tfc.ops.create_range_encoder

Creates range encoder objects to be used by EntropyEncode* ops.

The output handle has the shape specified by the input shape. Each element in handle is an independent range encoder object, and EntropyEncode* processes as many concurrent code streams as contained in handle.

This op expects lookup to be either a concatenation (1-D) or stack (2-D) of CDFs, where each CDF is preceded by a corresponding precision value. In case of a stack:

   lookup[..., 0] = precision in [1, 16],
   lookup[..., 1] / 2^precision = Pr(X < 0) = 0,
   lookup[..., 2] / 2^precision = Pr(X < 1),
   lookup[..., 3] / 2^precision = Pr(X < 2),
   ...
   lookup[..., -1] / 2^precision = 1,

Subsequent values in each CDF may be equal, indicating a symbol with zero probability. Attempting to encode such a symbol will result in undefined behavior. However, any number of trailing zero-probability symbols will be interpreted as padding, and attempting to use those will result in an encoding error (unless overflow functionality is used).

Overflow functionality can be enabled by negating the precision value in lookup. In that case, the last non-zero probability symbol in the CDF is used as an escape code, allowing negative integers and integers greater or equal to the last non-zero probability symbol to be encoded using an Elias gamma code, which is interleaved into the code stream. Attempting to encode a zero-probability symbol within the valid range still causes undefined behavior.

shape A Tensor of type int32.
lookup A Tensor of type int32.
name A name for the operation (optional).

A Tensor of type variant.