function sigmoidCrossEntropyWithLogits(target, output) {
var maxOutput = tfc.maximum(output, tfc.zerosLike(output));
var outputXTarget = tfc.mul(output, target);
var sigmoidOutput = tfc.log(tfc.add(getScalar(1), tfc.exp(tfc.neg(tfc.abs(output)))));
var result = tfc.add(tfc.sub(maxOutput, outputXTarget), sigmoidOutput);
return result;
}
function decodeBoxesLayer(x0, x1) {
var _a = getCenterCoordinatesAndSizesLayer(x0), sizes = _a.sizes, centers = _a.centers;
var vec = tf.unstack(tf.transpose(x1, [1, 0]));
var div0_out = tf.div(tf.mul(tf.exp(tf.div(vec[2], tf.scalar(5))), sizes[0]), tf.scalar(2));
var add0_out = tf.add(tf.mul(tf.div(vec[0], tf.scalar(10)), sizes[0]), centers[0]);
var div1_out = tf.div(tf.mul(tf.exp(tf.div(vec[3], tf.scalar(5))), sizes[1]), tf.scalar(2));
var add1_out = tf.add(tf.mul(tf.div(vec[1], tf.scalar(10)), sizes[1]), centers[1]);
return tf.transpose(tf.stack([
tf.sub(add0_out, div0_out),
tf.sub(add1_out, div1_out),
tf.add(add0_out, div0_out),
tf.add(add1_out, div1_out)
]), [1, 0]);
}
function getCenterCoordinatesAndSizesLayer(x) {
var vec = tf.unstack(tf.transpose(x, [1, 0]));
var sizes = [
tf.sub(vec[2], vec[0]),
tf.sub(vec[3], vec[1])
];
var centers = [
tf.add(vec[0], tf.div(sizes[0], tf.scalar(2))),
tf.add(vec[1], tf.div(sizes[1], tf.scalar(2)))
];
return {
sizes: sizes,
centers: centers
};
}
function residual(x, params) {
var out = convLayer_1.conv(x, params.conv1);
out = convLayer_1.convNoRelu(out, params.conv2);
out = tf.add(out, x);
out = tf.relu(out);
return out;
}
function convLayer(x, params, strides, withRelu, padding) {
if (padding === void 0) { padding = 'same'; }
var _a = params.conv, filters = _a.filters, bias = _a.bias;
var out = tf.conv2d(x, filters, strides, padding);
out = tf.add(out, bias);
out = scaleLayer_1.scale(out, params.scale);
return withRelu ? tf.relu(out) : out;
}
function biasAdd(x, bias, dataFormat) {
common_1.checkDataFormat(dataFormat);
if (ndim(bias) !== 1 && ndim(bias) !== ndim(x)) {
throw new errors_1.ValueError('Unexpected bias dimensions: ' + ndim(bias) +
'; expected it to be 1 or ' + ndim(x));
}
if (dataFormat) {
throw new errors_1.NotImplementedError('dataFormat logic is not yet implemented.');
}
return tfc.add(x, bias);
}
function dropout(x, level, noiseShape, seed) {
if (noiseShape != null && !_.isEqual(x.shape, noiseShape)) {
throw new errors_1.NotImplementedError('Non-default noise shape is not implemented yet: ' +
JSON.stringify(noiseShape));
}
if (seed != null) {
throw new errors_1.NotImplementedError('seed is not implemented for dropout yet.');
}
var multiplier = tfc.step(tfc.add(neg(level), randomUniform(x.shape, 0, 1, types_1.DType.float32)));
multiplier = tfc.mul(divide(getScalar(1), subtract(getScalar(1), level)), multiplier);
return tfc.mul(x, multiplier);
}
function residualDown(x, params) {
var out = convLayer_1.convDown(x, params.conv1);
out = convLayer_1.convNoRelu(out, params.conv2);
var pooled = tf.avgPool(x, 2, 2, 'valid');
var zeros = tf.zeros(pooled.shape);
var isPad = pooled.shape[3] !== out.shape[3];
var isAdjustShape = pooled.shape[1] !== out.shape[1] || pooled.shape[2] !== out.shape[2];
if (isAdjustShape) {
var padShapeX = out.shape.slice();
padShapeX[1] = 1;
var zerosW = tf.zeros(padShapeX);
out = tf.concat([out, zerosW], 1);
var padShapeY = out.shape.slice();
padShapeY[2] = 1;
var zerosH = tf.zeros(padShapeY);
out = tf.concat([out, zerosH], 2);
}
pooled = isPad ? tf.concat([pooled, zeros], 3) : pooled;
out = tf.add(pooled, out);
out = tf.relu(out);
return out;
}
function softsign(x) {
return tfc.div(x, tfc.add(getScalar(1), tfc.abs(x)));
}
function softplus(x) {
return tfc.log(tfc.add(getScalar(1), tfc.exp(x)));
}
function updateAdd(x, increment) {
return x.write(tfc.add(x.read(), increment));
}
function scalarPlusArray(c, x) {
return tfc.add(c, x);
}
function add(x, y) {
return tfc.add(x, y);
}
return tf.tidy(function () {
var out = tf.conv2d(x, params.filters, strides, 'same');
out = tf.add(out, params.batch_norm_offset);
return tf.clipByValue(out, 0, 6);
});