x_shape = [label_to_size.get(x_label) for x_label in x_labels]
    y_shape = [label_to_size.get(y_label) for y_label in y_labels]
    bias_shape = None
    if use_bias:
      bias_shape = [label_to_size.get(out_label) for out_label in out_labels]
      bias_shape = bias_shape[-1:]
    contracting_dims = set()

    x_signature = list(x_shape)
    y_signature = list(y_shape)

      self,
      equation: str,
  ):
    _, y_shape, bias_shape, x_signature, y_signature = (
        self._prepare_sample_einsum_datashapes(equation, use_bias=True)
    )

    model = self._create_einsum_model(
        self._input_saved_model_path,
        equation,
        y_shape,
        x_signature,
        y_signature,
        bias_shape,
    )

    # Generate model input data.

  for (size_t i = 0; i < out_channel_size; i++) {
    const float bias_scale = disable_per_channel_quantization_for_dense_
                                 ? bias_scales[0]
                                 : bias_scales[i];
    auto dequantized_value = bias_values[i] * bias_scale;
    EXPECT_THAT(dequantized_value,
                FloatNear(bias_float_buffer[i], bias_scale / 2));
  }