def __init__(self, num_classes, use_xyz=True, mode='TRAIN'):
super().__init__()
assert cfg.RPN.ENABLED or cfg.RCNN.ENABLED
if cfg.PSP.ENABLED:
self.psp = PSPNet(n_classes=1)
# self.psp = PSPNet()
if cfg.RPN.ENABLED:
self.rpn = RPN(use_xyz=use_xyz, mode=mode)
# merge down the xyz features and image features from PSPNet
feature_channel = cfg.RPN.FP_MLPS[0][-1]
self.merge_down = pt_utils.SharedMLP(
[feature_channel * 2, feature_channel], bn=cfg.RPN.USE_BN)
# self.merge_down = pt_utils.SharedMLP([feature_channel + 512, feature_channel], bn=cfg.RPN.USE_BN)
if cfg.RCNN.ENABLED:
rcnn_input_channels = 128 # channels of 128 merged rpn and PSPnet feature
if cfg.RCNN.BACKBONE == 'pointnet':
self.rcnn_net = RCNNNet(num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
elif cfg.RCNN.BACKBONE == 'pointsift':
pass
else:
raise NotImplementedError
def __init__(self, num_classes, use_xyz=True, mode='TRAIN'):
super().__init__()
assert cfg.RPN.ENABLED or cfg.RCNN.ENABLED
if cfg.RPN.ENABLED:
self.rpn = RPN(use_xyz=use_xyz, mode=mode)
if cfg.RCNN.ENABLED:
rcnn_input_channels = 128 # channels of rpn features
if cfg.RCNN.BACKBONE == 'pointnet':
self.rcnn_net = RCNNNet(num_classes=num_classes, input_channels=rcnn_input_channels, use_xyz=use_xyz)
elif cfg.RCNN.BACKBONE == 'pointsift':
pass
else:
raise NotImplementedError
def __init__(self,
num_classes,
num_point=512,
use_xyz=True,
mode='TRAIN',
old_model=False):
super().__init__()
self.mode = mode
assert cfg.RPN.ENABLED or cfg.RCNN.ENABLED or cfg.IOUN.ENABLED
if cfg.RPN.ENABLED:
self.rpn = RPN(use_xyz=use_xyz, mode=mode, old_model=old_model)
if cfg.RCNN.ENABLED or cfg.IOUN.ENABLED:
rcnn_input_channels = 128 # channels of rpn features x,y,z,r,mask
#self.rcnn_net = RCNNNet(num_point=num_point, num_classes=num_classes, input_channels=rcnn_input_channels, use_xyz=use_xyz)
self.rcnn_net = RCNNNet(num_point=num_point,
num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
ans_dict[key] = [batch[k][key] for k in range(batch_size)]
if isinstance(batch[0][key], int):
ans_dict[key] = np.array(ans_dict[key], dtype=np.int32)
elif isinstance(batch[0][key], float):
ans_dict[key] = np.array(ans_dict[key], dtype=np.float32)
input_data = ans_dict
input_data['pts_input'] = torch.FloatTensor(input_data['pts_input']).cuda()
input_data['pts_rect'] = torch.FloatTensor(input_data['pts_rect']).cuda()
input_data['gt_boxes3d'] = torch.FloatTensor(
input_data['gt_boxes3d']).cuda()
#ipdb.set_trace()
training = True
output = {}
rpn = RPN(use_xyz=True, mode=mode).cuda()
rcnn_net = RCNNNet(num_classes=2, input_channels=128, use_xyz=True).cuda()
rpn_output = rpn(input_data)
"""
rpn_cls:B,N=16384,1
rpn_reg:B,N,76
backbone_xyz:B,N,3
backbone_features:B,128,N
"""
output.update(rpn_output)
with torch.no_grad():
rpn_cls, rpn_reg = rpn_output['rpn_cls'], rpn_output['rpn_reg']
backbone_xyz, backbone_features = rpn_output[
'backbone_xyz'], rpn_output['backbone_features']
rpn_scores_raw = rpn_cls[:, :, 0]
def __init__(self, num_classes, use_xyz=True, mode='TRAIN'):
super().__init__()
assert cfg.RPN.ENABLED or cfg.RCNN.ENABLED
if cfg.RPN.ENABLED:
self.rpn = RPN(use_xyz=use_xyz, mode=mode)
if cfg.RCNN.ENABLED:
rcnn_input_channels = 128 # channels of rpn features
if cfg.RCNN.BACKBONE == 'pointnet':
self.rcnn_net = RCNNNet(num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
elif cfg.RCNN.BACKBONE == 'pointsift':
pass
elif cfg.RCNN.BACKBONE == 'edgeconv':
rcnn_input_channels = cfg.RCNN.GCN_CONFIG.FILTERS[-1] * 2
self.rcnn_net = GCNNet(num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
elif cfg.RCNN.BACKBONE == 'rotnet':
rcnn_input_channels = get_num_rot(
cfg.RCNN.ROT_CONFIG.DEGREE_RES)
self.rcnn_net = RotRCNN(num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
elif cfg.RCNN.BACKBONE == 'deepgcn':
# rcnn_input_channels = 256
self.rcnn_net = DenseRCNN(num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
elif cfg.RCNN.BACKBONE == 'refine':
# rcnn_input_channels = 256
self.rcnn_net = RefineRCNNNet(
num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
elif cfg.RCNN.BACKBONE == 'deeprefine':
# rcnn_input_channels = 256
self.rcnn_net = RefineDeepRCNNNet(
num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
elif cfg.RCNN.BACKBONE == 'deepfeats':
# rcnn_input_channels = 256
self.rcnn_net = DenseFeatRCNN(
num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
elif cfg.RCNN.BACKBONE == 'deepfeatsrefine':
# rcnn_input_channels = 256
self.rcnn_net = DenseFeatRefineRCNN(
num_classes=num_classes,
input_channels=rcnn_input_channels,
use_xyz=use_xyz)
else:
raise NotImplementedError
if cfg.CGCN.ENABLED:
self.cgcn_net = None
print(sum(p.numel() for p in self.rcnn_net.parameters()))
