def test_visual_advanced_sac(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_TF_CONFIG.network_settings,
vis_encode_type=EncoderType(vis_encode_type))
new_hyperparams = attr.evolve(
SAC_TF_CONFIG.hyperparameters,
batch_size=16,
learning_rate=3e-4,
buffer_init_steps=0,
)
config = attr.evolve(
SAC_TF_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_networksettings,
max_steps=100,
framework=FrameworkType.TENSORFLOW,
)
# The number of steps is pretty small for these encoders
_check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.5)
def test_load_policy_different_hidden_units(tmp_path, vis_encode_type):
path1 = os.path.join(tmp_path, "runid1")
trainer_params = TrainerSettings()
trainer_params.network_settings = NetworkSettings(
hidden_units=12, vis_encode_type=EncoderType(vis_encode_type))
policy = create_policy_mock(trainer_params, use_visual=True)
conv_params = [
mod for mod in policy.actor.parameters() if len(mod.shape) > 2
]
model_saver = TorchModelSaver(trainer_params, path1)
model_saver.register(policy)
model_saver.initialize_or_load(policy)
policy.set_step(2000)
mock_brain_name = "MockBrain"
model_saver.save_checkpoint(mock_brain_name, 2000)
# Try load from this path
trainer_params2 = TrainerSettings()
trainer_params2.network_settings = NetworkSettings(
hidden_units=10, vis_encode_type=EncoderType(vis_encode_type))
model_saver2 = TorchModelSaver(trainer_params2, path1, load=True)
policy2 = create_policy_mock(trainer_params2, use_visual=True)
conv_params2 = [
mod for mod in policy2.actor.parameters() if len(mod.shape) > 2
]
# asserts convolutions have different parameters before load
for conv1, conv2 in zip(conv_params, conv_params2):
assert not torch.equal(conv1, conv2)
# asserts layers still have different dimensions
for mod1, mod2 in zip(policy.actor.parameters(),
policy2.actor.parameters()):
if mod1.shape[0] == 12:
assert mod2.shape[0] == 10
model_saver2.register(policy2)
model_saver2.initialize_or_load(policy2)
# asserts convolutions have same parameters after load
for conv1, conv2 in zip(conv_params, conv_params2):
assert torch.equal(conv1, conv2)
# asserts layers still have different dimensions
for mod1, mod2 in zip(policy.actor.parameters(),
policy2.actor.parameters()):
if mod1.shape[0] == 12:
assert mod2.shape[0] == 10
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=(36, 36, 3),
)
new_networksettings = attr.evolve(
SAC_CONFIG.network_settings, vis_encode_type=EncoderType(vis_encode_type)
)
new_hyperparams = attr.evolve(PPO_CONFIG.hyperparameters, learning_rate=3.0e-4)
config = attr.evolve(
PPO_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_networksettings,
max_steps=500,
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_visual_advanced_ppo(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_TF_CONFIG.network_settings,
vis_encode_type=EncoderType(vis_encode_type))
new_hyperparams = attr.evolve(PPO_TF_CONFIG.hyperparameters,
learning_rate=3.0e-4)
config = attr.evolve(
PPO_TF_CONFIG,
hyperparameters=new_hyperparams,
network_settings=new_networksettings,
max_steps=400,
summary_freq=100,
framework=FrameworkType.TENSORFLOW,
)
# The number of steps is pretty small for these encoders
_check_environment_trains(env, {BRAIN_NAME: config}, success_threshold=0.5)
