def advs(self,
ro,
lambd=None,
use_is=None,
ref_policy=None): # advantage function
""" Compute adv (evaluated at ro) wrt to ref_policy.
ro: a list of Rollout isinstances
Note `ref_policy` argument is only considered when `self.use_is`
is True; in this case, if `ref_policy` is None, it is wrt to
`self.ref_policy`. Otherwise, when `self.use_is`_is is False, the
adv is biased toward the behavior policy that collected the data.
"""
use_is = use_is or self.use_is
vfns = self.vfns(ro)
if use_is is 'multi':
ws = self.weights(ro, ref_policy) # importance weight
advs = [
self._pe.adv(rollout.rws, vf, rollout.done, w=w, lambd=lambd)
for rollout, vf, w in zipsame(ro, vfns, ws)
]
else:
advs = [
self._spe.adv(rollout.rws,
vf,
rollout.done,
w=1.0,
lambd=lambd)
for rollout, vf in zipsame(ro, vfns)
]
return advs, vfns
def _build_graph(self, **bg_kwargs):
ts_loss, ph_args = self._build_loss_op(**bg_kwargs)
# define compute_loss and compute_grad wrt loss
self._compute_loss = U.function(ph_args, ts_loss)
ts_grads = U.gradients(ts_loss, self._ts_vars)
# fill None with zeros; otherwise tf.run will attempt to fetch for None.
ts_grads = [g if g is not None else tf.zeros_like(v) for (v, g) in
zipsame(self._ts_vars, ts_grads)]
self._compute_grad = U.function(ph_args, ts_grads)
def ts_fvp0(self, ts_xs, ts_ys, ts_gs):
""" Computes F(self.pi)*g based on the expected outer product. """
dummy = tf.ones(shape=(len(ts_xs,)))
with tf.GradientTape(watch_accessed_variables=False) as gt:
gt.watch(dummy)
ts_sum_logp_grads = self.ts_logp_grad(ts_xs, ts_ys, dummy)
ts_pd = tf.math.accumulate_n([tf.reduce_sum(u*v) for (u, v) in zipsame(ts_sum_logp_grads, ts_gs)])
ts_fs = gt.gradient(ts_pd, dummy) # shape (N,)
N = tf.constant(len(dummy),dtype=tf_float)
return self.ts_logp_grad(ts_xs, ts_ys, ts_fs/N)
def ts_fvp(self, ts_xs, ts_gs):
""" Computes F(self.pi)*g based on the Hessian of the entropy. """
with tf.GradientTape(watch_accessed_variables=False) as gt:
gt.watch(self.ts_variables)
with tf.GradientTape(watch_accessed_variables=False) as gt2:
gt2.watch(self.ts_variables) # TODO add sample weight below??
ts_kl = self.ts_kl(self, ts_xs, p1_sg=True)
ts_kl_grads = gt2.gradient(ts_kl, self.ts_variables)
ts_pd = tf.math.accumulate_n([tf.reduce_sum(kg*v) for (kg, v) in zipsame(ts_kl_grads, ts_gs)])
ts_fvp = gt.gradient(ts_pd, self.ts_variables)
return ts_fvp
def ts_fvp(self, ts_xs, ts_gs):
""" Computes F(self.pi)*g, where F is the Fisher information matrix and
g is a np.ndarray in the same shape as self.variable """
with tf.GradientTape() as gt:
gt.watch(self.ts_variables)
with tf.GradientTape() as gt2:
gt2.watch(self.ts_variables) # TODO add sample weight below??
ts_kl = self.ts_kl(self, ts_xs, p1_sg=True)
ts_kl_grads = gt2.gradient(ts_kl, self.ts_variables)
ts_pd = tf.add_n([
tf.reduce_sum(kg * v)
for (kg, v) in zipsame(ts_kl_grads, ts_gs)
])
ts_fvp = gt.gradient(ts_pd, self.ts_variables)
return ts_fvp
def get_combs_and_keys(ranges):
keys = []
values = []
for r in ranges:
keys += r[::2]
values = [list(zipsame(*r[1::2])) for r in ranges]
cs = itertools.product(*values)
combs = []
for c in cs:
comb = []
for x in c:
comb += x
# print(comb)
combs.append(comb)
return combs, keys
def split(self, ro, policy_as_expert):
# Split ro into two phases
rollouts = ro.to_list()
ro_mix = [rollouts[i] for i in self._ind_ro_mix]
ro_pol = [rollouts[i] for i in self._ind_ro_pol]
assert (len(ro_mix) + len(ro_pol)) == len(rollouts)
ro_exps = [[] for _ in range(len(self.experts))]
for r, t, s, k in zipsame(ro_mix, self._t_switch, self._scale,
self._k_star):
assert len(r) >= t # because t >= 1
if not policy_as_expert or k < len(self.experts) - 1:
# we assume the last expert is the learner
r = r[t:]
r.weight = 1.0
ro_exps[k].append(r)
if policy_as_expert:
ro_pol += ro_exps[-1]
del ro_exps[-1]
ro_exps = [Dataset(ro_exp) for ro_exp in ro_exps]
ro_pol = Dataset(ro_pol)
return ro_exps, ro_pol
def _build_graph(self, **kwargs):
""" We treat tfFunctionApproximator as the stochastic map of the policy
(which inputs ph_x and outputs ts_yh) and build additional
attributes/methods required by Policy """
# build tf.Variables
# add attributes ph_x, ts_nor_x, ts_y, _yh, _sh_vars,
# ph_y, ts_pi, ts_logp, ts_pid
tfFunctionApproximator._build_graph(self, **kwargs)
# build additional graphs for Policy
# build conditional distribution
self._pi = self._yh
self._pid = U.function([self.ph_x], self.ts_pid)
self._logp = U.function([self.ph_x, self.ph_y], self.ts_logp)
# build fvp operator (this depends only on self)
ph_g, ts_grads = self._sh_vars.build_flat_ph()
ts_kl = self.build_kl(self, self, p1_sg=True)
ts_kl_grads = U.gradients(ts_kl, self.ts_vars) # grad to the 2nd arg of KL
ts_inner_prod = tf.add_n([tf.reduce_sum(kg * v) for (kg, v) in zipsame(ts_kl_grads, ts_grads)])
ts_fvp = U.gradients(ts_inner_prod, self.ts_vars) # Fisher (information matrix) and Vector Product
ts_fvp = tf.concat([tf.reshape(f, [-1]) for f in ts_fvp], axis=-1) # continuous vector
self._fvp = U.function([self.ph_x, ph_g], ts_fvp)
def mean_variable(self, val):
vals = unflatten(val, shapes=self.mean_var_shapes)
[var.assign(val) for var, val in zipsame(self.ts_mean_variables, vals)]
def variables(self, vals): # vals can be a list of nd.array or tf.Tensor
[var.assign(val) for var, val in zipsame(self.ts_variables, vals)]
def var_assign(ts_var, x):
# convert np.ndarray(s) to tf.Tensor(s)
if type(ts_var) is list:
return [vv.assign(xx) for vv, xx in zipsame(ts_var, x)]
else:
return ts_var(x)