'loss',
optimgain + meankl + entbonus + surrgain + meanent)
self.assign_old_eq_new = \
tf_util.function([], [], updates=[tf.assign(oldv, newv) for (oldv, newv) in
zipsame(tf_util.get_globals_vars("oldpi"),
tf_util.get_globals_vars("model"))])
self.compute_losses = tf_util.function(
[observation, old_policy.obs_ph, action, atarg],
losses)
self.compute_fvp = tf_util.function([
flat_tangent, observation, old_policy.obs_ph, action,
atarg
], fvp)
self.compute_vflossandgrad = tf_util.function(
[observation, old_policy.obs_ph, ret],
tf_util.flatgrad(vferr, vf_var_list))
@contextmanager
def timed(msg):
if self.rank == 0 and self.verbose >= 1:
print(colorize(msg, color='magenta'))
start_time = time.time()
yield
print(
colorize("done in {:.3f} seconds".format(
(time.time() - start_time)),
color='magenta'))
else:
yield
def allmean(arr):
assert isinstance(arr, np.ndarray)
out = np.empty_like(arr)
MPI.COMM_WORLD.Allreduce(arr, out, op=MPI.SUM)
out /= self.nworkers
return out
tf_util.initialize(sess=self.sess)
th_init = self.get_flat()
MPI.COMM_WORLD.Bcast(th_init, root=0)
self.set_from_flat(th_init)
with tf.variable_scope("Adam_mpi", reuse=False):
self.vfadam = MpiAdam(vf_var_list, sess=self.sess)
if self.using_gail:
self.d_adam = MpiAdam(
self.reward_giver.get_trainable_variables(),
sess=self.sess)
self.d_adam.sync()
self.vfadam.sync()
with tf.variable_scope("input_info", reuse=False):
tf.summary.scalar('discounted_rewards',
tf.reduce_mean(ret))
tf.summary.scalar('learning_rate',
tf.reduce_mean(self.vf_stepsize))
tf.summary.scalar('advantage', tf.reduce_mean(atarg))
tf.summary.scalar('kl_clip_range',
tf.reduce_mean(self.max_kl))
if self.full_tensorboard_log:
tf.summary.histogram('discounted_rewards', ret)
tf.summary.histogram('learning_rate', self.vf_stepsize)
tf.summary.histogram('advantage', atarg)
tf.summary.histogram('kl_clip_range', self.max_kl)
if tf_util.is_image(self.observation_space):
tf.summary.image('observation', observation)
else:
tf.summary.histogram('observation', observation)
self.timed = timed
self.allmean = allmean
self.step = self.policy_pi.step
self.proba_step = self.policy_pi.proba_step
self.initial_state = self.policy_pi.initial_state
self.params = find_trainable_variables("model")
if self.using_gail:
self.params.extend(
self.reward_giver.get_trainable_variables())
self.summary = tf.summary.merge_all()
self.compute_lossandgrad = \
tf_util.function([observation, old_policy.obs_ph, action, atarg, ret],
[self.summary, tf_util.flatgrad(optimgain, var_list)] + losses)
def general_actor_critic(input_shape_vec,
act_output_shape,
comm,
learn_rate=[0.001, 0.001],
trainable=True,
label=""):
sess = K.get_session()
np.random.seed(0)
tf.set_random_seed(0)
# network 1 (new policy)
with tf.variable_scope(label + "_pi_new", reuse=False):
inp = Input(shape=input_shape_vec) # [5,6,3]
# rc_lyr = Lambda(lambda x: ned_to_ripCoords_tf(x, 4000))(inp)
trunk_x = Reshape([input_shape_vec[0], input_shape_vec[1] * 3])(inp)
trunk_x = LSTM(128)(trunk_x)
dist, sample_action_op, action_ph, value_output = ppo_continuous(
3, trunk_x)
# network 2 (old policy)
with tf.variable_scope(label + "_pi_old", reuse=False):
inp_old = Input(shape=input_shape_vec) # [5,6,3]
# rc_lyr = Lambda(lambda x: ned_to_ripCoords_tf(x, 4000))(inp_old)
trunk_x = Reshape([input_shape_vec[0],
input_shape_vec[1] * 3])(inp_old)
trunk_x = LSTM(128)(trunk_x)
dist_old, sample_action_op_old, action_ph_old, value_output_old = ppo_continuous(
3, trunk_x)
# additional placeholders
adv_ph = tf.placeholder(tf.float32, [None], name="advantages_ph")
alpha_ph = tf.placeholder(tf.float32, shape=(), name="alpha_ph")
vtarg = tf.placeholder(tf.float32, [None]) # target value placeholder
# loss
loss = ppo_continuous_loss(dist, dist_old, value_output, action_ph,
alpha_ph, adv_ph, vtarg)
# gradient
with tf.variable_scope("grad", reuse=False):
gradient = tf_util.flatgrad(
loss, tf_util.get_trainable_vars(label + "_pi_new"))
adam = MpiAdam(tf_util.get_trainable_vars(label + "_pi_new"),
epsilon=0.00001,
sess=sess,
comm=comm)
# method for sync'ing the two policies
assign_old_eq_new = tf_util.function(
[], [],
updates=[
tf.assign(oldv, newv)
for (oldv,
newv) in zipsame(tf_util.get_globals_vars(label + "_pi_old"),
tf_util.get_globals_vars(label + "_pi_new"))
])
# initialize all the things
init_op = tf.global_variables_initializer()
sess.run(init_op)
# methods for interacting with this model
def sync_weights():
assign_old_eq_new(sess=sess)
def sample_action(states, logstd_override=None):
a = sess.run(sample_action_op, feed_dict={inp: states})
return a
def sample_value(states):
v = sess.run(value_output, feed_dict={inp: states})
return v
def train(states, actions, vtarget, advs, alpha):
alpha = max(alpha, 0.0)
adam_lr = learn_rate[0]
g = sess.run(
[gradient],
feed_dict={
inp: states,
inp_old: states,
action_ph: actions,
adv_ph: advs,
alpha_ph: alpha,
vtarg: vtarget
})
adam.update(g[0], adam_lr * alpha)
# initial sync
adam.sync()
sync_weights()
return sync_weights, sample_action, sample_value, train
