tf.random.gamma

Draws shape samples from each of the given Gamma distribution(s).

View aliases

Compat aliases for migration

See Migration guide for more details.

tf.compat.v1.random.gamma, tf.compat.v1.random_gamma

tf.random.gamma(
    shape, alpha, beta=None, dtype=tf.dtypes.float32, seed=None, name=None
)

alpha is the shape parameter describing the distribution(s), and beta is the inverse scale parameter(s).

Note: Because internal calculations are done using float64 and casting has floor semantics, we must manually map zero outcomes to the smallest possible positive floating-point value, i.e., np.finfo(dtype).tiny. This means that np.finfo(dtype).tiny occurs more frequently than it otherwise should. This bias can only happen for small values of alpha, i.e., alpha << 1 or large values of beta, i.e., beta >> 1.

The samples are differentiable w.r.t. alpha and beta. The derivatives are computed using the approach described in (Figurnov et al., 2018).

Example:

samples = tf.random.gamma([10], [0.5, 1.5])
# samples has shape [10, 2], where each slice [:, 0] and [:, 1] represents
# the samples drawn from each distribution

samples = tf.random.gamma([7, 5], [0.5, 1.5])
# samples has shape [7, 5, 2], where each slice [:, :, 0] and [:, :, 1]
# represents the 7x5 samples drawn from each of the two distributions

alpha = tf.constant([[1.],[3.],[5.]])
beta = tf.constant([[3., 4.]])
samples = tf.random.gamma([30], alpha=alpha, beta=beta)
# samples has shape [30, 3, 2], with 30 samples each of 3x2 distributions.

loss = tf.reduce_mean(tf.square(samples))
dloss_dalpha, dloss_dbeta = tf.gradients(loss, [alpha, beta])
# unbiased stochastic derivatives of the loss function
alpha.shape == dloss_dalpha.shape  # True
beta.shape == dloss_dbeta.shape  # True

References:

Implicit Reparameterization Gradients: Figurnov et al., 2018 (pdf)