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, ts_pir, ts_pi_given_r
tfFunctionApproximator._build_graph(self, **kwargs)
# r_dim: dimension of randomness in generating actions.
# build additional graphs for Policy
# build conditional distribution
self._pi = self._yh
self._pi_given_r = U.function([self.ph_x, self.ph_r],
self.ts_pi_given_r)
self._pid = U.function([self.ph_x],
self.ts_pid) # derandomized actions
# actions and the randomness used in generating actions concatenated.
self._pir = U.function([self.ph_x], self.ts_pir)
self._logp = U.function([self.ph_x, self.ph_y], self.ts_logp)
self._logp_grad = U.function([self.ph_x, self.ph_y],
tf.gradients(self.ts_logp, self.ts_vars))
# 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)
# build nabla logp f.
ts_loss = tf.reduce_sum(self.ph_f * self.ts_logp) # sum!!
ts_grads = U.gradients(ts_loss, self.ts_vars)
ts_grads = [
g if g is not None else tf.zeros_like(v)
for (v, g) in zipsame(self.ts_vars, ts_grads)
]
# need to flatten
compute_ts_grad = U.function([self.ph_x, self.ph_y, self.ph_f],
ts_grads)
self.nabla_logp_f = lambda x, y, f: flatten(compute_ts_grad(x, y, f))
def flatten(self, vs):
return flatten(vs)
def logp_grad(self, x, y):
# XXX Only support single instance, due to tf can not compute jacobian.
# Should throw error if more than one instances are provided.
# Return a flat np array.
return flatten(self._logp_grad(x[None], y[None]))
def _build_dist(self, ts_nor_x, ph_y):
# mean and std
self.ts_mean = cls._build_func_apprx(
self,
ts_nor_x) # use the tfFunctionApproximator to define mean
self._ts_logstd = tf.get_variable(
'logstd',
shape=[self.y_dim],
initializer=tf.constant_initializer(np.log(self._init_std)))
self._ts_stop_std_grad = tf.get_variable(
'stop_std_grad',
initializer=tf.constant(False),
trainable=False)
_ts_logstd = tf.cond(
self._ts_stop_std_grad, # whether to stop gradient
true_fn=lambda: tf.stop_gradient(self._ts_logstd),
false_fn=lambda: self._ts_logstd)
# make sure the distribution does not degenerate
self.ts_logstd = tf.maximum(tf.to_float(np.log(self._min_std)),
_ts_logstd)
ts_std = tf.exp(self.ts_logstd)
self._std = U.function([], ts_std)
self._set_logstd = U.build_set([self._ts_logstd])
self._set_stop_std_grad = U.build_set([self._ts_stop_std_grad])
# pi
# self.ts_noise = tf.random_normal(tf.shape(ts_std), stddev=ts_std, seed=self.seed)
rand = tf.random_normal(tf.shape(self.ts_mean), seed=self.seed)
ts_noise = ts_std * rand
ts_pi = self.ts_mean + ts_noise
ts_pid = self.ts_mean
# Need to broadcast noise to each row.
# n = tf.shape(self.ts_mean)[0]
# noise = tf.reshape(tf.tile(self.ts_noise, [n]), [n, -1])
# ts_pir = tf.concat([ts_pi, noise], 1)
ts_pir = tf.concat([ts_pi, rand], axis=1)
ts_pi_given_r = self.ts_mean + ts_std * self.ph_r
# logp
ts_logp = self._build_logp(self.y_dim, ph_y, self.ts_mean,
self.ts_logstd)
# XXX Full matrix
# expectation with a quadratic function 0.5 y^t A y + b^t y + c
ph_A = tf.placeholder(tf_float,
shape=(None, self.y_dim, self.y_dim),
name='A')
ph_b = tf.placeholder(tf_float,
shape=(None, self.y_dim, 1),
name='b')
ph_c = tf.placeholder(tf_float, shape=(None, ), name='c')
ph_w = tf.placeholder(tf_float, shape=(None, ), name='w')
ts_mean = tf.expand_dims(self.ts_mean, -1) # None * dim_y * 1
ts_var = tf.reshape(ts_std**2,
(1, self.y_dim, 1)) # None * dim_y * 1
ts_quad_exp = tf.squeeze(0.5*tf.matmul(ts_mean, tf.matmul(ph_A, ts_mean), adjoint_a=True)) \
+ tf.squeeze(tf.matmul(ph_b, ts_mean,adjoint_a=True)) \
+ ph_c + 0.5*tf.trace(ph_A*ts_var) # None
self._quad_exp = U.function([self.ph_x, ph_A, ph_b, ph_c],
ts_quad_exp)
ts_quad_exp_grads = U.gradients(ts_quad_exp * ph_w,
self.ts_vars) # just a vector
compute_ts_quad_exp_grad = U.function(
[self.ph_x, ph_A, ph_b, ph_c, ph_w], ts_quad_exp_grads)
self._nabla_quad_exp = lambda x, A, b, c, w: flatten(
compute_ts_quad_exp_grad(x, A, b, c, w))
# XXX Diagonal
ph_A_diag = tf.placeholder(tf_float,
shape=(None, self.y_dim, 1),
name='A_diag')
ts_quad_exp_diag = tf.squeeze(0.5*tf.matmul(ts_mean, ph_A_diag*ts_mean, adjoint_a=True)) \
+ tf.squeeze(tf.matmul(ph_b, ts_mean,adjoint_a=True)) \
+ ph_c + 0.5*tf.squeeze(tf.reduce_sum(ph_A_diag*ts_var, axis=1)) # None
self._quad_exp_diag = U.function(
[self.ph_x, ph_A_diag, ph_b, ph_c], ts_quad_exp_diag)
ts_quad_exp_grads_diag = U.gradients(ts_quad_exp_diag * ph_w,
self.ts_vars) # just a vector
compute_ts_quad_exp_grad_diag = U.function(
[self.ph_x, ph_A_diag, ph_b, ph_c, ph_w],
ts_quad_exp_grads_diag)
self._nabla_quad_exp_diag = lambda x, A, b, c, w: flatten(
compute_ts_quad_exp_grad_diag(x, A, b, c, w))
return ts_pi, ts_logp, ts_pid, ts_pir, ts_pi_given_r
def compute_grad(self):
if self._args is None:
raise ValueError('Oracle has not been initialized')
grads = self._compute_grad(*self._args)
return flatten(grads)