// CHECK-DAG: {{\[\[\[}}-87, -42]

// CHECK: %[[quantize:.*]] = "tf.PartitionedCall"(%arg0, %[[in_scale]], %[[in_zp]])
// CHECK-SAME: f = @quantize_i8
// CHECK: %[[conv_quant:.*]] = "tf.PartitionedCall"(%[[quantize]], %[[w_quant]], %[[b_quant]],
// CHECK-SAME: %[[in_scale]], %[[in_zp]], %[[w_scale]], %[[w_zp]],
// CHECK-SAME: %[[b_scale]], %[[w_zp]], %[[out_scale]], %[[out_zp]])
// CHECK-SAME: f = @quantized_conv2d_with_bias_and_relu6_fn_0

  // TFR_ConstantTensorOp (
  //   ConstantOp (ConstAttr<F32Attr (in_scale[0] * in_scale[1] /
  //   out_scale))
  // )
  // Currently, all decompositions using this pattern (Conv2D, FC) have the
  // following preconditions:
  // * out_scale: float scalar attribute
  // * in_scale[0] (input scale): float scalar, given by tf.Const -> tfr.cast
  // * in_scale[1] (filter scale): float scalar/vector

// CHECK-DAG: %[[w0_quantized:.*]] = "tf.Const"() <{value = dense<{{\[\[\[\[}}-42, 18
// CHECK-DAG: %[[in_scale:.*]] = "tf.Const"() <{value = dense<0.0039215642> : tensor<f32>}> : () -> tensor<f32>
// CHECK-DAG: %[[in_out_zp:.*]] = "tf.Const"() <{value = dense<-128> : tensor<i32>}> : () -> tensor<i32>
// CHECK-DAG: %[[w0_scale:.*]] = "tf.Const"() <{value = dense<0.0150353173> : tensor<f32>}> : () -> tensor<f32>

  std::vector<int64_t> in_shape{input_type.getShape().vec()};
  std::vector<int64_t> out_shape{output_type.getShape().vec()};

  // If the reshape changes the number of dimensions so it cannot be interpreted
  // as a transpose.
  if (in_shape.size() != out_shape.size()) {
    return false;
  }

  in_shape.erase(std::remove(in_shape.begin(), in_shape.end(), 1),
                 in_shape.end());

  let touchid = null;    // Current touch identifier

  // State needed for pinch zooming
  let touchid2 = null;     // Second id for pinch zooming
  let initGap = 1.0;       // Starting gap between two touches
  let initScale = 1.0;     // currentScale when pinch zoom started
  let centerPoint = null;  // Center point for scaling

  // Convert event coordinates to svg coordinates.
  function toSvg(x, y) {
    const p = svg.createSVGPoint();

    ```
  }];

  let arguments = (ins
      F32:$out_scale,
      TFR_TensorListType:$in_scales);

  let results = (outs TFR_TensorType:$scale_factor);

  let assemblyFormat = [{
    `(` $out_scale `,` $in_scales `)` attr-dict `:` functional-type(operands, results)
  }];

  let hasCanonicalizer = 1;
}

	if _, ok := amount.SetString(value); !ok {
		return Quantity{}, ErrNumeric
	}

	// So that no one but us has to think about suffixes, remove it.
	if base == 10 {
		amount.SetScale(amount.Scale() + Scale(exponent).infScale())
	} else if base == 2 {
		// numericSuffix = 2 ** exponent
		numericSuffix := big.NewInt(1).Lsh(bigOne, uint(exponent))
		ub := amount.UnscaledBig()
		amount.SetUnscaledBig(ub.Mul(ub, numericSuffix))
	}

				_, err = p.output.Write(aNewline)
			case tabwriter.Escape:
				_, err = p.output.Write(p.space)
				p.state = inEscape
				m = n + 1 // +1: skip tabwriter.Escape
			default:
				_, err = p.output.Write(p.space)
				p.state = inText
				m = n
			}
		case inEscape:
			if b == tabwriter.Escape {
				_, err = p.output.Write(data[m:n])
				p.resetSpace()
			}
		case inText:

    "'h_scale' type: 'float'} attr : { name: 'max_classes_per_detection' "
    "type: 'int'} attr : { name: 'max_detections' type: 'int'} attr : { "
    "name: 'nms_iou_threshold' type: 'float'} attr : { name: "
    "'nms_score_threshold' type: 'float'} attr : { name: 'num_classes' type: "
    "'int'} attr : { name: 'w_scale' type: 'float'} attr : { name: 'x_scale' "

    Como esse arquivo **não muda com frequência**, o Docker irá detectá-lo e usar o **cache** para esse passo, habilitando o cache para o próximo passo também.

4. Instale as dependências de pacote vindas do arquivo de requisitos.