def create_rnn_params_symmetric(input_size, state_size, output_size,
g, random_key):
input_factor = 1.0 / np.sqrt(input_size)
hidden_scale = 1.0
hidden_factor = g / np.sqrt(state_size+1)
predict_factor = 1.0 / np.sqrt(state_size+1)
normal_mat = random.normal(random_key,(state_size, state_size)) * hidden_factor
return {'hidden unit': np.array(random.normal(random_key,(1, state_size)) * hidden_scale),
'change': np.concatenate((random.normal(random_key,(input_size, state_size)) * input_factor,
np.tril(normal_mat)+np.transpose(np.tril(normal_mat,-1)),
random.normal(random_key,(1,state_size))*hidden_factor),
axis=0),#hidden weights
'predict': random.normal(random_key,(state_size+1, output_size)) * predict_factor}#readout weights
def sdcorr_params_to_sds_and_corr_jax(sdcorr_params):
dim = number_of_triangular_elements_to_dimension_jax(len(sdcorr_params))
sds = jnp.array(sdcorr_params[:dim])
corr = jnp.eye(dim)
corr = index_update(corr, index[jnp.tril_indices(dim, k=-1)], sdcorr_params[dim:])
corr += jnp.tril(corr, k=-1).T
return sds, corr
def crossover(parent_1, parent_2, offspring_size):
all_offspring = []
for o in range(offspring_size):
lower_1 = np.tril(parent_1)
upper_2 = np.triu(parent_2)
offspring = lower_1 + upper_2
all_offspring.append(offspring)
def custom_assert(tst, result_jax, result_tf, *, tol, err_msg, **_):
# cholesky_p returns garbage in the strictly upper triangular part of the
# result, so we can safely ignore that part.
tst.assertAllClose(jnp.tril(result_jax),
result_tf,
atol=tol,
err_msg=err_msg)
def test_correlated_mvn(regularize):
# This requires dense mass matrix estimation.
D = 5
warmup_steps, num_samples = 5000, 8000
true_mean = 0.0
a = jnp.tril(
0.5 * jnp.fliplr(jnp.eye(D))
+ 0.1 * jnp.exp(random.normal(random.PRNGKey(0), shape=(D, D)))
)
true_cov = jnp.dot(a, a.T)
true_prec = jnp.linalg.inv(true_cov)
def potential_fn(z):
return 0.5 * jnp.dot(z.T, jnp.dot(true_prec, z))
init_params = jnp.zeros(D)
kernel = NUTS(
potential_fn=potential_fn, dense_mass=True, regularize_mass_matrix=regularize
)
mcmc = MCMC(kernel, warmup_steps, num_samples)
mcmc.run(random.PRNGKey(0), init_params=init_params)
samples = mcmc.get_samples()
assert_allclose(jnp.mean(samples), true_mean, atol=0.02)
assert np.sum(np.abs(np.cov(samples.T) - true_cov)) / D ** 2 < 0.02
def custom_assert(tst, result_jax, result_tf, *, args, tol, err_msg):
operand, = args
lu, pivots, perm = result_tf
batch_dims = operand.shape[:-2]
m, n = operand.shape[-2], operand.shape[-1]
def _make_permutation_matrix(perm):
result = []
for idx in itertools.product(*map(range, operand.shape[:-1])):
result += [0 if c != perm[idx] else 1 for c in range(m)]
result = np.reshape(np.array(result, dtype=dtype),
[*batch_dims, m, m])
return result
k = min(m, n)
l = jnp.tril(lu, -1)[..., :, :k] + jnp.eye(m, k, dtype=dtype)
u = jnp.triu(lu)[..., :k, :]
p_mat = _make_permutation_matrix(perm)
tst.assertArraysEqual(
lax.linalg.lu_pivots_to_permutation(pivots, m), perm)
tst.assertAllClose(jnp.matmul(p_mat, operand),
jnp.matmul(l, u),
atol=tol,
rtol=tol,
err_msg=err_msg)
def setup(self):
config = self.config
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.embed_dim // self.num_heads
if self.head_dim * self.num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and "
f"`num_heads`: {self.num_heads}).")
self.attn_dropout = nn.Dropout(config.attention_dropout)
self.resid_dropout = nn.Dropout(config.resid_dropout)
dense = partial(
nn.Dense,
self.embed_dim,
dtype=self.dtype,
kernel_init=jax.nn.initializers.normal(
self.config.initializer_range),
)
self.q_proj, self.k_proj, self.v_proj = dense(use_bias=False), dense(
use_bias=False), dense(use_bias=False)
self.out_proj = dense()
self.causal_mask = make_causal_mask(jnp.ones(
(1, config.max_position_embeddings), dtype="bool"),
dtype="bool")
if self.attention_type == "local":
self.causal_mask = self.causal_mask ^ jnp.tril(
self.causal_mask, -config.window_size)
def __call__(self, x):
tril = jnp.tril(x)
lower_triangular = jnp.all(jnp.reshape(tril == x, x.shape[:-2] + (-1,)), axis=-1)
positive_diagonal = jnp.all(jnp.diagonal(x, axis1=-2, axis2=-1) > 0, axis=-1)
x_norm = jnp.linalg.norm(x, axis=-1)
unit_norm_row = jnp.all((x_norm <= 1) & (x_norm > 1 - 1e-6), axis=-1)
return lower_triangular & positive_diagonal & unit_norm_row
def gen_values_within_bounds(constraint, size, key=random.PRNGKey(11)):
eps = 1e-6
if isinstance(constraint, constraints._Boolean):
return random.bernoulli(key, shape=size)
elif isinstance(constraint, constraints._GreaterThan):
return np.exp(random.normal(key, size)) + constraint.lower_bound + eps
elif isinstance(constraint, constraints._IntegerInterval):
lower_bound = np.broadcast_to(constraint.lower_bound, size)
upper_bound = np.broadcast_to(constraint.upper_bound, size)
return random.randint(key, size, lower_bound, upper_bound + 1)
elif isinstance(constraint, constraints._IntegerGreaterThan):
return constraint.lower_bound + poisson(key, 5, shape=size)
elif isinstance(constraint, constraints._Interval):
lower_bound = np.broadcast_to(constraint.lower_bound, size)
upper_bound = np.broadcast_to(constraint.upper_bound, size)
return random.uniform(key, size, minval=lower_bound, maxval=upper_bound)
elif isinstance(constraint, constraints._Real):
return random.normal(key, size)
elif isinstance(constraint, constraints._Simplex):
return osp.dirichlet.rvs(alpha=np.ones((size[-1],)), size=size[:-1])
elif isinstance(constraint, constraints._Multinomial):
n = size[-1]
return multinomial(key, p=np.ones((n,)) / n, n=constraint.upper_bound, shape=size[:-1])
elif isinstance(constraint, constraints._CorrCholesky):
return signed_stick_breaking_tril(
random.uniform(key, size[:-2] + (size[-1] * (size[-1] - 1) // 2,),
minval=-1, maxval=1))
elif isinstance(constraint, constraints._LowerCholesky):
return np.tril(random.uniform(key, size))
elif isinstance(constraint, constraints._PositiveDefinite):
x = random.normal(key, size)
return np.matmul(x, np.swapaxes(x, -2, -1))
else:
raise NotImplementedError('{} not implemented.'.format(constraint))
def __call__(self, x):
tril = np.tril(x)
lower_triangular = np.all(np.reshape(tril == x, x.shape[:-2] + (-1, )),
axis=-1)
positive_diagonal = np.all(np.diagonal(x, axis1=-2, axis2=-1) > 0,
axis=-1)
return lower_triangular & positive_diagonal
def testTriangularSolveGradPrecision(self):
rng = jtu.rand_default()
a = np.tril(rng((3, 3), onp.float32))
b = rng((1, 3), onp.float32)
jtu.assert_dot_precision(lax.Precision.HIGHEST,
partial(jvp, lax_linalg.triangular_solve),
(a, b), (a, b))
def _band_part(input, num_lower, num_upper, name=None): # pylint: disable=redefined-builtin
del name
result = input
if num_lower > -1:
result = np.triu(result, -num_lower)
if num_upper > -1:
result = np.tril(result, num_upper)
return result
def testSolveTriangularGrad(self, lower, transpose_a, unit_diagonal,
lhs_shape, rhs_shape, dtype, rng):
_skip_if_unsupported_type(dtype)
A = np.tril(rng(lhs_shape, dtype) + 5 * onp.eye(lhs_shape[-1], dtype=dtype))
A = A if lower else T(A)
B = rng(rhs_shape, dtype)
f = partial(jsp.linalg.solve_triangular, lower=lower, trans=transpose_a,
unit_diagonal=unit_diagonal)
jtu.check_grads(f, (A, B), 2, rtol=2e-2, eps=1e-3)
def testSolveTriangularGrad(self, lower, transpose_a, lhs_shape, rhs_shape,
dtype, rng):
A = np.tril(
rng(lhs_shape, dtype) + 5 * onp.eye(lhs_shape[-1], dtype=dtype))
A = A if lower else T(A)
B = rng(rhs_shape, dtype)
f = partial(jsp.linalg.solve_triangular,
lower=lower,
trans=transpose_a)
jtu.check_grads(f, (A, B), 2, rtol=2e-2, eps=1e-3)
def inverse_fun(params, inputs, **kwargs):
L, U, S = params
L = np.tril(L, -1) + identity
U = np.triu(U, 1)
W = P @ L @ (U + np.diag(S))
outputs = inputs @ linalg.inv(W)
log_det_jacobian = np.full(inputs.shape[:1],
-np.log(np.abs(S)).sum())
return outputs, log_det_jacobian
def testSolveTriangularGrad(self, lower, transpose_a, lhs_shape,
rhs_shape, dtype, rng):
# TODO(frostig): change ensemble to support a bigger rtol
self.skipTest("rtol does not cover all devices and precision modes")
A = np.tril(rng(lhs_shape, dtype) + 5 * onp.eye(lhs_shape[-1], dtype=dtype))
A = A if lower else T(A)
B = rng(rhs_shape, dtype)
f = partial(jsp.linalg.solve_triangular, lower=lower,
trans=1 if transpose_a else 0)
jtu.check_grads(f, (A, B), 2, rtol=1e-3)
def house_qr_j_lt_m(H, j):
m, n = H.shape
Htri = jnp.tril(H)
v, thisbeta = house_padded(Htri[:, j], j)
# Hjj = jax.lax.dynamic_slice(H, (j, j), (m-j, n-j)) # H[j:, j:]
# H_update = house_leftmult(Hjj, v, thisbeta)
# H = index_update(H, index[:, :],
# jax.lax.dynamic_update_slice(H, H_update, [j, j]))
# H = index_update(H, index[:, :],
# jax.lax.dynamic_update_slice(H, v[1:], [j+1, j]))
return H, thisbeta
def testSolveTriangularBlockedGrad(self, lower, transpose_a, lhs_shape,
rhs_shape, dtype, rng):
# TODO(frostig): change ensemble to support a bigger rtol
A = np.tril(
rng(lhs_shape, dtype) + 5 * onp.eye(lhs_shape[-1], dtype=dtype))
A = A if lower else T(A)
B = rng(rhs_shape, dtype)
f = partial(scipy.linalg.solve_triangular,
lower=lower,
trans=1 if transpose_a else 0)
jtu.check_grads(f, (A, B), 2, rtol=1e-3)
def testTriangularSolveBatching(self, left_side, a_shape, b_shape, bdims):
rng = jtu.rand_default()
A = np.tril(rng(a_shape, onp.float32)
+ 5 * onp.eye(a_shape[-1], dtype=onp.float32))
B = rng(b_shape, onp.float32)
solve = partial(lax_linalg.triangular_solve, lower=True,
transpose_a=False, conjugate_a=False,
unit_diagonal=False, left_side=left_side)
X = vmap(solve, bdims)(A, B)
matmul = partial(np.matmul, precision=lax.Precision.HIGHEST)
Y = matmul(A, X) if left_side else matmul(X, A)
onp.testing.assert_allclose(Y - B, 0, atol=1e-5)
def build(self, sc: OrderedDict):
gp_matrices = OrderedDict()
gp_matrices["l_u_beta"] = aux_math.matrix_diag_transform(np.tril(sc["S_u_beta"]), nn.softplus)
gp_matrices["l_u_gamma"] = aux_math.matrix_diag_transform(np.tril(sc["S_u_gamma"]), nn.softplus)
# Kernel
gp_matrices["k_beta_beta"] = self.kernel.matrix(sc["X_u_beta"], sc["X_u_beta"])
gp_matrices["l_beta_beta"] = scipy.linalg.cholesky(gp_matrices["k_beta_beta"] +
self.jitter * np.eye(self.n_inducing_beta),
lower=True)
k_gamma_gamma = \
self.kernel.matrix(sc["X_u_gamma"], sc["X_u_gamma"]) + self.jitter * np.eye(self.n_inducing_gamma)
k_beta_gamma = self.kernel.matrix(sc["X_u_beta"], sc["X_u_gamma"])
l_beta_inv_k_beta_gamma = scipy.linalg.solve_triangular(gp_matrices["l_beta_beta"], k_beta_gamma,
lower=True)
gp_matrices["l_beta_inv_k_beta_gamma"] = l_beta_inv_k_beta_gamma
c_gamma_gamma = k_gamma_gamma - np.matmul(
np.transpose(gp_matrices["l_beta_inv_k_beta_gamma"], (0, 2, 1)), gp_matrices["l_beta_inv_k_beta_gamma"])
gp_matrices["l_gamma_gamma"] = scipy.linalg.cholesky(c_gamma_gamma +
self.jitter * np.eye(self.n_inducing_gamma), lower=True)
# U_beta_dists
gp_matrices["q_u_beta_mean"] = sc["mu_u_beta"]
gp_matrices["q_u_beta_tril"] = gp_matrices["l_u_beta"]
gp_matrices["p_u_beta_mean"] = np.zeros([self.ndims_out, self.n_inducing_beta])
gp_matrices["p_u_beta_tril"] = gp_matrices["l_beta_beta"]
# U_gamma_dists
gp_matrices["q_u_gamma_mean"] = sc["mu_u_gamma"]
gp_matrices["q_u_gamma_tril"] = gp_matrices["l_u_gamma"]
gp_matrices["p_u_gamma_mean"] = np.zeros([self.ndims_out, self.n_inducing_gamma])
gp_matrices["p_u_gamma_tril"] = gp_matrices["l_gamma_gamma"]
return gp_matrices
def K_net(x):
k=1024
mlp = hk.Sequential([
hk.Linear(k), jax.nn.swish,
hk.Linear(k), jax.nn.swish,
hk.Linear(k), jax.nn.swish,
hk.Linear(k), jax.nn.swish,
hk.Linear(9)])
y = radial_hyperbolic_compactification(x)
out = mlp(y)
B = jnp.tril(out.reshape((3,3)))
K = B+B.T
return 0*.05*K
def cov_params_to_matrix_jax(cov_params):
"""Build covariance matrix from 1d array with its lower triangular elements.
Args:
cov_params (np.array): 1d array with the lower triangular elements of a
covariance matrix (in C-order)
Returns:
cov (np.array): a covariance matrix
"""
lower = chol_params_to_lower_triangular_matrix_jax(cov_params)
cov = lower + jnp.tril(lower, k=-1).T
return cov
def gamma_net(x):
k=1024
mlp = hk.Sequential([
hk.Linear(k), jax.nn.swish,
hk.Linear(k), jax.nn.swish,
hk.Linear(k), jax.nn.swish,
hk.Linear(k), jax.nn.swish,
hk.Linear(9)])
y = radial_hyperbolic_compactification(x)
out = mlp(y)
A = jnp.tril(out.reshape((3,3)))
L = jnp.eye(3) + .3*(1-jnp.linalg.norm(y))*A
gamma = [email protected]
return gamma
def testTriangularSolveGrad(
self, lower, transpose_a, conjugate_a, unit_diagonal, left_side, a_shape,
b_shape, dtype, rng_factory):
_skip_if_unsupported_type(dtype)
rng = rng_factory()
# Test lax_linalg.triangular_solve instead of scipy.linalg.solve_triangular
# because it exposes more options.
A = np.tril(rng(a_shape, dtype) + 5 * onp.eye(a_shape[-1], dtype=dtype))
A = A if lower else T(A)
B = rng(b_shape, dtype)
f = partial(lax_linalg.triangular_solve, lower=lower,
transpose_a=transpose_a, conjugate_a=conjugate_a,
unit_diagonal=unit_diagonal, left_side=left_side)
jtu.check_grads(f, (A, B), 2, rtol=4e-2, eps=1e-3)
def test_mask_arg(self):
seq_len = 3
embed_size = 2
model_size = 15
query = key = value = jnp.zeros((seq_len, embed_size))
causal_mask = jnp.tril(jnp.ones((seq_len, seq_len)))
causal_mask = causal_mask[None, :, :]
mha = attention.MultiHeadAttention(key_size=7,
num_heads=11,
value_size=13,
model_size=model_size,
w_init_scale=1.0)(query,
key,
value,
mask=causal_mask)
self.assertEqual(mha.shape, (seq_len, model_size))
def test_correlated_mvn():
# This requires dense mass matrix estimation.
D = 5
warmup_steps, num_samples = 5000, 8000
true_mean = 0.
a = np.tril(0.5 * np.fliplr(np.eye(D)) + 0.1 * np.exp(random.normal(random.PRNGKey(0), shape=(D, D))))
true_cov = np.dot(a, a.T)
true_prec = np.linalg.inv(true_cov)
def potential_fn(z):
return 0.5 * np.dot(z.T, np.dot(true_prec, z))
init_params = np.zeros(D)
samples = mcmc(warmup_steps, num_samples, init_params, potential_fn=potential_fn, dense_mass=True)
assert_allclose(np.mean(samples), true_mean, atol=0.02)
assert onp.sum(onp.abs(onp.cov(samples.T) - true_cov)) / D**2 < 0.02
def setup(self):
self.model = DeepNetwork(H=self.H, kernel_size=self.kernel_size)
if self.params is None:
self.params = self.model.init(
jax.random.PRNGKey(0),
jnp.expand_dims(jnp.ones([self.history_len]),
axis=(0, 2)))["params"]
# linear feature transform:
# errs -> [average of last h errs, ..., average of last 2 errs, last err]
# emulates low-pass filter bank
self.featurizer = jnp.tril(
jnp.ones((self.history_len, self.history_len)))
self.featurizer /= jnp.expand_dims(jnp.arange(self.history_len, 0, -1),
axis=0)
# TODO(dsuo): resolve jit load/unload
if self.use_model_apply_jit:
self.model_apply = jax.jit(self.model.apply)
else:
self.model_apply = self.model.apply
def _log_probs(self, X: np.ndarray, alpha: np.ndarray) -> np.ndarray:
"""Compute class log probabilities for X.
Args:
X: An array of shape ``(n_samples, n_features)`` containing the training
examples.
alpha: The SVM normal vector scales. Normally this should be
``self.alpha_``, but we leave it as an argument so we can differentiate
through this function when fitting.
Returns:
An array of shape ``(n_samples, n_classes)`` containing the predicted log
probabilities for each class.
"""
n = alpha.shape[0]
L = jnp.tril(np.ones((n, n)))
A = jnp.diag(alpha)
likelihoods = (L @ A @ (self.coefs_ @ X.T + self.b_[:, None])).T
return logsoftmax(likelihoods)
def setup(self, waveform=None):
self.model = ActorCritic()
if self.params is None:
self.params = self.model.init(
jax.random.PRNGKey(0),
jnp.expand_dims(jnp.ones([self.history_len]), axis=(0, 1)))['params']
# linear feature transform:
# errs -> [average of last h errs, ..., average of last 2 errs, last err]
# emulates low-pass filter bank
self.featurizer = jnp.tril(jnp.ones((self.history_len, self.history_len)))
self.featurizer /= jnp.expand_dims(
jnp.arange(self.history_len, 0, -1), axis=0)
if waveform is None:
self.waveform = BreathWaveform.create()
if self.normalize:
self.u_scaler = u_scaler
self.p_scaler = p_scaler
def _validate_butcher_tableau(alphas: jnp.array, betas: jnp.array,
gammas: jnp.array) -> None:
_error_msg = []
if len(alphas) != len(gammas):
_error_msg.append(
"Alpha and gamma vectors must have the same length")
if betas.shape[0] != betas.shape[1]:
_error_msg.append("Betas must be a quadratic matrix with the same "
"dimension as the alphas/gammas arrays")
# for an explicit method, betas must be lower triangular
if not jnp.allclose(betas, jnp.tril(betas, k=-1)):
_error_msg.append("The beta matrix has to be lower triangular for "
"an explicit Runge-Kutta method, i.e. "
"b_ij = 0 for i <= j")
if _error_msg:
raise ValueError("An error occurred while validating the Input "
"Butcher tableau. More information: "
"{}.".format(",".join(_error_msg)))
