def test_simple_cmd(self):
logging.disable(logging.INFO)
echo_params = ParamGrid([
('p1', [3.14, 2.71]),
('p2', ['a', 'b', 'c']),
('p3', list(np.arange(3))),
])
experiments = [
Experiment('test_echo1', 'echo', echo_params.generate_params(randomize=True)),
Experiment('test_echo2', 'echo', echo_params.generate_params(randomize=False)),
]
train_dir = ensure_dir_exists(join(project_tmp_dir(), 'tests'))
root_dir_name = '__test_run__'
rd = RunDescription(root_dir_name, experiments)
args = runner_argparser().parse_args([])
args.max_parallel = 8
args.pause_between = 0
args.train_dir = train_dir
run(rd, args)
rd2 = RunDescription(root_dir_name, experiments, experiment_dirs_sf_format=False, experiment_arg_name='--experiment_tst', experiment_dir_arg_name='--dir')
run(rd2, args)
logging.disable(logging.NOTSET)
shutil.rmtree(join(train_dir, root_dir_name))
def test_experiment(self):
params = ParamGrid([('p1', [3.14, 2.71]), ('p2', ['a', 'b', 'c'])])
cmd = 'python super_rl.py'
ex = Experiment('test', cmd, params.generate_params(randomize=False))
cmds = ex.generate_experiments('train_dir', customize_experiment_name=True, param_prefix='--')
for index, value in enumerate(cmds):
command, name = value
self.assertTrue(command.startswith(cmd))
self.assertTrue(name.startswith(f'0{index}_test'))
def test_descr(self):
params = ParamGrid([('p1', [3.14, 2.71]), ('p2', ['a', 'b', 'c'])])
experiments = [
Experiment('test1', 'python super_rl1.py', params.generate_params(randomize=False)),
Experiment('test2', 'python super_rl2.py', params.generate_params(randomize=False)),
]
rd = RunDescription('test_run', experiments)
cmds = rd.generate_experiments('train_dir')
for command, name, root_dir, env_vars in cmds:
exp_name = split(root_dir)[-1]
self.assertIn('--experiment', command)
self.assertIn('--experiments_root', command)
self.assertTrue(exp_name in name)
self.assertTrue(root_dir.startswith('test_run'))
('ppo_clip_value', [10]),
('with_vtrace', ['False']),
('learning_rate', [0.0001]),
('max_grad_norm', [100.0]),
('use_rnn', ['False']),
('recurrence', [1]),
('num_minibatches_to_accumulate', [0]),
('device', ['gpu']),
('actor_critic_share_weights', ['False']),
('max_policy_lag', [1000000]),
('adaptive_stddev', ['False']),
('ppo_epochs', [20]),
('ppo_clip_ratio', [0.3]),
('batch_size', [1024]),
('num_batches_per_iteration', [10]),
('rollout', [128]),
('nonlinearity', ['tanh']),
('exploration_loss_coeff', [0.0]),
])
_experiment = Experiment(
'mujoco_hopper',
'python -m sample_factory.run_algorithm --env=mujoco_hopper --train_for_env_steps=7000000 --algo=APPO --num_workers=16 --num_envs_per_worker=4 --benchmark=False --with_pbt=False',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('mujoco_hopper_v94',
experiments=[_experiment])
# python -m runner.run --run=mujoco_halfcheetah_grid_search --runner=processes --max_parallel=8 --pause_between=1 --experiments_per_gpu=10000 --num_gpus=1
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('use_rnn', [True, False]),
('mem_size', [4, 0]),
])
cmd = 'python -m train_pytorch --algo=PPO --rollout=64 --recurrence=32 --num_envs=96 --num_workers=96 --train_for_env_steps=1000000000 --normalize_advantage=False --prior_loss_coeff=0.005 '
# IMPORTANT: for DMLAB number of workers better be equal to the number of envs, otherwise spurious crashes may occur!
_experiment_nm = Experiment(
'mem_dmlab_nm',
cmd + '--reward_scale=0.1 --env=dmlab_nonmatch',
_params.generate_params(randomize=False),
)
_experiment_wm = Experiment(
'mem_dmlab_wm',
cmd + '--reward_scale=1.0 --env=dmlab_watermaze',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('mem_dmlab_v24', experiments=[_experiment_nm, _experiment_wm])
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('seed', [1111, 2222, 3333, 4444, 5555]),
('env', ['doom_health_gathering_supreme']),
])
_experiments = [
Experiment(
'health_0_255',
'python -m sample_factory.algorithms.appo.train_appo --train_for_env_steps=40000000 --algo=APPO --env_frameskip=4 --use_rnn=True --num_workers=20 --num_envs_per_worker=12 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --wide_aspect_ratio=False',
_params.generate_params(randomize=False),
),
Experiment(
'health_128_128',
'python -m sample_factory.algorithms.appo.train_appo --train_for_env_steps=40000000 --algo=APPO --env_frameskip=4 --use_rnn=True --num_workers=20 --num_envs_per_worker=12 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --wide_aspect_ratio=False --obs_subtract_mean=128.0 --obs_scale=128.0',
_params.generate_params(randomize=False),
),
]
RUN_DESCRIPTION = RunDescription('doom_health_gathering_v97_fs4',
experiments=_experiments)
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params_earlystop = ParamGrid([
('seed', [0000, 1111, 2222, 3333, 4444]),
])
_experiment_earlystop = Experiment(
'lunar_lander_cont',
'python -m sample_factory_examples.train_gym_env --train_for_env_steps=500000000 --algo=APPO --num_workers=20 --num_envs_per_worker=6 --seed 0 --gae_lambda 0.99 --experiment=lunar_lander_2 --env=gym_LunarLanderContinuous-v2 --exploration_loss_coeff=0.0 --max_grad_norm=0.0 --encoder_type=mlp --encoder_subtype=mlp_mujoco --encoder_extra_fc_layers=0 --hidden_size=128 --policy_initialization=xavier_uniform --actor_critic_share_weights=False --adaptive_stddev=False --recurrence=1 --use_rnn=False --batch_size=256 --ppo_epochs=4 --with_vtrace=False --reward_scale=0.05 --max_policy_lag=100000 --save_every_sec=15 --experiment_summaries_interval=10',
_params_earlystop.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('lunar_lander_cont_v100', experiments=[_experiment_earlystop])