View source on GitHub

Draws samples from a multinomial distribution. (deprecated)

    logits, num_samples, seed=None, name=None, output_dtype=None


# samples has shape [1, 5], where each value is either 0 or 1 with equal
# probability.
samples = tf.random.categorical(tf.math.log([[0.5, 0.5]]), 5)


  • logits: 2-D Tensor with shape [batch_size, num_classes]. Each slice [i, :] represents the unnormalized log-probabilities for all classes.
  • num_samples: 0-D. Number of independent samples to draw for each row slice.
  • seed: A Python integer. Used to create a random seed for the distribution. See tf.compat.v1.set_random_seed for behavior.
  • name: Optional name for the operation.
  • output_dtype: integer type to use for the output. Defaults to int64.


The drawn samples of shape [batch_size, num_samples].