def make_env(env_name, env_type, seed, wrapper_kwargs=None):
if env_type == 'atari':
env = wrap_deepmind(make_atari("{}NoFrameskip-v4".format(env_name)), wrapper_kwargs)
else:
env = gym.make(env_name)
env.seed(seed)
return env
def prep_env(env_name, video_path):
if env_name.lower() == "cartpole":
env = gym.make("CartPole-v0")
env.record_start = lambda: None
env.record_end = lambda: None
else:
env = wrap_deepmind(
make_atari(env_name +
"NoFrameskip-v4")) # make sure to add NoFrameskip-v4
env = Monitor(env=env, directory=video_path, force=True)
return env
def invoke_agent_env(params, alg):
"""Returns the wrapped env and string name of agent, then Use `eval(agent)` to activate it from main script
"""
if params.mode == "Atari":
env = wrap_deepmind(make_atari("{}NoFrameskip-v4".format(
params.env_name, skip_frame_k=params.skip_frame_k)),
skip_frame_k=params.skip_frame_k)
if params.debug_flg:
agent = "{}_debug".format(alg)
else:
agent = "{}".format(alg)
else:
agent = "{}".format(alg)
if params.mode == "CartPole":
env = MyWrapper(gym.make("CartPole-v0"))
elif params.mode == "CartPole-p":
env = CartPole_Pixel(gym.make("CartPole-v0"))
return agent, env
import tensorflow as tf
from tf_rl.common.memory_tf import ReplayBuffer
from tf_rl.common.wrappers import wrap_deepmind, make_atari
env = wrap_deepmind(make_atari("PongNoFrameskip-v4"))
state = env.reset()
memory = ReplayBuffer(capacity=100,
n_step=0,
act_shape=(),
obs_shape=state.shape,
obs_dtype=tf.int8,
checkpoint_dir="./tmp")
done = False
for t in range(100):
action = env.action_space.sample()
next_state, reward, done, info = env.step(action)
memory.add(state, action, reward, next_state, done)
state = next_state
env.close()
print(len(memory))
obs, action, next_obs, reward, done = memory.sample(batch_size=10)
print(obs.shape, action.shape, next_obs.shape, reward.shape, done.shape)
path = memory.save()
# recover phase
print("=== Recover Phase ===")
del memory
memory = ReplayBuffer(capacity=100,
n_step=0,
act_shape=(),
import itertools
from tf_rl.common.monitor import Monitor
from tf_rl.common.wrappers import wrap_deepmind, make_atari
env_name = "PongNoFrameskip-v4"
env = wrap_deepmind(make_atari(env_name))
env = Monitor(env=env, directory="./video/{}".format(env_name), force=True)
env.record_start()
state = env.reset()
for t in itertools.count():
action = env.action_space.sample()
next_state, reward, done, info = env.step(action)
state = next_state
if done:
break
print("End at {}".format(t + 1))
env.record_end()
env.close()
policy = EpsilonGreedyPolicy_eager(Epsilon_fn=anneal_ep)
optimizer = tf.compat.v1.train.RMSPropOptimizer(learning_rate=0.00025,
decay=0.95,
momentum=0.0,
epsilon=0.00001,
centered=True)
replay_buffer = ReplayBuffer(params.memory_size)
reward_buffer = deque(maxlen=params.reward_buffer_ep)
loss_fn = create_loss_func(params.loss_fn)
grad_clip_fn = gradient_clip_fn(flag=params.grad_clip_flg)
# create a directory for log/model
params = create_log_model_directory(params, get_alg_name())
summary_writer = tf.contrib.summary.create_file_writer(params.log_dir)
# choose env and instantiate the agent correspondingly
env = wrap_deepmind(make_atari("{}NoFrameskip-v4".format(
params.env_name, skip_frame_k=params.skip_frame_k)),
skip_frame_k=params.skip_frame_k)
if params.debug_flg:
agent = DQN_debug(Model, optimizer, loss_fn, grad_clip_fn,
env.action_space.n, params)
else:
agent = DQN(Model, optimizer, loss_fn, grad_clip_fn, env.action_space.n,
params)
# set seed
env.seed(params.seed)
tf.compat.v1.random.set_random_seed(params.seed)
train_DQN(agent, env, policy, replay_buffer, reward_buffer, summary_writer)
