realstate = tf.placeholder(tf.float32, shape=([num_processing_steps,batch_size_tr*grounding_size,1])) action_filled = tf.placeholder(tf.float32, shape=([num_processing_steps,batch_size_tr,dimension])) action_filled_test = tf.placeholder(tf.float32, shape=([num_processing_steps,1,dimension])) # get graphs first_action = np.random.uniform(-6 / np.sqrt(dimension), 6 / np.sqrt(dimension), dimension) first_action = np.array(first_action, dtype='float32') nodes = [1] * ((grounding_size + 1) * dimension) nodes = np.reshape(nodes, (grounding_size + 1,dimension)) nodes = np.array(nodes, dtype='float32') graph_place = [] for temp_index in range(batch_size_tr): graph_place.append(base_graph(np.ones(grounding_size), grounding_size, first_action, dimension, nodes)) graphs_tuple_ph = utils_tf.placeholders_from_data_dicts(graph_place) graphs_tuple_test = utils_tf.placeholders_from_data_dicts([base_graph(np.ones(grounding_size), grounding_size, first_action, dimension, nodes)]) model = new_model.GraphProcess(edge_output_size=1, dimension=dimension) output_ops_tr, sta_vecs = model(graphs_tuple_ph, action_filled, num_processing_steps) output_ops_test, sta_vecs_test = model(graphs_tuple_test, action_filled_test, num_processing_steps) # Training Loss
action_filled_index = tf.placeholder( tf.int32, shape=([num_processing_steps, batch_size_tr])) action_filled_plan_onestep = tf.placeholder(tf.float32, shape=([1, dimension])) # get graphs first_action = np.random.uniform(-6 / np.sqrt(dimension), 6 / np.sqrt(dimension), dimension) first_action = np.array(first_action, dtype='float32') nodes = [1] * ((grounding_size + 1) * dimension) nodes = np.reshape(nodes, (grounding_size + 1, dimension)) nodes = np.array(nodes, dtype='float32') graph_place = [] for temp_index in range(batch_size_tr): graph_place.append( base_graph(np.ones(grounding_size), grounding_size, first_action, dimension, nodes)) graphs_tuple_ph = utils_tf.placeholders_from_data_dicts(graph_place) graphs_tuple_goal = utils_tf.placeholders_from_data_dicts(graph_place) model = new_model.GraphProcess(edge_output_size=1, dimension=dimension) output_ops_tr, sta_vecs = model(graphs_tuple_ph, action_filled, num_processing_steps) action_size = len(action_dict) + 1 correct_action_output = tf.one_hot(indices=action_filled_index, depth=action_size) goal_graph = model.getInitialVecForState(graphs_tuple_goal) training_input, training_output = datas.getTrainData(sta_vecs, goal_graph, correct_action_output,