private NDArrayOrSymbol F(NDArray a, NDArray b)
{
var _tup_1 = this._pre.Call(a);
NDArray al = _tup_1[0];
NDArray at = _tup_1[1];
NDArray ar = _tup_1[2];
NDArray ab = _tup_1[3];
var _tup_2 = this._pre.Call(b);
NDArray bl = _tup_2[0];
NDArray bt = _tup_2[1];
NDArray br = _tup_2[2];
NDArray bb = _tup_2[3];
// (B, N, M)
var left = nd.BroadcastMaximum(al.ExpandDims(-1), bl.ExpandDims(-2));
var right = nd.BroadcastMinimum(ar.ExpandDims(-1), br.ExpandDims(-2));
var top = nd.BroadcastMaximum(at.ExpandDims(-1), bt.ExpandDims(-2));
var bot = nd.BroadcastMinimum(ab.ExpandDims(-1), bb.ExpandDims(-2));
// clip with (0, float16.max)
var iw = nd.Clip(right - left + this._offset, a_min: 0, a_max: 65504);
var ih = nd.Clip(bot - top + this._offset, a_min: 0, a_max: 65504);
var i = iw * ih;
// areas
var area_a = ((ar - al + this._offset) * (ab - at + this._offset)).ExpandDims(-1);
var area_b = ((br - bl + this._offset) * (bb - bt + this._offset)).ExpandDims(-2);
var union = nd.BroadcastAdd(area_a, area_b) - i;
return(i / (union + this._eps));
}
private NDArrayOrSymbol F(NDArray x, NDArray gamma, NDArray beta)
{
NDArray y = null;
// normalization
using (var ag = Autograd.TrainMode())
{
y = x.ExpandDims(0).Reshape(0, 0, this.Ngroups, -1);
y = y.Reshape(1, -3, -1);
var batch = x.Shape[0];
y = nd.BatchNorm(y, nd.Ones(new Shape(batch * this.Ngroups), ctx: x.Context), nd.Zeros(new Shape(batch * this.Ngroups), ctx: x.Context), nd.Zeros(new Shape(batch * this.Ngroups), ctx: x.Context), nd.Ones(new Shape(batch * this.Ngroups), ctx: x.Context), eps: Epsilon, axis: Axis);
}
// scale and shift
y = nd.ReshapeLike(y, x).Reshape(0, 0, -1);
y = y * gamma.Reshape(1, -1, 1) + beta.Reshape(1, -1, 1);
return(nd.ReshapeLike(y, x));
}
private NDArrayOrSymbol F(NDArray samples, NDArray matches, NDArray anchors, NDArray labels, NDArray refs)
{
var _tup_2 = this.class_agnostic_encoder.Call(samples, matches, anchors, refs);
NDArray targets = _tup_2[0];
NDArray masks = _tup_2[1];
var ref_labels = nd.BroadcastLike(labels.Reshape(0, 1, -1), matches, lhs_axes: new Shape(1), rhs_axes: new Shape(1));
// labels [B, N, M] -> pick from matches [B, N] -> [B, N, 1]
ref_labels = nd.Pick(ref_labels, matches, axis: 2).Reshape(0, -1).ExpandDims(2);
// boolean array [B, N, C]
var same_cids = nd.BroadcastEqual(ref_labels, nd.Reshape(nd.Arange(this._num_class), shape: new Shape(1, 1, -1)));
// reduce box targets to positive samples only
var indices = nd.SliceAxis(nd.Reshape(nd.Argsort(nd.SliceAxis(masks, axis: -1, begin: 0, end: 1), axis: 1, is_ascend: false), new Shape(this._batch_size, -1)), axis: 1, begin: 0, end: this._max_pos);
var targets_tmp = new List <NDArray>();
var masks_tmp = new List <NDArray>();
var same_cids_tmp = new List <NDArray>();
foreach (var i in Enumerable.Range(0, this._batch_size))
{
var ind = nd.SliceAxis(indices, axis: 0, begin: i, end: i + 1).Squeeze(axis: 0);
var target = nd.SliceAxis(targets, axis: 0, begin: i, end: i + 1).Squeeze(axis: 0);
var mask = nd.SliceAxis(masks, axis: 0, begin: i, end: i + 1).Squeeze(axis: 0);
var same_cid = nd.SliceAxis(same_cids, axis: 0, begin: i, end: i + 1).Squeeze(axis: 0);
targets_tmp.Add(nd.Take(target, ind).ExpandDims(axis: 0));
masks_tmp.Add(nd.Take(mask, ind).ExpandDims(axis: 0));
same_cids_tmp.Add(nd.Take(same_cid, ind).ExpandDims(axis: 0));
}
targets = nd.Concat(targets_tmp, dim: 0);
masks = nd.Concat(masks_tmp, dim: 0);
same_cids = nd.Concat(same_cids_tmp, dim: 0).ExpandDims(3);
// targets, masks [B, N_pos, C, 4]
var all_targets = nd.BroadcastAxis(targets.ExpandDims(2), axis: new Shape(2), size: new Shape(this._num_class));
var all_masks = nd.BroadcastMul(masks.ExpandDims(2), nd.BroadcastAxis(same_cids, axis: new Shape(3), size: new Shape(4)));
return(new NDArrayOrSymbol(all_targets, all_masks, indices));
}