def setUpClass(cls):
cls.generative_model = GenerativeModel(
ex.priors.dm_prior, ex.simulators.dm_batch_simulator)
cls.network = InvariantNetwork()
params, sim_data = cls.generative_model(64, 128)
cls.network(sim_data) # one prediction to initialize layers
cls.optimizer = tf.keras.optimizers.Adam()
def test_no_generative_model(self):
D = 6
J = 6
bf_meta = build_meta_dict({'n_params': D, 'n_models': J}, DEFAULT_SETTING_INVARIANT_BAYES_FLOW)
amortizer = ex.amortizers.InvariantBayesFlow(bf_meta)
trainer = MetaTrainer(amortizer,
loss=kl_latent_space,
learning_rate=.0003
)
generative_model = GenerativeModel(
model_prior,
[TPrior(D // 2, 1.0, 5.0)] * J,
[MultivariateTSimulator(df) for df in np.arange(1, J + 1, 1)]
)
model_indices, params, sim_data = generative_model(64, 128)
_losses = trainer.train_offline(2, 16, model_indices, params, sim_data)
with self.assertRaises(OperationNotSupportedError):
_losses = trainer.train_online(epochs=2, iterations_per_epoch=20, batch_size=32, n_obs=110)
with self.assertRaises(OperationNotSupportedError):
_losses = trainer.train_rounds(epochs=1, rounds=5, sim_per_round=100, batch_size=32, n_obs=110)
with self.assertRaises(OperationNotSupportedError):
_losses = trainer.simulate_and_train_offline(n_sim=100, epochs=2, batch_size=16, n_obs=110)
def test_no_generative_model(self):
summary_net = SequenceNet()
evidential_meta = {
'n_models': 3,
'out_activation': 'softplus',
'n_dense': 3,
'dense_args': {'kernel_initializer': 'glorot_uniform', 'activation': 'relu', 'units': 128}
}
evidential_net = EvidentialNetwork(evidential_meta)
amortizer = MultiModelAmortizer(evidential_net, summary_net)
trainer = ModelComparisonTrainer(amortizer, n_models=3)
generative_model = GenerativeModel(
model_prior,
[ex.priors.model1_params_prior, ex.priors.model2_params_prior, ex.priors.model3_params_prior],
[ex.simulators.forward_model1, ex.simulators.forward_model2, ex.simulators.forward_model3]
)
model_indices, params, sim_data = generative_model(64, 128)
_losses = trainer.train_offline(2, 16, model_indices, sim_data)
_losses = trainer.train_offline(2, 16, np.random.randint(0, 3, (64,)), sim_data) # expect message
with self.assertRaises(OperationNotSupportedError):
_losses = trainer.train_online(epochs=2, iterations_per_epoch=20, batch_size=32, n_obs=100)
with self.assertRaises(OperationNotSupportedError):
_losses = trainer.train_rounds(epochs=2, rounds=2, sim_per_round=50, batch_size=32, n_obs=110)
def setUpClass(cls):
summary_net = InvariantNetwork()
inference_net = InvertibleNetwork({'n_params': 5})
amortizer = SingleModelAmortizer(inference_net, summary_net)
generative_model = GenerativeModel(ex.priors.dm_prior, ex.simulators.dm_batch_simulator)
trainer = ParameterEstimationTrainer(amortizer, generative_model)
cls.trainer = trainer
def setUpClass(cls):
summary_net = SequenceNet()
evidential_meta = {
'n_models': 3,
'out_activation': 'softplus',
'n_dense': 3,
'dense_args': {
'kernel_initializer': 'glorot_uniform',
'activation': 'relu',
'units': 128
}
}
evidential_net = EvidentialNetwork(evidential_meta)
amortizer = MultiModelAmortizer(evidential_net, summary_net)
generative_model = GenerativeModel(ex.priors.model_prior, [
ex.priors.model1_params_prior, ex.priors.model2_params_prior,
ex.priors.model3_params_prior
], [
ex.simulators.forward_model1, ex.simulators.forward_model2,
ex.simulators.forward_model3
])
trainer = ModelComparisonTrainer(amortizer, generative_model)
cls.trainer = trainer
def test_meta_generative_model(self):
M = 10
D = 100
prior = TPrior(D // 2, mu_scale=1.0, scale_scale=5.0)
priors = [prior] * M
simulators = [
MultivariateTSimulator(df) for df in np.arange(1, 101, M)
]
generative_model = GenerativeModel(model_prior, priors, simulators)
_model_indices, _params, _sim_data = generative_model(n_sim=N_SIM,
n_obs=N_OBS)
def test_structural_param_transform_tuple_to_numpy(self):
D = 5
def param_transform_mvn(theta):
means, cov = theta
var = np.diagonal(cov, axis1=1, axis2=2)
return np.concatenate([means, var], axis=1)
prior = GaussianMeanCovPrior(D=D, a0=10, b0=1, m0=0, beta0=1)
simulator = GaussianMeanCovSimulator()
generative_model = GenerativeModel(prior,
simulator,
param_transform=param_transform_mvn)
def setUpClass(cls):
D = 10
J = 10
bf_meta = build_meta_dict({'n_params': D, 'n_models': J}, DEFAULT_SETTING_INVARIANT_BAYES_FLOW)
amortizer = ex.amortizers.InvariantBayesFlow(bf_meta)
generative_model = GenerativeModel(
model_prior,
[TPrior(D // 2, 1.0, 5.0)] * J,
[MultivariateTSimulator(df) for df in np.arange(1, J + 1, 1)]
)
trainer = MetaTrainer(amortizer,
generative_model,
loss=kl_latent_space,
learning_rate=.0003
)
cls.trainer = trainer
def test_no_generative_model(self):
summary_net = InvariantNetwork()
inference_net = InvertibleNetwork({'n_params': 5})
amortizer = SingleModelAmortizer(inference_net, summary_net)
generative_model = GenerativeModel(ex.priors.dm_prior, ex.simulators.dm_batch_simulator)
trainer = ParameterEstimationTrainer(amortizer)
params, sim_data = generative_model(64, 128)
_losses = trainer.train_offline(2, 16, params, sim_data)
with self.assertRaises(OperationNotSupportedError):
_losses = trainer.train_online(epochs=2, iterations_per_epoch=20, batch_size=32, n_obs=100)
with self.assertRaises(OperationNotSupportedError):
_losses = trainer.train_experience_replay(epochs=2, batch_size=32, iterations_per_epoch=20,
capacity=100, n_obs=np.random.randint(106, 301))
with self.assertRaises(OperationNotSupportedError):
_losses = trainer.train_rounds(epochs=2, rounds=2, sim_per_round=20, batch_size=32, n_obs=150)
with self.assertRaises(OperationNotSupportedError):
_losses = trainer.simulate_and_train_offline(n_sim=100, epochs=2, batch_size=32, n_obs=100)
def setUpClass(cls):
cls.generative_model = GenerativeModel(
ex.priors.dm_prior, ex.simulators.dm_batch_simulator)
cls.invariant_network = InvariantNetwork()
cls.network = InvertibleNetwork({'n_params': 5})
cls.optimizer = tf.keras.optimizers.Adam()
def test_simple_generative_model(self):
generative_model = GenerativeModel(ex.priors.dm_prior,
ex.simulators.dm_batch_simulator)
_params, _sim_data = generative_model(n_sim=N_SIM, n_obs=N_OBS)
