tfc.ops.stochastic_round

Rounds inputs / step_size stochastically.

tfc.ops.stochastic_round(
    inputs, step_size, seed, name=None
)

This op computes the elementwise function:

output = { floor(x) with prob. p = x - floor(x) floor(x) + 1 with prob. 1 - p } where x = input / step_size.

inputs A Tensor. Must be one of the following types: bfloat16, half, float32. Floating point tensor to be rounded.
step_size A Tensor of type float32. Scalar tensor. Step size for rounding.
seed A Tensor of type int32. Arbitrary shape tensor. Seed for random number generator. If it has no elements, seeding is attempted from system time.
name A name for the operation (optional).

A Tensor of type int32. Integer tensor of same shape as inputs, containing rounded values.