def __compute_reward(self, action, preact_interp_prd_logits, episode_step_count, action_shape):
interp_action = layers.Interp(x=action,
shape=self.inputs_shape_spatial,
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
interp_action_prd = tf.stop_gradient(interp_action * tf.expand_dims(
(tf.argmax(preact_interp_prd_logits, axis=-1) + 1),
axis=-1
)
)
certain_region_mask = x=tf.to_float(tf.not_equal(action, 0))
uncertain_region_mask = x=tf.to_float(tf.equal(action, 0))
uncertain_region_sum = uncertain_region_mask
pos_mask = x=tf.to_float(tf.equal(interp_action_prd, self.ground_truth + 1))
neg_mask = 1.0 - pos_mask
pos_mask = layers.Interp(x=pos_mask, shape=action_shape)
neg_mask = layers.Interp(x=neg_mask, shape=action_shape)
unignored_mask = layers.Interp(x=tf.to_float(tf.not_equal(self.ground_truth, 255)),
shape=action_shape,
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
pos_certreg_sum = pos_mask
neg_certreg_sum = unignored_mask * certain_region_mask * neg_mask
reward_dict_idx = episode_step_count if episode_step_count < (len(self.reward_dict)-1) else len(self.reward_dict)-1
#reward = 1.0 * pos_certreg_sum - 1.0 * neg_certreg_sum - 0.1 * uncertain_region_sum
reward = pos_certreg_sum - neg_certreg_sum + self.reward_dict[reward_dict_idx] * uncertain_region_sum
return tf.stop_gradient(tf.squeeze(reward, axis=-1))
def __inference_once(self):
"""
Predict segementation for an image.
Arguments:
img: must be rowsxcolsx3
"""
state_value, _policys = layers.pspnet_top_plc(inputs=self.features_priors,
dynamic_envs=self.dynamic_envs,
inputs_shape_spatial=self.inputs_shape_spatial,
is_training=self.is_training)
_preact_prd_logits = layers.pspnet_top_cls(inputs=self.features_priors,
dynamic_envs=self.dynamic_envs,
nb_classes=self.nb_classes,
activation='softmax',
is_training=self.is_training)
_preact_policy, policy, preact_interp_policy, interp_policy = _policys
preact_interp_prd_logits = layers.Interp(x=_preact_prd_logits,
shape=self.inputs_shape_spatial)
action = self.__gen_action(policy)
NS_action = tf.expand_dims(tf.argmax(policy, axis=-1), axis=-1)
NS_action_interp_smooth = tf.expand_dims(
tf.argmax(
layers.Interp(x=policy,
shape=self.inputs_shape_spatial,
method=tf.image.ResizeMethod.BILINEAR),
axis=-1),
axis=-1)
_action_prd = tf.stop_gradient(action * tf.expand_dims(
(tf.argmax(_preact_prd_logits, axis=-1) + 1),
axis=-1
)
)
NS_action_prd = tf.stop_gradient(NS_action * tf.expand_dims(
(tf.argmax(_preact_prd_logits, axis=-1) + 1),
axis=-1
)
)
self.dynamic_envs = self.__new_envs(_action_prd)
return [state_value,
policy,
preact_interp_policy,
interp_policy,
action,
NS_action,
preact_interp_prd_logits,
NS_action_interp_smooth]
def __new_envs_soft(self, prd_softmax):
return tf.stop_gradient(
tf.to_float(
layers.Interp(x=prd_softmax,
shape=self.dynamic_envs_shape_spatial,
method=tf.image.ResizeMethod.BILINEAR)))
def __new_knowledges(self, action_prd):
return tf.stop_gradient(
tf.to_float(
tf.one_hot(indices=tf.squeeze(layers.Interp(
x=action_prd,
shape=self.knowledges_shape_spatial,
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR),
axis=-1),
depth=self.total_classes)))
def __new_envs(self, prd):
return tf.stop_gradient(
tf.to_float(
tf.one_hot(indices=tf.squeeze(layers.Interp(
x=prd,
shape=self.dynamic_envs_shape_spatial,
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR),
axis=-1),
depth=self.nb_classes)))
def __inference_once(self, step):
"""
Predict segementation for an image.
Arguments:
img: must be rowsxcolsx3
"""
'''
state_value, _policys = self._get_stval_plc()
_preact_prd_logits = self._get_preact_prd_logits()
_preact_policy, policy, preact_interp_policy, interp_policy = _policys
preact_interp_prd_logits = layers.Interp(x=_preact_prd_logits,
shape=self.original_inputs_shape_spatial)
action = self.__gen_action(policy) #input softmax policy to explore more.
NS_action = tf.expand_dims(tf.argmax(policy, axis=-1), axis=-1)
NS_action_interp_smooth = tf.expand_dims(
tf.argmax(
layers.Interp(x=policy,
shape=self.original_inputs_shape_spatial,
method=tf.image.ResizeMethod.BILINEAR),
axis=-1),
axis=-1)
_action_prd = tf.stop_gradient(action * tf.expand_dims(
(tf.argmax(_preact_prd_logits, axis=-1) + 1),
axis=-1
)
)
NS_action_prd = tf.stop_gradient(NS_action * tf.expand_dims(
(tf.argmax(_preact_prd_logits, axis=-1) + 1),
axis=-1
)
)
self.dynamic_envs = self.__new_envs(_action_prd)
'''
_preact_prd_logits = self._get_preact_prd_logits(
is_knowledge_empty=True if step == 0 else False)
preact_interp_prd_logits = layers.Interp(
x=_preact_prd_logits, shape=self.original_inputs_shape_spatial)
_preact_prd_logits_argmax = tf.expand_dims(tf.argmax(
_preact_prd_logits, axis=-1),
axis=-1)
#_prd = tf.stop_gradient(_preact_prd_logits_argmax)
#self.dynamic_envs = self.__new_envs(prd=_prd)
self.dynamic_envs = self.__new_envs_soft(
prd_softmax=tf.nn.softmax(_preact_prd_logits, axis=-1))
state_value, _policys = self._get_stval_plc()
_preact_policy, policy, preact_interp_policy, interp_policy = _policys
action = self.__gen_action(
policy) # input softmax policy to explore more.
NS_action = tf.expand_dims(tf.argmax(policy, axis=-1), axis=-1)
NS_action_interp_smooth = tf.expand_dims(tf.argmax(layers.Interp(
x=policy,
shape=self.original_inputs_shape_spatial,
method=tf.image.ResizeMethod.BILINEAR),
axis=-1),
axis=-1)
_action_prd = tf.stop_gradient(action *
(_preact_prd_logits_argmax + 1))
self.knowledges = self.__new_knowledges(_action_prd)
element_wise_comp = tf.equal(
_preact_prd_logits_argmax,
layers.Interp(
x=self.ground_truth,
shape=_preact_prd_logits_argmax.shape.as_list()[1:-1],
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR))
return [
state_value, policy, preact_interp_policy, interp_policy, action,
NS_action, preact_interp_prd_logits, NS_action_interp_smooth,
tf.stop_gradient(element_wise_comp)
]
