def _uniform_correlation_like_matrix(num_rows, batch_shape, dtype, seed): """Returns a uniformly random `Tensor` of "correlation-like" matrices. A "correlation-like" matrix is a symmetric square matrix with all entries between -1 and 1 (inclusive) and 1s on the main diagonal. Of these, the ones that are positive semi-definite are exactly the correlation matrices. Args: num_rows: Python `int` dimension of the correlation-like matrices. batch_shape: `Tensor` or Python `tuple` of `int` shape of the batch to return. dtype: `dtype` of the `Tensor` to return. seed: Random seed. Returns: matrices: A `Tensor` of shape `batch_shape + [num_rows, num_rows]` and dtype `dtype`. Each entry is in [-1, 1], and each matrix along the bottom two dimensions is symmetric and has 1s on the main diagonal. """ num_entries = num_rows * (num_rows + 1) / 2 ones = tf.ones(shape=[num_entries], dtype=dtype) # It seems wasteful to generate random values for the diagonal since # I am going to throw them away, but `fill_triangular` fills the # diagonal, so I probably need them. # It's not impossible that it would be more efficient to just fill # the whole matrix with random values instead of messing with # `fill_triangular`. Then would need to filter almost half out with # `matrix_band_part`. unifs = uniform.Uniform(-ones, ones).sample(batch_shape, seed=seed) tril = util.fill_triangular(unifs) symmetric = tril + tf.linalg.transpose(tril) diagonal_ones = tf.ones(shape=util.pad(batch_shape, axis=0, back=True, value=num_rows), dtype=dtype) return tf.linalg.set_diag(symmetric, diagonal_ones)
def _uniform_correlation_like_matrix(num_rows, batch_shape, dtype, seed): """Returns a uniformly random `Tensor` of "correlation-like" matrices. A "correlation-like" matrix is a symmetric square matrix with all entries between -1 and 1 (inclusive) and 1s on the main diagonal. Of these, the ones that are positive semi-definite are exactly the correlation matrices. Args: num_rows: Python `int` dimension of the correlation-like matrices. batch_shape: `Tensor` or Python `tuple` of `int` shape of the batch to return. dtype: `dtype` of the `Tensor` to return. seed: Random seed. Returns: matrices: A `Tensor` of shape `batch_shape + [num_rows, num_rows]` and dtype `dtype`. Each entry is in [-1, 1], and each matrix along the bottom two dimensions is symmetric and has 1s on the main diagonal. """ num_entries = num_rows * (num_rows + 1) / 2 ones = tf.ones(shape=[num_entries], dtype=dtype) # It seems wasteful to generate random values for the diagonal since # I am going to throw them away, but `fill_triangular` fills the # diagonal, so I probably need them. # It's not impossible that it would be more efficient to just fill # the whole matrix with random values instead of messing with # `fill_triangular`. Then would need to filter almost half out with # `matrix_band_part`. unifs = uniform.Uniform(-ones, ones).sample(batch_shape, seed=seed) tril = util.fill_triangular(unifs) symmetric = tril + tf.matrix_transpose(tril) diagonal_ones = tf.ones( shape=util.pad(batch_shape, axis=0, back=True, value=num_rows), dtype=dtype) return tf.matrix_set_diag(symmetric, diagonal_ones)
def _forward(self, x): return distribution_util.fill_triangular(x, upper=self._upper)
def _random_scale_tril(self, event_size): n = np.int32(event_size * (event_size + 1) // 2) p = 2. * self._rng.random_sample(n).astype(np.float32) - 1. return distribution_util.fill_triangular(0.25 * p)
def _forward(self, x): return distribution_util.fill_triangular(x, upper=self._upper)
def _random_scale_tril(self, event_size): n = np.int32(event_size * (event_size + 1) // 2) p = 2. * self._rng.random_sample(n).astype(np.float32) - 1. return distribution_util.fill_triangular(0.25 * p)