# 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()
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
episode_reward += reward
all_rewards.append(episode_reward)
tf.contrib.summary.scalar("Evaluation Score",
episode_reward,
step=agent.index_timestep)
print("| Ep: {}/{} | Score: {} |".format(ep + 1, params.n_trial,
episode_reward))
env.record_end()
# 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
"""
if global_timestep.numpy() > agent.params.learning_start:
states, actions, rewards, next_states, dones = replay_buffer.sample(
agent.params.batch_size)
loss = agent.update(states, actions, rewards, next_states,
dones)
soft_target_model_update_eager(
agent.target_actor,
agent.actor,
tau=agent.params.soft_update_tau)
soft_target_model_update_eager(
agent.target_critic,
import gym
from tf_rl.common.monitor import Monitor
ENVS = [
"Ant-v2",
"HalfCheetah-v2",
"Hopper-v2",
"Humanoid-v2",
# "Reacher-v2",
# "Swimmer-v2",
"Walker2d-v2"
]
DEFAULT = 250
for env_name in ENVS:
env = gym.make(env_name)
env = Monitor(env, "./video/{}".format(env_name), force=True)
print(env_name)
env.record_start()
env.reset()
done = False
while not done:
# env.render(mode="human", annotation_flg=False)
s, r, done, i = env.step(env.action_space.sample())
env.record_end()
env.close()
import gym
from tf_rl.common.monitor import Monitor
env = gym.make('CartPole-v0')
env = Monitor(env=env, directory="./video/cartpole", force=True)
for ep in range(20):
if ep == 0: env.record_start()
state = env.reset()
for t in range(1000):
action = env.action_space.sample()
state, reward, done, info = env.step(action)
if done:
if ep == 0: env.record_end()
break
env.close()