dnnlowp::Tanh< T > Class Template Reference

We use the 3-region approach described in "Efficient VLSI Implementation of Neural Networks with Hyperbolic Tangent Activation Function", IEEE Transactions on Very Large Scale Integration Systems, Zamanlooy and Mirhassani. More...

#include <tanh.h>

Public Member Functions
	Tanh (double max_abs_err=DEFAULT_MAX_ABS_ERR)
T	Compute (T x) const
TensorQuantizationParams	GetInputQuantizationParams () const
TensorQuantizationParams	GetOutputQuantizationParams () const
int	GetPassRegionEnd () const
float	GetPassRegionEndDequantized () const
float	GetSaturationRegionBegin () const

Static Public Attributes
static constexpr double	DEFAULT_MAX_ABS_ERR = 0.02
static constexpr int	DEFAULT_NUM_IN_BITS = 8
static constexpr int	DEFAULT_NUM_OUT_BITS = 8

Detailed Description

template<typename T>
class dnnlowp::Tanh< T >

We use the 3-region approach described in "Efficient VLSI Implementation of Neural Networks with Hyperbolic Tangent Activation Function", IEEE Transactions on Very Large Scale Integration Systems, Zamanlooy and Mirhassani.

The pass region (x < x_pq) is approximated as x. The saturation region (x >= x_sq) is approximated as 1. The processing region (x_pq <= x < x_sq) is divided into sub-ranges and the average value of tanh(x) is used per sub-range.

Definition at line 21 of file tanh.h.

---

The documentation for this class was generated from the following files:

• caffe2/quantization/server/tanh.h
• caffe2/quantization/server/tanh.cc