evaluation_data_collector=eval_path_collector,
replay_buffer=replay_buffer,
**variant['algo_kwargs']
)
algorithm.to(ptu.device)
algorithm.train()
if __name__ == "__main__":
variant = dict(
algo_kwargs=dict(
num_epochs=100,
max_path_length=50,
num_eval_steps_per_epoch=1000,
num_expl_steps_per_train_loop=1000,
num_trains_per_train_loop=1000,
min_num_steps_before_training=1000,
batch_size=128,
),
trainer_kwargs=dict(
discount=0.99,
),
replay_buffer_kwargs=dict(
max_size=100000,
fraction_goals_rollout_goals=0.2, # equal to k = 4 in HER paper
fraction_goals_env_goals=0.0,
),
)
setup_logger('her-dqn-gridworld-experiment', variant=variant)
experiment(variant)
policy = TanhGaussianPolicy(
hidden_sizes=[100, 100],
obs_dim=obs_dim,
action_dim=action_dim,
)
algorithm = SoftActorCritic(env=env,
policy=policy,
qf=qf,
vf=vf,
**variant['algo_params'])
algorithm.to(ptu.device)
with torch.autograd.profiler.profile() as prof:
algorithm.train()
prof.export_chrome_trace("tmp-torch-chrome-trace.prof")
if __name__ == "__main__":
# noinspection PyTypeChecker
variant = dict(algo_params=dict(
num_epochs=10,
num_steps_per_epoch=1000,
num_steps_per_eval=300,
batch_size=64,
max_path_length=30,
reward_scale=0.3,
discount=0.99,
soft_target_tau=0.001,
), )
setup_logger("11-24-profile")
experiment(variant)
algorithm = SoftActorCritic(env=env,
policy=policy,
qf=qf,
vf=vf,
**variant['algo_params'])
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
if __name__ == "__main__":
# noinspection PyTypeChecker
variant = dict(
algo_params=dict(
num_epochs=1000,
num_steps_per_epoch=1000,
num_steps_per_eval=1000,
batch_size=128,
max_path_length=999,
discount=0.99,
soft_target_tau=0.001,
policy_lr=3E-4,
qf_lr=3E-4,
vf_lr=3E-4,
use_automatic_entropy_tuning=True,
),
net_size=300,
)
setup_logger('test', variant=variant)
experiment(variant)
num_steps_per_eval=500,
max_path_length=200,
discount=0.99,
replay_buffer_size=int(1E4),
),
cnn_params=dict(
kernel_sizes=[3, 3],
n_channels=[16, 16],
strides=[2, 2],
pool_sizes=[1, 1],
hidden_sizes=[128, 64],
paddings=[0, 0],
use_layer_norm=False,
),
env_id='InvertedPendulum-v2',
)
setup_logger('name-of-td3-experiment', variant=variant)
experiment(variant)
for i in range(2):
run_experiment(
experiment,
variant=variant,
exp_id=0,
exp_prefix="TD3-images-inverted-pendulum",
mode='local',
# exp_prefix="double-vs-dqn-huber-sweep-cartpole",
# mode='local',
#use_gpu=True,
)
variant = dict(
algorithm='HER-SAC',
version='normal',
algo_kwargs=dict(
batch_size=128,
num_epochs=100,
num_eval_steps_per_epoch=5000,
num_expl_steps_per_train_loop=1000,
num_trains_per_train_loop=1000,
min_num_steps_before_training=1000,
max_path_length=50,
),
sac_trainer_kwargs=dict(
discount=0.99,
soft_target_tau=5e-3,
target_update_period=1,
policy_lr=3E-4,
qf_lr=3E-4,
reward_scale=1,
use_automatic_entropy_tuning=True,
),
replay_buffer_kwargs=dict(
max_size=int(1E6),
fraction_goals_rollout_goals=0.2, # equal to k = 4 in HER paper
fraction_goals_env_goals=0,
),
qf_kwargs=dict(hidden_sizes=[400, 300], ),
policy_kwargs=dict(hidden_sizes=[400, 300], ),
)
setup_logger('her-sac-fetch-experiment', variant=variant)
experiment(variant)
'Reacher-v2',
],
'bins': [9],
'algo_class': [
DoubleDQN,
],
'learning_rate': [1e-3, 1e-4],
'qf_criterion_class': [
HuberLoss,
],
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
setup_logger('dqn-images-experiment', variant=variant)
experiment(variant)
for exp_id, variant in enumerate(sweeper.iterate_hyperparameters()):
#for i in range(2):
run_experiment(
experiment,
variant=variant,
exp_id=exp_id,
exp_prefix="dqn-Pusher2D-test",
mode='ec2',
# use_gpu=False,
# exp_prefix="double-vs-dqn-huber-sweep-cartpole",
# mode='local',
# use_gpu=True,
)
trainer_kwargs=dict(
discount=0.99,
),
replay_buffer_kwargs=dict(
max_size=100000,
fraction_goals_rollout_goals=0.2,
fraction_goals_env_goals=0.0,
),
qf_kwargs=dict(
hidden_sizes=[400, 300],
),
policy_kwargs=dict(
hidden_sizes=[400, 300],
),
)
setup_logger('her-td3-pusher-0', variant=variant)
experiment(variant)
# search_space = {
# }
# sweeper = hyp.DeterministicHyperparameterSweeper(
# search_space, default_parameters=variant,
# )
#
# n_seeds = 1
# mode = 'local'
# exp_prefix = 'dev-{}'.format(
# __file__.replace('/', '-').replace('_', '-').split('.')[0]
# )
#
# n_seeds = 5
# mode = 'sss'
object_high=(x_high - t - t, y_high - t, 0.02),
),
algo_kwargs=dict(
num_epochs=2000,
max_path_length=20,
batch_size=128,
num_eval_steps_per_epoch=1000,
num_expl_steps_per_train_loop=1000,
num_trains_per_train_loop=100,
min_num_steps_before_training=1000,
),
sac_trainer_kwargs=dict(
discount=0.99,
soft_target_tau=5e-3,
target_update_period=1,
policy_lr=3E-4,
qf_lr=3E-4,
reward_scale=1,
use_automatic_entropy_tuning=True,
),
replay_buffer_kwargs=dict(
max_size=int(1E6),
fraction_goals_rollout_goals=0.2, # equal to k = 4 in HER paper
fraction_goals_env_goals=0,
),
qf_kwargs=dict(hidden_sizes=[400, 300], ),
policy_kwargs=dict(hidden_sizes=[400, 300], ),
)
setup_logger('her-sac-pusher', variant=variant)
experiment(variant)