Error when training pose_estimation about load pcd file.

[yikakabu]

Hi, Jimmy. I have made my own dataset that contains 4 objects and pcd files. When I run:

python train.py config/kinect1_mask.yml

There is an error occured:

Traceback (most recent call last): File "train.py", line 295, in main(cfg) File "train.py", line 215, in main criterion = PointsL1Loss(numpy_pcs).cuda() File "train.py", line 63, in init self._pcs = torch.from_numpy(np.array(numpy_pcs)).cuda() ValueError: could not broadcast input array from shape (4,38720) into shape (4)

There are 38720 points in my first pcd file, and the format is same with your pcd files. It's like:

VERSION 0.7 FIELDS x y z SIZE 4 4 4 TYPE F F F COUNT 1 1 1 WIDTH 38720 HEIGHT 1 VIEWPOINT 0 0 0 1 0 0 0 POINTS 38720 DATA ascii 0 0.00193 0 0 0 0 0.0019499999 0 0 0.0038999999 0 0 0.0058499998 0 0 0.0077999998 0 0 0.0097500002 0 0 ...

My config file is like :

data: root: data/kinect1_mask pcd_root: objects num_subsets: 10 val_subset_num: 101 objects: [ box, bottle, cola_can, power_bank ] batch_size: 32 workers: 4 arch: num_input_channels: 1 num_shared_fc_layers: 1 num_shared_fc_nodes: 256 num_position_fc_layers: 1 num_position_fc_nodes: 256 num_orientation_fc_layers: 1 num_orientation_fc_nodes: 256 loss: points # l1 | posecnn | points_simple | points optimizer: lr: 0.01 lr_decay_epochs: [700, 1400] momentum: 0.9 weight_decay: 0.0001 training: logs_dir: logs/ checkpoints_dir: checkpoints/ experiment_name: kinect1_mask print_freq: 10 checkpoint_epochs: 100 epochs: 2100 log_dir: resume:

Could you give me any suggestions? Thanks in advance.

[jimmyyhwu]

The number of points must match in all of the pcd files. See here for my pcd preprocessing.

[yikakabu]

That's the problem! THANK you very much.