def test_fit_batch_online_atari_with_dqn():
import d4rl_atari
make_env = lambda: ChannelFirst(DummyAtari())
env = AsyncBatchEnv([make_env for _ in range(2)])
eval_env = ChannelFirst(DummyAtari())
algo = DQN(n_frames=4)
buffer = BatchReplayBuffer(1000, env)
explorer = LinearDecayEpsilonGreedy()
algo.fit_batch_online(
env,
buffer,
explorer,
n_epochs=1,
n_steps_per_epoch=500,
n_updates_per_epoch=1,
eval_env=eval_env,
logdir="test_data",
)
assert algo.impl.observation_shape == (4, 84, 84)
def train(params):
# setup algorithm
dqn = DQN(batch_size=params.get("batch_size"),
learning_rate=params.get("learning_rate"),
target_update_interval=params.get("target_update_interval"),
q_func_factory=QRQFunctionFactory(
n_quantiles=params.get("n_quantiles")),
n_steps=params.get("train_freq"),
gamma=params.get("gamma"),
n_critics=1,
target_reduction_type="min",
use_gpu=True)
# setup replay buffer
buffer = ReplayBuffer(maxlen=params.get("buffer_size"), env=env)
# setup explorers
explorer = LinearDecayEpsilonGreedy(
start_epsilon=1.0,
end_epsilon=params.get("exploration_final_eps"),
duration=100000)
# start training
dqn.fit_online(
env,
buffer,
n_steps=params.get("train_steps"),
explorer=
explorer, # you don't need this with probablistic policy algorithms
tensorboard_dir=log_dir,
eval_env=eval_env)
dqn.save_model(exp_name)
def test_fit_batch_online_atari_with_dqn():
import d4rl_atari
make_env = lambda: gym.make("breakout-mixed-v0", stack=False)
env = AsyncBatchEnv([make_env for _ in range(2)])
eval_env = gym.make("breakout-mixed-v0", stack=False)
algo = DQN(n_frames=4)
buffer = BatchReplayBuffer(1000, env)
explorer = LinearDecayEpsilonGreedy()
algo.fit_batch_online(
env,
buffer,
explorer,
n_epochs=1,
n_steps_per_epoch=500,
n_updates_per_epoch=1,
eval_env=eval_env,
logdir="test_data",
tensorboard=False,
)
assert algo.impl.observation_shape == (4, 84, 84)
def test_collect_atari_with_dqn():
import d4rl_atari
env = ChannelFirst(DummyAtari())
algo = DQN(n_frames=4)
explorer = LinearDecayEpsilonGreedy()
buffer = algo.collect(env, explorer=explorer, n_steps=100)
assert algo.impl.observation_shape == (4, 84, 84)
assert buffer.size() > 90 and buffer.size() < 100
def test_train_with_dqn():
env = gym.make('CartPole-v0')
eval_env = gym.make('CartPole-v0')
algo = DQN(n_epochs=1)
buffer = ReplayBuffer(1000, env)
explorer = LinearDecayEpsilonGreedy()
train(env,
algo,
buffer,
explorer,
eval_env=eval_env,
logdir='test_data',
tensorboard=False)
def test_fit_online_cartpole_with_dqn():
env = gym.make('CartPole-v0')
eval_env = gym.make('CartPole-v0')
algo = DQN()
buffer = ReplayBuffer(1000, env)
explorer = LinearDecayEpsilonGreedy()
algo.fit_online(env,
buffer,
explorer,
n_epochs=1,
eval_env=eval_env,
logdir='test_data',
tensorboard=False)
def test_fit_online_cartpole_with_dqn():
env = gym.make("CartPole-v0")
eval_env = gym.make("CartPole-v0")
algo = DQN()
buffer = ReplayBuffer(1000, env)
explorer = LinearDecayEpsilonGreedy()
algo.fit_online(
env,
buffer,
explorer,
n_steps=100,
eval_env=eval_env,
logdir="test_data",
)
def test_fit_batch_online_cartpole_with_dqn():
make_env = lambda: gym.make("CartPole-v0")
env = AsyncBatchEnv([make_env for _ in range(5)])
eval_env = gym.make("CartPole-v0")
algo = DQN()
buffer = BatchReplayBuffer(1000, env)
explorer = LinearDecayEpsilonGreedy()
algo.fit_batch_online(
env,
buffer,
explorer,
n_epochs=1,
n_steps_per_epoch=500,
n_updates_per_epoch=1,
eval_env=eval_env,
logdir="test_data",
)
def test_train_atari_with_dqn():
import d4rl_atari
env = gym.make('breakout-mixed-v0', stack=False)
eval_env = gym.make('breakout-mixed-v0', stack=False)
algo = DQN(n_frames=4)
buffer = ReplayBuffer(1000, env)
explorer = LinearDecayEpsilonGreedy()
train(env,
algo,
buffer,
explorer,
n_steps=100,
eval_env=eval_env,
logdir='test_data',
tensorboard=False)
assert algo.impl.observation_shape == (4, 84, 84)
def test_fit_online_atari_with_dqn():
import d4rl_atari
env = ChannelFirst(DummyAtari())
eval_env = ChannelFirst(DummyAtari())
algo = DQN(n_frames=4)
buffer = ReplayBuffer(1000, env)
explorer = LinearDecayEpsilonGreedy()
algo.fit_online(
env,
buffer,
explorer,
n_steps=100,
eval_env=eval_env,
logdir="test_data",
)
assert algo.impl.observation_shape == (4, 84, 84)
def test_fit_online_atari_with_dqn():
import d4rl_atari
env = ChannelFirst(gym.make("breakout-mixed-v0"))
eval_env = ChannelFirst(gym.make("breakout-mixed-v0"))
algo = DQN(n_frames=4)
buffer = ReplayBuffer(1000, env)
explorer = LinearDecayEpsilonGreedy()
algo.fit_online(
env,
buffer,
explorer,
n_steps=100,
eval_env=eval_env,
logdir="test_data",
tensorboard=False,
)
assert algo.impl.observation_shape == (4, 84, 84)
def test_linear_decay_epsilon_greedy(action_size, observation_shape,
start_epsilon, end_epsilon, duration):
explorer = LinearDecayEpsilonGreedy(start_epsilon, end_epsilon, duration)
# check epsilon
assert explorer.compute_epsilon(0) == start_epsilon
assert explorer.compute_epsilon(duration) == end_epsilon
base = start_epsilon - end_epsilon
ref_epsilon = end_epsilon + base * (1.0 - 1.0 / duration)
assert explorer.compute_epsilon(1) == ref_epsilon
ref_x = np.random.random((1, ) + observation_shape)
ref_y = np.random.randint(action_size, size=(1, ))
class DummyAlgo:
def predict(self, x):
assert np.all(x == ref_x)
return ref_y
@property
def impl(self):
return self
@property
def action_size(self):
return action_size
algo = DummyAlgo()
# check sample
for i in range(10):
action = np.random.randint(action_size)
if action != explorer.sample(algo, ref_x, 0):
break
elif i == 9:
assert False
terminate_on_life_loss=False)
# setup algorithm
dqn = DoubleDQN(batch_size=32,
learning_rate=2.5e-4,
optim_factory=AdamFactory(eps=1e-2 / 32),
target_update_interval=10000,
q_func_factory='mean',
scaler='pixel',
n_frames=4,
use_gpu=True)
# replay buffer for experience replay
buffer = ReplayBuffer(maxlen=1000000, env=env)
# epilon-greedy explorer
explorer = LinearDecayEpsilonGreedy(start_epsilon=1.0,
end_epsilon=0.1,
duration=1000000)
# start training
dqn.fit_online(env,
buffer,
explorer,
eval_env=eval_env,
eval_epsilon=0.01,
n_steps=50000000,
n_steps_per_epoch=100000,
update_interval=4,
update_start_step=50000)
def train(params):
# setup algorithm
if pretrain:
dqn = DQN(batch_size=params.get("batch_size"),
learning_rate=params.get("learning_rate"),
target_update_interval=params.get("target_update_interval"),
q_func_factory=QRQFunctionFactory(
n_quantiles=params.get("n_quantiles")),
n_steps=params.get("train_freq"),
gamma=params.get("gamma"),
n_critics=1,
target_reduction_type="min",
use_gpu=True)
# setup replay buffer
buffer = ReplayBuffer(maxlen=params.get("buffer_size"), env=env)
# setup explorers
explorer = LinearDecayEpsilonGreedy(
start_epsilon=1.0,
end_epsilon=params.get("exploration_final_eps"),
duration=100000)
# start training
dqn.fit_online(
env,
buffer,
n_steps=params.get("train_steps"),
explorer=
explorer, # you don't need this with probablistic policy algorithms
tensorboard_dir=log_dir,
eval_env=eval_env)
print("Saving Model")
dqn.save_model(exp_name)
print("convert buffer to dataset")
dataset = buffer.to_mdp_dataset()
# save MDPDataset
dataset.dump('{0}.h5'.format(exp_name))
print("Loading Dataset for Offline Training")
dataset = d3rlpy.dataset.MDPDataset.load('{0}.h5'.format(exp_name))
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
# The dataset can then be used to train a d3rlpy model
cql = DiscreteCQL(learning_rate=6.25e-05,
encoder_factory='default',
q_func_factory='mean',
batch_size=32,
n_frames=1,
n_steps=1,
gamma=0.99,
n_critics=1,
bootstrap=False,
share_encoder=False,
target_reduction_type='min',
target_update_interval=8000,
use_gpu=True,
scaler=None,
augmentation=None,
generator=None,
impl=None)
cql_exp = params.get("model_name") + "_offline_" + params.get(
"environment")
cql_log = '../../../logs/' + cql_exp
cql.fit(dataset.episodes,
eval_episodes=test_episodes,
n_epochs=1000,
scorers={
'environment': evaluate_on_environment(env, epsilon=0.05),
'td_error': td_error_scorer,
'discounted_advantage': discounted_sum_of_advantage_scorer,
'value_scale': average_value_estimation_scorer,
},
tensorboard_dir=cql_log)
cql.save_model(cql_exp)
