AttributeError: 'Tensor' object has no attribute 'append'

Question

I can't figure out why this code isn't working. When I make rewards into a list, I get an error telling me that the dimensions are incorrect. I'm not sure what to do.

I am implementing a reinforcement deep q network. r is a numpy 2d array giving 1 divided by the distance between stops. This is so that closer stops have a higher reward.

No matter what I do, I can't get rewards to run smoothly. I am new to Tensorflow, so it may just be a result of my inexperience with things like Tensorflow placeholders and feed dicts.

Thanks in advance for your help.

observations = tf.placeholder('float32', shape=[None, num_stops])

game states : r[stop], r[next_stop], r[third_stop]

actions = tf.placeholder('int32',shape=[None]) 

rewards = tf.placeholder('float32',shape=[None])  # +1, -1 with discounts

Y = tf.layers.dense(observations, 200, activation=tf.nn.relu)
Ylogits = tf.layers.dense(Y, num_stops)

sample_op = tf.random.categorical(logits=Ylogits, num_samples=1)

cross_entropies = tf.losses.softmax_cross_entropy(onehot_labels=tf.one_hot  (actions,num_stops), logits=Ylogits)

loss = tf.reduce_sum(rewards * cross_entropies)


optimizer = tf.train.RMSPropOptimizer(learning_rate=0.001, decay=.99)
train_op = optimizer.minimize(loss)




visited_stops = []
steps = 0

with tf.Session() as sess:

    sess.run(tf.global_variables_initializer())

    # Start at a random stop, initialize done to false
    current_stop = random.randint(0, len(r) - 1)
    done = False

    # reset everything    
    while not done: # play a game in x steps   

        observations_list = []
        actions_list = []
        rewards_list = []

        # List all stops and their scores
        observation = r[current_stop]

        # Add the stop to a list of non-visited stops if it isn't
        # already there
        if current_stop not in visited_stops:
            visited_stops.append(current_stop)

        # decide where to go
        action = sess.run(sample_op, feed_dict={observations: [observation]})

        # play it, output next state, reward if we got a point, and whether the game is over
        #game_state, reward, done, info = pong_sim.step(action)
        new_stop = int(action)


        reward = r[current_stop][action]

        if len(visited_stops) == num_stops:
            done = True

        if steps >= BATCH_SIZE:
            done = True

        steps += 1

        observations_list.append(observation)
        actions_list.append(action)
        rewards.append(reward)



        #rewards_list = np.reshape(rewards, [-1, 25])
        current_stop = new_stop

    #processed_rewards = discount_rewards(rewards, args.gamma)
    #processed_rewards = normalize_rewards(rewards, args.gamma)

    print(rewards)
    sess.run(train_op, feed_dict={observations: [observations_list],
                             actions: [actions_list],
                             rewards: [rewards_list]})

Answer 1

the row rewards.append(reward) causes the error, an it is because your rewards variable is a Tensor, as you defined it in rewards = tf.placeholder('float32',shape=[None]) and you can not append values to tensor like that. You probably wanted to call rewards_list.append(reward).

Also, you are initializing variables

observations_list = []
actions_list = []
rewards_list = []

inside the loop, so in each iteration, ols values will be overwritten by empty list. You probably want to have those 3 lines before the while not done: line.