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_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([
('seed', [1111, 2222, 3333]),
('env', ['doom_defend_the_center_flat_actions']),
('num_envs_per_worker', [16]),
])
_experiments = [
Experiment(
'basic_envs_fs4',
'python -m sample_factory.algorithms.appo.train_appo --train_for_env_steps=100000000 --algo=APPO --env_frameskip=4 --use_rnn=True --rnn_type=lstm --num_workers=72 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --wide_aspect_ratio=False --policy_workers_per_policy=3 --experiment_summaries_interval=5 --ppo_clip_value=10.0 --nonlinearity=relu',
_params.generate_params(randomize=False),
),
]
RUN_DESCRIPTION = RunDescription('paper_doom_wall_time_v97_fs4',
experiments=_experiments)
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
from swarm_rl.runs.quad_multi_mix_baseline import QUAD_BASELINE_CLI
_params = ParamGrid([
('quads_collision_falloff_radius', [4.0]),
('quads_collision_reward', [5.0]),
('quads_collision_smooth_max_penalty', [10.0]),
('quads_neighbor_encoder_type', ['attention']),
('replay_buffer_sample_prob', [0.75]),
])
PBT_CLI = QUAD_BASELINE_CLI + (
' --pbt_replace_reward_gap=0.2 --pbt_replace_reward_gap_absolute=200.0 --pbt_period_env_steps=10000000 --pbt_start_mutation=50000000 --with_pbt=True --num_policies=8'
' --pbt_mix_policies_in_one_env=False'
' --num_workers=72 --num_envs_per_worker=10')
_experiment = Experiment(
'quad_mix_baseline-8_pbt',
PBT_CLI,
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('quads_multi_baseline_pbt8_v116',
experiments=[_experiment])
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([])
_experiment = Experiment(
'basic',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_basic --train_for_env_steps=3000000 --algo=APPO --env_frameskip=4 --use_rnn=True --ppo_epochs=1 --rollout=32 --recurrence=32 --wide_aspect_ratio=False --num_workers=20 --num_envs_per_worker=20 --experiment=doom_basic',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('doom_basic', experiments=[_experiment])
('quads_obstacle_num', [1]),
('quads_obstacle_type', ['sphere']),
('quads_obstacle_traj', ['mix']),
('quads_collision_obstacle_reward', [5.0]),
('quads_obstacle_obs_mode', ['absolute']),
('quads_collision_obst_smooth_max_penalty', [10.0]),
('quads_obstacle_hidden_size', [256]),
('replay_buffer_sample_prob', [0.0]),
('quads_obst_penalty_fall_off', [10.0]),
])
_experiment = Experiment(
'quad_mix_baseline_obst_mix-8a',
QUAD_BASELINE_CLI,
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('quads_multi_mix_obst_mix_8a_v116',
experiments=[_experiment])
# On Brain server, when you use num_workers = 72, if the system reports: Resource temporarily unavailable,
# then, try to use two commands below
# export OMP_NUM_THREADS=1
# export USE_SIMPLE_THREADED_LEVEL3=1
# Command to use this script on server:
# xvfb-run python -m runner.run --run=quad_multi_mix_baseline_obstacle_mix --runner=processes --max_parallel=4 --pause_between=1 --experiments_per_gpu=1 --num_gpus=4
# Command to use this script on local machine:
# Please change num_workers to the physical cores of your local machine
# python -m runner.run --run=quad_multi_mix_baseline_obstacle_mix --runner=processes --max_parallel=4 --pause_between=1 --experiments_per_gpu=1 --num_gpus=4
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([])
_experiments = [
Experiment(
'battle_fs4',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_battle --train_for_env_steps=4000000000 --algo=APPO --env_frameskip=4 --use_rnn=True --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --wide_aspect_ratio=False --num_workers=72 --num_envs_per_worker=32 --num_policies=8 --with_pbt=True',
_params.generate_params(randomize=False),
),
]
RUN_DESCRIPTION = RunDescription('paper_doom_battle_appo_pbt_v98_fs4',
experiments=_experiments)
ACTOR_GPUS = '0' # replace with '0 1 2 3 4 5 6 7' for 8-GPU server
NUM_POLICIES = 1
_basic_cli = f'python -m sample_factory.algorithms.appo.train_appo --train_for_seconds={TIMEOUT_SECONDS} --train_for_env_steps=20000000000 --algo=APPO --gamma=0.997 --use_rnn=True --rnn_num_layers=2 --num_workers={NUM_WORKERS} --num_envs_per_worker=16 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --num_policies={NUM_POLICIES} --with_pbt=False --max_grad_norm=0.0 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --policy_workers_per_policy=2 --learner_main_loop_num_cores=4 --reward_clip=30'
_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_cli = f'python -m sample_factory.algorithms.appo.train_appo --train_for_seconds={TIMEOUT_SECONDS} --train_for_env_steps=20000000000 --algo=APPO --gamma=0.997 --use_rnn=True --rnn_num_layers=2 --num_workers={NUM_WORKERS_VOXEL_ENV} --num_envs_per_worker=2 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --actor_worker_gpus {ACTOR_GPUS} --num_policies={NUM_POLICIES} --with_pbt=False --max_grad_norm=0.0 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --voxel_num_simulation_threads=2 --voxel_use_vulkan=True --policy_workers_per_policy=2 --learner_main_loop_num_cores=4 --reward_clip=30 --voxel_num_envs_per_instance=36 --voxel_num_agents_per_env=1 --pbt_mix_policies_in_one_env=False'
_params_voxel_env = ParamGrid([
('env', ['voxel_env_obstacleshard']),
('voxel_use_vulkan', [True, False]),
])
_experiment_voxel_env = Experiment(
'benchmark_voxel_env',
_voxel_cli,
_params_voxel_env.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription(
'voxel_train_benchmark',
experiments=[_experiment_basic_envs, _experiment_voxel_env])
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 = ParamGrid([
('seed', [0000, 3333]),
('quads_obs_repr', ['xyz_vxyz_R_omega', 'xyz_vxyz_R_omega_wall']),
])
_experiment = Experiment(
'grid_search_add_wall_collision_func-8_mixed',
'python -m swarm_rl.train --env=quadrotor_multi --train_for_env_steps=5000000000 --algo=APPO --use_rnn=False --num_workers=72 --num_envs_per_worker=4 --learning_rate=0.0001 --ppo_clip_value=5.0 --recurrence=1 --nonlinearity=tanh --actor_critic_share_weights=False --policy_initialization=xavier_uniform --adaptive_stddev=False --hidden_size=256 --quads_neighbor_hidden_size=256 --with_vtrace=False --max_policy_lag=100000000 --gae_lambda=1.00 --max_grad_norm=5.0 --exploration_loss_coeff=0.0 --rollout=128 --batch_size=1024 --quads_use_numba=True --quads_num_agents=8 --quads_episode_duration=15.0 --quads_mode=mix --quads_formation_size=0.0 --encoder_custom=quad_multi_encoder --with_pbt=False --quads_neighbor_encoder_type=attention --quads_collision_reward=5.0 --neighbor_obs_type=pos_vel',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('quads_multi_mixed_pvg_v112',
experiments=[_experiment])
# On Brain server, when you use num_workers = 72, if the system reports: Resource temporarily unavailable,
# then, try to use two commands below
# export OMP_NUM_THREADS=1
# export USE_SIMPLE_THREADED_LEVEL3=1
# Command to use this script on server:
# xvfb-run python -m runner.run --run=quad_multi_mix_obs_add_walls --runner=processes --max_parallel=4 --pause_between=1 --experiments_per_gpu=1 --num_gpus=4
# Command to use this script on local machine:
# Please change num_workers to the physical cores of your local machine
# python -m runner.run --run=quad_multi_mix_obs_add_walls --runner=processes --max_parallel=4 --pause_between=1 --experiments_per_gpu=1 --num_gpus=4
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('ppo_epochs', [1]),
])
_experiments = [
Experiment(
'battle_fs4_100',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_battle --train_for_env_steps=1000000000 --algo=APPO --env_frameskip=4 --use_rnn=True --num_workers=10 --num_envs_per_worker=10 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --macro_batch=2048 --batch_size=2048 --wide_aspect_ratio=False',
_params.generate_params(randomize=False),
),
Experiment(
'battle_fs4_400',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_battle --train_for_env_steps=1000000000 --algo=APPO --env_frameskip=4 --use_rnn=True --num_workers=20 --num_envs_per_worker=20 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --macro_batch=2048 --batch_size=2048 --wide_aspect_ratio=False',
_params.generate_params(randomize=False),
),
Experiment(
'battle_fs4_800',
'algorithms.appo.train_appo --env=doom_battle --train_for_env_steps=1000000000 --algo=APPO --env_frameskip=4 --use_rnn=True --num_workers=20 --num_envs_per_worker=40 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --macro_batch=2048 --batch_size=2048 --wide_aspect_ratio=False',
_params.generate_params(randomize=False),
),
]
RUN_DESCRIPTION = RunDescription('paper_policy_lag_v66_fs4', experiments=_experiments)
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('env', ['MiniGrid-MemoryS7-v0', 'MiniGrid-RedBlueDoors-8x8-v0', 'MiniGrid-MemoryS17Random-v0']),
])
_experiment = Experiment(
'mem_minigrid_obs',
'python -m train_pytorch --algo=PPO --rollout=64 --num_envs=96 --recurrence=1 --use_rnn=False --train_for_env_steps=200000000 --prior_loss_coeff=0.005 --obs_mem=True',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('mem_minigrid_obs_v26', experiments=[_experiment], pause_between_experiments=5, use_gpus=2, experiments_per_gpu=4, max_parallel=12)
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
from swarm_rl.runs.quad_multi_mix_baseline import QUAD_BASELINE_CLI
_params = ParamGrid([
('quads_neighbor_encoder_type', ['mlp']),
('seed', [0000, 1111, 2222, 3333]),
])
_experiment = Experiment(
'quad_mix_baseline-8_mixed_mlp',
QUAD_BASELINE_CLI,
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('paper_quads_multi_mix_baseline_8a_mlp_v116',
experiments=[_experiment])
# On Brain server, when you use num_workers = 72, if the system reports: Resource temporarily unavailable,
# then, try to use two commands below
# export OMP_NUM_THREADS=1
# export USE_SIMPLE_THREADED_LEVEL3=1
# Command to use this script on server:
# xvfb-run python -m runner.run --run=quad_multi_mix_baseline --runner=processes --max_parallel=3 --pause_between=1 --experiments_per_gpu=1 --num_gpus=3
# Command to use this script on local machine:
# Please change num_workers to the physical cores of your local machine
# python -m runner.run --run=quad_multi_mix_baseline --runner=processes --max_parallel=3 --pause_between=1 --experiments_per_gpu=1 --num_gpus=3
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('ppo_epochs', [1]),
])
_experiment = Experiment(
'static_goal-agents_8',
'python -m swarm_rl.train --env=quadrotor_multi --train_for_env_steps=1000000000 --algo=APPO --use_rnn=False --num_workers=72 --num_envs_per_worker=4 --learning_rate=0.0001 --ppo_clip_value=5.0 --recurrence=1 --nonlinearity=tanh --actor_critic_share_weights=False --policy_initialization=xavier_uniform --adaptive_stddev=False --hidden_size=256 --with_vtrace=False --max_policy_lag=100000000 --gae_lambda=1.00 --max_grad_norm=5.0 --exploration_loss_coeff=0.0 --rollout=128 --batch_size=1024 --extend_obs=True --quads_use_numba=True --quads_num_agents=8 --quads_episode_duration=10.0 --quads_mode=static_goal --quads_dist_between_goals=0.0 --quads_collision_reward=1.0 --encoder_custom=quad_multi_encoder --pbt_replace_reward_gap=0.1 --pbt_replace_reward_gap_absolute=200.0 --pbt_period_env_steps=1000000 --pbt_start_mutation=20000000 --with_pbt=True --num_policies=8',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('quads_multi_same_goal_pbt_v112',
experiments=[_experiment])
# this is just a placeholder for command line. You can run it through runner or just copy-paste to command line and run
'--encoder_custom=quad_multi_encoder --with_pbt=False --quads_collision_reward=5.0 '
'--quads_neighbor_hidden_size=256 --neighbor_obs_type=pos_vel '
'--quads_settle_reward=0.0 --quads_collision_hitbox_radius=2.0 --quads_collision_falloff_radius=4.0 '
'--quads_local_obs=6 --quads_local_metric=dist '
'--quads_local_coeff=1.0 --quads_num_agents=8 '
'--quads_collision_reward=5.0 '
'--quads_collision_smooth_max_penalty=10.0 '
'--quads_neighbor_encoder_type=attention '
'--replay_buffer_sample_prob=0.75 '
'--anneal_collision_steps=300000000')
_experiment = Experiment(
'quad_mix_baseline-8_mixed',
QUAD_BASELINE_CLI,
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('quads_multi_mix_baseline_8a_local_v116',
experiments=[_experiment])
# On Brain server, when you use num_workers = 72, if the system reports: Resource temporarily unavailable,
# then, try to use two commands below
# export OMP_NUM_THREADS=1
# export USE_SIMPLE_THREADED_LEVEL3=1
# Command to use this script on server:
# xvfb-run python -m runner.run --run=quad_multi_mix_baseline --runner=processes --max_parallel=3 --pause_between=1 --experiments_per_gpu=1 --num_gpus=3
# Command to use this script on local machine:
# Please change num_workers to the physical cores of your local machine
# python -m runner.run --run=quad_multi_mix_baseline --runner=processes --max_parallel=3 --pause_between=1 --experiments_per_gpu=1 --num_gpus=3
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('ppo_epochs', [1]),
])
_experiment = Experiment(
'bots_freedm_fs2',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_freedm --train_for_seconds=360000 --algo=APPO --gamma=0.995 --env_frameskip=2 --use_rnn=True --reward_scale=0.5 --num_workers=20 --num_envs_per_worker=4 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --macro_batch=2048 --batch_size=2048 --benchmark=False --start_bot_difficulty=150',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('doom_freedm_v64_fs2',
experiments=[_experiment],
pause_between_experiments=100,
use_gpus=1,
experiments_per_gpu=-1,
max_parallel=1)
from sample_factory.runner.run_description import RunDescription
from sample_factory.runner.runs.voxel_base_experiments import EXPERIMENT_1AGENT
RUN_DESCRIPTION = RunDescription('voxel_env_v115_single_v55', experiments=[EXPERIMENT_1AGENT])
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('ppo_epochs', [1]),
])
_experiment = Experiment(
'quads_pbt',
'python -m swarm_rl.train --env=quadrotor_single --train_for_seconds=3600000 --algo=APPO --gamma=0.99 --use_rnn=False --num_workers=72 --num_envs_per_worker=4 --num_policies=8 --ppo_epochs=1 --rollout=128 --recurrence=1 --batch_size=512 --benchmark=False --pbt_replace_reward_gap=0.1 --pbt_replace_reward_gap_absolute=200.0 --pbt_period_env_steps=1000000 --pbt_start_mutation=20000000 --with_pbt=True --adam_eps=1e-8 --nonlinearity=tanh --actor_critic_share_weights=False --policy_initialization=xavier_uniform --adaptive_stddev=False --hidden_size=64 --with_vtrace=False --max_policy_lag=100000000 --gae_lambda=1.00 --max_grad_norm=0.0 --ppo_clip_value=5.0 --exploration_loss_coeff=0.00001 --learning_rate=5e-4',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('quads_single_pbt_v96_v2',
experiments=[_experiment])
Experiment(
'bots_128_fs2_wide',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_dwango5_bots_experimental --train_for_seconds=3600000 --algo=APPO --use_rnn=True --gamma=0.995 --env_frameskip=2 --rollout=32 --reward_scale=0.5 --num_workers=18 --num_envs_per_worker=20 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --macro_batch=2048 --batch_size=2048 --res_w=128 --res_h=72 --wide_aspect_ratio=True',
_params.generate_params(randomize=False),
dict(DOOM_DEFAULT_UDP_PORT=35300),
),
Experiment(
'bots_128_fs2_narrow',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_dwango5_bots_experimental --train_for_seconds=3600000 --algo=APPO --use_rnn=True --gamma=0.995 --env_frameskip=2 --rollout=32 --reward_scale=0.5 --num_workers=18 --num_envs_per_worker=20 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --macro_batch=2048 --batch_size=2048 --res_w=128 --res_h=72 --wide_aspect_ratio=False',
_params.generate_params(randomize=False),
dict(DOOM_DEFAULT_UDP_PORT=40300),
),
Experiment(
'bots_128_fs2_wide_adam0.5',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_dwango5_bots_experimental --train_for_seconds=3600000 --algo=APPO --use_rnn=True --gamma=0.995 --env_frameskip=2 --rollout=32 --reward_scale=0.5 --num_workers=18 --num_envs_per_worker=20 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --macro_batch=2048 --batch_size=2048 --res_w=128 --res_h=72 --wide_aspect_ratio=True --adam_beta1=0.5',
_params.generate_params(randomize=False),
dict(DOOM_DEFAULT_UDP_PORT=45300),
),
Experiment(
'bots_128_fs2_narrow_adam0.5',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_dwango5_bots_experimental --train_for_seconds=3600000 --algo=APPO --use_rnn=True --gamma=0.995 --env_frameskip=2 --rollout=32 --reward_scale=0.5 --num_workers=18 --num_envs_per_worker=20 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --macro_batch=2048 --batch_size=2048 --res_w=128 --res_h=72 --wide_aspect_ratio=False --adam_beta1=0.5',
_params.generate_params(randomize=False),
dict(DOOM_DEFAULT_UDP_PORT=50300),
),
]
RUN_DESCRIPTION = RunDescription('doom_bots_v60_sweep', experiments=_experiments, pause_between_experiments=120, use_gpus=4, experiments_per_gpu=1, max_parallel=4)
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('seed', [42]),
])
_experiments = [
Experiment(
'bots_128_fs2_narrow',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_deathmatch_bots --train_for_seconds=3600000 --algo=APPO --use_rnn=True --gamma=0.995 --env_frameskip=2 --rollout=32 --num_workers=80 --num_envs_per_worker=24 --num_policies=8 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --res_w=128 --res_h=72 --wide_aspect_ratio=False --with_pbt=True --pbt_period_env_steps=5000000',
_params.generate_params(randomize=False),
dict(DOOM_DEFAULT_UDP_PORT=35300),
),
]
RUN_DESCRIPTION = RunDescription('doom_bots_v100_pbt',
experiments=_experiments)
from sample_factory.runner.run_description import RunDescription
from sample_factory.runner.runs.voxel_base_experiments import EXPERIMENT_2AGENTS, EXPERIMENT_4AGENTS
RUN_DESCRIPTION = RunDescription(
'voxel_env_v115_multi_agent_v55',
experiments=[EXPERIMENT_2AGENTS, EXPERIMENT_4AGENTS])
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('ppo_epochs', [1]),
])
_experiment = Experiment(
'bots_ssl2_fs2',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_duel_bots --train_for_seconds=360000 --algo=APPO --gamma=0.995 --env_frameskip=2 --use_rnn=True --reward_scale=0.5 --num_workers=72 --num_envs_per_worker=32 --num_policies=8 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --benchmark=False --res_w=128 --res_h=72 --wide_aspect_ratio=False --pbt_replace_reward_gap=0.2 --pbt_replace_reward_gap_absolute=3.0 --pbt_period_env_steps=5000000 --save_milestones_sec=1800 --with_pbt=True',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('paper_doom_duel_bots_v98_fs2', experiments=[_experiment])
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('ppo_epochs', [1]),
])
_experiments = [
Experiment(
'battle_d4_fs4_pbt',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_battle_d4 --train_for_env_steps=50000000000 --algo=APPO --env_frameskip=4 --use_rnn=True --reward_scale=0.5 --num_workers=24 --num_envs_per_worker=30 --num_policies=4 --ppo_epochs=1 --rollout=32 --recurrence=32 --macro_batch=2048 --batch_size=2048 --wide_aspect_ratio=False --pbt_period_env_steps=5000000',
_params.generate_params(randomize=False),
),
]
RUN_DESCRIPTION = RunDescription('doom_battle_d4_appo_v64_fs4_pbt',
experiments=_experiments,
use_gpus=2,
experiments_per_gpu=-1,
max_parallel=1)
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('seed', [1111, 2222, 3333]),
])
_experiments = [
Experiment(
'battle2_fs4',
'python -m sample_factory.algorithms.appo.train_appo --env=doom_battle2 --train_for_env_steps=3000000000 --algo=APPO --env_frameskip=4 --use_rnn=True --reward_scale=0.5 --num_workers=20 --num_envs_per_worker=20 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --wide_aspect_ratio=False',
_params.generate_params(randomize=False),
),
]
RUN_DESCRIPTION = RunDescription('doom_battle2_appo_v1.119.0_fs4',
experiments=_experiments)
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),
)
RUN_DESCRIPTION = RunDescription(
'voxel_bench_sampling',
experiments=[_experiment_basic_envs, _experiment_voxel_env])
from swarm_rl.runs.quad_multi_mix_baseline import QUAD_BASELINE_CLI
_params = ParamGrid([
('seed', [0000, 1111, 2222, 3333]),
])
_experiment_no_replay = Experiment(
'quad_mix_baseline-8_mixed_noreplay',
QUAD_BASELINE_CLI + ' --replay_buffer_sample_prob=0.00',
_params.generate_params(randomize=False),
)
_experiment_no_anneal = Experiment(
'quad_mix_baseline-8_mixed_noannealing',
QUAD_BASELINE_CLI + ' --anneal_collision_steps=0',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('paper_quads_multi_mix_baseline_8a_ablation_v116', experiments=[_experiment_no_replay, _experiment_no_anneal])
# On Brain server, when you use num_workers = 72, if the system reports: Resource temporarily unavailable,
# then, try to use two commands below
# export OMP_NUM_THREADS=1
# export USE_SIMPLE_THREADED_LEVEL3=1
# Command to use this script on server:
# xvfb-run python -m runner.run --run=quad_multi_mix_baseline --runner=processes --max_parallel=3 --pause_between=1 --experiments_per_gpu=1 --num_gpus=3
# Command to use this script on local machine:
# Please change num_workers to the physical cores of your local machine
# python -m runner.run --run=quad_multi_mix_baseline --runner=processes --max_parallel=3 --pause_between=1 --experiments_per_gpu=1 --num_gpus=3
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('seed', [0, 1111, 2222, 3333, 4444, 5555, 6666, 7777, 8888, 9999]),
('env', [
'doom_my_way_home', 'doom_deadly_corridor', 'doom_defend_the_center',
'doom_defend_the_line', 'doom_health_gathering',
'doom_health_gathering_supreme'
]),
])
_experiments = [
Experiment(
'basic_envs_fs4',
'python -m sample_factory.algorithms.appo.train_appo --train_for_env_steps=500000000 --algo=APPO --env_frameskip=4 --use_rnn=True --num_workers=36 --num_envs_per_worker=8 --num_policies=1 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --wide_aspect_ratio=False',
_params.generate_params(randomize=False),
),
]
RUN_DESCRIPTION = RunDescription('paper_doom_basic_envs_appo_v97_fs4',
experiments=_experiments)
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('env', ['doom_two_colors_easy', 'doom_two_colors_hard']),
('use_rnn', [True, False]),
('mem_size', [4, 0]),
])
_experiment = Experiment(
'mem_doom',
'python -m train_pytorch --algo=PPO --train_for_env_steps=1000000000 --prior_loss_coeff=0.005 --reward_scale=0.5',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('mem_doom_v39',
experiments=[_experiment],
pause_between_experiments=10,
use_gpus=2,
experiments_per_gpu=2,
max_parallel=4)
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])
from sample_factory.runner.run_description import RunDescription, Experiment, ParamGrid
_params = ParamGrid([
('ppo_epochs', [1]),
])
_experiment = Experiment(
'voxel_env_pbt',
'python -m sample_factory.algorithms.appo.train_appo --env=voxel_env_v23_v --train_for_seconds=360000000 --algo=APPO --gamma=0.997 --use_rnn=True --num_workers=28 --num_envs_per_worker=2 --ppo_epochs=1 --rollout=32 --recurrence=32 --batch_size=2048 --actor_worker_gpus 0 1 2 3 4 5 6 7 --num_policies=8 --with_pbt=True --max_grad_norm=0.0 --pbt_replace_reward_gap_absolute=0.3 --exploration_loss=symmetric_kl --exploration_loss_coeff=0.001 --pbt_mix_policies_in_one_env=False --experiment=voxel_env_v23_v --voxel_num_envs_per_instance=48 --voxel_num_agents_per_env=4 --voxel_num_simulation_threads=4 --voxel_vertical_look_limit=0.2 --voxel_use_vulkan=True --policy_workers_per_policy=2 --learner_main_loop_num_cores=4',
_params.generate_params(randomize=False),
)
RUN_DESCRIPTION = RunDescription('voxel_env_pbt_v112_env_v23_8p',
experiments=[_experiment])
