def update_loss_trainer(progress):
# Update all coef regarding progression
pg_coef = np.clip(self.start_pg_coef * (1.0 - 2.0 * progress), 0.0,
1.0)
st_coef = np.clip(self.start_st_coef * (1.0 - 2.0 * progress), 0.0,
1.0)
pg_lt_coef = np.clip(self.start_pg_lt_coef + 2.0 * progress,
self.start_pg_lt_coef, 1.0)
loss = pg_coef * pg_loss - entropy * ent_coef + (
vf_loss + vf_loss_lt
) * vf_coef + st_coef * st_loss + pg_lt_coef * pg_loss_lt
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
if self.max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(
grads, self.max_grad_norm)
grads = list(zip(grads, params))
_train = trainer.apply_gradients(grads)
return pg_coef, st_coef, pg_lt_coef
def add_noise(self, mean=0.0, stddev=0.01):
sess = tf_util.get_session()
params = find_trainable_variables(self.model_name)
for param in params:
variables_shape = tf.shape(param)
noise = tf.random_normal(
variables_shape,
mean=mean,
stddev=stddev,
dtype=tf.float32,
)
sess.run(tf.assign_add(param, noise))
def get_train_op(loss_op):
params = find_trainable_variables("model")
# Switch from GATE_NONE to GATE_GRAPH to enhance reproducibility.
#grads = tf.gradients(loss, params)
grads_and_params = trainer.compute_gradients(
loss=loss_op,
var_list=params,
gate_gradients=tf.train.RMSPropOptimizer.GATE_GRAPH)
grads = [x[0] for x in grads_and_params]
params = [x[1] for x in grads_and_params]
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
return trainer.apply_gradients(grads)
def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, nstack,
num_procs, ent_coef, q_coef, gamma, max_grad_norm, lr,
rprop_alpha, rprop_epsilon, total_timesteps, lrschedule, c,
trust_region, alpha, delta):
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=num_procs,
inter_op_parallelism_threads=num_procs)
sess = tf.Session(config=config)
nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch]) # actions
D = tf.placeholder(tf.float32, [nbatch]) # dones
R = tf.placeholder(tf.float32, [nbatch]) # rewards, not returns
MU = tf.placeholder(tf.float32, [nbatch, nact]) # mu's
LR = tf.placeholder(tf.float32, [])
eps = 1e-6
step_model = policy(sess,
ob_space,
ac_space,
nenvs,
1,
nstack,
reuse=False)
train_model = policy(sess,
ob_space,
ac_space,
nenvs,
nsteps + 1,
nstack,
reuse=True)
params = find_trainable_variables("model")
print("Params {}".format(len(params)))
for var in params:
print(var)
# create polyak averaged model
ema = tf.train.ExponentialMovingAverage(alpha)
ema_apply_op = ema.apply(params)
def custom_getter(getter, *args, **kwargs):
v = ema.average(getter(*args, **kwargs))
print(v.name)
return v
with tf.variable_scope("", custom_getter=custom_getter, reuse=True):
polyak_model = policy(sess,
ob_space,
ac_space,
nenvs,
nsteps + 1,
nstack,
reuse=True)
# Notation: (var) = batch variable, (var)s = seqeuence variable, (var)_i = variable index by action at step i
v = tf.reduce_sum(train_model.pi * train_model.q,
axis=-1) # shape is [nenvs * (nsteps + 1)]
# strip off last step
f, f_pol, q = map(lambda var: strip(var, nenvs, nsteps),
[train_model.pi, polyak_model.pi, train_model.q])
# Get pi and q values for actions taken
f_i = get_by_index(f, A)
q_i = get_by_index(q, A)
# Compute ratios for importance truncation
rho = f / (MU + eps)
rho_i = get_by_index(rho, A)
# Calculate Q_retrace targets
qret = q_retrace(R, D, q_i, v, rho_i, nenvs, nsteps, gamma)
# Calculate losses
# Entropy
entropy = tf.reduce_mean(cat_entropy_softmax(f))
# Policy Graident loss, with truncated importance sampling & bias correction
v = strip(v, nenvs, nsteps, True)
check_shape([qret, v, rho_i, f_i], [[nenvs * nsteps]] * 4)
check_shape([rho, f, q], [[nenvs * nsteps, nact]] * 2)
# Truncated importance sampling
adv = qret - v
logf = tf.log(f_i + eps)
gain_f = logf * tf.stop_gradient(
adv * tf.minimum(c, rho_i)) # [nenvs * nsteps]
loss_f = -tf.reduce_mean(gain_f)
# Bias correction for the truncation
adv_bc = (q - tf.reshape(v, [nenvs * nsteps, 1])
) # [nenvs * nsteps, nact]
logf_bc = tf.log(f + eps) # / (f_old + eps)
check_shape([adv_bc, logf_bc], [[nenvs * nsteps, nact]] * 2)
gain_bc = tf.reduce_sum(
logf_bc *
tf.stop_gradient(adv_bc * tf.nn.relu(1.0 - (c / (rho + eps))) * f),
axis=1) #IMP: This is sum, as expectation wrt f
loss_bc = -tf.reduce_mean(gain_bc)
loss_policy = loss_f + loss_bc
# Value/Q function loss, and explained variance
check_shape([qret, q_i], [[nenvs * nsteps]] * 2)
ev = q_explained_variance(tf.reshape(q_i, [nenvs, nsteps]),
tf.reshape(qret, [nenvs, nsteps]))
loss_q = tf.reduce_mean(tf.square(tf.stop_gradient(qret) - q_i) * 0.5)
# Net loss
check_shape([loss_policy, loss_q, entropy], [[]] * 3)
loss = loss_policy + q_coef * loss_q - ent_coef * entropy
if trust_region:
g = tf.gradients(-(loss_policy - ent_coef * entropy) * nsteps *
nenvs, f) #[nenvs * nsteps, nact]
# k = tf.gradients(KL(f_pol || f), f)
k = -f_pol / (
f + eps
) #[nenvs * nsteps, nact] # Directly computed gradient of KL divergence wrt f
k_dot_g = tf.reduce_sum(k * g, axis=-1)
adj = tf.maximum(0.0, (tf.reduce_sum(k * g, axis=-1) - delta) /
(tf.reduce_sum(tf.square(k), axis=-1) +
eps)) #[nenvs * nsteps]
# Calculate stats (before doing adjustment) for logging.
avg_norm_k = avg_norm(k)
avg_norm_g = avg_norm(g)
avg_norm_k_dot_g = tf.reduce_mean(tf.abs(k_dot_g))
avg_norm_adj = tf.reduce_mean(tf.abs(adj))
g = g - tf.reshape(adj, [nenvs * nsteps, 1]) * k
grads_f = -g / (
nenvs * nsteps
) # These are turst region adjusted gradients wrt f ie statistics of policy pi
grads_policy = tf.gradients(f, params, grads_f)
grads_q = tf.gradients(loss_q * q_coef, params)
grads = [
gradient_add(g1, g2, param)
for (g1, g2, param) in zip(grads_policy, grads_q, params)
]
avg_norm_grads_f = avg_norm(grads_f) * (nsteps * nenvs)
norm_grads_q = tf.global_norm(grads_q)
norm_grads_policy = tf.global_norm(grads_policy)
else:
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, norm_grads = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=rprop_alpha,
epsilon=rprop_epsilon)
_opt_op = trainer.apply_gradients(grads)
# so when you call _train, you first do the gradient step, then you apply ema
with tf.control_dependencies([_opt_op]):
_train = tf.group(ema_apply_op)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
# Ops/Summaries to run, and their names for logging
run_ops = [
_train, loss, loss_q, entropy, loss_policy, loss_f, loss_bc, ev,
norm_grads
]
names_ops = [
'loss', 'loss_q', 'entropy', 'loss_policy', 'loss_f', 'loss_bc',
'explained_variance', 'norm_grads'
]
if trust_region:
run_ops = run_ops + [
norm_grads_q, norm_grads_policy, avg_norm_grads_f, avg_norm_k,
avg_norm_g, avg_norm_k_dot_g, avg_norm_adj
]
names_ops = names_ops + [
'norm_grads_q', 'norm_grads_policy', 'avg_norm_grads_f',
'avg_norm_k', 'avg_norm_g', 'avg_norm_k_dot_g', 'avg_norm_adj'
]
def train(obs, actions, rewards, dones, mus, states, masks, steps):
cur_lr = lr.value_steps(steps)
td_map = {
train_model.X: obs,
polyak_model.X: obs,
A: actions,
R: rewards,
D: dones,
MU: mus,
LR: cur_lr
}
if states != []:
td_map[train_model.S] = states
td_map[train_model.M] = masks
td_map[polyak_model.S] = states
td_map[polyak_model.M] = masks
return names_ops, sess.run(run_ops, td_map)[1:] # strip off _train
def save(save_path):
ps = sess.run(params)
make_path(osp.dirname(save_path))
joblib.dump(ps, save_path)
self.train = train
self.save = save
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
total_timesteps,
nprocs=32,
nscripts=16,
nsteps=20,
nstack=4,
ent_coef=0.1,
vf_coef=0.5,
vf_fisher_coef=1.0,
lr=0.25,
max_grad_norm=0.001,
kfac_clip=0.001,
lrschedule='linear',
alpha=0.99,
epsilon=1e-5):
config = tf.ConfigProto(
allow_soft_placement=True,
intra_op_parallelism_threads=nprocs,
inter_op_parallelism_threads=nprocs)
config.gpu_options.allow_growth = True
self.sess = sess = tf.Session(config=config)
nsml.bind(sess=sess)
#nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch])
XY0 = tf.placeholder(tf.int32, [nbatch])
XY1 = tf.placeholder(tf.int32, [nbatch])
# ADV == TD_TARGET - values
ADV = tf.placeholder(tf.float32, [nbatch])
TD_TARGET = tf.placeholder(tf.float32, [nbatch])
PG_LR = tf.placeholder(tf.float32, [])
VF_LR = tf.placeholder(tf.float32, [])
self.model = step_model = policy(
sess, ob_space, ac_space, nenvs, 1, nstack, reuse=False)
self.model2 = train_model = policy(
sess, ob_space, ac_space, nenvs, nsteps, nstack, reuse=True)
# Policy 1 : Base Action : train_model.pi label = A
script_mask = tf.concat(
[
tf.zeros([nscripts * nsteps, 1]),
tf.ones([(nprocs - nscripts) * nsteps, 1])
],
axis=0)
pi = train_model.pi
pac_weight = script_mask * (tf.nn.softmax(pi) - 1.0) + 1.0
pac_weight = tf.reduce_sum(pac_weight * tf.one_hot(A, depth=3), axis=1)
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=pi, labels=A)
neglogpac *= tf.stop_gradient(pac_weight)
inv_A = 1.0 - tf.cast(A, tf.float32)
xy0_mask = tf.cast(A, tf.float32)
xy1_mask = tf.cast(A, tf.float32)
condition0 = tf.equal(xy0_mask, 2)
xy0_mask = tf.where(condition0, tf.ones(tf.shape(xy0_mask)), xy0_mask)
xy0_mask = 1.0 - xy0_mask
condition1 = tf.equal(xy1_mask, 2)
xy1_mask = tf.where(condition1, tf.zeros(tf.shape(xy1_mask)), xy1_mask)
# One hot representation of chosen marine.
# [batch_size, 2]
pi_xy0 = train_model.pi_xy0
pac_weight = script_mask * (tf.nn.softmax(pi_xy0) - 1.0) + 1.0
pac_weight = tf.reduce_sum(
pac_weight * tf.one_hot(XY0, depth=1024), axis=1)
logpac_xy0 = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=pi_xy0, labels=XY0)
logpac_xy0 *= tf.stop_gradient(pac_weight)
logpac_xy0 *= tf.cast(xy0_mask, tf.float32)
pi_xy1 = train_model.pi_xy1
pac_weight = script_mask * (tf.nn.softmax(pi_xy1) - 1.0) + 1.0
pac_weight = tf.reduce_sum(
pac_weight * tf.one_hot(XY0, depth=1024), axis=1)
# 1D? 2D?
logpac_xy1 = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=pi_xy1, labels=XY1)
logpac_xy1 *= tf.stop_gradient(pac_weight)
logpac_xy1 *= tf.cast(xy1_mask, tf.float32)
pg_loss = tf.reduce_mean(ADV * neglogpac)
pg_loss_xy0 = tf.reduce_mean(ADV * logpac_xy0)
pg_loss_xy1 = tf.reduce_mean(ADV * logpac_xy1)
vf_ = tf.squeeze(train_model.vf)
vf_r = tf.concat(
[
tf.ones([nscripts * nsteps, 1]),
tf.zeros([(nprocs - nscripts) * nsteps, 1])
],
axis=0) * TD_TARGET
vf_masked = vf_ * script_mask + vf_r
#vf_mask[0:nscripts * nsteps] = R[0:nscripts * nsteps]
vf_loss = tf.reduce_mean(mse(vf_masked, TD_TARGET))
entropy_a = tf.reduce_mean(cat_entropy(train_model.pi))
entropy_xy0 = tf.reduce_mean(cat_entropy(train_model.pi_xy0))
entropy_xy1 = tf.reduce_mean(cat_entropy(train_model.pi_xy1))
entropy = entropy_a + entropy_xy0 + entropy_xy1
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, _ = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(
learning_rate=lr, decay=alpha, epsilon=epsilon)
_train = trainer.apply_gradients(grads)
self.logits = logits = train_model.pi
# xy0
self.params_common = params_common = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='model/common')
self.params_xy0 = params_xy0 = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES,
scope='model/xy0') + params_common
train_loss_xy0 = pg_loss_xy0 - entropy * ent_coef + vf_coef * vf_loss
self.grads_check_xy0 = grads_xy0 = tf.gradients(
train_loss_xy0, params_xy0)
if max_grad_norm is not None:
grads_xy0, _ = tf.clip_by_global_norm(grads_xy0, max_grad_norm)
grads_xy0 = list(zip(grads_xy0, params_xy0))
trainer_xy0 = tf.train.RMSPropOptimizer(
learning_rate=lr, decay=alpha, epsilon=epsilon)
_train_xy0 = trainer_xy0.apply_gradients(grads_xy0)
# xy1
self.params_xy1 = params_xy1 = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES,
scope='model/xy1') + params_common
train_loss_xy1 = pg_loss_xy1 - entropy * ent_coef + vf_coef * vf_loss
self.grads_check_xy1 = grads_xy1 = tf.gradients(
train_loss_xy1, params_xy1)
if max_grad_norm is not None:
grads_xy1, _ = tf.clip_by_global_norm(grads_xy1, max_grad_norm)
grads_xy1 = list(zip(grads_xy1, params_xy1))
trainer_xy1 = tf.train.RMSPropOptimizer(
learning_rate=lr, decay=alpha, epsilon=epsilon)
_train_xy1 = trainer_xy1.apply_gradients(grads_xy1)
self.lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, td_targets, masks, actions, xy0, xy1, values):
advs = td_targets - values
for step in range(len(obs)):
cur_lr = self.lr.value()
td_map = {
train_model.X: obs,
A: actions,
XY0: xy0,
XY1: xy1,
ADV: advs,
TD_TARGET: td_targets,
PG_LR: cur_lr
}
if states != []:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _, \
policy_loss_xy0, policy_entropy_xy0, _, \
policy_loss_xy1, policy_entropy_xy1, _ = sess.run(
[pg_loss, vf_loss, entropy, _train,
pg_loss_xy0, entropy_xy0, _train_xy0,
pg_loss_xy1, entropy_xy1, _train_xy1],
td_map)
return policy_loss, value_loss, policy_entropy, \
policy_loss_xy0, policy_entropy_xy0, \
policy_loss_xy1, policy_entropy_xy1
def save(save_path):
ps = sess.run(params)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
sess.run(restores)
self.train = train
self.save = save
self.load = load
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
print("global_variables_initializer start")
tf.global_variables_initializer().run(session=sess)
print("global_variables_initializer complete")
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
nsteps,
nstack,
num_procs,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear'):
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=num_procs,
inter_op_parallelism_threads=num_procs)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
nact = ac_space.n
nbatch = nenvs * nsteps
writter = tf.summary.FileWriter(
"/tmp/a2c_demo/1") # Change for SAT: this is to use tensorBoard
A = tf.placeholder(
tf.int32, [nbatch]) # Comments by Fei: this must be the action
ADV = tf.placeholder(
tf.float32,
[nbatch]) # Comments by Fei: this must be the advantage
R = tf.placeholder(
tf.float32, [nbatch]) # Comments by Fei: this must be the reward
LR = tf.placeholder(
tf.float32, []) # Comments by Fei: this must be the learning rate
step_model = policy(sess,
ob_space,
ac_space,
nenvs,
1,
nstack,
reuse=False)
train_model = policy(sess,
ob_space,
ac_space,
nenvs,
nsteps,
nstack,
reuse=True)
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=train_model.pi,
labels=A) # Comments by Fei: pi is nbatch * nact
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
entropy = tf.reduce_mean(cat_entropy(train_model.pi))
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
if states != []:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
# writter.add_graph(sess.graph)
return policy_loss, value_loss, policy_entropy
def save(save_path):
ps = sess.run(params)
make_path(save_path)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
ps = sess.run(restores)
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = save
self.load = load
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, nstack, num_procs,
ent_coef=0.01, vf_coef=0.5, max_grad_norm=0.5, lr=7e-4,
alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6), lrschedule='linear', logdir=None):
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=num_procs,
inter_op_parallelism_threads=num_procs)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
nact = ac_space.n
nbatch = nenvs*nsteps
ADV = tf.placeholder(tf.float32, [None])
R = tf.placeholder(tf.float32, [None])
LR = tf.placeholder(tf.float32, [])
step_model = policy(sess, ob_space, ac_space, nenvs, 1, nstack, reuse=False)
train_model = policy(sess, ob_space, ac_space, nenvs, nsteps, nstack, reuse=True)
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=train_model.pi, labels=train_model.a0)
entropy = tf.reduce_sum(cat_entropy(train_model.pi))
params = find_trainable_variables("model")
tf.summary.histogram("vf", train_model.vf)
tf.summary.histogram("R", R)
if train_model.relaxed:
pg_loss = tf.constant(0.0)
oh_A = tf.one_hot(train_model.a0, ac_space.n)
params = find_trainable_variables("model")
policy_params = [v for v in params if "pi" in v.name]
vf_params = [v for v in params if "vf" in v.name]
entropy_grads = tf.gradients(entropy, policy_params)
ddiff_loss = tf.reduce_sum(train_model.vf - train_model.vf_t)
ddiff_grads = tf.gradients(ddiff_loss, policy_params)
sm = tf.nn.softmax(train_model.pi)
dlogp_dpi = oh_A * (1. - sm) + (1. - oh_A) * (-sm)
pi_grads = -((tf.expand_dims(R, 1) - train_model.vf_t) * dlogp_dpi)
pg_grads = tf.gradients(train_model.pi, policy_params, grad_ys=pi_grads)
pg_grads = [pg - dg for pg, dg in zip(pg_grads, ddiff_grads)]
pi_param_grads = tf.gradients(train_model.pi, policy_params, grad_ys=pi_grads)
cv_grads = tf.concat([tf.reshape(p, [-1]) for p in pg_grads], 0)
cv_grad_splits = tf.reduce_sum(tf.square(cv_grads))
vf_loss = cv_grad_splits * vf_coef
cv_grads = tf.gradients(vf_loss, vf_params)
policy_grads = []
for e_grad, p_grad, param in zip(entropy_grads, pg_grads, policy_params):
grad = -e_grad * ent_coef + p_grad
policy_grads.append(grad)
grad_dict = {}
for g, v in list(zip(policy_grads, policy_params))+list(zip(cv_grads, vf_params)):
grad_dict[v] = g
grads = [grad_dict[v] for v in params]
print(grads)
else:
pg_loss = tf.reduce_sum((tf.stop_gradient(R) - tf.stop_gradient(train_model.vf)) * neglogpac)
policy_params = [v for v in params if "pi" in v.name]
pg_grads = tf.gradients(pg_loss, policy_params)
vf_loss = tf.reduce_sum(mse(tf.squeeze(train_model.vf), R))
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
grads = tf.gradients(loss, params)
grads = list(zip(grads, params))
ema = tf.train.ExponentialMovingAverage(.99)
all_policy_grads = tf.concat([tf.reshape(g, [-1]) for g in pg_grads], 0)
all_policy_grads_sq = tf.square(all_policy_grads)
apply_mean_op = ema.apply([all_policy_grads, all_policy_grads_sq])
em_mean = ema.average(all_policy_grads)
em_mean_sq = ema.average(all_policy_grads_sq)
em_var = em_mean_sq - tf.square(em_mean)
em_log_var = tf.log(em_var + 1e-20)
mlgv = tf.reduce_mean(em_log_var)
for g, v in grads:
print(v.name, g)
tf.summary.histogram(v.name, v)
tf.summary.histogram(v.name+"_grad", g)
self.sum_op = tf.summary.merge_all()
self.writer = tf.summary.FileWriter(logdir)
trainer = tf.train.AdamOptimizer(learning_rate=LR, beta2=.99999)
with tf.control_dependencies([apply_mean_op]):
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
self._step = 0
def train(obs, states, rewards, masks, u1, u2, values, summary=False):
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X:obs, train_model.U1:u1, train_model.U2:u2,
ADV:advs, R:rewards, LR:cur_lr
}
if states != []:
td_map[train_model.S] = states
td_map[train_model.M] = masks
if summary:
sum_str, policy_loss, value_loss, policy_entropy, lv, _ = sess.run(
[self.sum_op, pg_loss, vf_loss, entropy, mlgv, _train],
td_map
)
self.writer.add_summary(sum_str, self._step)
else:
policy_loss, value_loss, policy_entropy, lv, _ = sess.run(
[pg_loss, vf_loss, entropy, mlgv, _train],
td_map
)
self._step += 1
return policy_loss, value_loss, policy_entropy, lv
def save(save_path):
ps = sess.run(params)
make_path(save_path)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
ps = sess.run(restores)
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = save
self.load = load
tf.global_variables_initializer().run(session=sess)
def load_variables_from_another_model(self, another_model_name):
sess = tf_util.get_session()
params = find_trainable_variables(self.model_name)
another_params = find_trainable_variables(another_model_name)
for pair in zip(params, another_params):
sess.run(tf.assign(pair[0], pair[1]))
def __init__(self, policy, ob_space, ac_space, nenvs,total_timesteps, nprocs=32, nsteps=20,
ent_coef=0.01, vf_coef=0.5, vf_fisher_coef=1.0, lr=0.25, max_grad_norm=0.5,
kfac_clip=0.001, lrschedule='linear'):
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=nprocs,
inter_op_parallelism_threads=nprocs)
config.gpu_options.allow_growth = True
self.sess = sess = tf.Session(config=config)
nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
PG_LR = tf.placeholder(tf.float32, [])
VF_LR = tf.placeholder(tf.float32, [])
self.model = step_model = policy(sess, ob_space, ac_space, nenvs, 1, reuse=False)
self.model2 = train_model = policy(sess, ob_space, ac_space, nenvs*nsteps, nsteps, reuse=True)
logpac = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=train_model.pi, labels=A)
self.logits = logits = train_model.pi
##training loss
pg_loss = tf.reduce_mean(ADV*logpac)
entropy = tf.reduce_mean(cat_entropy(train_model.pi))
pg_loss = pg_loss - ent_coef * entropy
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
train_loss = pg_loss + vf_coef * vf_loss
##Fisher loss construction
self.pg_fisher = pg_fisher_loss = -tf.reduce_mean(logpac)
sample_net = train_model.vf + tf.random_normal(tf.shape(train_model.vf))
self.vf_fisher = vf_fisher_loss = - vf_fisher_coef*tf.reduce_mean(tf.pow(train_model.vf - tf.stop_gradient(sample_net), 2))
self.joint_fisher = joint_fisher_loss = pg_fisher_loss + vf_fisher_loss
self.params=params = find_trainable_variables("model")
self.grads_check = grads = tf.gradients(train_loss,params)
with tf.device('/gpu:0'):
self.optim = optim = kfac.KfacOptimizer(learning_rate=PG_LR, clip_kl=kfac_clip,\
momentum=0.9, kfac_update=1, epsilon=0.01,\
stats_decay=0.99, async=1, cold_iter=10, max_grad_norm=max_grad_norm)
update_stats_op = optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
train_op, q_runner = optim.apply_gradients(list(zip(grads,params)))
self.q_runner = q_runner
self.lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = self.lr.value()
td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, PG_LR:cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, train_op],
td_map
)
return policy_loss, value_loss, policy_entropy
def save(save_path):
ps = sess.run(params)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
sess.run(restores)
self.train = train
self.save = save
self.load = load
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, ob_space, ac_space, nenvs, master_ts = 1, worker_ts = 30,
ent_coef=0.01, vf_coef=0.5, max_grad_norm=0.5, lr=7e-4, cell = 256,
alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6), lrschedule='linear',
algo='regular', beta=1e-3):
print('Create Session')
gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
nact = ac_space.n
nbatch = nenvs*master_ts*worker_ts
A = tf.placeholder(tf.int32, [nbatch])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
step_model = policy(sess, ob_space, ac_space, nenvs, 1, 1, cell = cell, model='step_model', algo=algo)
train_model = policy(sess, ob_space, ac_space, nbatch, master_ts, worker_ts, model='train_model', algo=algo)
print('model_setting_done')
#loss construction
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=train_model.wpi, labels=A)
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.wvf), R))
entropy = tf.reduce_mean(cat_entropy(train_model.wpi))
pg_loss = pg_loss - entropy * ent_coef
print('algo: ', algo, 'max_grad_norm: ', str(max_grad_norm))
try:
if algo == 'regular':
loss = pg_loss + vf_coef * vf_loss
elif algo == 'VIB':
'''
implement VIB here, apart from the vf_loss and pg_loss, there should be a third loss,
the kl_loss = ds.kl_divergence(model.encoding, prior), where prior is a Gaussian distribution with mu=0, std=1
the final loss should be pg_loss + vf_coef * vf_loss + beta*kl_loss
'''
prior = ds.Normal(0.0, 1.0)
kl_loss = tf.reduce_mean(ds.kl_divergence(train_model.encoding, prior))
loss = pg_loss + vf_coef * vf_loss + beta*kl_loss
# pass
else:
raise Exception('Algorithm not exists')
except Exception as e:
print(e)
grads, global_norm = grad_clip(loss, max_grad_norm, ['model'])
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=alpha, epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(wobs, whs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(whs)):
cur_lr = lr.value()
td_map = {train_model.wX:wobs, A:actions, ADV:advs, R:rewards, LR:cur_lr}
if states is not None:
td_map[train_model.wS] = states
td_map[train_model.wM] = masks
'''
you can add and run additional loss for VIB here for debugging, such as kl_loss
'''
tloss, value_loss, policy_loss, policy_entropy, _ = sess.run(
[loss, vf_loss, pg_loss, entropy, _train],
feed_dict=td_map
)
return tloss, value_loss, policy_loss, policy_entropy
params = find_trainable_variables("model")
def save(save_path):
ps = sess.run(params)
make_path(osp.dirname(save_path))
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
ps = sess.run(restores)
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.wvalue
self.get_wh = step_model.get_wh
self.initial_state = step_model.w_initial_state
self.train = train
self.save = save
self.load = load
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
env,
nsteps,
ent_coef=1,
vf_coef=0.5,
max_grad_norm=0.5,
fisher_matrix=None,
star_param=None,
lam=None,
batch_size=None,
fisher_matrix2=None,
star_param2=None):
sess = tf_util.get_session()
nbatch = batch_size
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# train_model is used to train our network
train_model = policy(nbatch, 1, sess)
eval_model = policy(1, 1, sess)
# Update parameters using lossa.
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
print(len(params))
ewc_loss = 0
ewc_per_layer = []
j = 0
#star_param是load进去的array|params是模型的参数，也是tensor
for v in range(len(params) - 4):
if v % 2 == 0:
C = star_param[v].shape[-1]
W = tf.transpose(tf.reshape(params[v] - star_param[v],
[-1, C]))
right = tf.reshape(
tf.transpose(
tf.matmul(tf.matmul(fisher_matrix[j + 1], W),
fisher_matrix[j])), [-1, 1])
W_ = tf.reshape(params[v] - star_param[v], [1, -1])
ewc_loss += tf.matmul(W_, right)
ewc_per_layer.append(tf.matmul(W_, right))
j = j + 2
else:
B = tf.reshape(params[v] - star_param[v], [-1, 1])
right_B = tf.matmul(fisher_matrix[j], B)
B_ = tf.reshape(params[v] - star_param[v], [1, -1])
ewc_loss += tf.matmul(B_, right_B)
ewc_per_layer.append(tf.matmul(B_, right_B))
j += 1
ewc_loss2 = 0
j = 0
for v in range(len(params) - 4):
if v % 2 == 0:
C2 = star_param2[v].shape[-1]
W2 = tf.transpose(
tf.reshape(params[v] - star_param2[v], [-1, C2]))
right2 = tf.reshape(
tf.transpose(
tf.matmul(tf.matmul(fisher_matrix2[j + 1], W2),
fisher_matrix2[j])), [-1, 1])
W_2 = tf.reshape(params[v] - star_param2[v], [1, -1])
ewc_loss2 += tf.matmul(W_2, right2)
j = j + 2
else:
B2 = tf.reshape(params[v] - star_param2[v], [-1, 1])
right_B2 = tf.matmul(fisher_matrix2[j], B2)
B_2 = tf.reshape(params[v] - star_param2[v], [1, -1])
ewc_loss2 += tf.matmul(B_2, right_B2)
j += 1
loss1 = ewc_loss * (lam / 2)
loss2 = ewc_loss2 * (lam / 2)
loss = loss1 + loss2
# 2. Calculate the gradients
trainer = tf.train.AdamOptimizer(learning_rate=1e-3)
# grads_and_var = trainer.compute_gradients(loss, params)
# grads, var = zip(*grads_and_var)
# grads, _grad_norm = tf.clip_by_global_norm(grads, 0.5)
#
# grads_and_var = list(zip(grads, var))
grads = tf.gradients(loss, params)
grads_and_var = list(zip(grads, params))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
# trainer = tf.train.RMSPropOptimizer(learning_rate=7e-4, decay=alpha, epsilon=epsilon)
# trainer = tf.train.MomentumOptimizer(learning_rate=0.001,momentum=0.9,use_nesterov=True)
_train = trainer.apply_gradients(grads_and_var)
#
#
# grads2 = tf.gradients(loss2,params)
# grads2 = list(zip(grads2, params))
# trainer2 = tf.train.RMSPropOptimizer(learning_rate=7e-4, decay=alpha, epsilon=epsilon)
# trainer2 = tf.train.AdamOptimizer(learning_rate=1e-4)
# trainer2 = tf.train.MomentumOptimizer(learning_rate=0.001, momentum=0.9, use_nesterov=True)
# _train2 = trainer2.apply_gradients(grads2)
def train():
td_map = {train_model.keep_prob: 1.0}
ewc1, ewc2, l, _ = sess.run([loss1, loss2, loss, _train], td_map)
return ewc1, ewc2, l
# KL,ewc,_ = sess.run([KL_loss,ewc_loss,_train],td_map)
# return KL, ewc
def creat_star_param():
star_list = []
for i in params[:-2]:
star_list.append(star_param[params[i].name])
return star_list
self.creat_star_param = creat_star_param
self.train = train
self.train_model = train_model
self.act = eval_model.step
self.act2 = eval_model.step2
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
nenvs,
total_timesteps,
nprocs=32,
nsteps=20,
ent_coef=0.01,
vf_coef=0.5,
vf_fisher_coef=1.0,
lr=0.25,
max_grad_norm=0.5,
kfac_clip=0.001,
lrschedule='linear',
is_async=True):
sess = get_session()
nbatch = nenvs * nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
eval_model = policy(1, 1, sess)
# A = train_model.pdtype.sample_placeholder([None])
# A = tf.placeholder(step_model.action.dtype, step_model.action.shape)
probs = tf.nn.softmax(step_model.pi)
class_ind = tf.to_int32(tf.multinomial(tf.log(probs), 1)[0][0])
self.pg_fisher = pg_fisher_loss = tf.log(probs[0, class_ind])
##Fisher loss construction
# self.pg_fisher = pg_fisher_loss = -tf.reduce_mean(neglogpac)
# sample_net = train_model.vf + tf.random_normal(tf.shape(train_model.vf))
# self.vf_fisher = vf_fisher_loss = - vf_fisher_coef*tf.reduce_mean(tf.pow(train_model.vf - tf.stop_gradient(sample_net), 2))
# self.joint_fisher = joint_fisher_loss = pg_fisher_loss + vf_fisher_loss
self.joint_fisher = joint_fisher_loss = pg_fisher_loss
self.params = params = find_trainable_variables("a2c_model")
with tf.device('/gpu:0'):
self.optim = optim = kfac.KfacOptimizer()
# update_stats_op = optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
stats = optim.compute_and_apply_stats(joint_fisher_loss,
var_list=params[:-4])
def compute_fisher(obs):
# action = action[:, np.newaxis]
td_map = {step_model.X: obs, step_model.keep_prob: 1.0}
fisher = sess.run(stats, td_map)
return fisher
self.compute_fisher = compute_fisher
self.save = functools.partial(save_variables, sess=sess)
self.load = functools.partial(load_variables, sess=sess)
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, ob_space, ac_space, nenvs,total_timesteps, nprocs=32, nsteps=20,
ent_coef=0.01, vf_coef=0.5, vf_fisher_coef=1.0, lr=0.25, max_grad_norm=0.5,
kfac_clip=0.001, lrschedule='linear', is_async=True):
self.sess = sess = get_session()
nbatch = nenvs * nsteps
with tf.variable_scope('acktr_model', reuse=tf.AUTO_REUSE):
self.model = step_model = policy(nenvs, 1, sess=sess)
self.model2 = train_model = policy(nenvs*nsteps, nsteps, sess=sess)
A = train_model.pdtype.sample_placeholder([None])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
PG_LR = tf.placeholder(tf.float32, [])
VF_LR = tf.placeholder(tf.float32, [])
neglogpac = train_model.pd.neglogp(A)
self.logits = train_model.pi
##training loss
pg_loss = tf.reduce_mean(ADV*neglogpac)
entropy = tf.reduce_mean(train_model.pd.entropy())
pg_loss = pg_loss - ent_coef * entropy
vf_loss = tf.losses.mean_squared_error(tf.squeeze(train_model.vf), R)
train_loss = pg_loss + vf_coef * vf_loss
##Fisher loss construction
self.pg_fisher = pg_fisher_loss = -tf.reduce_mean(neglogpac)
sample_net = train_model.vf + tf.random_normal(tf.shape(train_model.vf))
self.vf_fisher = vf_fisher_loss = - vf_fisher_coef*tf.reduce_mean(tf.pow(train_model.vf - tf.stop_gradient(sample_net), 2))
self.joint_fisher = joint_fisher_loss = pg_fisher_loss + vf_fisher_loss
self.params=params = find_trainable_variables("acktr_model")
self.grads_check = grads = tf.gradients(train_loss,params)
with tf.device('/gpu:0'):
self.optim = optim = kfac.KfacOptimizer(learning_rate=PG_LR, clip_kl=kfac_clip,\
momentum=0.9, kfac_update=1, epsilon=0.01,\
stats_decay=0.99, is_async=is_async, cold_iter=10, max_grad_norm=max_grad_norm)
# update_stats_op = optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
train_op, q_runner = optim.apply_gradients(list(zip(grads,params)))
self.q_runner = q_runner
self.lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = self.lr.value()
td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, PG_LR:cur_lr, VF_LR:cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, train_op],
td_map
)
return policy_loss, value_loss, policy_entropy
self.train = train
self.save = functools.partial(save_variables, sess=sess)
self.load = functools.partial(load_variables, sess=sess)
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, nstack, num_procs,
ent_coef=0.01, vf_coef=0.5, max_grad_norm=0.5, lr=7e-4,
alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6), lrschedule='linear'):
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=num_procs,
inter_op_parallelism_threads=num_procs)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
nact = ac_space.n
nbatch = nenvs*nsteps
A = tf.placeholder(tf.int32, [nbatch])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
step_model = policy(sess, ob_space, ac_space, nenvs, 1, nstack, reuse=False)
train_model = policy(sess, ob_space, ac_space, nenvs, nsteps, nstack, reuse=True)
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=train_model.pi, labels=A)
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
entropy = tf.reduce_mean(cat_entropy(train_model.pi))
loss = pg_loss - entropy*ent_coef + vf_loss * vf_coef
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=alpha, epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, LR:cur_lr}
if states != []:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train],
td_map
)
return policy_loss, value_loss, policy_entropy
def save(save_path):
ps = sess.run(params)
make_path(save_path)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
ps = sess.run(restores)
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = save
self.load = load
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear',
param=None):
sess = tf_util.make_session()
nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
step_model = policy(sess,
ob_space,
ac_space,
nenvs,
1,
reuse=False,
param=param)
train_model = policy(sess,
ob_space,
ac_space,
nenvs * nsteps,
nsteps,
reuse=True,
param=param)
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=train_model.pi, labels=A)
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
entropy = tf.reduce_mean(cat_entropy(train_model.pi))
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
def save(save_path):
ps = sess.run(params)
make_path(save_path)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
ps = sess.run(restores)
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = save
self.load = load
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
total_timesteps,
nprocs=32,
nsteps=20,
ent_coef=0.01,
vf_coef=0.5,
vf_fisher_coef=1.0,
lr=0.25,
max_grad_norm=0.5,
kfac_clip=0.001,
lrschedule='linear',
is_async=True):
self.sess = sess = get_session()
nbatch = nenvs * nsteps
with tf.variable_scope('acktr_model', reuse=tf.AUTO_REUSE):
self.model = step_model = policy(nenvs, 1, sess=sess)
self.model2 = train_model = policy(nenvs * nsteps,
nsteps,
sess=sess)
A = train_model.pdtype.sample_placeholder([None])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
PG_LR = tf.placeholder(tf.float32, [])
VF_LR = tf.placeholder(tf.float32, [])
neglogpac = train_model.pd.neglogp(A)
self.logits = train_model.pi
##training loss
pg_loss = tf.reduce_mean(ADV * neglogpac)
entropy = tf.reduce_mean(train_model.pd.entropy())
pg_loss = pg_loss - ent_coef * entropy
vf_loss = tf.losses.mean_squared_error(tf.squeeze(train_model.vf), R)
train_loss = pg_loss + vf_coef * vf_loss
##Fisher loss construction
self.pg_fisher = pg_fisher_loss = -tf.reduce_mean(neglogpac)
sample_net = train_model.vf + tf.random_normal(tf.shape(
train_model.vf))
self.vf_fisher = vf_fisher_loss = -vf_fisher_coef * tf.reduce_mean(
tf.pow(train_model.vf - tf.stop_gradient(sample_net), 2))
self.joint_fisher = joint_fisher_loss = pg_fisher_loss + vf_fisher_loss
self.params = params = find_trainable_variables("acktr_model")
self.grads_check = grads = tf.gradients(train_loss, params)
with tf.device('/gpu:0'):
self.optim = optim = kfac.KfacOptimizer(learning_rate=PG_LR, clip_kl=kfac_clip,\
momentum=0.9, kfac_update=1, epsilon=0.01,\
stats_decay=0.99, is_async=is_async, cold_iter=10, max_grad_norm=max_grad_norm)
# update_stats_op = optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
optim.compute_and_apply_stats(joint_fisher_loss, var_list=params)
train_op, q_runner = optim.apply_gradients(list(zip(grads,
params)))
self.q_runner = q_runner
self.lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = self.lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
PG_LR: cur_lr,
VF_LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, train_op], td_map)
return policy_loss, value_loss, policy_entropy
self.train = train
self.save = functools.partial(save_variables, sess=sess)
self.load = functools.partial(load_variables, sess=sess)
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, p, has_state):
"""
policy : Internal Policy model such as SnakeModel.CNNPolicy
p : Hyperparameters required for training
"""
sess = tf_util.make_session()
# Tensorflow model initiallization
step_model = policy(sess=sess,
p=p,
train_phase=False,
has_state=has_state) # Deploy model settings
train_model = policy(sess=sess,
p=p,
train_phase=True,
has_state=has_state) # Training model settings
saver = tf.train.Saver()
#Step 2 : Initialize the training parameters
A = tf.placeholder(tf.int32, [p.N_BATCH])
ADV = tf.placeholder(tf.float32, [p.N_BATCH])
R = tf.placeholder(tf.float32, [p.N_BATCH])
LR = tf.placeholder(tf.float32, [])
#Step 3 : Define the loss Function
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=train_model.pi, labels=A) #
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
entropy = tf.reduce_mean(cat_entropy(train_model.pi))
loss = pg_loss - entropy * p.ENTROPY_COEFF + vf_loss * p.VALUE_FUNC_COEFF
#Step 4 : Define the loss optimizer
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if p.MAX_GRAD_NORM is not None:
grads, grad_norm = tf.clip_by_global_norm(
grads, p.MAX_GRAD_NORM
) # Clipping the gradients to protect learned weights
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=p.RMS_DECAY,
epsilon=p.EPSILON)
_train = trainer.apply_gradients(
grads) # This is the variable which will be used
lr = Scheduler(v=p.LEARNING_RATE,
nvalues=p.N_TIMESTEPS,
schedule=p.LEARNING_RATE_SCHEDULE
) # Learning rate changes linearly or as per arguments
# Step 5 : Write down the summary parameters to be used
writer = tf.summary.FileWriter(p.LOG_PATH) #summary writer
def train(obs, rewards, masks, actions, values, states):
"""
obs : batch x n x m x 1 snake matrix
rewards : batch x 1 rewards corrosponding to action
actions : batch x 1 discrete action taken
values : batch x 1 output of value function during the training process
"""
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
train_model.S: states,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
def save(save_path):
#ps = sess.run(params)
#make_path(save_path)
#joblib.dump(ps, save_path)
saver.save(sess, save_path)
def load(load_path):
#loaded_params = joblib.load(load_path)
#restores = []
#for p, loaded_p in zip(params, loaded_params):
# restores.append(p.assign(loaded_p))
#ps = sess.run(restores)
saver.restore(sess, load_path)
def add_scalar_summary(tag, value, step):
summary = tf.Summary(
value=[tf.Summary.Value(tag=tag, simple_value=value)])
writer.add_summary(summary, step)
# Expose the user to closure functions
self.train = train
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.hidden_value = step_model.hidden_value
self.initial_state = step_model.initial_state
self.add_scalar_summary = add_scalar_summary
self.save = save
self.load = load
# Initialize global variables and add tf graph
tf.global_variables_initializer().run(session=sess)
writer.add_graph(tf.get_default_graph()) #write graph
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
lambda_dist=0.01,
total_timesteps=None,
lrschedule='linear'):
sess = tf.get_default_session()
nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
config = Config()
act_model = policy(config)
config.reuse = True
train_model = policy(config)
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=train_model.logits, labels=A)
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
entropy = tf.reduce_mean(cat_entropy(train_model.logits))
aux_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=train_model.rp_logits, labels=A)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef + aux_loss * lambda_dist
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
saver = tf.train.Saver()
def train(obs, rs, rr, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr,
train_model.inputs_s: rs,
train_model.inputs_r: rr
}
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
def save(save_path):
saver.save(sess, save_path + 'model.ckpt')
def load(load_path):
saver.restore(sess, load_path + 'model.ckpt')
self.train = train
self.train_model = train_model
self.act_model = act_model
self.act = act_model.act
self.value = act_model.value
self.save = save
self.load = load
def __init__(self,
policy,
env,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear',
network='cnn',
prio_args=None):
self.prio_args = prio_args
sess = tf_util.get_session()
nenvs = self.get_active_envs(env)
nbatch = nenvs * nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
# our TD evaluating network
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# Value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
# TD loss
# td_loss = losses.mean_squared_error(tf.squeeze(train_model.dt), TD)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
"""prio model"""
with tf.variable_scope('a2c_model_prio', reuse=tf.AUTO_REUSE):
# prio_model = policy(nbatch, nsteps, sess)
prio_model = MyNN(env, nbatch, network)
P_R = tf.placeholder(tf.float32, [nbatch])
PRIO = tf.placeholder(tf.float32, [nbatch])
P_LR = tf.placeholder(tf.float32, [])
# prio_model_loss = losses.mean_squared_error(tf.squeeze(prio_model.out), P_R) # Reward
prio_model_loss = losses.mean_squared_error(tf.squeeze(prio_model.out),
PRIO) # TD Error
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
params_prio = find_trainable_variables("a2c_model_prio")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
prio_grads = tf.gradients(prio_model_loss, params_prio)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
prio_grads, prio_grad_norm = tf.clip_by_global_norm(
prio_grads, max_grad_norm)
grads = list(zip(grads, params))
prio_grads = list(zip(prio_grads, params_prio))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
prio_trainer = tf.train.RMSPropOptimizer(learning_rate=P_LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
_prio_train = prio_trainer.apply_gradients(prio_grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
prio_loss = 0
if self.prio_args is not None:
prio_values = GetValuesForPrio(self.prio_args['prio_type'],
self.prio_args['prio_param'],
advs, rewards)
prio_td_map = {
prio_model.X: obs,
P_R: rewards,
P_LR: cur_lr,
PRIO: prio_values
}
prio_loss, _, p_td = sess.run(
[prio_model_loss, _prio_train, PRIO], prio_td_map)
# mb aranged as 1D-vector = [[env_1: n1, ..., n_nstep],...,[env_n_active]]
# need to take last value of each env's buffer
self.prio_score = prio_values[list(
filter(lambda x: x % nsteps == (nsteps - 1),
range(len(prio_values))))]
return policy_loss, value_loss, policy_entropy, prio_loss
self.train = train
self.train_model = train_model
self.step_model = step_model
self.prio_model = prio_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
env,
nsteps,
ent_coef=1,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear',
fisher_matrix=None,
star_param=None,
lam=None,
batch_size=None):
sess = tf_util.get_session()
nbatch = batch_size
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# train_model is used to train our network
train_model = policy(nbatch, 1, sess)
eval_model = policy(1, 1, sess)
OUTPUT = tf.placeholder(tf.float32, [None, 6], name='sample_action')
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# constant = tf.constant(0.01,shape=train_model.pi.get_shape())
pi = tf.nn.softmax(OUTPUT)
# u = tf.argmax(pi,axis=-1)
# x = tf.one_hot(u, pi.get_shape().as_list()[-1])
# cross_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=train_model.pi,labels=x))
train_model_pi = tf.nn.softmax(train_model.pi3)
KL = pi * tf.log(pi / (train_model_pi + 1e-4))
KL_loss = tf.reduce_mean(tf.reduce_sum(KL, 1))
# Update parameters using lossa.
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
print(len(params))
ewc_loss = 0
#star_param是load进去的array|params是模型的参数，也是tensor
# for i in range(len(fisher_matrix)):
# j = 0
# for v in range(len(params)-8):
# if v % 2 == 0:
# C = star_param[v].shape[-1]
# W = tf.transpose(tf.reshape(params[v] - star_param[v],[-1,C]))
# right = tf.reshape(tf.transpose(tf.matmul(tf.matmul(fisher_matrix[i][j+1], W), fisher_matrix[i][j])),[-1,1])
# W_ = tf.reshape(params[v] - star_param[v], [1,-1])
# ewc_loss += tf.matmul(W_, right)
# j = j + 2
# else:
# B = tf.reshape(params[v] - star_param[v], [-1,1])
# right_B = tf.matmul(fisher_matrix[i][j], B)
# B_ = tf.reshape(params[v] - star_param[v], [1,-1])
# ewc_loss += tf.matmul(B_, right_B)
# j += 1
# for v in range(len(params)-4):
# if v % 2 == 0:
# C = int(star_param[v].shape[-1])
# W_hat = tf.concat([tf.reshape(params[v], [-1, C]), tf.reshape(params[v + 1], [1, -1])], 0)
# W_hat_fixed = tf.concat(
# [tf.reshape(star_param[v], [-1, C]), tf.reshape(star_param[v + 1], [1, -1])], 0)
#
# W = tf.transpose(W_hat - W_hat_fixed)
# right = tf.reshape(tf.transpose(tf.matmul(tf.matmul(fisher_matrix[v + 1], W), fisher_matrix[v])),[-1, 1])
# W_ = tf.reshape(W_hat - W_hat_fixed, [1, -1])
# ewc_loss += tf.matmul(W_, right)
for i in range(len(fisher_matrix)):
for v in range(len(params) - 6):
# if v == 6:
ewc_loss += tf.reduce_sum(
tf.multiply(fisher_matrix[i][v].astype(np.float32),
tf.square(params[v] - star_param[v])))
loss1 = KL_loss * ent_coef
loss2 = ewc_loss * (lam / 2)
loss = loss1 + loss2
# 2. Calculate the gradients
trainer = tf.train.AdamOptimizer(learning_rate=1e-3)
# trainer = tf.train.RMSPropOptimizer(learning_rate=7e-4, decay=alpha, epsilon=epsilon)
# grads_and_var = trainer.compute_gradients(loss, params)
# grads, var = zip(*grads_and_var)
# grads, _grad_norm = tf.clip_by_global_norm(grads, 0.5)
#
# grads_and_var = list(zip(grads, var))
grads = tf.gradients(loss, params)
grads_and_var = list(zip(grads, params))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
# trainer = tf.train.RMSPropOptimizer(learning_rate=7e-4, decay=alpha, epsilon=epsilon)
# trainer = tf.train.MomentumOptimizer(learning_rate=0.001,momentum=0.9,use_nesterov=True)
_train = trainer.apply_gradients(grads_and_var)
#
#
# grads2 = tf.gradients(loss2,params)
# grads2 = list(zip(grads2, params))
# trainer2 = tf.train.RMSPropOptimizer(learning_rate=7e-4, decay=alpha, epsilon=epsilon)
# trainer2 = tf.train.AdamOptimizer(learning_rate=1e-4)
# trainer2 = tf.train.MomentumOptimizer(learning_rate=0.001, momentum=0.9, use_nesterov=True)
# _train2 = trainer2.apply_gradients(grads2)
def train(obs, actions):
td_map = {
train_model.X: obs,
OUTPUT: actions,
train_model.keep_prob: 1.0
}
kl, ewc, l, _ = sess.run([KL_loss, ewc_loss, loss, _train], td_map)
return kl, ewc, l
# KL,ewc,_ = sess.run([KL_loss,ewc_loss,_train],td_map)
# return KL, ewc
def creat_star_param():
star_list = []
for i in params[:-2]:
star_list.append(star_param[params[i].name])
return star_list
self.creat_star_param = creat_star_param
self.train = train
self.train_model = train_model
self.act = eval_model.step
self.act2 = eval_model.step2
self.act3 = eval_model.step3
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def learn_vanilla_a2c(
network,
env,
optimiser,
seed=None,
nsteps=5,
total_timesteps=int(80e6),
vf_coef=0.5,
ent_coef=0.01,
max_grad_norm=0.5,
lr=7e-4,
lrschedule='linear',
epsilon=1e-5,
alpha=0.99,
gamma=0.99,
log_interval=100,
load_path=None,
**network_kwargs):
'''
Main entrypoint for A2C algorithm. Train a policy with given network architecture on a given environment using a2c algorithm.
Parameters:
-----------
network: policy network architecture. Either string (mlp, lstm, lnlstm, cnn_lstm, cnn, cnn_small, conv_only - see baselines.common/models.py for full list)
specifying the standard network architecture, or a function that takes tensorflow tensor as input and returns
tuple (output_tensor, extra_feed) where output tensor is the last network layer output, extra_feed is None for feed-forward
neural nets, and extra_feed is a dictionary describing how to feed state into the network for recurrent neural nets.
See baselines.common/policies.py/lstm for more details on using recurrent nets in policies
env: RL environment. Should implement interface similar to VecEnv (baselines.common/vec_env) or be wrapped with DummyVecEnv (baselines.common/vec_env/dummy_vec_env.py)
seed: seed to make random number sequence in the alorightm reproducible. By default is None which means seed from system noise generator (not reproducible)
nsteps: int, number of steps of the vectorized environment per update (i.e. batch size is nsteps * nenv where
nenv is number of environment copies simulated in parallel)
total_timesteps: int, total number of timesteps to train on (default: 80M)
vf_coef: float, coefficient in front of value function loss in the total loss function (default: 0.5)
ent_coef: float, coeffictiant in front of the policy entropy in the total loss function (default: 0.01)
max_gradient_norm: float, gradient is clipped to have global L2 norm no more than this value (default: 0.5)
lr: float, learning rate for RMSProp (current implementation has RMSProp hardcoded in) (default: 7e-4)
lrschedule: schedule of learning rate. Can be 'linear', 'constant', or a function [0..1] -> [0..1] that takes fraction of the training progress as input and
returns fraction of the learning rate (specified as lr) as output
epsilon: float, RMSProp epsilon (stabilizes square root computation in denominator of RMSProp update) (default: 1e-5)
alpha: float, RMSProp decay parameter (default: 0.99)
gamma: float, reward discounting parameter (default: 0.99)
log_interval: int, specifies how frequently the logs are printed out (default: 100)
**network_kwargs: keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
For instance, 'mlp' network architecture has arguments num_hidden and num_layers.
'''
set_global_seeds(seed)
# Get the nb of env
nenvs = env.num_envs
policy = build_policy(env, network, **network_kwargs)
# Instantiate the model object (that creates step_model and train_model)
model = Model(optimiser=optimiser, policy=policy, env=env, nsteps=nsteps, ent_coef=ent_coef, vf_coef=vf_coef,
max_grad_norm=max_grad_norm, lr=lr, alpha=alpha, epsilon=epsilon, total_timesteps=total_timesteps, lrschedule=lrschedule)
if load_path is not None:
model.load(load_path)
# Instantiate the runner object
runner = VanillaRunner(env, model, nsteps=nsteps, gamma=gamma)
epinfobuf = deque(maxlen=100)
# Calculate the batch_size
nbatch = nenvs*nsteps
def kl(new_mean, new_sd, old_mean, old_sd): # added for logging
approx_kl = np.log(new_sd/old_sd) + (old_sd**2 + (old_mean - new_mean)**2)/(2.0*new_sd**2 + 10**-8) - 0.5
approx_kl = np.sum(approx_kl, axis=1)
approx_kl = np.mean(approx_kl)
return approx_kl
# model helper functions
model_params = find_trainable_variables("a2c_model")
get_flat = U.GetFlat(model_params)
# Start total timer
tstart = time.time()
for update in range(1, int(total_timesteps//nbatch+1)):
old_params = get_flat()
# Get mini batch of experiences
obs, states, rewards, masks, actions, values, epinfos = runner.run()
epinfobuf.extend(epinfos)
old_mean, old_sd = model.get_mean_std(obs) # added
policy_loss, value_loss, policy_entropy = model.train(obs, states, rewards, masks, actions, values)
nseconds = time.time()-tstart
#added
new_mean, new_sd = model.get_mean_std(obs)
approx_kl = kl(new_mean, new_sd, old_mean, old_sd)
new_params = get_flat()
grads = new_params - old_params
# Calculate the fps (frame per second)
fps = int((update*nbatch)/nseconds)
if update % log_interval == 0 or update == 1:
# Calculates if value function is a good predicator of the returns (ev > 1)
# or if it's just worse than predicting nothing (ev =< 0)
ev = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps", update*nbatch)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy", float(policy_entropy))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance", float(ev))
logger.record_tabular("approx_kl", float(approx_kl))
logger.record_tabular("grad_norm", np.linalg.norm(grads))
logger.record_tabular("eprewmean", safemean([epinfo['r'] for epinfo in epinfobuf]))
logger.record_tabular("eplenmean", safemean([epinfo['l'] for epinfo in epinfobuf]))
logger.dump_tabular()
return model
def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, nstack,
num_procs, ent_coef, vf_coef, max_grad_norm, lr, rprop_alpha,
rprop_epsilon, total_timesteps, lrschedule):
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=num_procs,
inter_op_parallelism_threads=num_procs)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
nbatch = nenvs * nsteps
step_model = policy(sess,
ob_space,
ac_space,
nenvs,
1,
nstack,
reuse=False)
train_model = policy(sess,
ob_space,
ac_space,
nenvs,
nsteps,
nstack,
reuse=True)
A = train_model.pdtype.sample_placeholder([nbatch])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
eps = 1e-6
#nadv = ADV / (train_model.ret_rms.std + eps)
#nr = (R - train_model.ret_rms.mean) / (train_model.ret_rms.std + eps)
nadv = (ADV - train_model.ret_rms.mean) / (train_model.ret_rms.std +
eps)
nr = (R - train_model.ret_rms.mean) / (train_model.ret_rms.std + eps)
nlogpac = -train_model.pd.logp(A)
pg_loss = tf.reduce_mean(nadv * nlogpac)
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), nr))
#vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vnorm), nr))
entropy = tf.reduce_mean(train_model.pd.entropy())
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=rprop_alpha,
epsilon=rprop_epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
avg_norm_ret = tf.reduce_mean(tf.abs(train_model.ret_rms.mean))
avg_norm_obs = tf.reduce_mean(tf.abs(train_model.ob_rms.mean))
def train(obs, states, returns, masks, actions, values):
advs = returns - values
#advs = (advs - np.mean(advs)) / (np.std(advs) + eps)
for step in range(len(obs)):
cur_lr = lr.value()
if hasattr(train_model, "ob_rms"):
train_model.ob_rms.update(
sess,
obs) # update running mean/std for observations of policy
if hasattr(train_model, "ret_rms"):
train_model.ret_rms.update(
sess, returns) # # update running mean/std for returns
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: returns,
LR: cur_lr
}
if states != []:
td_map[train_model.S] = states
td_map[train_model.M] = masks
ravg_norm_obs, policy_loss, value_loss, policy_entropy, _ = sess.run(
[avg_norm_obs, pg_loss, vf_loss, entropy, _train], td_map)
return ravg_norm_obs, policy_loss, value_loss, policy_entropy
def save(save_path):
ps = sess.run(params)
make_path(save_path)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
ps = sess.run(restores)
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = save
self.load = load
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, ent_coef, q_coef, gamma, max_grad_norm, lr,
rprop_alpha, rprop_epsilon, total_timesteps, lrschedule,
c, trust_region, alpha, delta, icm ):
sess = get_session()
nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch]) # actions
D = tf.placeholder(tf.float32, [nbatch]) # dones
R = tf.placeholder(tf.float32, [nbatch]) # rewards, not returns
MU = tf.placeholder(tf.float32, [nbatch, nact]) # mu's
LR = tf.placeholder(tf.float32, [])
eps = 1e-6
step_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs,) + ob_space.shape)
train_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs*(nsteps+1),) + ob_space.shape)
with tf.variable_scope('acer_model', reuse=tf.AUTO_REUSE):
step_model = policy(nbatch=nenvs , nsteps=1,observ_placeholder=step_ob_placeholder, sess=sess)
train_model = policy(nbatch=nbatch , nsteps=nsteps, observ_placeholder=train_ob_placeholder, sess=sess)
params = find_trainable_variables("acer_model")
print("Params {}".format(len(params)))
# for var in params:
# print(var)
# create polyak averaged model
ema = tf.train.ExponentialMovingAverage(alpha)
ema_apply_op = ema.apply(params)
def custom_getter(getter, *args, **kwargs):
v = ema.average(getter(*args, **kwargs))
print(v.name)
return v
with tf.variable_scope("acer_model", custom_getter=custom_getter, reuse=True):
polyak_model = policy(nbatch=nbatch, nsteps=nsteps, observ_placeholder=train_ob_placeholder, sess=sess)
# Notation: (var) = batch variable, (var)s = seqeuence variable, (var)_i = variable index by action at step i
# shape is [n_envs * (n_steps + 1)]
# action probability distributions according to train_model, polyak_model and step_model
# poilcy.pi is probability distribution parameters; to obtain distribution that sums to 1 need to take softmax
train_model_p = tf.nn.softmax(train_model.pi)
polyak_model_p = tf.nn.softmax(polyak_model.pi)
step_model_p = tf.nn.softmax(step_model.pi)
# train model policy probility and train model q value
v = tf.reduce_sum(train_model_p * train_model.q, axis = -1) # shape is [nenvs * (nsteps + 1)]
# strip off last step
# dictribution_f , f_polyak, q_value
f, f_pol, q = map(lambda var: strip(var, nenvs, nsteps), [train_model_p, polyak_model_p, train_model.q])
# Get pi and q values for actions taken
f_i = get_by_index(f, A)
q_i = get_by_index(q, A)
# Compute ratios for importance truncation
rho = f / (MU + eps)
rho_i = get_by_index(rho, A)
# Calculate Q_retrace targets
# passed
# R = reward , D = done_ph , v = value ,... rest is same
qret = q_retrace(R, D, q_i, v, rho_i, nenvs, nsteps, gamma)
# Calculate losses
# Entropy
# entropy = tf.reduce_mean(strip(train_model.pd.entropy(), nenvs, nsteps))
entropy = tf.reduce_mean(cat_entropy_softmax(f)) # f is distribution here
# Policy Graident loss, with truncated importance sampling & bias correction
v = strip(v, nenvs, nsteps, True) # v is value here
check_shape([qret, v, rho_i, f_i], [[nenvs * nsteps]] * 4)
check_shape([rho, f, q], [[nenvs * nsteps, nact]] * 2)
# Truncated importance sampling
if icm is None :
adv = qret - v # v is value here
else :
# print("Adv Normalization")
# > Advantage Normalization
adv = qret - v
# m , s = get_mean_and_std(icm_adv)
# advs = (icm_adv - m) / (s + 1e-7)
# > Advantage Normalization
logf = tf.log(f_i + eps)
# c is correction term
# importance weight clipping factor : c
gain_f = logf * tf.stop_gradient(adv * tf.minimum(c, rho_i)) # [nenvs * nsteps]
loss_f = -tf.reduce_mean(gain_f)
# Bias correction for the truncation
adv_bc = (q - tf.reshape(v, [nenvs * nsteps, 1])) # [nenvs * nsteps, nact]
logf_bc = tf.log(f + eps) # / (f_old + eps)
check_shape([adv_bc, logf_bc], [[nenvs * nsteps, nact]]*2)
gain_bc = tf.reduce_sum(logf_bc * tf.stop_gradient(adv_bc * tf.nn.relu(1.0 - (c / (rho + eps))) * f), axis = 1) #IMP: This is sum, as expectation wrt f
loss_bc= -tf.reduce_mean(gain_bc)
# IMP: This is sum, as expectation wrt f
loss_policy = loss_f + loss_bc
# Value/Q function loss, and explained variance
check_shape([qret, q_i], [[nenvs * nsteps]]*2)
ev = q_explained_variance(tf.reshape(q_i, [nenvs, nsteps]), tf.reshape(qret, [nenvs, nsteps]))
loss_q = tf.reduce_mean(tf.square(tf.stop_gradient(qret) - q_i)*0.5)
# Net loss
check_shape([loss_policy, loss_q, entropy], [[]] * 3)
loss = loss_policy + q_coef * loss_q - ent_coef * entropy
if trust_region:
g = tf.gradients(- (loss_policy - ent_coef * entropy) * nsteps * nenvs, f) #[nenvs * nsteps, nact]
# k = tf.gradients(KL(f_pol || f), f)
k = - f_pol / (f + eps) #[nenvs * nsteps, nact] # Directly computed gradient of KL divergence wrt f
k_dot_g = tf.reduce_sum(k * g, axis=-1)
adj = tf.maximum(0.0, (tf.reduce_sum(k * g, axis=-1) - delta) / (tf.reduce_sum(tf.square(k), axis=-1) + eps)) #[nenvs * nsteps]
# Calculate stats (before doing adjustment) for logging.
avg_norm_k = avg_norm(k)
avg_norm_g = avg_norm(g)
avg_norm_k_dot_g = tf.reduce_mean(tf.abs(k_dot_g))
avg_norm_adj = tf.reduce_mean(tf.abs(adj))
g = g - tf.reshape(adj, [nenvs * nsteps, 1]) * k
grads_f = -g/(nenvs*nsteps) # These are turst region adjusted gradients wrt f ie statistics of policy pi
grads_policy = tf.gradients(f, params, grads_f)
grads_q = tf.gradients(loss_q * q_coef, params)
grads = [gradient_add(g1, g2, param) for (g1, g2, param) in zip(grads_policy, grads_q, params)]
avg_norm_grads_f = avg_norm(grads_f) * (nsteps * nenvs)
norm_grads_q = tf.global_norm(grads_q)
norm_grads_policy = tf.global_norm(grads_policy)
else:
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, norm_grads = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
if icm is not None :
# print("with ICM")
grads = grads + icm.pred_grads_and_vars
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=rprop_alpha, epsilon=rprop_epsilon)
_opt_op = trainer.apply_gradients(grads)
# so when you call _train, you first do the gradient step, then you apply ema
with tf.control_dependencies([_opt_op]):
_train = tf.group(ema_apply_op)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
# Ops/Summaries to run, and their names for logging
if icm is not None :
# print("With ICM")
run_ops = [_train, loss, loss_q, entropy, loss_policy, loss_f, loss_bc, ev, norm_grads ,
icm.forw_loss , icm.inv_loss, icm.icm_loss]
names_ops = ['loss', 'loss_q', 'entropy', 'loss_policy', 'loss_f', 'loss_bc', 'explained_variance',
'norm_grads' ,'icm.forw_loss' , 'icm.inv_loss', 'icm.icm_loss' ]
if trust_region:
run_ops = run_ops + [norm_grads_q, norm_grads_policy, avg_norm_grads_f, avg_norm_k, avg_norm_g, avg_norm_k_dot_g,
avg_norm_adj ]
names_ops = names_ops + ['norm_grads_q', 'norm_grads_policy', 'avg_norm_grads_f', 'avg_norm_k', 'avg_norm_g',
'avg_norm_k_dot_g', 'avg_norm_adj' ]
else :
run_ops = [_train, loss, loss_q, entropy, loss_policy, loss_f, loss_bc, ev, norm_grads ]
names_ops = ['loss', 'loss_q', 'entropy', 'loss_policy', 'loss_f', 'loss_bc', 'explained_variance',
'norm_grads' ]
if trust_region:
run_ops = run_ops + [norm_grads_q, norm_grads_policy, avg_norm_grads_f, avg_norm_k, avg_norm_g, avg_norm_k_dot_g,
avg_norm_adj ]
names_ops = names_ops + ['norm_grads_q', 'norm_grads_policy', 'avg_norm_grads_f', 'avg_norm_k', 'avg_norm_g',
'avg_norm_k_dot_g', 'avg_norm_adj' ]
def train(obs, actions, rewards, dones, mus, states, masks, steps, next_states, icm_actions ):
cur_lr = lr.value_steps(steps)
if icm is not None :
print("with ICM ")
td_map = {train_model.X: obs, polyak_model.X: obs, A: actions, R: rewards, D: dones, MU: mus, LR: cur_lr ,
icm.state_:obs, icm.next_state_ : next_states , icm.action_ : icm_actions}
else :
td_map = {train_model.X: obs, polyak_model.X: obs, A: actions, R: rewards, D: dones, MU: mus, LR: cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
td_map[polyak_model.S] = states
td_map[polyak_model.M] = masks
return names_ops, sess.run(run_ops, td_map)[1:] # strip off _train
def _step(observation, **kwargs):
return step_model._evaluate([step_model.action, step_model_p, step_model.state], observation, **kwargs)
self.train = train
self.save = functools.partial(save_variables, sess=sess, variables=params)
self.train_model = train_model
self.step_model = step_model
self._step = _step
self.step = self.step_model.step
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, ob_space, ac_space, nenvs, nsteps, ent_coef, q_coef, gamma, max_grad_norm, lr,
rprop_alpha, rprop_epsilon, total_timesteps, lrschedule,
c, trust_region, alpha, delta):
sess = get_session()
nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch]) # actions
D = tf.placeholder(tf.float32, [nbatch]) # dones
R = tf.placeholder(tf.float32, [nbatch]) # rewards, not returns
MU = tf.placeholder(tf.float32, [nbatch, nact]) # mu's
LR = tf.placeholder(tf.float32, [])
eps = 1e-6
step_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs,) + ob_space.shape)
train_ob_placeholder = tf.placeholder(dtype=ob_space.dtype, shape=(nenvs*(nsteps+1),) + ob_space.shape)
with tf.variable_scope('acer_model', reuse=tf.AUTO_REUSE):
step_model = policy(nbatch=nenvs, nsteps=1, observ_placeholder=step_ob_placeholder, sess=sess)
train_model = policy(nbatch=nbatch, nsteps=nsteps, observ_placeholder=train_ob_placeholder, sess=sess)
params = find_trainable_variables("acer_model")
print("Params {}".format(len(params)))
for var in params:
print(var)
# create polyak averaged model
ema = tf.train.ExponentialMovingAverage(alpha)
ema_apply_op = ema.apply(params)
def custom_getter(getter, *args, **kwargs):
v = ema.average(getter(*args, **kwargs))
print(v.name)
return v
with tf.variable_scope("acer_model", custom_getter=custom_getter, reuse=True):
polyak_model = policy(nbatch=nbatch, nsteps=nsteps, observ_placeholder=train_ob_placeholder, sess=sess)
# Notation: (var) = batch variable, (var)s = seqeuence variable, (var)_i = variable index by action at step i
# action probability distributions according to train_model, polyak_model and step_model
# poilcy.pi is probability distribution parameters; to obtain distribution that sums to 1 need to take softmax
train_model_p = tf.nn.softmax(train_model.pi)
polyak_model_p = tf.nn.softmax(polyak_model.pi)
step_model_p = tf.nn.softmax(step_model.pi)
v = tf.reduce_sum(train_model_p * train_model.q, axis = -1) # shape is [nenvs * (nsteps + 1)]
# strip off last step
f, f_pol, q = map(lambda var: strip(var, nenvs, nsteps), [train_model_p, polyak_model_p, train_model.q])
# Get pi and q values for actions taken
f_i = get_by_index(f, A)
q_i = get_by_index(q, A)
# Compute ratios for importance truncation
rho = f / (MU + eps)
rho_i = get_by_index(rho, A)
# Calculate Q_retrace targets
qret = q_retrace(R, D, q_i, v, rho_i, nenvs, nsteps, gamma)
# Calculate losses
# Entropy
# entropy = tf.reduce_mean(strip(train_model.pd.entropy(), nenvs, nsteps))
entropy = tf.reduce_mean(cat_entropy_softmax(f))
# Policy Graident loss, with truncated importance sampling & bias correction
v = strip(v, nenvs, nsteps, True)
check_shape([qret, v, rho_i, f_i], [[nenvs * nsteps]] * 4)
check_shape([rho, f, q], [[nenvs * nsteps, nact]] * 2)
# Truncated importance sampling
adv = qret - v
logf = tf.log(f_i + eps)
gain_f = logf * tf.stop_gradient(adv * tf.minimum(c, rho_i)) # [nenvs * nsteps]
loss_f = -tf.reduce_mean(gain_f)
# Bias correction for the truncation
adv_bc = (q - tf.reshape(v, [nenvs * nsteps, 1])) # [nenvs * nsteps, nact]
logf_bc = tf.log(f + eps) # / (f_old + eps)
check_shape([adv_bc, logf_bc], [[nenvs * nsteps, nact]]*2)
gain_bc = tf.reduce_sum(logf_bc * tf.stop_gradient(adv_bc * tf.nn.relu(1.0 - (c / (rho + eps))) * f), axis = 1) #IMP: This is sum, as expectation wrt f
loss_bc= -tf.reduce_mean(gain_bc)
loss_policy = loss_f + loss_bc
# Value/Q function loss, and explained variance
check_shape([qret, q_i], [[nenvs * nsteps]]*2)
ev = q_explained_variance(tf.reshape(q_i, [nenvs, nsteps]), tf.reshape(qret, [nenvs, nsteps]))
loss_q = tf.reduce_mean(tf.square(tf.stop_gradient(qret) - q_i)*0.5)
# Net loss
check_shape([loss_policy, loss_q, entropy], [[]] * 3)
loss = loss_policy + q_coef * loss_q - ent_coef * entropy
if trust_region:
g = tf.gradients(- (loss_policy - ent_coef * entropy) * nsteps * nenvs, f) #[nenvs * nsteps, nact]
# k = tf.gradients(KL(f_pol || f), f)
k = - f_pol / (f + eps) #[nenvs * nsteps, nact] # Directly computed gradient of KL divergence wrt f
k_dot_g = tf.reduce_sum(k * g, axis=-1)
adj = tf.maximum(0.0, (tf.reduce_sum(k * g, axis=-1) - delta) / (tf.reduce_sum(tf.square(k), axis=-1) + eps)) #[nenvs * nsteps]
# Calculate stats (before doing adjustment) for logging.
avg_norm_k = avg_norm(k)
avg_norm_g = avg_norm(g)
avg_norm_k_dot_g = tf.reduce_mean(tf.abs(k_dot_g))
avg_norm_adj = tf.reduce_mean(tf.abs(adj))
g = g - tf.reshape(adj, [nenvs * nsteps, 1]) * k
grads_f = -g/(nenvs*nsteps) # These are turst region adjusted gradients wrt f ie statistics of policy pi
grads_policy = tf.gradients(f, params, grads_f)
grads_q = tf.gradients(loss_q * q_coef, params)
grads = [gradient_add(g1, g2, param) for (g1, g2, param) in zip(grads_policy, grads_q, params)]
avg_norm_grads_f = avg_norm(grads_f) * (nsteps * nenvs)
norm_grads_q = tf.global_norm(grads_q)
norm_grads_policy = tf.global_norm(grads_policy)
else:
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, norm_grads = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=rprop_alpha, epsilon=rprop_epsilon)
_opt_op = trainer.apply_gradients(grads)
# so when you call _train, you first do the gradient step, then you apply ema
with tf.control_dependencies([_opt_op]):
_train = tf.group(ema_apply_op)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
# Ops/Summaries to run, and their names for logging
run_ops = [_train, loss, loss_q, entropy, loss_policy, loss_f, loss_bc, ev, norm_grads]
names_ops = ['loss', 'loss_q', 'entropy', 'loss_policy', 'loss_f', 'loss_bc', 'explained_variance',
'norm_grads']
if trust_region:
run_ops = run_ops + [norm_grads_q, norm_grads_policy, avg_norm_grads_f, avg_norm_k, avg_norm_g, avg_norm_k_dot_g,
avg_norm_adj]
names_ops = names_ops + ['norm_grads_q', 'norm_grads_policy', 'avg_norm_grads_f', 'avg_norm_k', 'avg_norm_g',
'avg_norm_k_dot_g', 'avg_norm_adj']
def train(obs, actions, rewards, dones, mus, states, masks, steps):
cur_lr = lr.value_steps(steps)
td_map = {train_model.X: obs, polyak_model.X: obs, A: actions, R: rewards, D: dones, MU: mus, LR: cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
td_map[polyak_model.S] = states
td_map[polyak_model.M] = masks
return names_ops, sess.run(run_ops, td_map)[1:] # strip off _train
def _step(observation, **kwargs):
return step_model._evaluate([step_model.action, step_model_p, step_model.state], observation, **kwargs)
self.train = train
self.save = functools.partial(save_variables, sess=sess, variables=params)
self.train_model = train_model
self.step_model = step_model
self._step = _step
self.step = self.step_model.step
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
ob_space,
ac_space,
n_envs,
total_timesteps,
nprocs=32,
n_steps=20,
ent_coef=0.01,
vf_coef=0.25,
vf_fisher_coef=1.0,
learning_rate=0.25,
max_grad_norm=0.5,
kfac_clip=0.001,
lr_schedule='linear'):
"""
The ACKTR (Actor Critic using Kronecker-Factored Trust Region) model class, https://arxiv.org/abs/1708.05144
:param policy: (Object) The policy model to use (MLP, CNN, LSTM, ...)
:param ob_space: (Gym Space) The observation space
:param ac_space: (Gym Space) The action space
:param n_envs: (int) The number of environments
:param total_timesteps: (int) The total number of timesteps for training the model
:param nprocs: (int) The number of threads for TensorFlow operations
:param n_steps: (int) The number of steps to run for each environment
:param ent_coef: (float) The weight for the entropic loss
:param vf_coef: (float) The weight for the loss on the value function
:param vf_fisher_coef: (float) The weight for the fisher loss on the value function
:param learning_rate: (float) The initial learning rate for the RMS prop optimizer
:param max_grad_norm: (float) The clipping value for the maximum gradient
:param kfac_clip: (float) gradient clipping for Kullback leiber
:param lr_schedule: (str) The type of scheduler for the learning rate update ('linear', 'constant',
'double_linear_con', 'middle_drop' or 'double_middle_drop')
"""
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=nprocs,
inter_op_parallelism_threads=nprocs)
config.gpu_options.allow_growth = True
self.sess = sess = tf.Session(config=config)
n_batch = n_envs * n_steps
action_ph = tf.placeholder(tf.int32, [n_batch])
advs_ph = tf.placeholder(tf.float32, [n_batch])
rewards_ph = tf.placeholder(tf.float32, [n_batch])
pg_lr_ph = tf.placeholder(tf.float32, [])
self.model = step_model = policy(sess,
ob_space,
ac_space,
n_envs,
1,
reuse=False)
self.model2 = train_model = policy(sess,
ob_space,
ac_space,
n_envs * n_steps,
n_steps,
reuse=True)
logpac = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=train_model.policy, labels=action_ph)
self.logits = train_model.policy
# training loss
pg_loss = tf.reduce_mean(advs_ph * logpac)
entropy = tf.reduce_mean(calc_entropy(train_model.policy))
pg_loss = pg_loss - ent_coef * entropy
vf_loss = mse(tf.squeeze(train_model.value_fn), rewards_ph)
train_loss = pg_loss + vf_coef * vf_loss
# Fisher loss construction
self.pg_fisher = pg_fisher_loss = -tf.reduce_mean(logpac)
sample_net = train_model.value_fn + tf.random_normal(
tf.shape(train_model.value_fn))
self.vf_fisher = vf_fisher_loss = -vf_fisher_coef * tf.reduce_mean(
tf.pow(train_model.value_fn - tf.stop_gradient(sample_net), 2))
self.joint_fisher = pg_fisher_loss + vf_fisher_loss
self.params = params = find_trainable_variables("model")
self.grads_check = grads = tf.gradients(train_loss, params)
with tf.device('/gpu:0'):
self.optim = optim = kfac.KfacOptimizer(
learning_rate=pg_lr_ph,
clip_kl=kfac_clip,
momentum=0.9,
kfac_update=1,
epsilon=0.01,
stats_decay=0.99,
async=1,
cold_iter=10,
max_grad_norm=max_grad_norm)
optim.compute_and_apply_stats(self.joint_fisher, var_list=params)
train_op, q_runner = optim.apply_gradients(list(zip(grads,
params)))
self.q_runner = q_runner
self.learning_rate = Scheduler(initial_value=learning_rate,
n_values=total_timesteps,
schedule=lr_schedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for _ in range(len(obs)):
cur_lr = self.learning_rate.value()
td_map = {
train_model.obs_ph: obs,
action_ph: actions,
advs_ph: advs,
rewards_ph: rewards,
pg_lr_ph: cur_lr
}
if states is not None:
td_map[train_model.states_ph] = states
td_map[train_model.masks_ph] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, train_op], td_map)
return policy_loss, value_loss, policy_entropy
def save(save_path):
session_params = sess.run(params)
joblib.dump(session_params, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for param, loaded_p in zip(params, loaded_params):
restores.append(param.assign(loaded_p))
sess.run(restores)
self.train = train
self.save = save
self.load = load
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
ent_coef=1,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear',
batch_size=None):
sess = tf_util.get_session()
nbatch = batch_size
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# train_model is used to train our network
train_model = policy(nbatch, 1, sess)
eval_model = policy(1, 1, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
OUTPUT = tf.placeholder(tf.float32, [None, 6])
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
pi = tf.nn.softmax(OUTPUT)
train_model_pi = tf.nn.softmax(train_model.pi2)
# Policy loss
# neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
# entropy_loss = tf.reduce_mean(neglogpac)
KL = pi * tf.log(pi / train_model_pi)
KL_loss = tf.reduce_mean(tf.reduce_sum(KL, 1))
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
# regularization_loss = tf.contrib.layers.apply_regularization(regularizer=tf.contrib.layers.l2_regularizer(1e-4),
# weights_list=params)
loss = KL_loss * ent_coef
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
grads = list(zip(grads, params))
# 3. Make op for one policy and value update step of A2C
# trainer = tf.train.RMSPropOptimizer(learning_rate=7e-4, decay=alpha, epsilon=epsilon)
trainer = tf.train.AdamOptimizer(learning_rate=1e-3)
_train = trainer.apply_gradients(grads)
def train(obs, actions):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
td_map = {
train_model.X: obs,
OUTPUT: actions,
train_model.keep_prob: 1.0
}
l, _ = sess.run([loss, _train], td_map)
return l
self.train = train
self.train_model = train_model
self.act = eval_model.step
self.act2 = eval_model.step2
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def learn_ent_hoof_a2c(network,
env,
optimiser,
seed=None,
nsteps=5,
total_timesteps=int(1e6),
lr_upper_bound=None,
ent_upper_bound=None,
num_lr=None,
num_ent_coeff=None,
gamma=0.99,
max_kl=None,
max_grad_norm=0.5,
log_interval=100,
load_path=None,
**network_kwargs):
'''
Main entrypoint for A2C algorithm. Train a policy with given network architecture on a given environment using a2c algorithm.
Parameters:
-----------
network: policy network architecture. Either string (mlp, lstm, lnlstm, cnn_lstm, cnn, cnn_small, conv_only - see baselines.common/models.py for full list)
specifying the standard network architecture, or a function that takes tensorflow tensor as input and returns
tuple (output_tensor, extra_feed) where output tensor is the last network layer output, extra_feed is None for feed-forward
neural nets, and extra_feed is a dictionary describing how to feed state into the network for recurrent neural nets.
See baselines.common/policies.py/lstm for more details on using recurrent nets in policies
env: RL environment. Should implement interface similar to VecEnv (baselines.common/vec_env) or be wrapped with DummyVecEnv (baselines.common/vec_env/dummy_vec_env.py)
seed: seed to make random number sequence in the alorightm reproducible. By default is None which means seed from system noise generator (not reproducible)
nsteps: int, number of steps of the vectorized environment per update (i.e. batch size is nsteps * nenv where
nenv is number of environment copies simulated in parallel)
total_timesteps: int, total number of timesteps to train on (default: 80M)
max_gradient_norm: float, gradient is clipped to have global L2 norm no more than this value (default: 0.5)
lr: float, learning rate for RMSProp (current implementation has RMSProp hardcoded in) (default: 7e-4)
gamma: float, reward discounting parameter (default: 0.99)
log_interval: int, specifies how frequently the logs are printed out (default: 100)
**network_kwargs: keyword arguments to the policy / network builder. See baselines.common/policies.py/build_policy and arguments to a particular type of network
For instance, 'mlp' network architecture has arguments num_hidden and num_layers.
'''
set_global_seeds(seed)
# Get the nb of env
nenvs = env.num_envs
policy = build_policy(env, network, **network_kwargs)
# Instantiate the model object (that creates step_model and train_model)
model = Ent_HOOF_Model(optimiser=optimiser,
policy=policy,
env=env,
nsteps=nsteps,
total_timesteps=total_timesteps,
max_grad_norm=max_grad_norm)
runner = HOOF_Runner(env, model, nsteps=nsteps, gamma=gamma)
epinfobuf = deque(maxlen=100)
# Calculate the batch_size
nbatch = nenvs * nsteps
# model helper functions
model_params = find_trainable_variables("a2c_model")
get_flat = U.GetFlat(model_params)
set_from_flat = U.SetFromFlat(model_params)
def kl(new_mean, new_sd, old_mean, old_sd):
approx_kl = np.log(new_sd / old_sd) + (
old_sd**2 +
(old_mean - new_mean)**2) / (2.0 * new_sd**2 + 10**-8) - 0.5
approx_kl = np.sum(approx_kl, axis=1)
approx_kl = np.mean(approx_kl)
return approx_kl
if max_kl is None: # set max kl to a high val in case there is no constraint
max_kl = 10**8
# Start total timer
tstart = time.time()
for update in range(1, int(total_timesteps // nbatch + 1)):
opt_pol_val = -10**8
approx_kl = np.zeros((num_ent_coeff, num_lr))
epv = np.zeros((num_ent_coeff, num_lr))
rand_lr = lr_upper_bound * np.random.rand(num_lr)
rand_lr = np.sort(rand_lr)
rand_ent_coeff = ent_upper_bound * np.random.rand(num_ent_coeff)
old_params = get_flat()
rms_weights_before_upd = model.get_opt_state()
obs, states, rewards, masks, actions, values, undisc_rwds, epinfos = runner.run(
)
epinfobuf.extend(epinfos)
old_mean, old_sd, old_neg_ll = model.get_mean_std_neg_ll(obs, actions)
for nec in range(num_ent_coeff):
# reset policy and rms prop optimiser
set_from_flat(old_params)
model.set_opt_state(rms_weights_before_upd)
# get grads for loss fn with given entropy coeff
policy_loss, value_loss, policy_entropy = model.train(
obs, states, rewards, masks, actions, values,
rand_ent_coeff[nec])
new_params = get_flat()
ent_grads = new_params - old_params
# enumerate over different LR
for nlr in range(num_lr):
new_params = old_params + rand_lr[nlr] * ent_grads
set_from_flat(new_params)
new_mean, new_sd, new_neg_ll = model.get_mean_std_neg_ll(
obs, actions)
lik_ratio = np.exp(-new_neg_ll + old_neg_ll)
est_pol_val = wis_estimate(nenvs, nsteps, undisc_rwds,
lik_ratio)
approx_kl[nec, nlr] = kl(new_mean, new_sd, old_mean, old_sd)
epv[nec, nlr] = est_pol_val
if (nec == 0
and nlr == 0) or (est_pol_val > opt_pol_val
and approx_kl[nec, nlr] < max_kl):
opt_pol_val = est_pol_val
opt_pol_params = get_flat()
opt_rms_wts = model.get_opt_state()
opt_lr = rand_lr[nlr]
opt_ent_coeff = rand_ent_coeff[nec]
opt_kl = approx_kl[nec, nlr]
# update policy and rms prop to optimal wts
set_from_flat(opt_pol_params)
model.set_opt_state(opt_rms_wts)
# Shrink LR search space if too many get rejected
rejections = np.sum(approx_kl > max_kl) / num_lr
if rejections > 0.8:
lr_upper_bound *= 0.8
if rejections == 0:
lr_upper_bound *= 1.25
nseconds = time.time() - tstart
# Calculate the fps (frame per second)
fps = int((update * nbatch) / nseconds)
if update % log_interval == 0 or update == 1:
# Calculates if value function is a good predicator of the returns (ev > 1)
# or if it's just worse than predicting nothing (ev =< 0)
ev = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps", update * nbatch)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy", float(policy_entropy))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance", float(ev))
logger.record_tabular("opt_lr", float(opt_lr))
logger.record_tabular("ent_coeff", float(opt_ent_coeff))
logger.record_tabular("approx_kl", float(opt_kl))
logger.record_tabular("rejections", rejections)
logger.record_tabular("lr_ub", lr_upper_bound)
logger.record_tabular(
"eprewmean", safemean([epinfo['r'] for epinfo in epinfobuf]))
logger.record_tabular(
"eplenmean", safemean([epinfo['l'] for epinfo in epinfobuf]))
logger.dump_tabular()
return model
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
total_timesteps,
nprocs=32,
nsteps=20,
ent_coef=0.01,
vf_coef=0.5,
vf_fisher_coef=1.0,
lr=0.25,
max_grad_norm=0.5,
kfac_clip=0.001,
lrschedule='linear'):
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=nprocs,
inter_op_parallelism_threads=nprocs)
config.gpu_options.allow_growth = True
self.sess = sess = tf.Session(config=config)
nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
PG_LR = tf.placeholder(tf.float32, [])
VF_LR = tf.placeholder(tf.float32, [])
self.model = step_model = policy(sess,
ob_space,
ac_space,
nenvs,
1,
reuse=False)
self.model2 = train_model = policy(sess,
ob_space,
ac_space,
nenvs * nsteps,
nsteps,
reuse=True)
logpac = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=train_model.pi, labels=A)
self.logits = logits = train_model.pi
##training loss
pg_loss = tf.reduce_mean(ADV * logpac)
entropy = tf.reduce_mean(cat_entropy(train_model.pi))
pg_loss = pg_loss - ent_coef * entropy
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
train_loss = pg_loss + vf_coef * vf_loss
##Fisher loss construction
self.pg_fisher = pg_fisher_loss = -tf.reduce_mean(logpac)
sample_net = train_model.vf + tf.random_normal(tf.shape(
train_model.vf))
self.vf_fisher = vf_fisher_loss = -vf_fisher_coef * tf.reduce_mean(
tf.pow(train_model.vf - tf.stop_gradient(sample_net), 2))
self.joint_fisher = joint_fisher_loss = pg_fisher_loss + vf_fisher_loss
self.params = params = find_trainable_variables("model")
self.grads_check = grads = tf.gradients(train_loss, params)
with tf.device('/gpu:0'):
self.optim = optim = kfac.KfacOptimizer(learning_rate=PG_LR, clip_kl=kfac_clip,\
momentum=0.9, kfac_update=1, epsilon=0.01,\
stats_decay=0.99, async=1, cold_iter=10, max_grad_norm=max_grad_norm)
update_stats_op = optim.compute_and_apply_stats(joint_fisher_loss,
var_list=params)
train_op, q_runner = optim.apply_gradients(list(zip(grads,
params)))
self.q_runner = q_runner
self.lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = self.lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
PG_LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, train_op], td_map)
return policy_loss, value_loss, policy_entropy
def save(save_path):
ps = sess.run(params)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
sess.run(restores)
self.train = train
self.save = save
self.load = load
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
tf.global_variables_initializer().run(session=sess)
def __init__(self,
optimiser,
policy,
env,
nsteps,
vf_coef=0.5,
max_grad_norm=0.5,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6)):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs * nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
Ent_Coeff = tf.placeholder(tf.float32, []) # for Entropy
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# Value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy * Ent_Coeff + vf_loss * vf_coef
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
if optimiser == 'RMSProp':
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
elif optimiser == 'SGD':
trainer = tf.train.GradientDescentOptimizer(learning_rate=LR)
_train = trainer.apply_gradients(grads)
#https://stackoverflow.com/a/45624533
_slot_vars = [
trainer.get_slot(var, name) for name in trainer.get_slot_names()
for var in params
]
SLOTS = [tf.placeholder(tf.float32, slot.shape) for slot in _slot_vars]
_set_slots = [var.assign(SLOTS[i]) for i, var in enumerate(_slot_vars)]
def get_opt_state():
return sess.run(_slot_vars)
def set_opt_state(state):
feed = {k: v for k, v in zip(SLOTS, state)}
return sess.run(_set_slots, feed)
def train(obs, states, rewards, masks, actions, values, ent_coeff):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
Ent_Coeff: ent_coeff,
LR: 1.0
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
# Only this bit added
def get_mean_std_neg_ll(obs, actions):
td_map = {train_model.X: obs, A: actions}
vals = sess.run(
[train_model.pd.mean, train_model.pd.std, neglogpac], td_map)
return vals
self.get_mean_std_neg_ll = get_mean_std_neg_ll
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
self.get_opt_state = get_opt_state
self.set_opt_state = set_opt_state
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
env,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear'):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs * nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
step_model = policy(nenvs, 1, sess)
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
neglogpac = train_model.pd.neglogp(A)
entropy = tf.reduce_mean(train_model.pd.entropy())
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables("a2c_model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, ob_space, ac_space, nenvs, nsteps,
ent_coef=0.01, vf_coef=0.5, max_grad_norm=0.5, lr=7e-4,
alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6),
lrschedule='linear', replay_lambda=1, ss_rate=1,
replay_loss=None):
sess = tf_util.make_session()
nact = ac_space.n
nbatch = nenvs*nsteps
# If we have replay_loss, create replay buffer and stage buffer
# Use this to enforce replay loss lower
if replay_loss is not None:
self.replay_buffer = [] # holds all past data
A = tf.placeholder(tf.int32, [nbatch])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
step_model = policy(sess, ob_space, ac_space, nenvs, 1, reuse=False)
train_model = policy(sess, ob_space, ac_space, nenvs*nsteps, nsteps, reuse=True)
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=train_model.pi, labels=A)
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
entropy = tf.reduce_mean(cat_entropy(train_model.pi))
# Introduce replay_loss if given
if replay_loss == "L2":
# Replace train_model.pi with whatever is predicted label
# Replace A with whatever is recorded label
re_loss = tf.nn.l2_loss(tf.nn.softmax(train_model.pi) - A) / nbatch
elif replay_loss == "Distillation":
# Replace y_donor with whatever is recorded label
# Replace y_acceptor with whatever is predicted label
re_loss = tf.reduce_mean( - tf.reduce_sum(tf.stop_gradient(y_donor)
* tf.log(y_acceptor),
reduction_indices=1))
loss = pg_loss - entropy*ent_coef + vf_loss * vf_coef
if replay_loss is not None:
loss = loss + replay_lambda*re_loss
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=alpha, epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, LR:cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train],
td_map
)
return policy_loss, value_loss, policy_entropy
def save(save_path):
ps = sess.run(params)
make_path(save_path)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
ps = sess.run(restores)
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = save
self.load = load
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, env, nsteps,
ent_coef=0.01, vf_coef=0.5, max_grad_norm=0.5, lr=7e-4,
alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6), lrschedule='linear'):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs*nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# Value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy*ent_coef + vf_loss * vf_coef
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=alpha, epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, LR:cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train],
td_map
)
return policy_loss, value_loss, policy_entropy
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def __init__(self,
optim,
policy,
ob_dim,
ac_dim,
num_procs,
max_grad_norm=0.5,
lr=7e-4,
vf_lr=0.001,
cv_lr=0.001,
cv_num=25,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear',
logdir=None):
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=num_procs,
inter_op_parallelism_threads=num_procs)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
A = tf.placeholder(tf.float32, [None, ac_dim], name="A")
ADV = tf.placeholder(tf.float32, [None], name="ADV")
R = tf.placeholder(tf.float32, [None], name="R")
train_model = policy(sess, ob_dim, ac_dim, vf_lr, cv_lr, reuse=False)
step_model = policy(sess, ob_dim, ac_dim, vf_lr, cv_lr, reuse=True)
params = find_trainable_variables("model")
tf.summary.histogram("vf", train_model.vf)
pi_params = [v for v in params if "pi" in v.name]
vf_params = [v for v in params if "vf" in v.name]
logpac = train_model.logprob_n
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
pg_loss = -tf.reduce_mean(ADV * logpac)
tf.summary.scalar("vf_loss", vf_loss)
if train_model.relaxed:
ddiff_loss = tf.reduce_mean(train_model.cv)
ddiff_grads_mean = tf.gradients(ddiff_loss, pi_params)
ddiff_grads_std = tf.gradients(ddiff_loss, train_model.logstd_1a)
dlogp_dmean = (A - train_model.mean) / tf.square(
train_model.std_na)
dlogp_dstd = -1 / train_model.std_na + 1 / tf.pow(
train_model.std_na, 3) * tf.square(A - train_model.mean)
pi_grads_mean = -((tf.expand_dims(ADV, 1) - train_model.cv) *
dlogp_dmean) / tf.to_float(tf.shape(ADV)[0])
pg_grads_mean = tf.gradients(train_model.mean,
pi_params,
grad_ys=pi_grads_mean)
pg_grads_mean = [
pg - dg for pg, dg in zip(pg_grads_mean, ddiff_grads_mean)
]
pi_grads_std = -((tf.expand_dims(ADV, 1) - train_model.cv) *
dlogp_dstd) / tf.to_float(tf.shape(ADV)[0])
pg_grads_std = tf.gradients(train_model.std_na,
train_model.logstd_1a,
grad_ys=pi_grads_std)
pg_grads_std = [
pg - dg for pg, dg in zip(pg_grads_std, ddiff_grads_std)
]
pg_grads = pg_grads_mean + pg_grads_std
cv_loss = tf.concat([tf.reshape(p, [-1]) for p in pg_grads], 0)
cv_loss = tf.squeeze(tf.reduce_sum(tf.square(cv_loss)))
tf.summary.scalar("cv_loss", cv_loss)
cv_params = [v for v in params if "cv" in v.name]
cv_grads = tf.gradients(cv_loss, cv_params)
cv_gradvars = list(zip(cv_grads, cv_params))
else:
pg_grads = tf.gradients(pg_loss, pi_params) + tf.gradients(
pg_loss, train_model.logstd_1a)
all_policy_grads = tf.concat([tf.reshape(pg, [-1]) for pg in pg_grads],
0)
# policy gradients
policy_gradvars = list(
zip(pg_grads, pi_params + [train_model.logstd_1a]))
vf_grads = tf.gradients(vf_loss, vf_params)
vf_gradvars = list(zip(vf_grads, vf_params))
grads_list = policy_gradvars + vf_gradvars
if train_model.relaxed: grads_list += cv_gradvars
for g, v in grads_list:
tf.summary.histogram(v.name, v)
tf.summary.histogram(v.name + "_grad", g)
sum_op = tf.summary.merge_all()
writer = tf.summary.FileWriter(logdir)
trainer = optim
_train = trainer.apply_gradients(policy_gradvars)
_vf_train = train_model.vf_optim.apply_gradients(vf_gradvars)
self._step = 0
def get_cv_grads(obs, old_actions, advs, rewards, vf_in, values):
advs = rewards - values
td_map = {
train_model.ob_no: obs,
train_model.oldac_na: old_actions,
train_model.X: vf_in,
A: old_actions,
ADV: advs,
R: rewards
}
cv_gs = sess.run(cv_grads, td_map)
return cv_gs
def update_cv(mean_cv_gs):
cv_gvs = list(zip(mean_cv_gs, cv_params))
train_model.cv_optim.apply_gradients(cv_gvs)
def update_policy_and_value(obs,
old_actions,
advs,
rewards,
vf_in,
values,
summary=False):
advs = rewards - values
td_map = {
train_model.ob_no: obs,
train_model.oldac_na: old_actions,
train_model.X: vf_in,
A: old_actions,
ADV: advs,
R: rewards
}
for _ in range(25):
sess.run(
_vf_train, {
train_model.ob_no: obs,
train_model.oldac_na: old_actions,
train_model.X: vf_in,
A: old_actions,
ADV: advs,
R: rewards
})
if summary:
sum_str, policy_loss, value_loss, _, = sess.run(
[sum_op, pg_loss, vf_loss, _train], td_map)
writer.add_summary(sum_str, self._step)
else:
policy_loss, value_loss, _ = sess.run(
[pg_loss, vf_loss, _train], td_map)
self._step += 1
return policy_loss, value_loss
def get_grads(obs, old_actions, advs, rewards, vf_in, value):
advs = rewards - value
td_map = {
train_model.ob_no: obs,
train_model.oldac_na: old_actions,
train_model.X: vf_in,
A: old_actions,
ADV: advs,
R: rewards
}
_g = all_policy_grads # seems to already have happened? / tf.to_float(tf.shape(rewards)[0])
pg = sess.run(_g, td_map)
return pg
def save(save_path):
ps = sess.run(params)
make_path(save_path)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
ps = sess.run(restores)
self.sess = sess
self.get_cv_grads = get_cv_grads
self.update_cv = update_cv
self.update_policy_and_value = update_policy_and_value
self.train_model = train_model
self.step_model = step_model
self.value = train_model.value
self.get_grads = get_grads
self.save = save
self.load = load
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(20e6),
lrschedule='linear'):
sess = tf.get_default_session()
nbatch = nenvs * nsteps
step_model = policy(sess, ob_space, ac_space, nenvs, 1, reuse=False)
train_model = policy(sess,
ob_space,
ac_space,
nenvs * nsteps,
nsteps,
reuse=True)
A = train_model.pdtype.sample_placeholder([nbatch])
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
neglogpac = train_model.pd.neglogp(A)
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = tf.reduce_mean(mse(tf.squeeze(train_model.vf), R))
entropy = tf.reduce_mean(train_model.pd.entropy())
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, discounted_rewards, rewards, masks,
prev_actions, actions, values, dones):
advs = discounted_rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
# reshape actions, rewards, and dones to have first dimension of size nenvs*nsteps, existing second dimension
# this is already done for obs
rews = np.reshape(rewards, (nbatch, 1))
ds = np.reshape(np.asarray(dones, dtype=np.float32), (nbatch, 1))
if len(ac_space.shape) == 0:
prev_actions = np.reshape(prev_actions, (nbatch, ))
one_hot = np.eye(ac_space.n)[prev_actions]
for i in range(nbatch):
if prev_actions[i] == -1:
one_hot[i, :] = np.zeros((ac_space.n, ), dtype=np.int)
x = np.concatenate((obs, one_hot, rews, ds), axis=1)
actions = np.reshape(actions, (nbatch, ))
else:
prev_actions = np.reshape(prev_actions,
(nbatch, ac_space.shape[0]))
x = np.concatenate((obs, prev_actions, rews, ds), axis=1)
td_map = {
train_model.X: x,
A: actions,
ADV: advs,
R: discounted_rewards,
LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
def save(save_path):
ps = sess.run(params)
make_path(osp.dirname(save_path))
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
sess.run(restores)
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = save
self.load = load
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
total_timesteps,
nprocs=32,
nscripts=16,
nsteps=20,
nstack=4,
ent_coef=0.1,
vf_coef=0.5,
vf_fisher_coef=1.0,
lr=0.25,
max_grad_norm=0.001,
kfac_clip=0.001,
lrschedule='linear',
alpha=0.99,
epsilon=1e-5):
config = tf.ConfigProto(
allow_soft_placement=True,
intra_op_parallelism_threads=nprocs,
inter_op_parallelism_threads=nprocs)
config.gpu_options.allow_growth = True
self.sess = sess = tf.Session(config=config)
nsml.bind(sess=sess)
#nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch])
XY0 = tf.placeholder(tf.int32, [nbatch])
XY1 = tf.placeholder(tf.int32, [nbatch])
# ADV == TD_TARGET - values
ADV = tf.placeholder(tf.float32, [nbatch])
TD_TARGET = tf.placeholder(tf.float32, [nbatch])
PG_LR = tf.placeholder(tf.float32, [])
VF_LR = tf.placeholder(tf.float32, [])
self.model = step_model = policy(
sess, ob_space, ac_space, nenvs, 1, nstack, reuse=False)
self.model2 = train_model = policy(
sess, ob_space, ac_space, nenvs, nsteps, nstack, reuse=True)
# Policy 1 : Base Action : train_model.pi label = A
script_mask = tf.concat(
[
tf.zeros([nscripts * nsteps, 1]),
tf.ones([(nprocs - nscripts) * nsteps, 1])
],
axis=0)
pi = train_model.pi
pac_weight = script_mask * (tf.nn.softmax(pi) - 1.0) + 1.0
pac_weight = tf.reduce_sum(pac_weight * tf.one_hot(A, depth=3), axis=1)
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=pi, labels=A)
neglogpac *= tf.stop_gradient(pac_weight)
inv_A = 1.0 - tf.cast(A, tf.float32)
xy0_mask = tf.cast(A, tf.float32)
xy1_mask = tf.cast(A, tf.float32)
condition0 = tf.equal(xy0_mask, 2)
xy0_mask = tf.where(condition0, tf.ones(tf.shape(xy0_mask)), xy0_mask)
xy0_mask = 1.0 - xy0_mask
condition1 = tf.equal(xy1_mask, 2)
xy1_mask = tf.where(condition1, tf.zeros(tf.shape(xy1_mask)), xy1_mask)
# One hot representation of chosen marine.
# [batch_size, 2]
pi_xy0 = train_model.pi_xy0
pac_weight = script_mask * (tf.nn.softmax(pi_xy0) - 1.0) + 1.0
pac_weight = tf.reduce_sum(
pac_weight * tf.one_hot(XY0, depth=1024), axis=1)
logpac_xy0 = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=pi_xy0, labels=XY0)
logpac_xy0 *= tf.stop_gradient(pac_weight)
logpac_xy0 *= tf.cast(xy0_mask, tf.float32)
pi_xy1 = train_model.pi_xy1
pac_weight = script_mask * (tf.nn.softmax(pi_xy1) - 1.0) + 1.0
pac_weight = tf.reduce_sum(
pac_weight * tf.one_hot(XY0, depth=1024), axis=1)
# 1D? 2D?
logpac_xy1 = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=pi_xy1, labels=XY1)
logpac_xy1 *= tf.stop_gradient(pac_weight)
logpac_xy1 *= tf.cast(xy1_mask, tf.float32)
pg_loss = tf.reduce_mean(ADV * neglogpac)
pg_loss_xy0 = tf.reduce_mean(ADV * logpac_xy0)
pg_loss_xy1 = tf.reduce_mean(ADV * logpac_xy1)
vf_ = tf.squeeze(train_model.vf)
vf_r = tf.concat(
[
tf.ones([nscripts * nsteps, 1]),
tf.zeros([(nprocs - nscripts) * nsteps, 1])
],
axis=0) * TD_TARGET
vf_masked = vf_ * script_mask + vf_r
#vf_mask[0:nscripts * nsteps] = R[0:nscripts * nsteps]
vf_loss = tf.reduce_mean(mse(vf_masked, TD_TARGET))
entropy_a = tf.reduce_mean(cat_entropy(train_model.pi))
entropy_xy0 = tf.reduce_mean(cat_entropy(train_model.pi_xy0))
entropy_xy1 = tf.reduce_mean(cat_entropy(train_model.pi_xy1))
entropy = entropy_a + entropy_xy0 + entropy_xy1
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, _ = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(
learning_rate=lr, decay=alpha, epsilon=epsilon)
_train = trainer.apply_gradients(grads)
self.logits = logits = train_model.pi
# xy0
self.params_common = params_common = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='model/common')
self.params_xy0 = params_xy0 = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES,
scope='model/xy0') + params_common
train_loss_xy0 = pg_loss_xy0 - entropy * ent_coef + vf_coef * vf_loss
self.grads_check_xy0 = grads_xy0 = tf.gradients(
train_loss_xy0, params_xy0)
if max_grad_norm is not None:
grads_xy0, _ = tf.clip_by_global_norm(grads_xy0, max_grad_norm)
grads_xy0 = list(zip(grads_xy0, params_xy0))
trainer_xy0 = tf.train.RMSPropOptimizer(
learning_rate=lr, decay=alpha, epsilon=epsilon)
_train_xy0 = trainer_xy0.apply_gradients(grads_xy0)
# xy1
self.params_xy1 = params_xy1 = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES,
scope='model/xy1') + params_common
train_loss_xy1 = pg_loss_xy1 - entropy * ent_coef + vf_coef * vf_loss
self.grads_check_xy1 = grads_xy1 = tf.gradients(
train_loss_xy1, params_xy1)
if max_grad_norm is not None:
grads_xy1, _ = tf.clip_by_global_norm(grads_xy1, max_grad_norm)
grads_xy1 = list(zip(grads_xy1, params_xy1))
trainer_xy1 = tf.train.RMSPropOptimizer(
learning_rate=lr, decay=alpha, epsilon=epsilon)
_train_xy1 = trainer_xy1.apply_gradients(grads_xy1)
self.lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, td_targets, masks, actions, xy0, xy1, values):
advs = td_targets - values
for step in range(len(obs)):
cur_lr = self.lr.value()
td_map = {
train_model.X: obs,
A: actions,
XY0: xy0,
XY1: xy1,
ADV: advs,
TD_TARGET: td_targets,
PG_LR: cur_lr
}
if states != []:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _, \
policy_loss_xy0, policy_entropy_xy0, _, \
policy_loss_xy1, policy_entropy_xy1, _ = sess.run(
[pg_loss, vf_loss, entropy, _train,
pg_loss_xy0, entropy_xy0, _train_xy0,
pg_loss_xy1, entropy_xy1, _train_xy1],
td_map)
return policy_loss, value_loss, policy_entropy, \
policy_loss_xy0, policy_entropy_xy0, \
policy_loss_xy1, policy_entropy_xy1
def save(save_path):
ps = sess.run(params)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
sess.run(restores)
self.train = train
self.save = save
self.load = load
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
print("global_variables_initializer start")
tf.global_variables_initializer().run(session=sess)
print("global_variables_initializer complete")