scopes = [network_scope, scene_scope, task_scope]
for i_episode in range(NUM_EVAL_EPISODES):
env.reset()
terminal = False
ep_reward = 0
ep_collision = 0
ep_t = 0
while not terminal:
usf_s_g = global_network.run_usf(sess, env.s_t, env.target,
scopes)
pi_values = global_network.run_policy(
sess, env.s_t, env.target, usf_s_g, scopes)
action = sample_action(pi_values)
env.step(action)
env.update()
terminal = env.terminal
if ep_t == 500: break
if env.collided: ep_collision += 1
ep_reward += env.reward
ep_t += 1
ep_lengths.append(ep_t)
ep_rewards.append(ep_reward)
ep_collisions.append(ep_collision)
if VERBOSE:
print("episode #{} ends after {} steps".format(
for i_episode in range(NNN):
optimal = env.reset()
terminal = False
ep_reward = 0
ep_collision = 0
ep_t = 0
f_path.write(str(counter * NNN + i_episode) + ': [')
while not terminal:
#viewer.imshow(env.observation)
#time.sleep(0.3)
#print(env.isCheckpoint)
pi_values = global_network.run_policy(
sess, env.s_t, env.s_position, env.checkpoint,
env.s_a_t, env.s_c_t, env.isCheckpoint, env.s_aux_cl,
scopes)
#print(pi_values)
action = sample_action(pi_values)
env.step(action)
env.update()
terminal = env.terminal
if ep_t == 500: break
if env.collided: ep_collision += 1
ep_reward += env.reward
ep_t += 1
#if not terminal:
# f_path.write('['+str(int(env.x*2))+', '+str(int(env.z*2))+'], ')
#else:
h = target[3]*300
xa = int((x+w)/2)
xm = int(x-xa)
ya = int((y+h)/2)
ym = int(y-ya)
#print(env.observation)
env.observation[ym:ym+2,xm:xa] = 255
env.observation[ya:ya+2,xm:xa] = 255
env.observation[ym:ya,xm:xm+2] = 255
env.observation[ym:ya,xa:xa+2] = 255
'''
#viewer.imshow(env.observation)
#time.sleep(0.5)
pi_values = global_network.run_policy(
sess, env.s_t, env.s_position, env.checkpoint, scopes)
action = sample_action(pi_values)
env.step(action)
env.update()
terminal = env.terminal
if ep_t == 500: break
if env.collided: ep_collision += 1
ep_reward += env.reward
ep_t += 1
if not terminal:
f.write('[' + str(int(env.x * 2)) + ', ' +
str(int(env.z * 2)) + '], ')
else:
f.write('[' + str(int(env.x * 2)) + ', ' +
def main():
# disable all v2 behavior
tf.disable_v2_behavior()
tf.disable_eager_execution()
device = "/cpu:0" # use CPU for display tool
network_scope = TASK_TYPE # Always 'navigation'
list_of_tasks = TASK_LIST
scene_scopes = list_of_tasks.keys()
global_network = ActorCriticFFNetwork(action_size=ACTION_SIZE,
device=device,
network_scope=network_scope,
scene_scopes=scene_scopes)
sess = tf.Session()
# sess = tf.coSession()
init = tf.global_variables_initializer()
sess.run(init)
saver = tf.train.Saver()
checkpoint = tf.train.get_checkpoint_state(CHECKPOINT_DIR)
# see if we saved a checkpoint from past training?
if checkpoint and checkpoint.model_checkpoint_path:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("checkpoint loaded: {}".format(checkpoint.model_checkpoint_path))
else:
print("Could not find old checkpoint")
scene_stats = dict()
for scene_scope in scene_scopes:
# TODO: remove
scene_scope = "FloorPlan402"
scene_stats[scene_scope] = []
for task_scope in list_of_tasks[scene_scope]:
# tasks are positions!!!
# env = ai2thor.controller.Controller(scene="FloorPlan227", gridSize=0.25, width=1000, height=1000)
with open(GOAL_FILE, 'r') as f:
GOAL_DATA = json.load(f)
GOAL_POS = GOAL_DATA["agent_position"]
env = RLController({
'scene': scene_scope,
'terminal_state_id': int(task_scope),
'goal_pos': GOAL_POS,
'goal_image_fpath': "data/FP402_goal_towel.png"
})
env.docker_enabled = True
ep_rewards = []
ep_lengths = []
ep_collisions = []
scopes = [network_scope, scene_scope]
for i_episode in range(NUM_EVAL_EPISODES):
env.reset()
terminal = False
ep_reward = 0
ep_collision = 0
ep_t = 0
while not terminal:
# mirrors actions taken in paper
# NOTE: rearranged these to mirror code in scene_loader
list_of_actions = [
"MoveAhead", "RotateRight", "RotateLeft", "MoveBack"
]
pi_values = global_network.run_policy(
sess, env.curr_state, env.target, scopes)
# action returned is an integer -- critical that the list_of_actions is in correct order
action = sample_action(pi_values)
print(
"Ep_t: {} \n\tCollided?: {} \n\tAction: {} \n\tValue: {} \n\tAll Action Values: {}"
.format(ep_t, env.collided, list_of_actions[action],
pi_values[action], pi_values))
env.step(list_of_actions[action])
env.update()
terminal = env.terminal
if ep_t == 10000: break
if env.collided: ep_collision += 1
ep_reward += env.reward
ep_t += 1
ep_lengths.append(ep_t)
ep_rewards.append(ep_reward)
ep_collisions.append(ep_collision)
print('evaluation: %s %s' % (scene_scope, task_scope))
print('mean episode reward: %.2f' % np.mean(ep_rewards))
print('mean episode length: %.2f' % np.mean(ep_lengths))
print('mean episode collision: %.2f' % np.mean(ep_collisions))