Dequantize¶
Versioned name: Dequantize-1
Category: lower_precision
Short description: Dequantize converts a quantized(int8 or uint8) tensor to a fp32 tensor. It supports both per tensor and per channel asymmetric linear dequantization. Nearest round is used in this OP.
Attributes
qtype
Description: specifies which dequantization type is used.
Range of values: “per_tensor” or “per_channel”
Type: string
Default value: “per_tensor”
Required: no
axis
Description: specifies dimension on which apply per-channel dequantization. Only valid if qtype is “per_channel”.
Range of values: integers in [-d, d-1] where d = input_tensor.shape().size()
Type: int
Default value: 1
Required: no
scales
Description: apply in quantization formula.
Range of values: float values
Type: float[]
Default value: None
Required: yes
zps
Description: offset value that maps to float zero.
Range of values: integer values
Type: int[]
Default value: None
Required: yes
Inputs:
1:
input- quantized tensor to be dequantized. Required.
Outputs:
1:
output– dequantized tensor. Data type is fp32.