env = gym.make(params.env_name)
env = Monitor(env, params.video_dir, force=True)
random_process = GaussianNoise(mu=0.0, sigma=0.0)
agent = DDPG(Actor, Critic, env.action_space.shape[0], random_process, params)
global_timestep = tf.compat.v1.train.get_or_create_global_step()
all_distances, all_rewards, all_actions = list(), list(), list()
distance_func = get_distance(
agent.params.env_name) # create the distance measure func
print("=== Evaluation Mode ===")
for ep in range(params.n_trial):
env.record_start()
obs = env.reset()
state = obs["flat_obs"]
done = False
episode_reward = 0
while not done:
action = agent.eval_predict(state)
# action = env.action_space.sample()
# scale for execution in env (in DDPG, every action is clipped between [-1, 1] in agent.predict)
obs, reward, done, info = env.step(action * env.action_space.high)
# print(action, reward)
next_flat_state, next_graph_state = obs["flat_obs"], obs["graph_obs"]
distance = distance_func(action, reward, info)
all_actions.append(action.mean()**2) # Mean Squared of action values
all_distances.append(distance)
state = next_flat_state
# run this from the terminal and make sure you are loading appropriate environment variables
# $ echo $LD_LIBRARY_PATH
import gym
from tf_rl.common.monitor import Monitor
import environments.register as register
video_dir = "./video/"
temp = 5
env = gym.make("CentipedeSix-v1")
env = Monitor(env, video_dir, force=True)
for ep in range(10):
if ep % temp == 0:
print("recording")
env.record_start()
env.reset()
done = False
while not done:
# env.render()
action = env.action_space.sample()
s, r, done, info = env.step(action) # take a random action
if ep % temp == 0:
env.record_end()
# Invoke the agent
agent = DDPG(GGNN, Critic, node_info, env.action_space.shape[0], params)
""" === Training Phase === """
get_ready(agent.params)
global_timestep = tf.compat.v1.train.get_or_create_global_step()
time_buffer = deque(maxlen=agent.params.reward_buffer_ep)
log = logger(agent.params)
action_buffer, distance_buffer, eval_epochs = list(), list(), list()
with summary_writer.as_default():
# for summary purpose, we put all codes in this context
with tf.contrib.summary.always_record_summaries():
for i in itertools.count():
state = env.reset()
total_reward = 0
start = time.time()
done = False
episode_len = 0
while not done:
if global_timestep.numpy() < agent.params.learning_start:
action = env.action_space.sample()
else:
action = agent.predict(state)
# scale for execution in env (in DDPG, every action is clipped between [-1, 1] in agent.predict)
next_state, reward, done, info = env.step(
action * env.action_space.high)
replay_buffer.add(state, action, reward, next_state, done)
"""
=== Update the models