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_param_grid(self): grid = ParamGrid([ ('p1', [0, 1]), ('p2', ['a', 'b', 'c']), ('p3', [None, {}]), ]) param_combinations = grid.generate_params(randomize=True) for p in param_combinations: for key in ('p1', 'p2', 'p3'): self.assertIn(key, p) param_combinations = list(grid.generate_params(randomize=False)) self.assertEqual(param_combinations[0], {'p1': 0, 'p2': 'a', 'p3': None}) self.assertEqual(param_combinations[1], {'p1': 0, 'p2': 'a', 'p3': {}}) self.assertEqual(param_combinations[-2], {'p1': 1, 'p2': 'c', 'p3': None}) self.assertEqual(param_combinations[-1], {'p1': 1, 'p2': 'c', 'p3': {}})
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'))
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid _params = ParamGrid([ ('env', ['voxel_env_multitask_Obstacles']), ('use_cpc', ['True']), ('seed', [11111, 22222, 33333, 44444, 55555]), ]) _cli = 'python -m sample_factory.algorithms.appo.train_appo --train_for_seconds=360000000 --train_for_env_steps=10000000000 --algo=APPO --gamma=0.997 --use_rnn=True --rnn_num_layers=2 --num_workers=12 --num_envs_per_worker=2 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --actor_worker_gpus 0 --num_policies=1 --with_pbt=False --max_grad_norm=0.0 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --voxel_num_simulation_threads=1 --voxel_use_vulkan=True --policy_workers_per_policy=2 --learner_main_loop_num_cores=2 --reward_clip=30 --pbt_mix_policies_in_one_env=False' EXPERIMENT_1AGENT = Experiment( 'voxel_env_multitask_obs', _cli + ' --voxel_num_envs_per_instance=36 --voxel_num_agents_per_env=1', _params.generate_params(randomize=False), ) RUN_DESCRIPTION = RunDescription('voxel_env_v115_multitask_obstacles_v55', experiments=[EXPERIMENT_1AGENT])
NUM_WORKERS = 20 # typically num logical cores NUM_WORKERS_VOXEL_ENV = 10 # typically num logical cores / 2, limited by the num of available Vulkan contexts TIMEOUT_SECONDS = 180 SAMPLER_GPUS = '0' # replace with '0 1 2 3 4 5 6 7' for 8-GPU server _basic_cli = f'python -m sample_factory.run_algorithm --algo=DUMMY_SAMPLER --num_workers={NUM_WORKERS} --num_envs_per_worker=1 --experiment=benchmark --timeout_seconds={TIMEOUT_SECONDS}' _params_basic_envs = ParamGrid([ ('env', ['doom_benchmark', 'atari_breakout', 'dmlab_benchmark']), ]) _experiment_basic_envs = Experiment( 'benchmark_basic_envs', _basic_cli, _params_basic_envs.generate_params(randomize=False), ) _voxel_env_cli = f'python -m sample_factory.run_algorithm --algo=DUMMY_SAMPLER --num_workers={NUM_WORKERS_VOXEL_ENV} --num_envs_per_worker=1 --experiment=benchmark --sampler_worker_gpus {SAMPLER_GPUS} --voxel_num_envs_per_instance=64 --voxel_num_agents_per_env=2 --voxel_num_simulation_threads=2 --timeout_seconds={TIMEOUT_SECONDS}' _params_voxel_env = ParamGrid([ ('env', ['voxel_env_obstacleshard']), ('voxel_use_vulkan', [True, False]), ]) _experiment_voxel_env = Experiment( 'benchmark_voxel_env', _voxel_env_cli, _params_voxel_env.generate_params(randomize=False), )
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid NUM_WORKERS_VOXEL_ENV = 48 # typically num logical cores / 2, limited by the num of available Vulkan contexts TIMEOUT_SECONDS = 180 SAMPLER_GPUS = '0 1 2 3 4 5 6 7' # replace with '0 1 2 3 4 5 6 7' for 8-GPU server _voxel_env_cli = f'python -m sample_factory.run_algorithm --algo=DUMMY_SAMPLER --num_workers={NUM_WORKERS_VOXEL_ENV} --num_envs_per_worker=1 --experiment=benchmark --sampler_worker_gpus {SAMPLER_GPUS} --voxel_num_envs_per_instance=64 --voxel_num_agents_per_env=2 --voxel_num_simulation_threads=2 --timeout_seconds={TIMEOUT_SECONDS}' _params_voxel_env = ParamGrid([ ('env', [ 'voxel_env_TowerBuilding', 'voxel_env_ObstaclesEasy', 'voxel_env_ObstaclesHard', 'voxel_env_Collect', 'voxel_env_Sokoban', 'voxel_env_HexMemory', 'voxel_env_HexExplore', 'voxel_env_Rearrange' ]), ('voxel_use_vulkan', [True]), ]) _experiment_voxel_env = Experiment( 'benchmark_voxel_env_8', _voxel_env_cli, _params_voxel_env.generate_params(randomize=False), ) RUN_DESCRIPTION = RunDescription('voxel_bench_sampling_all_envs', experiments=[_experiment_voxel_env])
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])