torch.fake_quantize_per_tensor_affine#

torch.fake_quantize_per_tensor_affine(input,scale,zero_point,quant_min,quant_max)→Tensor#

Returns a new tensor with the data ininput fake quantized usingscale,zero_point,quant_min andquant_max.

$$\text{output}=(min(\text{quant\_max},max(\text{quant\_min},\text{std::nearby\_int}(\text{input}\mathrm{/}\text{scale})+\text{zero\_point}))-\text{zero\_point})\times \text{scale}$$output=(min(quant_max,max(quant_min,std::nearby_int(input/scale)+zero_point))−zero_point)×scale

Parameters

• input (Tensor) – the input value(s),torch.float32 tensor

• scale (double scalar orfloat32 Tensor) – quantization scale

• zero_point (int64 scalar orint32 Tensor) – quantization zero_point

• quant_min (int64) – lower bound of the quantized domain

• quant_max (int64) – upper bound of the quantized domain

Returns

A newly fake_quantizedtorch.float32 tensor

Return type

Tensor

Example:

>>>x=torch.randn(4)>>>xtensor([ 0.0552,  0.9730,  0.3973, -1.0780])>>>torch.fake_quantize_per_tensor_affine(x,0.1,0,0,255)tensor([0.1000, 1.0000, 0.4000, 0.0000])>>>torch.fake_quantize_per_tensor_affine(x,torch.tensor(0.1),torch.tensor(0),0,255)tensor([0.1000, 1.0000, 0.4000, 0.0000])