org.tensorflow.op.quantization

Class Dequantize<U extends TNumber>

java.lang.Object

org.tensorflow.op.RawOp

org.tensorflow.op.quantization.Dequantize<U>

Type Parameters:

U - data type for output output

All Implemented Interfaces:

org.tensorflow.ndarray.Shaped, Op, Operand<U>

@Operator(group="quantization")
public final class Dequantize<U extends TNumber>
extends RawOp
implements Operand<U>

Dequantize the 'input' tensor into a float or bfloat16 Tensor. [min_range, max_range] are scalar floats that specify the range for the output. The 'mode' attribute controls exactly which calculations are used to convert the float values to their quantized equivalents.

In 'MIN_COMBINED' mode, each value of the tensor will undergo the following:

 if T == qint8: in[i] += (range(T) + 1)/ 2.0
 out[i] = min_range + (in[i]* (max_range - min_range) / range(T))
 

here range(T) = numeric_limits<T>::max() - numeric_limits<T>::min()

MIN_COMBINED Mode Example

If the input comes from a QuantizedRelu6, the output type is quint8 (range of 0-255) but the possible range of QuantizedRelu6 is 0-6. The min_range and max_range values are therefore 0.0 and 6.0. Dequantize on quint8 will take each value, cast to float, and multiply by 6 / 255. Note that if quantizedtype is qint8, the operation will additionally add each value by 128 prior to casting.

If the mode is 'MIN_FIRST', then this approach is used:

 num_discrete_values = 1 << (# of bits in T)
 range_adjust = num_discrete_values / (num_discrete_values - 1)
 range = (range_max - range_min) * range_adjust
 range_scale = range / num_discrete_values
 const double offset_input = static_cast<double>(input) - lowest_quantized;
 result = range_min + ((input - numeric_limits<T>::min()) * range_scale)
 

If the mode is SCALED, dequantization is performed by multiplying each input value by a scaling_factor. (Thus an input of 0 always maps to 0.0).

The scaling_factor is determined from min_range, max_range, and narrow_range in a way that is compatible with QuantizeAndDequantize{V2|V3} and QuantizeV2, using the following algorithm:

   const int min_expected_T = std::numeric_limits<T>::min() +
     (narrow_range ? 1 : 0);
   const int max_expected_T = std::numeric_limits<T>::max();
   const float max_expected_T = std::numeric_limits<float>::max();

   const float scale_factor =
     (std::numeric_limits<T>::min() == 0) ? (max_range / max_expected_T)
                                          : std::max(min_range / min_expected_T,
                                                     max_range / max_expected_T);
 

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	Dequantize.Inputs
static class	Dequantize.Options Optional attributes for Dequantize

Field Summary

Fields

Modifier and Type	Field and Description
static String	OP_NAME The name of this op, as known by TensorFlow core engine

Fields inherited from class org.tensorflow.op.RawOp

operation

Constructor Summary

Constructors

Constructor and Description
Dequantize(Operation operation)

Method Summary

Modifier and Type	Method and Description
Output<U>	asOutput() Returns the symbolic handle of the tensor.
static Dequantize.Options	axis(Long axis) Sets the axis option.
static <U extends TNumber> Dequantize<U>	create(Scope scope, Operand<? extends TNumber> input, Operand<TFloat32> minRange, Operand<TFloat32> maxRange, Class<U> dtype, Dequantize.Options... options) Factory method to create a class wrapping a new Dequantize operation.
static Dequantize<TFloat32>	create(Scope scope, Operand<? extends TNumber> input, Operand<TFloat32> minRange, Operand<TFloat32> maxRange, Dequantize.Options[] options) Factory method to create a class wrapping a new Dequantize operation, with the default output types.
static Dequantize.Options	mode(String mode) Sets the mode option.
static Dequantize.Options	narrowRange(Boolean narrowRange) Sets the narrowRange option.
Output<U>	output() Gets output.

Methods inherited from class org.tensorflow.op.RawOp

equals, hashCode, op, toString

Methods inherited from class java.lang.Object

clone, finalize, getClass, notify, notifyAll, wait, wait, wait

Methods inherited from interface org.tensorflow.Operand

asTensor, shape, type

Methods inherited from interface org.tensorflow.op.Op

env, op

Methods inherited from interface org.tensorflow.ndarray.Shaped

rank, size

Field Detail

OP_NAME

public static final String OP_NAME

The name of this op, as known by TensorFlow core engine

See Also:

Constant Field Values

Constructor Detail

Dequantize

public Dequantize(Operation operation)

Method Detail

create

@Endpoint(describeByClass=true)
public static <U extends TNumber> Dequantize<U> create(Scope scope,
                                                                                       Operand<? extends TNumber> input,
                                                                                       Operand<TFloat32> minRange,
                                                                                       Operand<TFloat32> maxRange,
                                                                                       Class<U> dtype,
                                                                                       Dequantize.Options... options)

Factory method to create a class wrapping a new Dequantize operation.

Type Parameters:

U - data type for Dequantize output and operands

Parameters:

scope - current scope

input - The input value

minRange - The minimum scalar value possibly produced for the input.

maxRange - The maximum scalar value possibly produced for the input.

dtype - Type of the output tensor. Currently Dequantize supports float and bfloat16. If 'dtype' is 'bfloat16', it only supports 'MIN_COMBINED' mode.

options - carries optional attribute values

Returns:

a new instance of Dequantize

create

@Endpoint(describeByClass=true)
public static Dequantize<TFloat32> create(Scope scope,
                                                                          Operand<? extends TNumber> input,
                                                                          Operand<TFloat32> minRange,
                                                                          Operand<TFloat32> maxRange,
                                                                          Dequantize.Options[] options)

Factory method to create a class wrapping a new Dequantize operation, with the default output types.

Parameters:

scope - current scope

input - The input value

minRange - The minimum scalar value possibly produced for the input.

maxRange - The maximum scalar value possibly produced for the input.

options - carries optional attribute values

Returns:

a new instance of Dequantize, with default output types

mode

public static Dequantize.Options mode(String mode)

Sets the mode option.

Parameters:

mode - the mode option

Returns:

this Options instance.

narrowRange

public static Dequantize.Options narrowRange(Boolean narrowRange)

Sets the narrowRange option.

Parameters:

narrowRange - the narrowRange option

Returns:

this Options instance.

axis

public static Dequantize.Options axis(Long axis)

Sets the axis option.

Parameters:

axis - the axis option

Returns:

this Options instance.

output

public Output<U> output()

Gets output.

Returns:

output.

asOutput

public Output<U> asOutput()

Description copied from interface: Operand

Returns the symbolic handle of the tensor.

Inputs to TensorFlow operations are outputs of another TensorFlow operation. This method is used to obtain a symbolic handle that represents the computation of the input.

Specified by:

asOutput in interface Operand<U extends TNumber>

See Also:

OperationBuilder.addInput(Output)