def test_train_with_sac():
env = gym.make('Pendulum-v0')
eval_env = gym.make('Pendulum-v0')
algo = SAC(n_epochs=1)
buffer = ReplayBuffer(1000, env)
train(env,
algo,
buffer,
eval_env=eval_env,
logdir='test_data',
tensorboard=False)
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_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)
from d3rlpy.online.buffers import ReplayBuffer
from d3rlpy.online.explorers import LinearDecayEpsilonGreedy
from d3rlpy.online.iterators import train
env = gym.make('CartPole-v0')
eval_env = gym.make('CartPole-v0')
# setup algorithm
dqn = DQN(n_epochs=30,
batch_size=32,
learning_rate=2.5e-4,
target_update_interval=100,
use_gpu=False)
# replay buffer for experience replay
buffer = ReplayBuffer(maxlen=100000, env=env)
# epilon-greedy explorer
explorer = LinearDecayEpsilonGreedy(start_epsilon=1.0,
end_epsilon=0.1,
duration=10000)
# start training
train(env,
dqn,
buffer,
explorer,
eval_env=eval_env,
n_steps_per_epoch=1000,
n_updates_per_epoch=100)
import gym
from d3rlpy.algos import SAC
from d3rlpy.online.buffers import ReplayBuffer
from d3rlpy.online.iterators import train
env = gym.make('Pendulum-v0')
eval_env = gym.make('Pendulum-v0')
# setup algorithm
sac = SAC(n_epochs=100, batch_size=100, use_gpu=False)
# replay buffer for experience replay
buffer = ReplayBuffer(maxlen=100000, env=env)
# start training
# probablistic policies does not need explorers
train(env,
sac,
buffer,
eval_env=eval_env,
n_steps_per_epoch=1000,
n_updates_per_epoch=100)
# prepare dataset and environment
dataset, env = get_pybullet('hopper-bullet-random-v0')
_, eval_env = get_pybullet('hopper-bullet-random-v0')
train_episodes, test_episodes = train_test_split(dataset)
# setup algorithm
awac = AWAC(n_epochs=30,
encoder_params={'hidden_units': [256, 256, 256, 256]},
use_gpu=True)
## pretrain
awac.fit(train_episodes[:10000],
eval_episodes=test_episodes,
scorers={
'environment': evaluate_on_environment(env),
'advantage': discounted_sum_of_advantage_scorer,
'value_scale': average_value_estimation_scorer
})
# fine-tuning
awac.n_epochs = 1000
buffer = ReplayBuffer(1000000, env, train_episodes[:10000])
train(env,
awac,
buffer,
eval_env=eval_env,
eval_epsilon=0.0,
n_steps_per_epoch=1000,
n_updates_per_epoch=1000)