def main(args):
dataset, env = get_atari(args.dataset)
d3rlpy.seed(args.seed)
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
cql = DiscreteCQL(
n_frames=4, # frame stacking
q_func_type=args.q_func_type,
scaler='pixel',
use_gpu=args.gpu)
cql.fit(train_episodes,
eval_episodes=test_episodes,
n_epochs=100,
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
})
def main(args):
dataset, env = get_atari(args.dataset)
d3rlpy.seed(args.seed)
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
device = None if args.gpu is None else Device(args.gpu)
cql = DiscreteCQL(n_epochs=100,
q_func_type=args.q_func_type,
scaler='pixel',
use_batch_norm=False,
use_gpu=device)
cql.fit(train_episodes,
eval_episodes=test_episodes,
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
})
def save_policy(self, path, epoch, as_onnx):
params_path = os.path.join(self.get_log_path(), 'params.json')
model_path = os.path.join(self.get_log_path(), 'model_%d.pt' % epoch)
if not os.path.exists(model_path):
raise ValueError('%s does not exist.' % model_path)
# initialize algorithm from json file
if self.project.algorithm == 'cql':
if self.project.dataset.is_discrete:
algo = DiscreteCQL.from_json(params_path)
else:
algo = CQL.from_json(params_path)
else:
raise ValueError('unsupported algorithm.')
# load model parameters
algo.load_model(model_path)
# save TorchScript policy
algo.save_policy(path, as_onnx)
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)
from sklearn.model_selection import train_test_split
from d3rlpy.datasets import get_atari
from d3rlpy.algos import DiscreteCQL
from d3rlpy.ope import DiscreteFQE
from d3rlpy.metrics.scorer import evaluate_on_environment
from d3rlpy.metrics.scorer import initial_state_value_estimation_scorer
from d3rlpy.metrics.scorer import soft_opc_scorer
dataset, env = get_atari('breakout-expert-v0')
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
# train algorithm
cql = DiscreteCQL(n_epochs=100,
scaler='pixel',
q_func_factory='qr',
n_frames=4,
use_gpu=True)
cql.fit(train_episodes,
eval_episodes=test_episodes,
scorers={
'environment': evaluate_on_environment(env, epsilon=0.05),
'init_value': initial_state_value_estimation_scorer,
'soft_opc': soft_opc_scorer(70)
})
# or load the trained model
# cql = DiscreteCQL.from_json('<path-to-json>/params.json')
# cql.load_model('<path-to-model>/model.pt')
# evaluate the trained policy
from d3rlpy.datasets import get_atari
from d3rlpy.algos import DiscreteCQL
from d3rlpy.metrics.scorer import evaluate_on_environment
from d3rlpy.metrics.scorer import td_error_scorer
from d3rlpy.metrics.scorer import discounted_sum_of_advantage_scorer
from d3rlpy.metrics.scorer import average_value_estimation_scorer
from sklearn.model_selection import train_test_split
dataset, env = get_atari('breakout-expert-v0')
train_episodes, test_episodes = train_test_split(dataset, test_size=0.2)
cql = DiscreteCQL(scaler='pixel',
n_frames=4,
augmentation=['random_shift', 'intensity'],
use_gpu=True)
cql.fit(train_episodes,
eval_episodes=test_episodes,
n_epochs=100,
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
})
from d3rlpy.algos import DiscreteCQL
from d3rlpy.models.optimizers import AdamFactory
from d3rlpy.datasets import get_atari
from d3rlpy.metrics.scorer import evaluate_on_environment
from d3rlpy.metrics.scorer import average_value_estimation_scorer
from sklearn.model_selection import train_test_split
dataset, env = get_atari('breakout-medium-v0')
_, test_episodes = train_test_split(dataset, test_size=0.2)
cql = DiscreteCQL(optim_factory=AdamFactory(eps=1e-2 / 32),
scaler='pixel',
n_frames=4,
q_func_factory='qr',
use_gpu=True)
cql.fit(dataset.episodes,
eval_episodes=test_episodes,
n_epochs=2000,
scorers={
'environment': evaluate_on_environment(env, epsilon=0.001),
'value_scale': average_value_estimation_scorer
})