def test_recurrent_poca(action_sizes, is_multiagent):
if is_multiagent:
# This is not a recurrent environment, just check if LSTM doesn't crash
env = MultiAgentEnvironment([BRAIN_NAME],
action_sizes=action_sizes,
num_agents=2)
else:
# Actually test LSTM here
env = MemoryEnvironment([BRAIN_NAME], action_sizes=action_sizes)
new_network_settings = attr.evolve(
POCA_TORCH_CONFIG.network_settings,
memory=NetworkSettings.MemorySettings(memory_size=16),
)
new_hyperparams = attr.evolve(
POCA_TORCH_CONFIG.hyperparameters,
learning_rate=1.0e-3,
batch_size=64,
buffer_size=128,
)
config = attr.evolve(
POCA_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_network_settings,
max_steps=500 if is_multiagent else 6000,
)
check_environment_trains(env, {BRAIN_NAME: config},
success_threshold=None if is_multiagent else 0.9)
def test_visual_advanced_sac(vis_encode_type, num_visual):
env = SimpleEnvironment(
[BRAIN_NAME],
action_sizes=(0, 1),
num_visual=num_visual,
num_vector=0,
step_size=0.5,
vis_obs_size=(5, 5, 5) if vis_encode_type == "match3" else (36, 36, 3),
)
new_networksettings = attr.evolve(
SAC_TORCH_CONFIG.network_settings,
vis_encode_type=EncoderType(vis_encode_type))
new_hyperparams = attr.evolve(
SAC_TORCH_CONFIG.hyperparameters,
batch_size=16,
learning_rate=3e-4,
buffer_init_steps=0,
)
config = attr.evolve(
SAC_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_networksettings,
max_steps=100,
)
# The number of steps is pretty small for these encoders
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.5)
def test_gail_visual_sac(simple_record, action_sizes):
demo_path = simple_record(action_sizes, num_visual=1, num_vector=0)
env = SimpleEnvironment(
[BRAIN_NAME],
num_visual=1,
num_vector=0,
action_sizes=action_sizes,
step_size=0.2,
)
bc_settings = BehavioralCloningSettings(demo_path=demo_path, steps=1000)
reward_signals = {
RewardSignalType.GAIL: GAILSettings(encoding_size=32,
demo_path=demo_path)
}
hyperparams = attr.evolve(SAC_TORCH_CONFIG.hyperparameters,
learning_rate=3e-4,
batch_size=16)
config = attr.evolve(
SAC_TORCH_CONFIG,
reward_signals=reward_signals,
hyperparameters=hyperparams,
behavioral_cloning=bc_settings,
max_steps=500,
)
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.9)
def test_simple_asymm_ghost_fails(action_sizes):
# Make opponent for asymmetric case
brain_name_opp = BRAIN_NAME + "Opp"
env = SimpleEnvironment(
[BRAIN_NAME + "?team=0", brain_name_opp + "?team=1"],
action_sizes=action_sizes)
# This config should fail because the team that us not learning when both have reached
# max step should be executing the initial, untrained poliy.
self_play_settings = SelfPlaySettings(
play_against_latest_model_ratio=0.0,
save_steps=5000,
swap_steps=5000,
team_change=2000,
)
config = attr.evolve(PPO_TORCH_CONFIG,
self_play=self_play_settings,
max_steps=3000)
check_environment_trains(env, {
BRAIN_NAME: config,
brain_name_opp: config
},
success_threshold=None)
processed_rewards = [
default_reward_processor(rewards)
for rewards in env.final_rewards.values()
]
success_threshold = 0.9
assert any(reward > success_threshold
for reward in processed_rewards) and any(
reward < success_threshold for reward in processed_rewards)
def test_simple_ghost(action_sizes):
env = SimpleEnvironment(
[BRAIN_NAME + "?team=0", BRAIN_NAME + "?team=1"], action_sizes=action_sizes
)
self_play_settings = SelfPlaySettings(
play_against_latest_model_ratio=1.0, save_steps=2000, swap_steps=2000
)
config = attr.evolve(PPO_TORCH_CONFIG, self_play=self_play_settings, max_steps=2500)
check_environment_trains(env, {BRAIN_NAME: config})
def test_2d_ppo(action_sizes):
env = SimpleEnvironment([BRAIN_NAME], action_sizes=action_sizes, step_size=0.8)
new_hyperparams = attr.evolve(
PPO_TORCH_CONFIG.hyperparameters, batch_size=64, buffer_size=640
)
config = attr.evolve(
PPO_TORCH_CONFIG, hyperparameters=new_hyperparams, max_steps=10000
)
check_environment_trains(env, {BRAIN_NAME: config})
def test_hybrid_visual_ppo(num_visual):
env = SimpleEnvironment([BRAIN_NAME],
num_visual=num_visual,
num_vector=0,
action_sizes=(1, 1))
new_hyperparams = attr.evolve(PPO_TORCH_CONFIG.hyperparameters,
learning_rate=3.0e-4)
config = attr.evolve(PPO_TORCH_CONFIG, hyperparameters=new_hyperparams)
check_environment_trains(env, {BRAIN_NAME: config})
def test_visual_poca(num_visual):
env = MultiAgentEnvironment([BRAIN_NAME],
action_sizes=(0, 1),
num_agents=2,
num_visual=num_visual)
new_hyperparams = attr.evolve(POCA_TORCH_CONFIG.hyperparameters,
learning_rate=3.0e-4)
config = attr.evolve(POCA_TORCH_CONFIG, hyperparameters=new_hyperparams)
check_environment_trains(env, {BRAIN_NAME: config})
def test_2d_sac(action_sizes):
env = SimpleEnvironment([BRAIN_NAME], action_sizes=action_sizes, step_size=0.8)
new_hyperparams = attr.evolve(
SAC_TORCH_CONFIG.hyperparameters, buffer_init_steps=2000
)
config = attr.evolve(
SAC_TORCH_CONFIG, hyperparameters=new_hyperparams, max_steps=6000
)
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.8)
def test_visual_ppo(num_visual, use_discrete):
env = SimpleEnvironment(
[BRAIN_NAME],
use_discrete=use_discrete,
num_visual=num_visual,
num_vector=0,
step_size=0.2,
)
new_hyperparams = attr.evolve(PPO_TORCH_CONFIG.hyperparameters,
learning_rate=3.0e-4)
config = attr.evolve(PPO_TORCH_CONFIG, hyperparameters=new_hyperparams)
check_environment_trains(env, {BRAIN_NAME: config})
def test_var_len_obs_ppo(num_vis, num_vector, num_var_len, action_sizes):
env = SimpleEnvironment(
[BRAIN_NAME],
action_sizes=action_sizes,
num_visual=num_vis,
num_vector=num_vector,
num_var_len=num_var_len,
step_size=0.2,
)
new_hyperparams = attr.evolve(PPO_TORCH_CONFIG.hyperparameters,
learning_rate=3.0e-4)
config = attr.evolve(PPO_TORCH_CONFIG, hyperparameters=new_hyperparams)
check_environment_trains(env, {BRAIN_NAME: config})
def test_hybrid_ppo(action_size):
env = SimpleEnvironment([BRAIN_NAME], action_sizes=action_size, step_size=0.8)
new_network_settings = attr.evolve(PPO_TORCH_CONFIG.network_settings)
new_hyperparams = attr.evolve(
PPO_TORCH_CONFIG.hyperparameters, batch_size=64, buffer_size=1024
)
config = attr.evolve(
PPO_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_network_settings,
max_steps=10000,
)
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.9)
def test_visual_sac(num_visual, action_sizes):
env = SimpleEnvironment(
[BRAIN_NAME],
action_sizes=action_sizes,
num_visual=num_visual,
num_vector=0,
step_size=0.2,
)
new_hyperparams = attr.evolve(SAC_TORCH_CONFIG.hyperparameters,
batch_size=16,
learning_rate=3e-4)
config = attr.evolve(SAC_TORCH_CONFIG, hyperparameters=new_hyperparams)
check_environment_trains(env, {BRAIN_NAME: config})
def test_simple_asymm_ghost(action_sizes):
# Make opponent for asymmetric case
brain_name_opp = BRAIN_NAME + "Opp"
env = SimpleEnvironment(
[BRAIN_NAME + "?team=0", brain_name_opp + "?team=1"], action_sizes=action_sizes
)
self_play_settings = SelfPlaySettings(
play_against_latest_model_ratio=1.0,
save_steps=10000,
swap_steps=10000,
team_change=400,
)
config = attr.evolve(PPO_TORCH_CONFIG, self_play=self_play_settings, max_steps=4000)
check_environment_trains(env, {BRAIN_NAME: config, brain_name_opp: config})
def test_gail(simple_record, action_sizes, trainer_config):
demo_path = simple_record(action_sizes)
env = SimpleEnvironment([BRAIN_NAME], action_sizes=action_sizes, step_size=0.2)
bc_settings = BehavioralCloningSettings(demo_path=demo_path, steps=1000)
reward_signals = {
RewardSignalType.GAIL: GAILSettings(encoding_size=32, demo_path=demo_path)
}
config = attr.evolve(
trainer_config,
reward_signals=reward_signals,
behavioral_cloning=bc_settings,
max_steps=500,
)
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.9)
def test_hybrid_visual_sac(num_visual):
env = SimpleEnvironment(
[BRAIN_NAME], num_visual=num_visual, num_vector=0, action_sizes=(1, 1)
)
new_hyperparams = attr.evolve(
SAC_TORCH_CONFIG.hyperparameters,
buffer_size=50000,
batch_size=128,
learning_rate=3.0e-4,
)
config = attr.evolve(
SAC_TORCH_CONFIG, hyperparameters=new_hyperparams, max_steps=3000
)
check_environment_trains(env, {BRAIN_NAME: config})
def test_hybrid_sac(action_size):
env = SimpleEnvironment([BRAIN_NAME], action_sizes=action_size, step_size=0.8)
new_hyperparams = attr.evolve(
SAC_TORCH_CONFIG.hyperparameters,
buffer_size=50000,
batch_size=256,
buffer_init_steps=0,
)
config = attr.evolve(
SAC_TORCH_CONFIG, hyperparameters=new_hyperparams, max_steps=2200
)
check_environment_trains(
env, {BRAIN_NAME: config}, success_threshold=0.9, training_seed=1336
)
def test_hybrid_visual_ppo(num_visual, training_seed):
env = SimpleEnvironment([BRAIN_NAME],
num_visual=num_visual,
num_vector=0,
action_sizes=(1, 1))
new_hyperparams = attr.evolve(
PPO_TORCH_CONFIG.hyperparameters,
batch_size=64,
buffer_size=1024,
learning_rate=1e-4,
)
config = attr.evolve(PPO_TORCH_CONFIG,
hyperparameters=new_hyperparams,
max_steps=8000)
check_environment_trains(env, {BRAIN_NAME: config},
training_seed=training_seed)
def test_subprocess_failing_step(num_envs):
def failing_step_env_factory(_worker_id, _config):
env = UnexpectedExceptionEnvironment(
["1D"], use_discrete=True, to_raise=CustomTestOnlyException
)
return env
env_manager = SubprocessEnvManager(failing_step_env_factory, RunOptions())
# Expect the exception raised to be routed back up to the top level.
with pytest.raises(CustomTestOnlyException):
check_environment_trains(
failing_step_env_factory(0, []),
{"1D": ppo_dummy_config()},
env_manager=env_manager,
success_threshold=None,
)
env_manager.close()
def test_simple_ghost_fails(action_sizes):
env = SimpleEnvironment(
[BRAIN_NAME + "?team=0", BRAIN_NAME + "?team=1"], action_sizes=action_sizes
)
# This config should fail because the ghosted policy is never swapped with a competent policy.
# Swap occurs after max step is reached.
self_play_settings = SelfPlaySettings(
play_against_latest_model_ratio=1.0, save_steps=2000, swap_steps=4000
)
config = attr.evolve(PPO_TORCH_CONFIG, self_play=self_play_settings, max_steps=2500)
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=None)
processed_rewards = [
default_reward_processor(rewards) for rewards in env.final_rewards.values()
]
success_threshold = 0.9
assert any(reward > success_threshold for reward in processed_rewards) and any(
reward < success_threshold for reward in processed_rewards
)
def test_var_len_obs_and_goal_ppo(num_vis, num_vector, num_var_len,
action_sizes, conditioning_type):
env = SimpleEnvironment(
[BRAIN_NAME],
action_sizes=action_sizes,
num_visual=num_vis,
num_vector=num_vector,
num_var_len=num_var_len,
step_size=0.2,
goal_indices=[0],
)
new_network = attr.evolve(POCA_TORCH_CONFIG.network_settings,
goal_conditioning_type=conditioning_type)
new_hyperparams = attr.evolve(PPO_TORCH_CONFIG.hyperparameters,
learning_rate=3.0e-4)
config = attr.evolve(PPO_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_network)
check_environment_trains(env, {BRAIN_NAME: config})
def test_hybrid_recurrent_ppo():
env = MemoryEnvironment([BRAIN_NAME], action_sizes=(1, 1), step_size=0.5)
new_network_settings = attr.evolve(
PPO_TORCH_CONFIG.network_settings,
memory=NetworkSettings.MemorySettings(memory_size=16),
)
new_hyperparams = attr.evolve(
PPO_TORCH_CONFIG.hyperparameters,
learning_rate=1.0e-3,
batch_size=64,
buffer_size=512,
)
config = attr.evolve(
PPO_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_network_settings,
max_steps=3000,
)
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.9)
def test_hybrid_recurrent_sac():
env = MemoryEnvironment([BRAIN_NAME], action_sizes=(1, 1), step_size=0.5)
new_networksettings = attr.evolve(
SAC_TORCH_CONFIG.network_settings,
memory=NetworkSettings.MemorySettings(memory_size=16, sequence_length=16),
)
new_hyperparams = attr.evolve(
SAC_TORCH_CONFIG.hyperparameters,
batch_size=256,
learning_rate=1e-3,
buffer_init_steps=1000,
steps_per_update=2,
)
config = attr.evolve(
SAC_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_networksettings,
max_steps=4000,
)
check_environment_trains(env, {BRAIN_NAME: config})
def test_subprocess_env_endtoend(num_envs):
def simple_env_factory(worker_id, config):
env = SimpleEnvironment(["1D"], use_discrete=True)
return env
env_manager = SubprocessEnvManager(simple_env_factory,
EngineConfig.default_config(), num_envs)
# Run PPO using env_manager
check_environment_trains(
simple_env_factory(0, []),
{"1D": ppo_dummy_config()},
env_manager=env_manager,
success_threshold=None,
)
# Note we can't check the env's rewards directly (since they're in separate processes) so we
# check the StatsReporter's debug stat writer's last reward.
assert isinstance(StatsReporter.writers[0], DebugWriter)
assert all(val > 0.7
for val in StatsReporter.writers[0].get_last_rewards().values())
env_manager.close()
def test_var_len_obs_and_goal_poca(num_vis, num_vector, num_var_len,
conditioning_type):
env = MultiAgentEnvironment(
[BRAIN_NAME],
action_sizes=(0, 1),
num_visual=num_vis,
num_vector=num_vector,
num_var_len=num_var_len,
step_size=0.2,
num_agents=2,
goal_indices=[0],
)
new_network = attr.evolve(POCA_TORCH_CONFIG.network_settings,
goal_conditioning_type=conditioning_type)
new_hyperparams = attr.evolve(POCA_TORCH_CONFIG.hyperparameters,
learning_rate=3.0e-4)
config = attr.evolve(
POCA_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_network,
max_steps=5000,
)
check_environment_trains(env, {BRAIN_NAME: config})
def test_recurrent_sac(use_discrete):
step_size = 0.2 if use_discrete else 0.5
env = MemoryEnvironment(
[BRAIN_NAME], use_discrete=use_discrete, step_size=step_size
)
new_networksettings = attr.evolve(
SAC_TORCH_CONFIG.network_settings,
memory=NetworkSettings.MemorySettings(memory_size=16, sequence_length=16),
)
new_hyperparams = attr.evolve(
SAC_TORCH_CONFIG.hyperparameters,
batch_size=256,
learning_rate=1e-3,
buffer_init_steps=1000,
steps_per_update=2,
)
config = attr.evolve(
SAC_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_networksettings,
max_steps=2000,
)
check_environment_trains(env, {BRAIN_NAME: config})
def test_visual_advanced_ppo(vis_encode_type, num_visual):
env = SimpleEnvironment(
[BRAIN_NAME],
use_discrete=True,
num_visual=num_visual,
num_vector=0,
step_size=0.5,
vis_obs_size=(5, 5, 5) if vis_encode_type == "match3" else (36, 36, 3),
)
new_networksettings = attr.evolve(
SAC_TORCH_CONFIG.network_settings,
vis_encode_type=EncoderType(vis_encode_type))
new_hyperparams = attr.evolve(PPO_TORCH_CONFIG.hyperparameters,
learning_rate=3.0e-4)
config = attr.evolve(
PPO_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_networksettings,
max_steps=900,
summary_freq=100,
)
# The number of steps is pretty small for these encoders
check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.5)
def test_recurrent_sac(action_sizes):
step_size = 0.2 if action_sizes == (0, 1) else 0.5
env = MemoryEnvironment(
[BRAIN_NAME], action_sizes=action_sizes, step_size=step_size
)
new_networksettings = attr.evolve(
SAC_TORCH_CONFIG.network_settings,
memory=NetworkSettings.MemorySettings(memory_size=16, sequence_length=16),
)
new_hyperparams = attr.evolve(
SAC_TORCH_CONFIG.hyperparameters,
batch_size=256,
learning_rate=3e-4,
buffer_init_steps=1000,
steps_per_update=2,
)
config = attr.evolve(
SAC_TORCH_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_networksettings,
max_steps=4000,
)
check_environment_trains(env, {BRAIN_NAME: config}, training_seed=1337)
def test_simple_ppo(use_discrete):
env = SimpleEnvironment([BRAIN_NAME], use_discrete=use_discrete)
config = attr.evolve(PPO_TORCH_CONFIG)
check_environment_trains(env, {BRAIN_NAME: config})
def test_simple_ppo(action_sizes):
env = SimpleEnvironment([BRAIN_NAME], action_sizes=action_sizes)
config = attr.evolve(PPO_TORCH_CONFIG)
check_environment_trains(env, {BRAIN_NAME: config})
