n_ep += 1
cum_reward = 0.0
n_steps = 0
cum_td_loss = 0.0
img, rule_action = env.reset()
# print "state shape: {}".format(state.shape)
# print "state type: {}".format(type(state))
# resize maybe different from tf.resize
# tensor_state = tf.convert_to_tensor(state)
# img = np.array([img])
# tens_img = tf.image.resize_images(img, [224, 224])
# img = tf.image.convert_image_dtype(img, tf.float32)
# tens_img = initialD_input.preprocess_image(tens_img)
# np_img = sess.run(tens_img)
img = cv2.resize(img, (224, 224))
np_img = initialD_input.preprocess_image(img)
print "=========img shape: {}".format(img.shape) + "=========\n"
using_learning_agent = True
# actions, np_probs = sess.run([network_train.preds, probs], feed_dict={
# network_train._images:np.array([np_img]),
# network_train.is_train:False})
# action = actions[0]
np_probs = sess.run(pi, feed_dict={images: np.array([np_img])})
action = np.argmax(np_probs)
all_scenes.append([np.copy(img), action, np_probs])
next_state, reward, done, info = env.step(ACTIONS[action])
next_img, next_rule_action = next_state
while True:
n_steps += 1
try:
# config = tf.ConfigProto()
with sv.managed_session(config=config) as sess:
agent.set_session(sess)
n_steps = 0
while True:
n_ep += 1
cum_reward = 0.0
cum_td_loss = 0.0
cum_spg_loss = 0.0
all_scenes = []
state = env.reset()
print "========reset======\n" * 5
img = cv2.resize(state, (224, 224))
pr_img = initialD_input.preprocess_image(img)
while True:
action = agent.act(state=pr_img, evaluate=False, sess=sess)
print "action: ", action
all_scenes.append([np.copy(img), action])
next_state, reward, done, info = env.step(ACTIONS[action])
next_img = cv2.resize(next_state, (224, 224))
pr_next_img = initialD_input.preprocess_image(next_img)
info = agent.step(state=pr_img, action=action, reward=reward, next_state=pr_next_img, episode_done=done)
record(summary_writer, n_steps, info)
n_steps += 1
if done is True:
print "========Run Done=======\n" * 5
break
img = next_img
pr_img = pr_next_img
# print "========\n"*5
# lr = graph.get_operation_by_name('lr').outputs[0]
while True:
n_ep += 1
cum_reward = 0.0
n_steps = 0
cum_td_loss = 0.0
img, rule_action = env.reset()
# print "state shape: {}".format(state.shape)
# print "state type: {}".format(type(state))
# resize maybe different from tf.resize
# tensor_state = tf.convert_to_tensor(state)
# img = np.array([img])
tens_img = tf.image.resize_images(img, [224, 224])
# img = tf.image.convert_image_dtype(img, tf.float32)
tens_img = initialD_input.preprocess_image(tens_img)
np_img = sess.run(tens_img)
print "=========img shape: {}".format(img.shape)+"=========\n"
using_learning_agent = True
actions, np_probs = sess.run([network_train.preds, probs], feed_dict={
network_train._images:np.array([np_img]),
network_train.is_train:False})
action = actions[0]
all_scenes.append([np.copy(img),action,rule_action,np_probs])
if action != rule_action:
print "not equal, sl: ", action, " rule: ", rule_action