def find_best_model_initialization(self, num_kernel_samples: int) -> None:
"""
Test `num_kernel_samples` models with sampled kernel parameters. The model's kernel
parameters are then set to the sample achieving maximal likelihood.
:param num_kernel_samples: Number of randomly sampled kernels to evaluate.
"""
@tf.function
def evaluate_loss_of_model_parameters() -> tf.Tensor:
randomize_hyperparameters(self.model)
return self.model.training_loss()
squeeze_hyperparameters(self.model)
current_best_parameters = read_values(self.model)
min_loss = self.model.training_loss()
for _ in tf.range(num_kernel_samples):
try:
train_loss = evaluate_loss_of_model_parameters()
except tf.errors.InvalidArgumentError: # allow badly specified kernel params
train_loss = 1e100
if train_loss < min_loss: # only keep best kernel params
min_loss = train_loss
current_best_parameters = read_values(self.model)
multiple_assign(self.model, current_best_parameters)
def test_dict_utilities(model):
"""
Test both `parameter_dict()` and `read_values()`
"""
class SubModule(tf.Module):
def __init__(self):
self.parameter = gpflow.Parameter(1.0)
self.variable = tf.Variable(1.0)
class Module(tf.Module):
def __init__(self):
self.submodule = SubModule()
self.top_parameter = gpflow.Parameter(3.0)
m = Module()
params = gpflow.utilities.parameter_dict(m)
# {
# ".submodule.parameter": <parameter object>,
# ".submodule.variable": <variable object>
# }
assert list(params.keys()) == [
".submodule.parameter",
".submodule.variable",
".top_parameter",
]
assert list(params.values()) == [
m.submodule.parameter,
m.submodule.variable,
m.top_parameter,
]
for k, v in read_values(m).items():
assert params[k].numpy() == v