def test_load_save_policy(tmp_path):
path1 = os.path.join(tmp_path, "runid1")
path2 = os.path.join(tmp_path, "runid2")
trainer_params = TrainerSettings()
policy = create_policy_mock(trainer_params)
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)
assert len(os.listdir(tmp_path)) > 0
# Try load from this path
model_saver2 = TorchModelSaver(trainer_params, path1, load=True)
policy2 = create_policy_mock(trainer_params)
model_saver2.register(policy2)
model_saver2.initialize_or_load(policy2)
_compare_two_policies(policy, policy2)
assert policy2.get_current_step() == 2000
# Try initialize from path 1
trainer_params.init_path = path1
model_saver3 = TorchModelSaver(trainer_params, path2)
policy3 = create_policy_mock(trainer_params)
model_saver3.register(policy3)
model_saver3.initialize_or_load(policy3)
_compare_two_policies(policy2, policy3)
# Assert that the steps are 0.
assert policy3.get_current_step() == 0
def test_load_save_optimizer(tmp_path, optimizer):
OptimizerClass, HyperparametersClass = optimizer
trainer_settings = TrainerSettings()
trainer_settings.hyperparameters = HyperparametersClass()
policy = create_policy_mock(trainer_settings, use_discrete=False)
optimizer = OptimizerClass(policy, trainer_settings)
# save at path 1
path1 = os.path.join(tmp_path, "runid1")
model_saver = TorchModelSaver(trainer_settings, path1)
model_saver.register(policy)
model_saver.register(optimizer)
model_saver.initialize_or_load()
policy.set_step(2000)
model_saver.save_checkpoint("MockBrain", 2000)
# create a new optimizer and policy
policy2 = create_policy_mock(trainer_settings, use_discrete=False)
optimizer2 = OptimizerClass(policy2, trainer_settings)
# load weights
model_saver2 = TorchModelSaver(trainer_settings, path1, load=True)
model_saver2.register(policy2)
model_saver2.register(optimizer2)
model_saver2.initialize_or_load() # This is to load the optimizers
# Compare the two optimizers
_compare_two_optimizers(optimizer, optimizer2)
def test_reward_provider_save(tmp_path, optimizer):
OptimizerClass, HyperparametersClass = optimizer
trainer_settings = TrainerSettings()
trainer_settings.hyperparameters = HyperparametersClass()
trainer_settings.reward_signals = {
RewardSignalType.CURIOSITY: CuriositySettings(),
RewardSignalType.GAIL: GAILSettings(demo_path=DEMO_PATH),
RewardSignalType.RND: RNDSettings(),
}
policy = create_policy_mock(trainer_settings, use_discrete=False)
optimizer = OptimizerClass(policy, trainer_settings)
# save at path 1
path1 = os.path.join(tmp_path, "runid1")
model_saver = TorchModelSaver(trainer_settings, path1)
model_saver.register(policy)
model_saver.register(optimizer)
model_saver.initialize_or_load()
policy.set_step(2000)
model_saver.save_checkpoint("MockBrain", 2000)
# create a new optimizer and policy
optimizer2 = OptimizerClass(policy, trainer_settings)
policy2 = create_policy_mock(trainer_settings, use_discrete=False)
# load weights
model_saver2 = TorchModelSaver(trainer_settings, path1, load=True)
model_saver2.register(policy2)
model_saver2.register(optimizer2)
model_saver2.initialize_or_load() # This is to load the optimizers
# assert the models have the same weights
module_dict_1 = optimizer.get_modules()
module_dict_2 = optimizer2.get_modules()
assert "Module:GAIL" in module_dict_1
assert "Module:GAIL" in module_dict_2
assert "Module:Curiosity" in module_dict_1
assert "Module:Curiosity" in module_dict_2
assert "Module:RND-pred" in module_dict_1
assert "Module:RND-pred" in module_dict_2
assert "Module:RND-target" in module_dict_1
assert "Module:RND-target" in module_dict_2
for name, module1 in module_dict_1.items():
assert name in module_dict_2
module2 = module_dict_2[name]
if hasattr(module1, "parameters"):
for param1, param2 in zip(module1.parameters(),
module2.parameters()):
assert param1.data.ne(param2.data).sum() == 0
# Run some rewards
data = create_agent_buffer(policy.behavior_spec, 1)
for reward_name in optimizer.reward_signals.keys():
rp_1 = optimizer.reward_signals[reward_name]
rp_2 = optimizer2.reward_signals[reward_name]
assert np.array_equal(rp_1.evaluate(data), rp_2.evaluate(data))
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_load_different_reward_provider(caplog, tmp_path, optimizer):
OptimizerClass, HyperparametersClass = optimizer
trainer_settings = TrainerSettings()
trainer_settings.hyperparameters = HyperparametersClass()
trainer_settings.reward_signals = {
RewardSignalType.CURIOSITY: CuriositySettings(),
RewardSignalType.RND: RNDSettings(),
}
policy = create_policy_mock(trainer_settings, use_discrete=False)
optimizer = OptimizerClass(policy, trainer_settings)
# save at path 1
path1 = os.path.join(tmp_path, "runid1")
model_saver = TorchModelSaver(trainer_settings, path1)
model_saver.register(policy)
model_saver.register(optimizer)
model_saver.initialize_or_load()
assert len(optimizer.critic.value_heads.stream_names) == 2
policy.set_step(2000)
model_saver.save_checkpoint("MockBrain", 2000)
trainer_settings2 = TrainerSettings()
trainer_settings2.hyperparameters = HyperparametersClass()
trainer_settings2.reward_signals = {
RewardSignalType.GAIL: GAILSettings(demo_path=DEMO_PATH)
}
# create a new optimizer and policy
policy2 = create_policy_mock(trainer_settings2, use_discrete=False)
optimizer2 = OptimizerClass(policy2, trainer_settings2)
# load weights
model_saver2 = TorchModelSaver(trainer_settings2, path1, load=True)
model_saver2.register(policy2)
model_saver2.register(optimizer2)
assert len(optimizer2.critic.value_heads.stream_names) == 1
model_saver2.initialize_or_load() # This is to load the optimizers
messages = [
rec.message for rec in caplog.records if rec.levelno == WARNING
]
assert len(messages) > 0
def test_checkpoint_conversion(tmpdir, rnn, visual, discrete):
dummy_config = TrainerSettings()
model_path = os.path.join(tmpdir, "Mock_Brain")
policy = create_policy_mock(
dummy_config, use_rnn=rnn, use_discrete=discrete, use_visual=visual
)
trainer_params = TrainerSettings()
model_saver = TorchModelSaver(trainer_params, model_path)
model_saver.register(policy)
model_saver.save_checkpoint("Mock_Brain", 100)
assert os.path.isfile(model_path + "/Mock_Brain-100.onnx")
def test_register(tmp_path):
trainer_params = TrainerSettings()
model_saver = TorchModelSaver(trainer_params, tmp_path)
opt = mock.Mock(spec=TorchPPOOptimizer)
opt.get_modules = mock.Mock(return_value={})
model_saver.register(opt)
assert model_saver.policy is None
trainer_params = TrainerSettings()
policy = create_policy_mock(trainer_params)
opt.get_modules = mock.Mock(return_value={})
model_saver.register(policy)
assert model_saver.policy is not None
