@Operator(group="quantization") public final class Dequantize extends PrimitiveOp implements Operand<Float>
[min_range, max_range] are scalar floats that specify the range for the 'input' data. 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_limitsMIN_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)
SCALED mode Example
`SCALED` mode matches the quantization approach used in `QuantizeAndDequantize{V2|V3}`.
If the mode is `SCALED`, we do not use the full range of the output type, choosing to elide the lowest possible value for symmetry (e.g., output range is -127 to 127, not -128 to 127 for signed 8 bit quantization), so that 0.0 maps to 0.
We first find the range of values in our tensor. The range we use is always centered on 0, so we find m such that
m = max(abs(input_min), abs(input_max))
Our input tensor range is then `[-m, m]`.
Next, we choose our fixed-point quantization buckets, `[min_fixed, max_fixed]`. If T is signed, this is
num_bits = sizeof(T) * 8
[min_fixed, max_fixed] =
[-(1 << (num_bits - 1) - 1), (1 << (num_bits - 1)) - 1]
Otherwise, if T is unsigned, the fixed-point range is
[min_fixed, max_fixed] = [0, (1 << num_bits) - 1]
From this we compute our scaling factor, s:
s = (2 * m) / (max_fixed - min_fixed)
Now we can dequantize the elements of our tensor:
result = input * s
Modifier and Type | Class and Description |
---|---|
static class |
Dequantize.Options
Optional attributes for
Dequantize |
operation
Modifier and Type | Method and Description |
---|---|
Output<Float> |
asOutput()
Returns the symbolic handle of a tensor.
|
static <T> Dequantize |
create(Scope scope,
Operand<T> input,
Operand<Float> minRange,
Operand<Float> maxRange,
Dequantize.Options... options)
Factory method to create a class wrapping a new Dequantize operation.
|
static Dequantize.Options |
mode(String mode) |
Output<Float> |
output() |
equals, hashCode, op, toString
public static <T> Dequantize create(Scope scope, Operand<T> input, Operand<Float> minRange, Operand<Float> maxRange, Dequantize.Options... options)
scope
- current scopeinput
- minRange
- The minimum scalar value possibly produced for the input.maxRange
- The maximum scalar value possibly produced for the input.options
- carries optional attributes valuespublic static Dequantize.Options mode(String mode)
mode
- public Output<Float> asOutput()
Operand
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.
asOutput
in interface Operand<Float>
OperationBuilder.addInput(Output)
Copyright © 2022. All rights reserved.