def create_input_placeholders(self): with self.graph.as_default(): ( self.global_step, self.increment_step_op, self.steps_to_increment, ) = ModelUtils.create_global_steps() self.visual_in = ModelUtils.create_visual_input_placeholders( self.brain.camera_resolutions ) self.vector_in = ModelUtils.create_vector_input(self.vec_obs_size) if self.normalize: normalization_tensors = ModelUtils.create_normalizer(self.vector_in) self.update_normalization_op = normalization_tensors.update_op self.normalization_steps = normalization_tensors.steps self.running_mean = normalization_tensors.running_mean self.running_variance = normalization_tensors.running_variance self.processed_vector_in = ModelUtils.normalize_vector_obs( self.vector_in, self.running_mean, self.running_variance, self.normalization_steps, ) else: self.processed_vector_in = self.vector_in self.update_normalization_op = None self.batch_size_ph = tf.placeholder( shape=None, dtype=tf.int32, name="batch_size" ) self.sequence_length_ph = tf.placeholder( shape=None, dtype=tf.int32, name="sequence_length" ) self.mask_input = tf.placeholder( shape=[None], dtype=tf.float32, name="masks" ) # Only needed for PPO, but needed for BC module self.epsilon = tf.placeholder( shape=[None, self.act_size[0]], dtype=tf.float32, name="epsilon" ) self.mask = tf.cast(self.mask_input, tf.int32) tf.Variable( int(self.brain.vector_action_space_type == "continuous"), name="is_continuous_control", trainable=False, dtype=tf.int32, ) tf.Variable( self._version_number_, name="version_number", trainable=False, dtype=tf.int32, ) tf.Variable( self.m_size, name="memory_size", trainable=False, dtype=tf.int32 ) if self.brain.vector_action_space_type == "continuous": tf.Variable( self.act_size[0], name="action_output_shape", trainable=False, dtype=tf.int32, ) else: tf.Variable( sum(self.act_size), name="action_output_shape", trainable=False, dtype=tf.int32, )
def create_input_placeholders(self): with self.graph.as_default(): ( self.global_step, self.increment_step_op, self.steps_to_increment, ) = ModelUtils.create_global_steps() self.vector_in, self.visual_in = ModelUtils.create_input_placeholders( self.behavior_spec.observation_shapes) if self.normalize: normalization_tensors = ModelUtils.create_normalizer( self.vector_in) self.update_normalization_op = normalization_tensors.update_op self.normalization_steps = normalization_tensors.steps self.running_mean = normalization_tensors.running_mean self.running_variance = normalization_tensors.running_variance self.processed_vector_in = ModelUtils.normalize_vector_obs( self.vector_in, self.running_mean, self.running_variance, self.normalization_steps, ) else: self.processed_vector_in = self.vector_in self.update_normalization_op = None self.batch_size_ph = tf.placeholder(shape=None, dtype=tf.int32, name="batch_size") self.sequence_length_ph = tf.placeholder(shape=None, dtype=tf.int32, name="sequence_length") self.mask_input = tf.placeholder(shape=[None], dtype=tf.float32, name="masks") # Only needed for PPO, but needed for BC module self.epsilon = tf.placeholder(shape=[None, self.act_size[0]], dtype=tf.float32, name="epsilon") self.mask = tf.cast(self.mask_input, tf.int32) tf.Variable( int(self.behavior_spec.is_action_continuous()), name="is_continuous_control", trainable=False, dtype=tf.int32, ) int_version = TFPolicy._convert_version_string(__version__) major_ver_t = tf.Variable( int_version[0], name="trainer_major_version", trainable=False, dtype=tf.int32, ) minor_ver_t = tf.Variable( int_version[1], name="trainer_minor_version", trainable=False, dtype=tf.int32, ) patch_ver_t = tf.Variable( int_version[2], name="trainer_patch_version", trainable=False, dtype=tf.int32, ) self.version_tensors = (major_ver_t, minor_ver_t, patch_ver_t) tf.Variable( MODEL_FORMAT_VERSION, name="version_number", trainable=False, dtype=tf.int32, ) tf.Variable(self.m_size, name="memory_size", trainable=False, dtype=tf.int32) if self.behavior_spec.is_action_continuous(): tf.Variable( self.act_size[0], name="action_output_shape", trainable=False, dtype=tf.int32, ) else: tf.Variable( sum(self.act_size), name="action_output_shape", trainable=False, dtype=tf.int32, )