def test_ppo_bc(self):
# Train bc model
model_dir = self.temp_model_dir
params_to_override = {
"layouts": ['inverse_marshmallow_experiment'],
"data_path": None,
"epochs": 10
}
bc_params = get_bc_params(**params_to_override)
train_bc_model(model_dir, bc_params)
# Train rllib model
config_updates = {
"results_dir": self.temp_results_dir,
"bc_schedule": [(0.0, 0.0), (8e3, 1.0)],
"num_training_iters": 20,
"bc_model_dir": model_dir,
"evaluation_interval": 5,
"verbose": False
}
results = ex.run(config_updates=config_updates,
options={
'--loglevel': 'ERROR'
}).result
# Sanity check
self.assertGreaterEqual(results['average_total_reward'],
self.min_performance)
if self.compute_pickle:
self.expected['test_ppo_bc'] = results
# Reproducibility test
if self.strict:
self.assertDictEqual(results, self.expected['test_ppo_bc'])
def test_training(self, lstm=False):
for name, bc_params in self.bc_params_to_test.items():
bc_params["use_lstm"] = lstm
model = train_bc_model(self.model_dir, bc_params)
key_name = f'test_training_{name}'
if lstm: key_name = "lstm_" + key_name
self._compare_iterable_with_expected(key_name, self._model_forward(model, bc_params=bc_params, lstm=lstm))
def test_training(self):
model = train_bc_model(self.model_dir, self.bc_params)
if self.compute_pickle:
self.expected['test_training'] = model(self.dummy_input)
if self.strict:
self.assertTrue(
np.allclose(model(self.dummy_input),
self.expected["test_training"]))
def test_agent_metrics_evaluation(self, lstm=False):
for name, bc_params in self.bc_params_to_test.items():
bc_params["use_lstm"] = lstm
bc_params["training_params"]["epochs"] = 1 if lstm else 20
model = train_bc_model(self.model_dir, bc_params)
metrics = evaluate_bc_model_metrics(model, bc_params)
key_name = f'test_agent_metrics_evaluation_{name}'
if lstm: key_name = "lstm_" + key_name
self._compare_iterable_with_expected(key_name, metrics)
def test_agent_evaluation(self, lstm=False):
for name, bc_params in self.bc_params_to_test.items():
bc_params["use_lstm"] = lstm
bc_params["training_params"]["epochs"] = 1 if lstm else 20
model = train_bc_model(self.model_dir, bc_params)
results = evaluate_bc_model(model, bc_params)
key_name = f'test_agent_evaluation_{name}'
if lstm: key_name = "lstm_" + key_name
# Sanity Check
self.assertGreaterEqual(results, self.min_performance)
self._compare_iterable_with_expected(key_name, results)
def test_agent_evaluation(self):
self.bc_params["training_params"]["epochs"] = 20
model = train_bc_model(self.model_dir, self.bc_params)
results = evaluate_bc_model(model, self.bc_params)
# Sanity Check
self.assertGreaterEqual(results, self.min_performance)
if self.compute_pickle:
self.expected['test_agent_evaluation'] = results
if self.strict:
self.assertAlmostEqual(results,
self.expected['test_agent_evaluation'])
def test_behavior_cloning_policy(self, lstm=False):
for name, bc_params in self.bc_params_to_test.items():
bc_params["use_lstm"] = lstm
model = train_bc_model(self.model_dir, bc_params)
policy = BehaviorCloningPolicy.from_model(model, bc_params, stochastic=False)
key_name = f'test_behavior_cloning_policy_actions_{name}'
if lstm: key_name = "lstm_" + key_name
observations = get_dummy_input(bc_params)
actions_output = policy.compute_actions(observations, self.initial_states)
self._compare_iterable_with_expected(key_name, actions_output)
if bc_params["predict_orders"]:
key_name = f'test_behavior_cloning_policy_orders_{name}'
if lstm: key_name = "lstm_" + key_name
orders_output = policy.predict_orders(observations, self.initial_states)
self._compare_iterable_with_expected(key_name, orders_output)
else:
self.assertRaises(AssertionError, policy.predict_orders, observations, self.initial_states)
