export function residualDown(x: tf.Tensor4D, params: ResidualLayerParams): tf.Tensor4D {
let out = convDown(x, params.conv1)
out = convNoRelu(out, params.conv2)
let pooled = tf.avgPool(x, 2, 2, 'valid') as tf.Tensor4D
const zeros = tf.zeros<tf.Rank.R4>(pooled.shape)
const isPad = pooled.shape[3] !== out.shape[3]
const isAdjustShape = pooled.shape[1] !== out.shape[1] || pooled.shape[2] !== out.shape[2]
if (isAdjustShape) {
const padShapeX = [...out.shape] as [number, number, number, number]
padShapeX[1] = 1
const zerosW = tf.zeros<tf.Rank.R4>(padShapeX)
out = tf.concat([out, zerosW], 1)
const padShapeY = [...out.shape] as [number, number, number, number]
padShapeY[2] = 1
const zerosH = tf.zeros<tf.Rank.R4>(padShapeY)
out = tf.concat([out, zerosH], 2)
}
pooled = isPad ? tf.concat([pooled, zeros], 3) : pooled
out = tf.add(pooled, out) as tf.Tensor4D
out = tf.relu(out)
return out
}
(node: Node, tensorMap: NamedTensorsMap,
context: ExecutionContext): tfc.Tensor[] => {
switch (node.op) {
case 'conv1d': {
const stride =
getParamValue('stride', node, tensorMap, context) as number;
const pad = getParamValue('pad', node, tensorMap, context);
const dataFormat =
(getParamValue('dataFormat', node, tensorMap, context) as string)
.toUpperCase();
const dilation =
getParamValue('dilation', node, tensorMap, context) as number;
return [tfc.conv1d(
getParamValue('x', node, tensorMap, context) as tfc.Tensor3D,
getParamValue('filter', node, tensorMap, context) as tfc.Tensor3D,
stride, pad as 'valid' | 'same', dataFormat as 'NWC' | 'NCW',
dilation)];
}
case 'conv2d': {
const stride =
getParamValue('strides', node, tensorMap, context) as number[];
const pad = getParamValue('pad', node, tensorMap, context);
const dataFormat =
(getParamValue('dataFormat', node, tensorMap, context) as string)
.toUpperCase();
const dilations =
getParamValue('dilations', node, tensorMap, context) as number[];
return [tfc.conv2d(
getParamValue('x', node, tensorMap, context) as tfc.Tensor3D |
tfc.Tensor4D,
getParamValue('filter', node, tensorMap, context) as tfc.Tensor4D,
[stride[1], stride[2]], pad as 'valid' | 'same',
dataFormat as 'NHWC' | 'NCHW', [dilations[0], dilations[1]])];
}
case 'conv2dTranspose': {
const shape = getParamValue(
'outputShape', node, tensorMap,
context) as [number, number, number] |
[number, number, number, number];
const stride =
getParamValue('strides', node, tensorMap, context) as number[];
const pad = getParamValue('pad', node, tensorMap, context);
return [tfc.conv2dTranspose(
getParamValue('x', node, tensorMap, context) as tfc.Tensor3D |
tfc.Tensor4D,
getParamValue('filter', node, tensorMap, context) as tfc.Tensor4D,
shape, [stride[1], stride[2]], pad as 'valid' | 'same')];
}
case 'depthwiseConv2d': {
const stride =
getParamValue('strides', node, tensorMap, context) as number[];
const pad = getParamValue('pad', node, tensorMap, context);
const dilations =
getParamValue('dilations', node, tensorMap, context) as number[];
const dataFormat =
(getParamValue('dataFormat', node, tensorMap, context) as string)
.toUpperCase();
return [tfc.depthwiseConv2d(
getParamValue('input', node, tensorMap, context) as tfc.Tensor3D |
tfc.Tensor4D,
getParamValue('filter', node, tensorMap, context) as tfc.Tensor4D,
[stride[1], stride[2]], pad as 'valid' | 'same',
dataFormat as 'NHWC' | 'NCHW', [dilations[0], dilations[1]])];
}
case 'avgPool': {
const stride =
getParamValue('strides', node, tensorMap, context) as number[];
const pad = getParamValue('pad', node, tensorMap, context);
const kernelSize =
getParamValue('kernelSize', node, tensorMap, context) as number[];
return [tfc.avgPool(
getParamValue('x', node, tensorMap, context) as tfc.Tensor3D |
tfc.Tensor4D,
[kernelSize[1], kernelSize[2]], [stride[1], stride[2]],
pad as 'valid' | 'same')];
}
case 'maxPool': {
const stride =
getParamValue('strides', node, tensorMap, context) as number[];
const pad = getParamValue('pad', node, tensorMap, context);
const kernelSize =
getParamValue('kernelSize', node, tensorMap, context) as number[];
return [tfc.maxPool(
getParamValue('x', node, tensorMap, context) as tfc.Tensor3D |
tfc.Tensor4D,
[kernelSize[1], kernelSize[2]], [stride[1], stride[2]],
pad as 'valid' | 'same')];
}
default:
throw TypeError(`Node type ${node.op} is not implemented`);
}
};