This op simulates the precision loss from the quantized forward pass by:
Quantizing the tensor to fixed point numbers, which should match the target
quantization method when it is used in inference.
Dequantizing it back to floating point numbers for the following ops, most
likely matmul.
There are different ways to quantize. This version uses only scaling, so 0.0
maps to 0.
From the specified 'num_bits' in the quantized output type, it determines
minimum and maximum representable quantized values.
e.g.
[-128, 127] for signed, num_bits = 8, or
[0, 255] for unsigned, num_bits = 8.
If range_given == False, the initial input_min, input_max will be determined
automatically as the minimum and maximum values in the input tensor, otherwise
the specified values of input_min, input_max are used.
This op determines the maximum scale_factor that would map the initial
[input_min, input_max] range to a range that lies within the representable
quantized range.
It determines the scale from one of input_min and input_max, then updates the
other one to maximize the representable range.
e.g.
if the output is signed, num_bits = 8, [input_min, input_max] = [-10.0,
5.0]: it would use a scale_factor of -128 / -10.0 = 12.8 In this case, it
would update input_max to be 127 / 12.8 = 9.921875
if the output is signed, num_bits = 8, [input_min, input_max] = [-10.0,
10.0]: it would use a scale_factor of 127 / 10.0 = 12.7 In this case, it
would update input_min to be 128.0 / 12.7 = -10.07874
if the output is unsigned, input_min is forced to be 0, and only the
specified input_max is used.
After determining the scale_factor and updating the input range, it applies the
following to each value in the 'input' tensor.
The above round function rounds the value based on the given round_mode.
Args
input
A Tensor. Must be one of the following types: bfloat16, half, float32, float64.
Tensor to quantize and then dequantize.
input_min
A Tensor. Must have the same type as input.
If range_given == True, this specifies the minimum input value that needs to
be represented, otherwise it is determined from the min value of the input
tensor.
input_max
A Tensor. Must have the same type as input.
If range_given == True, this specifies the maximum input value that needs to
be represented, otherwise it is determined from the max value of the input
tensor.
signed_input
An optional bool. Defaults to True.
Whether the quantization is signed or unsigned. (actually this parameter should
have been called signed_output)
num_bits
An optional int. Defaults to 8.
The bitwidth of the quantization.
range_given
An optional bool. Defaults to False.
Whether the range is given or should be determined from the input tensor.
round_mode
An optional string from: "HALF_TO_EVEN", "HALF_UP". Defaults to "HALF_TO_EVEN".
The 'round_mode' attribute controls which rounding tie-breaking algorithm is
used when rounding float values to their quantized equivalents. The following
rounding modes are currently supported:
HALF_TO_EVEN: this is the default round_mode.
HALF_UP: round towards positive. In this mode 7.5 rounds up to 8 and -7.5
rounds up to -7.
narrow_range
An optional bool. Defaults to False.
If True, then the absolute value of the quantized minimum value is the same as
the quantized maximum value, instead of 1 greater.
i.e. for 8 bit quantization, the minimum value is -127 instead of -128.
axis
An optional int. Defaults to -1.
If specified, this axis is treated as a channel or slice axis, and a separate
quantization range is used for each channel or slice along this axis.
[null,null,["Last updated 2024-04-26 UTC."],[],[],null,["# tf.raw_ops.QuantizeAndDequantizeV2\n\n\u003cbr /\u003e\n\nQuantizes then dequantizes a tensor.\n\n#### View aliases\n\n\n**Compat aliases for migration**\n\nSee\n[Migration guide](https://www.tensorflow.org/guide/migrate) for\nmore details.\n\n[`tf.compat.v1.raw_ops.QuantizeAndDequantizeV2`](https://www.tensorflow.org/api_docs/python/tf/raw_ops/QuantizeAndDequantizeV2)\n\n\u003cbr /\u003e\n\n tf.raw_ops.QuantizeAndDequantizeV2(\n input,\n input_min,\n input_max,\n signed_input=True,\n num_bits=8,\n range_given=False,\n round_mode='HALF_TO_EVEN',\n narrow_range=False,\n axis=-1,\n name=None\n )\n\nThis op simulates the precision loss from the quantized forward pass by:\n\n1. Quantizing the tensor to fixed point numbers, which should match the target quantization method when it is used in inference.\n2. Dequantizing it back to floating point numbers for the following ops, most likely matmul.\n\nThere are different ways to quantize. This version uses only scaling, so 0.0\nmaps to 0.\n\nFrom the specified 'num_bits' in the quantized output type, it determines\nminimum and maximum representable quantized values.\n\ne.g.\n\n- \\[-128, 127\\] for signed, num_bits = 8, or\n- \\[0, 255\\] for unsigned, num_bits = 8.\n\nIf range_given == False, the initial input_min, input_max will be determined\nautomatically as the minimum and maximum values in the input tensor, otherwise\nthe specified values of input_min, input_max are used.\n| **Note:** If the input_min, input_max are specified, they do not need to equal the actual minimum and maximum values in the tensor. e.g. in some cases it may be beneficial to specify these values such that the low probability extremes of the input distribution are clipped.\n\nThis op determines the maximum scale_factor that would map the initial\n\\[input_min, input_max\\] range to a range that lies within the representable\nquantized range.\n\nIt determines the scale from one of input_min and input_max, then updates the\nother one to maximize the representable range.\n\ne.g.\n\n- if the output is signed, num_bits = 8, \\[input_min, input_max\\] = \\[-10.0, 5.0\\]: it would use a scale_factor of -128 / -10.0 = 12.8 In this case, it would update input_max to be 127 / 12.8 = 9.921875\n- if the output is signed, num_bits = 8, \\[input_min, input_max\\] = \\[-10.0, 10.0\\]: it would use a scale_factor of 127 / 10.0 = 12.7 In this case, it would update input_min to be 128.0 / 12.7 = -10.07874\n- if the output is unsigned, input_min is forced to be 0, and only the specified input_max is used.\n\nAfter determining the scale_factor and updating the input range, it applies the\nfollowing to each value in the 'input' tensor.\n\noutput = round(clamp(value, input_min, input_max) \\* scale_factor) / scale_factor.\n\nThe above round function rounds the value based on the given round_mode.\n\n\u003cbr /\u003e\n\n\u003cbr /\u003e\n\n\u003cbr /\u003e\n\n| Args ---- ||\n|----------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|\n| `input` | A `Tensor`. Must be one of the following types: `bfloat16`, `half`, `float32`, `float64`. Tensor to quantize and then dequantize. |\n| `input_min` | A `Tensor`. Must have the same type as `input`. If `range_given == True`, this specifies the minimum input value that needs to be represented, otherwise it is determined from the min value of the `input` tensor. |\n| `input_max` | A `Tensor`. Must have the same type as `input`. If `range_given == True`, this specifies the maximum input value that needs to be represented, otherwise it is determined from the max value of the `input` tensor. |\n| `signed_input` | An optional `bool`. Defaults to `True`. Whether the quantization is signed or unsigned. (actually this parameter should have been called **`signed_output`**) |\n| `num_bits` | An optional `int`. Defaults to `8`. The bitwidth of the quantization. |\n| `range_given` | An optional `bool`. Defaults to `False`. Whether the range is given or should be determined from the `input` tensor. |\n| `round_mode` | An optional `string` from: `\"HALF_TO_EVEN\", \"HALF_UP\"`. Defaults to `\"HALF_TO_EVEN\"`. The 'round_mode' attribute controls which rounding tie-breaking algorithm is used when rounding float values to their quantized equivalents. The following rounding modes are currently supported: \u003cbr /\u003e - HALF_TO_EVEN: this is the default round_mode. - HALF_UP: round towards positive. In this mode 7.5 rounds up to 8 and -7.5 rounds up to -7. |\n| `narrow_range` | An optional `bool`. Defaults to `False`. If True, then the absolute value of the quantized minimum value is the same as the quantized maximum value, instead of 1 greater. i.e. for 8 bit quantization, the minimum value is -127 instead of -128. |\n| `axis` | An optional `int`. Defaults to `-1`. If specified, this axis is treated as a channel or slice axis, and a separate quantization range is used for each channel or slice along this axis. |\n| `name` | A name for the operation (optional). |\n\n\u003cbr /\u003e\n\n\u003cbr /\u003e\n\n\u003cbr /\u003e\n\n| Returns ------- ||\n|---|---|\n| A `Tensor`. Has the same type as `input`. ||\n\n\u003cbr /\u003e"]]
