function fusedConv2d_()

in tfjs-core/src/ops/fused/conv2d.ts [96:267]

• 143 lines of code
• 12 McCabe index (conditional complexity)

function fusedConv2d_<T extends Tensor3D|Tensor4D>({
  x,
  filter,
  strides,
  pad,
  dataFormat = 'NHWC',
  dilations = [1, 1],
  dimRoundingMode,
  bias,
  activation = 'linear',
  preluActivationWeights,
  leakyreluAlpha
}: {
  x: T|TensorLike,
  filter: Tensor4D|TensorLike,
  strides: [number, number]|number,
  pad: 'valid'|'same'|number|conv_util.ExplicitPadding,
  dataFormat?: 'NHWC'|'NCHW',
  dilations?: [number, number]|number,
  dimRoundingMode?: 'floor'|'round'|'ceil',
  bias?: Tensor|TensorLike,
  activation?: Activation,
  preluActivationWeights?: Tensor,
  leakyreluAlpha?: number
}): T {
  activation = activation || 'linear';

  if (shouldFuse(ENGINE.state.gradientDepth, activation) === false) {
    let result = unfusedConv2d(
        x, filter, strides, pad, dataFormat, dilations, dimRoundingMode);
    if (bias != null) {
      result = add(result, bias);
    }

    return applyActivation(
               result, activation, preluActivationWeights, leakyreluAlpha) as T;
  }

  const $x = convertToTensor(x, 'x', 'conv2d', 'float32');
  const $filter = convertToTensor(filter, 'filter', 'conv2d', 'float32');

  let x4D = $x as Tensor4D;
  let reshapedTo4D = false;

  if ($x.rank === 3) {
    reshapedTo4D = true;
    x4D = reshape($x, [1, $x.shape[0], $x.shape[1], $x.shape[2]]);
  }
  util.assert(
      x4D.rank === 4,
      () => `Error in fused conv2d: input must be rank 4, but got rank ` +
          `${x4D.rank}.`);
  util.assert(
      $filter.rank === 4,
      () => `Error in fused conv2d: filter must be rank 4, but got rank ` +
          `${$filter.rank}.`);
  conv_util.checkPadOnDimRoundingMode('fused conv2d', pad, dimRoundingMode);
  util.assert(
      x4D.shape[3] === $filter.shape[2],
      () => `Error in conv2d: depth of input (${x4D.shape[3]}) must match ` +
          `input depth for filter ${$filter.shape[2]}.`);
  util.assert(
      conv_util.eitherStridesOrDilationsAreOne(strides, dilations),
      () => 'Error in conv2D: Either strides or dilations must be 1. ' +
          `Got strides ${strides} and dilations '${dilations}'`);
  util.assert(
      dataFormat === 'NHWC',
      () => `Error in conv2d: got dataFormat of ${
          dataFormat} but only NHWC is currently supported.`);

  const convInfo = conv_util.computeConv2DInfo(
      x4D.shape, $filter.shape, strides, dilations, pad, dimRoundingMode);

  let $bias: Tensor;
  if (bias != null) {
    $bias = convertToTensor(bias, 'bias', 'fused conv2d');
    [$bias] = makeTypesMatch($bias, $x);

    broadcast_util.assertAndGetBroadcastShape(convInfo.outShape, $bias.shape);
  }

  let $preluActivationWeights: Tensor;
  if (preluActivationWeights != null) {
    $preluActivationWeights = convertToTensor(
        preluActivationWeights, 'prelu weights', 'fused conv2d');
  }

  const grad = (dy: Tensor4D, saved: Tensor[]) => {
    const [$filter, x4D, y, $bias] =
        saved as [Tensor4D, Tensor4D, Tensor4D, Tensor];

    const dyActivation = getFusedDyActivation(dy, y, activation) as Tensor4D;

    util.assert(
        conv_util.tupleValuesAreOne(dilations),
        () => 'Error in gradient of fused conv2D: ' +
            `dilation rates greater than 1 ` +
            `are not yet supported in gradients. Got dilations '${dilations}'`);

    const xDer =
        conv2DBackpropInput(x4D.shape, dyActivation, $filter, strides, pad);
    const filterDer =
        conv2DBackpropFilter(x4D, dyActivation, $filter.shape, strides, pad);
    const der: Tensor[] = [xDer, filterDer];

    if ($bias != null) {
      const biasDer = getFusedBiasGradient($bias, dyActivation);
      der.push(biasDer);
    }
    return der;
  };

  const inputs: FusedConv2DInputs = {
    x: x4D,
    filter: $filter,
    bias: $bias,
    preluActivationWeights: $preluActivationWeights
  };

  const attrs: FusedConv2DAttrs = {
    strides,
    pad,
    dataFormat,
    dilations,
    dimRoundingMode,
    activation,
    leakyreluAlpha
  };

  // Depending on the the params passed in we will have different number of
  // inputs and thus a a different number of elements in the gradient.
  if (bias == null) {
    const customOp =
        customGrad((x4D: Tensor4D, filter: Tensor4D, save: GradSaveFunc) => {
          let res: Tensor4D|Tensor3D =
              // tslint:disable-next-line: no-unnecessary-type-assertion
              ENGINE.runKernel(
                  FusedConv2D, inputs as {} as NamedTensorMap,
                  attrs as {} as NamedAttrMap);

          save([filter, x4D, res]);

          if (reshapedTo4D) {
            // tslint:disable-next-line: no-unnecessary-type-assertion
            res = reshape(res, [res.shape[1], res.shape[2], res.shape[3]]) as
                Tensor3D;
          }

          return {value: res, gradFunc: grad};
        });
    return customOp(x4D, $filter) as T;
  } else {
    const customOpWithBias = customGrad(
        (x4D: Tensor4D, filter: Tensor4D, bias: Tensor, save: GradSaveFunc) => {
          let res: Tensor4D|Tensor3D = ENGINE.runKernel(
              FusedConv2D, inputs as {} as NamedTensorMap,
              attrs as {} as NamedAttrMap);

          save([filter, x4D, res, bias]);

          if (reshapedTo4D) {
            // tslint:disable-next-line: no-unnecessary-type-assertion
            res = reshape(res, [res.shape[1], res.shape[2], res.shape[3]]) as
                Tensor3D;
          }

          return {value: res, gradFunc: grad};
        });

    return customOpWithBias(x4D, $filter, $bias) as T;
  }
}