HistogramObserver

class torch.ao.quantization.observer.HistogramObserver(bins=2048, dtype=torch.quint8, qscheme=torch.per_tensor_affine, reduce_range=False, quant_min=None, quant_max=None, factory_kwargs=None, eps=1.1920928955078125e-07, is_dynamic=False, *, consider_full_range=False, **kwargs)

The module records the running histogram of tensor values along with min/max values. calculate_qparams will calculate scale and zero_point.

Parameters:

• bins (int) – Number of bins to use for the histogram

• dtype (dtype) – dtype argument to the quantize node needed to implement the reference model spec

• qscheme – Quantization scheme to be used

• reduce_range – Reduces the range of the quantized data type by 1 bit

• eps (Tensor) – Epsilon value for float32, Defaults to torch.finfo(torch.float32).eps.

• consider_full_range (bool) – Whether the min/max search may return the full observed range. If True, the chosen range never has a higher estimated L2 quantization error than the full observed [min_val, max_val] range. If False (default), the search always trims at least one histogram bin from one end of the range, even when that increases the estimated error, matching historical behavior.

The scale and zero point are computed as follows:

1. Create the histogram of the incoming inputs.

   The histogram is computed continuously, and the ranges per bin change with every new tensor observed.

2. Search the distribution in the histogram for optimal min/max values.

   The search for the min/max values ensures the minimization of the quantization error with respect to the floating point model.

3. Compute the scale and zero point the same way as in the

   MinMaxObserver