@test_util.deprecated_graph_mode_only
  def test_replace_tensors_by_numpy_ndarrays_is_noop_when_no_tensor(self):
    # Fill the representative dataset with np.ndarrays only.
    repr_ds: repr_dataset.RepresentativeDataset = [
        {
            'input_tensor': np.random.uniform(low=-1.0, high=1.0, size=(4, 3)),
        }
        for _ in range(8)
    ]

    with self.session() as sess:

        self.filters = np.stack(
            [
                np.random.uniform(
                    low=-(i + 1), high=(i + 1), size=filter_shape[:-1]
                ).astype('f4')
                for i in range(self.out_channel_size)
            ],
            axis=-1,
        )

        self.bias = np.random.uniform(
            low=0, high=10, size=(self.out_channel_size)

    first_mid = self._lower_bound + self._bin_width / 2
    last_mid = first_mid + (self._num_bins - 1) * self._bin_width
    self._hist_mids = np.linspace(first_mid, last_mid, self._num_bins)

  def _get_dequantized_hist_mids_after_quantize(
      self, quant_min: float, quant_max: float
  ) -> np.ndarray:
    """Quantizes and dequantizes hist_mids using quant_min and quant_max.

	// observable side effects, and keeping it in place allows better escape analysis.
	if !ir.Any(*np, func(n ir.Node) bool { return n.Op() != ir.ONEW && callOrChan(n) }) {
		return
	}

	tmp := TempAt(base.Pos, ir.CurFunc, (*np).Type())
	as := ir.NewAssignStmt(base.Pos, tmp, *np)
	as.PtrInit().Append(Stmt(ir.NewDecl(n.Pos(), ir.ODCL, tmp)))
	*np = tmp

	n.PtrInit().Append(Stmt(as))
}

	return d.verifyChecksum()
}

func (d *decoder) parsePLTE(length uint32) error {
	np := int(length / 3) // The number of palette entries.
	if length%3 != 0 || np <= 0 || np > 256 || np > 1<<uint(d.depth) {
		return FormatError("bad PLTE length")
	}
	n, err := io.ReadFull(d.r, d.tmp[:3*np])
	if err != nil {
		return err
	}
	d.crc.Write(d.tmp[:n])
	switch d.cb {
	case cbP1, cbP2, cbP4, cbP8:

"""Defines types required for representative datasets for quantization."""

from collections.abc import Collection, Sized
import os
from typing import Iterable, Mapping, Optional, Union

import numpy as np

from tensorflow.compiler.mlir.quantization.tensorflow import quantization_options_pb2
from tensorflow.core.protobuf import meta_graph_pb2
from tensorflow.python.client import session

	// argument list.
	var err error
	var n int
	for _, arg := range args {
		nv := &nvargs[n]
		if np, ok := arg.(NamedArg); ok {
			if err = validateNamedValueName(np.Name); err != nil {
				return nil, err
			}
			arg = np.Value
			nv.Name = np.Name
		}
		nv.Ordinal = n + 1
		nv.Value = arg

		// Checking sequence has four routes:
		// A: 1. Default

	if len(meta) > maxMetaLen {
		return nil, fmt.Errorf("counter: metadata too large")
	}
	np := round(len(hdrPrefix), 4)
	n := round(np+4+len(meta), 32)
	hdr := make([]byte, n)
	copy(hdr, hdrPrefix)
	*(*uint32)(unsafe.Pointer(&hdr[np])) = uint32(n)
	copy(hdr[np+4:], meta)
	return hdr, nil
}

func (m *mappedFile) place(limit uint32, name string) (start, end uint32) {
	if limit == 0 {

			// looking for any user provided values.
			np := findRequestedNodePort(int(servicePort.Port), service.Spec.Ports)
			if np != 0 {
				err := nodePortOp.Allocate(np)
				if err != nil {
					// TODO: when validation becomes versioned, this gets more complicated.
					el := field.ErrorList{field.Invalid(field.NewPath("spec", "ports").Index(i).Child("nodePort"), np, err.Error())}

	// package/function combinations. This will help catch bugs in the
	// counter file reader.
	d.pkm = make(map[uint32]uint32)
	np := uint32(mfr.NumPackages())
	payload := []byte{}
	for pkIdx := uint32(0); pkIdx < np; pkIdx++ {
		var pd *decodemeta.CoverageMetaDataDecoder
		var err error
		pd, payload, err = mfr.GetPackageDecoder(pkIdx, payload)
		if err != nil {