def from_simulator(
cls,
simulator_func: Callable,
theta_0: Union[Number, np.array],
theta_1: Union[Number, np.array],
n_samples_per_theta: int,
shuffle: bool = True,
include_log_probs: bool = False,
):
theta_0, theta_1 = map(ensure_array, [theta_0, theta_1])
assert theta_0.ndim == 1 == theta_1.ndim
sim0 = build_simulator(simulator_func, theta_0)
sim1 = build_simulator(simulator_func, theta_1)
x0 = sim0.sample(n_samples_per_theta).numpy()
x1 = sim1.sample(n_samples_per_theta).numpy()
y0 = np.zeros(len(x0))
y1 = np.ones_like(y0)
x = np.concatenate([x0, x1], axis=0).astype(np.float32)
y = np.concatenate([y0, y1], axis=0).astype(np.float32)
if include_log_probs:
log_prob_0 = sim0.log_prob(x).astype(np.float32)
log_prob_1 = sim1.log_prob(x).astype(np.float32)
else:
log_prob_0 = None
log_prob_1 = None
return cls(x=x,
y=y,
theta_0=theta_0,
theta_1=theta_1,
log_prob_0=log_prob_0,
log_prob_1=log_prob_1,
shuffle=shuffle)
def _simulate(theta_1, x0_i):
sim1 = build_simulator(simulator_func, theta_1)
x1_i = sim1.sample(n_samples_per_theta).numpy()
x1_i = ensure_2d(x1_i)
x0_i = ensure_2d(x0_i)
x1.append(x1_i)
if include_log_probs:
log_prob_0_x0.append(sim0.log_prob(x0_i))
log_prob_0_x1.append(sim0.log_prob(x1_i))
log_prob_1_x0.append(sim1.log_prob(x0_i))
log_prob_1_x1.append(sim1.log_prob(x1_i))
def exact_param_scan(simulator_func: Callable,
X_true: np.ndarray,
param_grid: ParamGrid,
theta_0: float,
to_meshgrid_shape: bool = True) -> Sequence[np.ndarray]:
p_0 = build_simulator(simulator_func=simulator_func, theta=theta_0)
log_prob_0 = p_0.log_prob(X_true)
@tf.function
def exact_nllr(theta):
p_theta = build_simulator(simulator_func=simulator_func, theta=theta)
return -tf.keras.backend.sum(p_theta.log_prob(X_true) - log_prob_0)
nllr = np.array(
[exact_nllr(theta.squeeze()).numpy() for theta in param_grid])
mle = param_grid[np.argmin(nllr)]
if to_meshgrid_shape:
nllr = _to_meshgrid_shape(nllr, param_grid)
return nllr, mle
def from_simulator(cls,
simulator_func: Callable,
theta_0: Union[Number, np.ndarray],
theta_1_iterator: ParamIterator,
n_samples_per_theta: int,
shuffle: bool = True,
include_log_probs: bool = False):
theta_0 = ensure_array(theta_0)
assert theta_0.ndim == 1
sim0 = build_simulator(simulator_func, theta_0)
x0 = sim0.sample(n_samples_per_theta * len(theta_1_iterator)).numpy()
x0 = ensure_2d(x0)
y0 = np.zeros(len(x0))
y1 = np.ones_like(y0)
x1 = []
theta_1s = np.zeros((len(x0), len(theta_0)))
if include_log_probs:
log_prob_0_x0 = []
log_prob_0_x1 = []
log_prob_1_x0 = []
log_prob_1_x1 = []
def _simulate(theta_1, x0_i):
sim1 = build_simulator(simulator_func, theta_1)
x1_i = sim1.sample(n_samples_per_theta).numpy()
x1_i = ensure_2d(x1_i)
x0_i = ensure_2d(x0_i)
x1.append(x1_i)
if include_log_probs:
log_prob_0_x0.append(sim0.log_prob(x0_i))
log_prob_0_x1.append(sim0.log_prob(x1_i))
log_prob_1_x0.append(sim1.log_prob(x0_i))
log_prob_1_x1.append(sim1.log_prob(x1_i))
for i, theta_1 in enumerate(theta_1_iterator):
start = i * n_samples_per_theta
stop = (i + 1) * n_samples_per_theta
theta_1s[start:stop, :] = theta_1
_simulate(theta_1, x0[start:stop, :])
if include_log_probs:
log_prob_0 = np.concatenate(log_prob_0_x0 + log_prob_0_x1,
axis=0).astype(np.float32)
log_prob_1 = np.concatenate(log_prob_1_x0 + log_prob_1_x1,
axis=0).astype(np.float32)
else:
log_prob_0 = None
log_prob_1 = None
x1 = ensure_2d(np.concatenate(x1, axis=0))
x = np.concatenate([x0, x1], axis=0).astype(np.float32)
y = np.concatenate([y0, y1], axis=0).astype(np.float32)
theta_1s = concat_repeat(theta_1s, 2, axis=0).astype(np.float32)
return cls(x=x,
y=y,
theta_0=theta_0,
theta_1s=theta_1s,
log_prob_0=log_prob_0,
log_prob_1=log_prob_1,
shuffle=shuffle)
def exact_nllr(theta):
p_theta = build_simulator(simulator_func=simulator_func, theta=theta)
return -tf.keras.backend.sum(p_theta.log_prob(X_true) - log_prob_0)