def _build_losses(self, json_data):
actor_weight_decay = 0 if (self.ACTOR_WEIGHT_DECAY_KEY not in json_data) else json_data[self.ACTOR_WEIGHT_DECAY_KEY]
critic_weight_decay = 0 if (self.CRITIC_WEIGHT_DECAY_KEY not in json_data) else json_data[self.CRITIC_WEIGHT_DECAY_KEY]
norm_val_diff = self.val_norm.normalize_tf(self.tar_val_tf) - self.val_norm.normalize_tf(self.critic_tf)
self.critic_loss_tf = 0.5 * tf.reduce_mean(tf.square(norm_val_diff))
if (critic_weight_decay != 0):
self.critic_loss_tf += critic_weight_decay * self._weight_decay_loss('main/critic')
norm_tar_a_tf = self.a_norm.normalize_tf(self.a_tf)
self._norm_a_mean_tf = self.a_norm.normalize_tf(self.a_mean_tf)
self.logp_tf = TFUtil.calc_logp_gaussian(norm_tar_a_tf, self._norm_a_mean_tf, self.norm_a_std_tf)
ratio_tf = tf.exp(self.logp_tf - self.old_logp_tf)
actor_loss0 = self.adv_tf * ratio_tf
actor_loss1 = self.adv_tf * tf.clip_by_value(ratio_tf, 1.0 - self.ratio_clip, 1 + self.ratio_clip)
self.actor_loss_tf = -tf.reduce_mean(tf.minimum(actor_loss0, actor_loss1))
norm_a_bound_min = self.a_norm.normalize(self.a_bound_min)
norm_a_bound_max = self.a_norm.normalize(self.a_bound_max)
a_bound_loss = TFUtil.calc_bound_loss(self._norm_a_mean_tf, norm_a_bound_min, norm_a_bound_max)
self.actor_loss_tf += a_bound_loss
if (actor_weight_decay != 0):
self.actor_loss_tf += actor_weight_decay * self._weight_decay_loss('main/actor')
# for debugging
self.clip_frac_tf = tf.reduce_mean(tf.to_float(tf.greater(tf.abs(ratio_tf - 1.0), self.ratio_clip)))
return
def _build_losses(self, json_data):
actor_weight_decay = 0 if (self.ACTOR_WEIGHT_DECAY_KEY not in json_data) else json_data[
self.ACTOR_WEIGHT_DECAY_KEY]
critic_weight_decay = 0 if (self.CRITIC_WEIGHT_DECAY_KEY not in json_data) else json_data[
self.CRITIC_WEIGHT_DECAY_KEY]
norm_val_diff = self.val_norm.normalize_tf(self.tar_val_tf) - self.val_norm.normalize_tf(self.critic_tf)
self.critic_loss_tf = 0.5 * tf.reduce_mean(tf.square(norm_val_diff))
if (critic_weight_decay != 0):
self.critic_loss_tf += critic_weight_decay * self._weight_decay_loss('main/critic')
norm_tar_a_tf = self.a_norm.normalize_tf(self.a_tf)
self._norm_a_mean_tf = self.a_norm.normalize_tf(self.a_mean_tf)
self.logp_tf = TFUtil.calc_logp_gaussian(norm_tar_a_tf, self._norm_a_mean_tf, self.norm_a_std_tf)
ratio_tf = tf.exp(self.logp_tf - self.old_logp_tf)
actor_loss0 = self.adv_tf * ratio_tf
actor_loss1 = self.adv_tf * tf.clip_by_value(ratio_tf, 1.0 - self.ratio_clip, 1 + self.ratio_clip)
self.actor_loss_tf = -tf.reduce_mean(tf.minimum(actor_loss0, actor_loss1))
norm_a_bound_min = self.a_norm.normalize(self.a_bound_min)
norm_a_bound_max = self.a_norm.normalize(self.a_bound_max)
a_bound_loss = TFUtil.calc_bound_loss(self._norm_a_mean_tf, norm_a_bound_min, norm_a_bound_max)
self.actor_loss_tf += a_bound_loss
if actor_weight_decay != 0:
self.actor_loss_tf += actor_weight_decay * self._weight_decay_loss('main/actor')
# for debugging
self.clip_frac_tf = tf.reduce_mean(tf.to_float(tf.greater(tf.abs(ratio_tf - 1.0), self.ratio_clip)))
return
def _build_nets(self, json_data):
assert self.ACTOR_NET_KEY in json_data
assert self.CRITIC_NET_KEY in json_data
actor_net_name = json_data[self.ACTOR_NET_KEY]
critic_net_name = json_data[self.CRITIC_NET_KEY]
actor_init_output_scale = 1 if (
self.ACTOR_INIT_OUTPUT_SCALE_KEY
not in json_data) else json_data[self.ACTOR_INIT_OUTPUT_SCALE_KEY]
s_size = self.get_state_size()
g_size = self.get_goal_size()
a_size = self.get_action_size()
# setup input tensors
self.s_tf = tf.placeholder(tf.float32, shape=[None, s_size], name="s")
self.a_tf = tf.placeholder(tf.float32, shape=[None, a_size], name="a")
self.tar_val_tf = tf.placeholder(tf.float32,
shape=[None],
name="tar_val")
self.adv_tf = tf.placeholder(tf.float32, shape=[None], name="adv")
self.g_tf = tf.placeholder(
tf.float32,
shape=([None, g_size] if self.has_goal() else None),
name="g")
self.old_logp_tf = tf.placeholder(tf.float32,
shape=[None],
name="old_logp")
self.exp_mask_tf = tf.placeholder(tf.float32,
shape=[None],
name="exp_mask")
with tf.variable_scope('main'):
with tf.variable_scope('actor'):
self.a_mean_tf = self._build_net_actor(
actor_net_name, actor_init_output_scale)
with tf.variable_scope('critic'):
self.critic_tf = self._build_net_critic(critic_net_name)
if (self.a_mean_tf != None):
Logger.print('Built actor net: ' + actor_net_name)
if (self.critic_tf != None):
Logger.print('Built critic net: ' + critic_net_name)
self.norm_a_std_tf = self.exp_params_curr.noise * tf.ones(a_size)
norm_a_noise_tf = self.norm_a_std_tf * tf.random_normal(
shape=tf.shape(self.a_mean_tf))
norm_a_noise_tf *= tf.expand_dims(self.exp_mask_tf, axis=-1)
self.sample_a_tf = self.a_mean_tf + norm_a_noise_tf * self.a_norm.std_tf
self.sample_a_logp_tf = TFUtil.calc_logp_gaussian(
x_tf=norm_a_noise_tf, mean_tf=None, std_tf=self.norm_a_std_tf)
return
def _build_losses(self, json_data):
actor_weight_decay = 0 if (self.ACTOR_WEIGHT_DECAY_KEY not in json_data) else json_data[self.ACTOR_WEIGHT_DECAY_KEY]
critic_weight_decay = 0 if (self.CRITIC_WEIGHT_DECAY_KEY not in json_data) else json_data[self.CRITIC_WEIGHT_DECAY_KEY]
gating_regularization_lambda = 0 if (self.GATING_REGULARIZATION_LAMBDA_KEY not in json_data) else json_data[self.GATING_REGULARIZATION_LAMBDA_KEY]
norm_val_diff = self.val_norm.normalize_tf(self.tar_val_tf) - self.val_norm.normalize_tf(self.critic_tf)
self.critic_loss_tf = 0.5 * tf.reduce_mean(tf.square(norm_val_diff))
if (critic_weight_decay != 0):
self.critic_loss_tf += critic_weight_decay * self._weight_decay_loss('main/critic')
norm_tar_a_tf = self.a_norm.normalize_tf(self.a_tf)
self._norm_a_mean_tf = self.a_norm.normalize_tf(self.a_mean_tf)
self.logp_tf = TFUtil.calc_logp_gaussian(norm_tar_a_tf, self._norm_a_mean_tf, self.norm_a_std_tf)
ratio_tf = tf.exp(self.logp_tf - self.old_logp_tf)
actor_loss0 = self.adv_tf * ratio_tf
actor_loss1 = self.adv_tf * tf.clip_by_value(ratio_tf, 1.0 - self.ratio_clip, 1 + self.ratio_clip)
self.actor_loss_tf = -tf.reduce_mean(tf.minimum(actor_loss0, actor_loss1))
norm_a_bound_min = self.a_norm.normalize(self.a_bound_min)
norm_a_bound_max = self.a_norm.normalize(self.a_bound_max)
a_bound_loss = TFUtil.calc_bound_loss(self._norm_a_mean_tf, norm_a_bound_min, norm_a_bound_max)
self.actor_loss_tf += a_bound_loss
self.regularization_loss_tf = None
if gating_regularization_lambda > 0:
vars_generator = []
vars_gating = []
for var in tf.trainable_variables():
if 'bias' in var.name: continue ## Ignore bias
if 'generator' in var.name:
vars_generator.append(var)
elif 'gating' in var.name:
vars_gating.append(var)
self.regularization_loss_tf = 0
for i in range(0, len(vars_gating)):
l1_loss = tf.reduce_mean(tf.keras.losses.MAE(vars_generator[i], vars_gating[i]))
self.regularization_loss_tf += l1_loss
self.actor_loss_tf += self.regularization_loss_tf * gating_regularization_lambda
if (actor_weight_decay != 0):
self.actor_loss_tf += actor_weight_decay * self._weight_decay_loss('main/actor')
# for debugging
self.clip_frac_tf = tf.reduce_mean(tf.to_float(tf.greater(tf.abs(ratio_tf - 1.0), self.ratio_clip)))
return
def _build_nets(self, json_data):
assert self.ACTOR_NET_KEY in json_data
assert self.CRITIC_NET_KEY in json_data
actor_net_name = json_data[self.ACTOR_NET_KEY]
critic_net_name = json_data[self.CRITIC_NET_KEY]
actor_init_output_scale = 1 if (self.ACTOR_INIT_OUTPUT_SCALE_KEY not in json_data) else json_data[self.ACTOR_INIT_OUTPUT_SCALE_KEY]
s_size = self.get_state_size()
g_size = self.get_goal_size()
a_size = self.get_action_size()
# setup input tensors
self.s_tf = tf.placeholder(tf.float32, shape=[None, s_size], name="s")
self.a_tf = tf.placeholder(tf.float32, shape=[None, a_size], name="a")
self.tar_val_tf = tf.placeholder(tf.float32, shape=[None], name="tar_val")
self.adv_tf = tf.placeholder(tf.float32, shape=[None], name="adv")
self.g_tf = tf.placeholder(tf.float32, shape=([None, g_size] if self.has_goal() else None), name="g")
self.old_logp_tf = tf.placeholder(tf.float32, shape=[None], name="old_logp")
self.exp_mask_tf = tf.placeholder(tf.float32, shape=[None], name="exp_mask")
with tf.variable_scope('main'):
with tf.variable_scope('actor'):
self.a_mean_tf = self._build_net_actor(actor_net_name, actor_init_output_scale)
with tf.variable_scope('critic'):
self.critic_tf = self._build_net_critic(critic_net_name)
if (self.a_mean_tf != None):
Logger.print2('Built actor net: ' + actor_net_name)
if (self.critic_tf != None):
Logger.print2('Built critic net: ' + critic_net_name)
self.norm_a_std_tf = self.exp_params_curr.noise * tf.ones(a_size)
norm_a_noise_tf = self.norm_a_std_tf * tf.random_normal(shape=tf.shape(self.a_mean_tf))
norm_a_noise_tf *= tf.expand_dims(self.exp_mask_tf, axis=-1)
self.sample_a_tf = self.a_mean_tf + norm_a_noise_tf * self.a_norm.std_tf
self.sample_a_logp_tf = TFUtil.calc_logp_gaussian(x_tf=norm_a_noise_tf, mean_tf=None, std_tf=self.norm_a_std_tf)
return