public override void Update(NDArray labels, NDArray preds)
{
CheckLabelShapes(labels, preds, true);
NDArray pred_label = null;
if (preds.Shape != labels.Shape)
{
pred_label = preds.Argmax(Axis);
}
else
{
pred_label = preds;
}
var label = labels.Ravel().AsType(DType.Int32);
pred_label = pred_label.Ravel().AsType(DType.Int32);
var num_correct = nd.Equal(pred_label, label).AsType(DType.Float32).Sum();
sum_metric += num_correct;
global_sum_metric += num_correct;
num_inst += pred_label.Shape.Size;
global_num_inst += pred_label.Shape.Size;
}
private NDArrayOrSymbol F(NDArray x)
{
var match = nd.Contrib.BipartiteMatching(x, threshold: this._threshold, is_ascend: this._is_ascend);
// make sure if iou(a, y) == iou(b, y), then b should also be a good match
// otherwise positive/negative samples are confusing
// potential argmax and max
var pargmax = x.Argmax(axis: -1, keepdims: true);
var maxs = x.Max(axis: -2, keepdims: true);
NDArray mask = null;
if (this._share_max)
{
mask = nd.BroadcastGreaterEqual(x + this._eps, maxs);
mask = nd.Sum(mask, axis: -1, keepdims: true);
}
else
{
var pmax = nd.Pick(x, pargmax, axis: -1, keepdims: true);
mask = nd.BroadcastGreaterEqual(pmax + this._eps, maxs);
mask = nd.Pick(mask, pargmax, axis: -1, keepdims: true);
}
var new_match = nd.Where(mask > 0, pargmax, nd.OnesLike(pargmax) * -1);
var result = nd.Where(match < 0, new_match.Squeeze(axis: -1), match);
return(result);
}
public static float Acc(NDArray output, NDArray label)
{
// output: (batch, num_output) float32 ndarray
// label: (batch, ) int32 ndarray
return(nd.Equal(output.Argmax(axis: 1), label.AsType(DType.Float32)).Mean());
}