05360171创建于 2022年3月18日历史提交
diff --git a/mmdet/core/bbox/coder/delta_xywh_bbox_coder.py b/mmdet/core/bbox/coder/delta_xywh_bbox_coder.py
index e9eb3579..e8b53dce 100644
--- a/mmdet/core/bbox/coder/delta_xywh_bbox_coder.py
+++ b/mmdet/core/bbox/coder/delta_xywh_bbox_coder.py
@@ -168,8 +168,13 @@ def delta2bbox(rois,
                 [0.0000, 0.3161, 4.1945, 0.6839],
                 [5.0000, 5.0000, 5.0000, 5.0000]])
     """
-    means = deltas.new_tensor(means).view(1, -1).repeat(1, deltas.size(1) // 4)
-    stds = deltas.new_tensor(stds).view(1, -1).repeat(1, deltas.size(1) // 4)
+    # fix shape for means and stds for onnx
+    if torch.onnx.is_in_onnx_export():
+        means = deltas.new_tensor(means).view(1, -1).repeat(1, deltas.size(1).numpy() // 4)
+        stds = deltas.new_tensor(stds).view(1, -1).repeat(1, deltas.size(1).numpy() // 4)
+    else:
+        means = deltas.new_tensor(means).view(1, -1).repeat(1, deltas.size(1) // 4)
+        stds = deltas.new_tensor(stds).view(1, -1).repeat(1, deltas.size(1) // 4)
     denorm_deltas = deltas * stds + means
     dx = denorm_deltas[:, 0::4]
     dy = denorm_deltas[:, 1::4]
@@ -178,12 +183,22 @@ def delta2bbox(rois,
     max_ratio = np.abs(np.log(wh_ratio_clip))
     dw = dw.clamp(min=-max_ratio, max=max_ratio)
     dh = dh.clamp(min=-max_ratio, max=max_ratio)
-    # Compute center of each roi
-    px = ((rois[:, 0] + rois[:, 2]) * 0.5).unsqueeze(1).expand_as(dx)
-    py = ((rois[:, 1] + rois[:, 3]) * 0.5).unsqueeze(1).expand_as(dy)
-    # Compute width/height of each roi
-    pw = (rois[:, 2] - rois[:, 0]).unsqueeze(1).expand_as(dw)
-    ph = (rois[:, 3] - rois[:, 1]).unsqueeze(1).expand_as(dh)
+    # improve gather performance on NPU
+    if torch.onnx.is_in_onnx_export():
+        rois_perf = rois.permute(1, 0)
+        # Compute center of each roi
+        px = ((rois_perf[0, :] + rois_perf[2, :]) * 0.5).unsqueeze(1).expand_as(dx)
+        py = ((rois_perf[1, :] + rois_perf[3, :]) * 0.5).unsqueeze(1).expand_as(dy)
+        # Compute width/height of each roi
+        pw = (rois_perf[2, :] - rois_perf[0, :]).unsqueeze(1).expand_as(dw)
+        ph = (rois_perf[3, :] - rois_perf[1, :]).unsqueeze(1).expand_as(dh)
+    else:
+        # Compute center of each roi
+        px = ((rois[:, 0] + rois[:, 2]) * 0.5).unsqueeze(1).expand_as(dx)
+        py = ((rois[:, 1] + rois[:, 3]) * 0.5).unsqueeze(1).expand_as(dy)
+        # Compute width/height of each roi
+        pw = (rois[:, 2] - rois[:, 0]).unsqueeze(1).expand_as(dw)
+        ph = (rois[:, 3] - rois[:, 1]).unsqueeze(1).expand_as(dh)
     # Use exp(network energy) to enlarge/shrink each roi
     gw = pw * dw.exp()
     gh = ph * dh.exp()
diff --git a/mmdet/core/post_processing/bbox_nms.py b/mmdet/core/post_processing/bbox_nms.py
index 463fe2e4..1f8ad5a8 100644
--- a/mmdet/core/post_processing/bbox_nms.py
+++ b/mmdet/core/post_processing/bbox_nms.py
@@ -4,6 +4,57 @@ from mmcv.ops.nms import batched_nms
 from mmdet.core.bbox.iou_calculators import bbox_overlaps
 
 
+class BatchNMSOp(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, bboxes, scores, score_threshold, iou_threshold, max_size_per_class, max_total_size):
+        """
+        boxes (torch.Tensor): boxes in shape (batch, N, C, 4).
+        scores (torch.Tensor): scores in shape (batch, N, C).
+        return:
+            nmsed_boxes: (1, N, 4)
+            nmsed_scores: (1, N)
+            nmsed_classes: (1, N)
+            nmsed_num: (1,)
+        """
+
+        # Phony implementation for onnx export
+        nmsed_boxes = bboxes[:, :max_total_size, 0, :]
+        nmsed_scores = scores[:, :max_total_size, 0]
+        nmsed_classes = torch.arange(max_total_size, dtype=torch.long)
+        nmsed_num = torch.Tensor([max_total_size])
+
+        return nmsed_boxes, nmsed_scores, nmsed_classes, nmsed_num
+
+    @staticmethod
+    def symbolic(g, bboxes, scores, score_thr, iou_thr, max_size_p_class, max_t_size):
+        nmsed_boxes, nmsed_scores, nmsed_classes, nmsed_num = g.op('BatchMultiClassNMS',
+            bboxes, scores, score_threshold_f=score_thr, iou_threshold_f=iou_thr,
+            max_size_per_class_i=max_size_p_class, max_total_size_i=max_t_size, outputs=4)
+        return nmsed_boxes, nmsed_scores, nmsed_classes, nmsed_num
+
+def batch_nms_op(bboxes, scores, score_threshold, iou_threshold, max_size_per_class, max_total_size):
+    """
+    boxes (torch.Tensor): boxes in shape (N, 4).
+    scores (torch.Tensor): scores in shape (N, ).
+    """
+
+    if bboxes.dtype == torch.float32:
+        bboxes = bboxes.reshape(1, bboxes.shape[0].numpy(), -1, 4).half()
+        scores = scores.reshape(1, scores.shape[0].numpy(), -1).half()
+    else:
+        bboxes = bboxes.reshape(1, bboxes.shape[0].numpy(), -1, 4)
+        scores = scores.reshape(1, scores.shape[0].numpy(), -1)
+
+    nmsed_boxes, nmsed_scores, nmsed_classes, nmsed_num = BatchNMSOp.apply(bboxes, scores,
+        score_threshold, iou_threshold, max_size_per_class, max_total_size)
+    nmsed_boxes = nmsed_boxes.float()
+    nmsed_scores = nmsed_scores.float()
+    nmsed_classes = nmsed_classes.long()
+    dets = torch.cat((nmsed_boxes.reshape((max_total_size, 4)), nmsed_scores.reshape((max_total_size, 1))), -1)
+    labels = nmsed_classes.reshape((max_total_size, ))
+    return dets, labels
+
+
 def multiclass_nms(multi_bboxes,
                    multi_scores,
                    score_thr,
@@ -36,13 +87,25 @@ def multiclass_nms(multi_bboxes,
     if multi_bboxes.shape[1] > 4:
         bboxes = multi_bboxes.view(multi_scores.size(0), -1, 4)
     else:
-        bboxes = multi_bboxes[:, None].expand(
-            multi_scores.size(0), num_classes, 4)
+        # export expand operator to onnx more nicely
+        if torch.onnx.is_in_onnx_export:
+            bbox_shape_tensor = torch.ones(multi_scores.size(0), num_classes, 4)
+            bboxes = multi_bboxes[:, None].expand_as(bbox_shape_tensor)
+        else:
+            bboxes = multi_bboxes[:, None].expand(
+                multi_scores.size(0), num_classes, 4)
+
 
     scores = multi_scores[:, :-1]
     if score_factors is not None:
         scores = scores * score_factors[:, None]
 
+    # npu
+    if torch.onnx.is_in_onnx_export():
+        dets, labels = batch_nms_op(bboxes, scores, score_thr, nms_cfg.get("iou_threshold"), max_num, max_num)
+        return dets, labels
+
+    # cpu and gpu
     labels = torch.arange(num_classes, dtype=torch.long)
     labels = labels.view(1, -1).expand_as(scores)
 
@@ -53,6 +116,8 @@ def multiclass_nms(multi_bboxes,
     # remove low scoring boxes
     valid_mask = scores > score_thr
     inds = valid_mask.nonzero(as_tuple=False).squeeze(1)
+    # vals, inds = torch.topk(scores, 1000)
+
     bboxes, scores, labels = bboxes[inds], scores[inds], labels[inds]
     if inds.numel() == 0:
         if torch.onnx.is_in_onnx_export():
@@ -76,6 +141,7 @@ def multiclass_nms(multi_bboxes,
         return dets, labels[keep]
 
 
+
 def fast_nms(multi_bboxes,
              multi_scores,
              multi_coeffs,