observation_last_labels.append(observation_last_label_t)
observation_images.append(observation_image_t)
action_t = tf.one_hot(tf.random_uniform([sim.batch_size_t], maxval=num_actions, dtype=tf.int32), num_actions)
actions.append(action_t)
reward_t, (observation_image_t, observation_last_label_t) = sim.do_step(action_t)
rewards.append(reward_t)
observation_images_t = tf.pack(observation_images, 1)
actions_t = tf.pack(actions, 1)
observation_last_labels_t = tf.pack(observation_last_labels, 1)
rewards_t = tf.pack(rewards, 1)
sess = tf.Session()
#sess.run(init_op)
images, labels, last_labels = data.get_batch_of_episodes(BATCH_SIZE, TIME_STEPS, CLASSES_PER_EPISODE)
observation_images, actions, observation_last_labels, rewards = sess.run((observation_images_t, actions_t, observation_last_labels_t, rewards_t),
{ images_t: images,
#labels_t: labels,
last_labels_t: last_labels })
b = random.choice(range(BATCH_SIZE))
for t in range(actions.shape[1]):
image = observation_images[b,t]
im = Image.fromarray(np.uint8(image*255))
print t
print 'response from last question: ', observation_last_labels[b,t]
print 'current image'
print 'action for current image: ', actions[b,t]
print 'reward for current image + current action: ', rewards[b,t]
im.show()
raw_input()
# accuracy_1st_summary_t: accuracy_1st,
# accuracy_2nd_summary_t: accuracy_2nd,
# accuracy_5th_summary_t: accuracy_5th,
# accuracy_10th_summary_t: accuracy_10th,
# question_ave_summary_t: question_ave,
# question_1st_summary_t: question_1st,
# question_2nd_summary_t: question_2nd,
# question_5th_summary_t: question_5th,
# question_10th_summary_t: question_10th
# }), e+1)
# train_writer.flush()
# process a test batch
images, labels, last_labels = data.get_batch_of_episodes(
BATCH_SIZE,
TIME_STEPS,
CLASSES_PER_EPISODE,
NUM_LABELS,
use_test_data=True)
true_labels = np.argmax(labels[:, :-1], axis=2)
predictions, rewards, loss = sess.run(
[predictions_t, rewards_t, loss_t], {
images_t: images,
labels_t: labels,
last_labels_t: last_labels,
epsilon_t: 0.0
})
reward_max = np.mean(rewards)
pred_labels = np.argmax(predictions, axis=2)
accuracy_max = np.mean(true_labels == pred_labels)
num_correct_predictions = np.sum(