def test_points_in_boxes_gpu():
if not torch.cuda.is_available():
pytest.skip('test requires GPU and torch+cuda')
boxes = torch.tensor(
[[[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 0.3]],
[[-10.0, 23.0, 16.0, 10, 20, 20, 0.5]]],
dtype=torch.float32).cuda(
) # boxes (b, t, 7) with bottom center in lidar coordinate
pts = torch.tensor(
[[[1, 2, 3.3], [1.2, 2.5, 3.0], [0.8, 2.1, 3.5], [1.6, 2.6, 3.6],
[0.8, 1.2, 3.9], [-9.2, 21.0, 18.2], [3.8, 7.9, 6.3],
[4.7, 3.5, -12.2]],
[[3.8, 7.6, -2], [-10.6, -12.9, -20], [-16, -18, 9], [-21.3, -52, -5],
[0, 0, 0], [6, 7, 8], [-2, -3, -4], [6, 4, 9]]],
dtype=torch.float32).cuda() # points (b, m, 3) in lidar coordinate
point_indices = points_in_boxes_gpu(points=pts, boxes=boxes)
expected_point_indices = torch.tensor(
[[0, 0, 0, 0, 0, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1]],
dtype=torch.int32).cuda()
assert point_indices.shape == torch.Size([2, 8])
assert (point_indices == expected_point_indices).all()
if torch.cuda.device_count() > 1:
pts = pts.to('cuda:1')
boxes = boxes.to('cuda:1')
expected_point_indices = expected_point_indices.to('cuda:1')
point_indices = points_in_boxes_gpu(points=pts, boxes=boxes)
assert point_indices.shape == torch.Size([2, 8])
assert (point_indices == expected_point_indices).all()
def calc(model, data_loader):
model.eval()
dataset = data_loader.dataset
prog_bar = mmcv.ProgressBar(len(dataset))
seg_eval = SegEvaluator(class_names=dataset.SEG_CLASSES)
det_eval = SegEvaluator(class_names=dataset.SEG_CLASSES)
print('\nStart Test Loop')
# print('batch_size:', data_loader.batch_size) # 1
for idx, data in enumerate(data_loader):
# print(type(data['img'][0])) # DataContainter
with torch.no_grad():
seg_res, box_res = model(return_loss=False, rescale=True, **data)
# len(box_res) == batch_size == 1
# box_res: [dict('pts_bbox'=dict(boxes_3d=bboxes, scores_3d=scores, labels_3d=labels))]
# handle seg
seg_label = data['seg_label'][0].data[0] # list of tensor
seg_pts_indices = data['seg_pts_indices'][0].data[0] # list of tensor
seg_points = data['seg_points'][0].data[0]
seg_pred = seg_res.argmax(1).cpu().numpy()
pred_list = []
gt_list = []
left_idx = 0
for i in range(len(seg_label)):
# num_points = len(seg_pts_indices[i])
assert len(seg_label[i]) == len(seg_pts_indices[i])
num_points = len(seg_label[i])
right_idx = left_idx + num_points
pred_list.append(seg_pred[left_idx:right_idx])
gt_list.append(seg_label[i].numpy())
left_idx = right_idx
seg_eval.batch_update(pred_list, gt_list)
# handle det
dic = box_res[0]['pts_bbox']
tensor_boxes = dic['boxes_3d'].tensor[:, :7].cuda()
labels = dic['labels_3d']
num_seg_pts = len(seg_points[0])
num_pred_boxes = len(tensor_boxes)
fake_labels = torch.tensor([4] * num_seg_pts)
box_idx = points_in_boxes_gpu(seg_points[0].cuda().unsqueeze(0),
tensor_boxes.unsqueeze(0)).squeeze(0)
for i in range(num_pred_boxes):
mask = box_idx == i # select points in i_th box
fake_labels[mask] = labels[i]
det_eval.update(fake_labels.numpy(), seg_label[0])
# progress bar
batch_size = len(box_res)
for _ in range(batch_size):
prog_bar.update()
print(seg_eval.print_table())
print('overall_acc:', seg_eval.overall_acc)
print('overall_iou:', seg_eval.overall_iou)
print(det_eval.print_table())
print('overall_acc:', det_eval.overall_acc)
print('overall_iou:', det_eval.overall_iou)
def points_in_boxes(self, points):
"""Find the box which the points are in.
Args:
points (torch.Tensor): Points in shape (N, 3).
Returns:
torch.Tensor: The index of box where each point are in.
"""
box_idx = points_in_boxes_gpu(
points.unsqueeze(0),
self.tensor.unsqueeze(0).to(points.device)).squeeze(0)
return box_idx
def forward_target(
self,
img=None,
seg_points=None,
seg_pts_indices=None,
seg_label=None, # just for debug
points=None,
pts_indices=None,
img_metas=None,
**kwargs):
'''
Args:
img: (4, 3, 225, 400)
seg_points: list of tensor (N, 4); len == batch_size
seg_pts_indices: list of tensor (N, 2); len == batch_size
seg_label: list of tensor (N, ); len == batch_size
points: list of tensor (M, 4); len == batch_size
pts_indices: list of tensor (M, 2); len == batch_size
img_metas:
Returns: losses
'''
# no aug on target; if aug, pay attention to seg_points
img_feats, pts_feats = self.extract_feat(points, pts_indices, img,
img_metas)
seg_logits = self.img_seg_head(img_feats=img_feats,
seg_pts=seg_points,
seg_pts_indices=seg_pts_indices)
# seg_logits.shape=(num_seg_pts_batch, num_classes=4+1)
outs = self.pts_bbox_head(pts_feats)
bbox_list = self.pts_bbox_head.get_bboxes(*outs, img_metas)
# bbox_list: list of tuple; len(bbox_list) = target_batch_size
# bbox_list[i]: (bboxes, scores, labels)
# bboxes: LiDARInstance3DBoxes of tensor (num_pred, 9); scores: tensor (num_pred, ); labels: tensor (num_pred, )
fake_labels = []
for batch_id, (bboxes, scores, labels) in enumerate(bbox_list):
seg_points_i = seg_points[batch_id]
num_seg_pts = len(seg_points_i)
num_pred_boxes = len(labels)
tensor_boxes = bboxes.tensor[:, :
7] # len(tensor_boxes) == len(scores) == len(labels) == num_pred_boxes
fake_labels_i = torch.tensor([self.img_seg_head.num_classes - 1] *
num_seg_pts).to(seg_points[0].device)
# fake_labels_i: tensor (num_seg_pts, ); filled with num_classes-1 (background == 4)
box_idx = points_in_boxes_gpu(seg_points_i.unsqueeze(0),
tensor_boxes.unsqueeze(0)).squeeze(0)
# box_idx: tensor (num_seg_pts, ); box_idx[i]==-1: not in any bounding box; box_idx[i]=k>=0: in k_th box
for i in range(num_pred_boxes):
mask = box_idx == i # select points in i_th box
fake_labels_i[mask] = labels[i]
fake_labels.append(fake_labels_i)
target_seg_loss = self.img_seg_head.loss(seg_logits,
fake_labels) # cross_entropy
return target_seg_loss
for idx in range(len(dataset)):
data = dataset[idx]
points = data['points'].data[:, :3]
seg_points = data['seg_points'].data[:, :3] # (N, 4) -> (N, 3)
seg_labels = data['seg_label'].data # (N, )
gt_bboxes_3d = data['gt_bboxes_3d'].data # tensor=(M, 9)
gt_labels_3d = data['gt_labels_3d'].data # (M, )
# LiDARInstance3DBoxes points_in_boxes
tensor_boxes = gt_bboxes_3d.tensor.cuda()
pts1 = seg_points.cuda()
start = time.time()
tensor_boxes = tensor_boxes[:, :7]
box_idx = points_in_boxes_gpu(pts1.unsqueeze(0), tensor_boxes.unsqueeze(0)).squeeze(0)
# boxes = LiDARInstance3DBoxes(gt_bboxes_3d.tensor[:, :7])
# box_idx = boxes.points_in_boxes(seg_points[:, :3].cuda())
# 1.
# fake_labels = torch.tensor([num_classes-1] * len(seg_labels))
# for i in range(len(box_idx)):
# if box_idx[i] != -1:
# fake_labels[i] = gt_labels_3d[box_idx[i]]
# 2.
# fake_labels = torch.tensor(list(map(lambda x: gt_labels_3d[x] if x >= 0 else num_classes-1, box_idx)))
# 3.
fake_labels = torch.tensor([num_classes-1] * len(seg_labels))
mask = box_idx != -1
