torch.fake_quantize_per_channel_affine — PyTorch 2.5 documentation (original) (raw)

torch.fake_quantize_per_channel_affine(input, scale, zero_point, axis, quant_min, quant_max) → Tensor¶

Returns a new tensor with the data in input fake quantized per channel using scale,zero_point, quant_min and quant_max, across the channel specified by axis.

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

Parameters

• input (Tensor) – the input value(s), in torch.float32
• scale (Tensor) – quantization scale, per channel in torch.float32
• zero_point (Tensor) – quantization zero_point, per channel in torch.int32 or torch.half or torch.float32
• axis (int32) – channel axis
• quant_min (int64) – lower bound of the quantized domain
• quant_max (int64) – upper bound of the quantized domain

Returns

A newly fake_quantized per channel torch.float32 tensor

Return type

Tensor

Example:

x = torch.randn(2, 2, 2) x tensor([[[-0.2525, -0.0466], [ 0.3491, -0.2168]],

    [[-0.5906,  1.6258],
     [ 0.6444, -0.0542]]])

scales = (torch.randn(2) + 1) * 0.05 scales tensor([0.0475, 0.0486]) zero_points = torch.zeros(2).to(torch.int32) zero_points tensor([0, 0]) torch.fake_quantize_per_channel_affine(x, scales, zero_points, 1, 0, 255) tensor([[[0.0000, 0.0000], [0.3405, 0.0000]],

    [[0.0000, 1.6134],
    [0.6323, 0.0000]]])