Thank you very much for sharing your code!
I encountered some problems during the reproduction process, and now I have summarized the problems and some of the solutions.
Resolved:
TensorFlow version issue:
Currently, Tensorflow has been updated to Tensorflow2, and almost all import Tensorflow as tf in the code need to be changed to:
Import tensorflow. compat. v1 as tf
To avoid some errors, in utils/tf_ In utils, a 'dial' statement needs to be added after 'import':
Import tensorflow. compat. v1 as tf
Tf.disable_ V2_ Behavior()
utils/tf_utils line 42, xavier seems to have been deprecated in TensorFlow2 and replaced with the same in else:
initializer = tf.truncated_normal_initializer(stddev=stddev)
Path and Call Issue:
in pointnet_part_seg.py,Some of the code in the import section should be changed to:
sys.path.append(os.path.join(BASE_DIR, 'utils'))
import group_conv as gconv
Unresolved
Network structure issues:
pointnet_part_seg line 104 group_vae error:
Expected list for 'values' argument to 'pack' Op, not <tf.Tensor 'Relu:0' shape=(32, 14, 14, 14, 48) dtype=float32>.
It seems that a list should be sent instead of a tensor. As I am not very familiar with TensorFlow, I have simply commented it out here in order to run the following sections. I hope you can provide some suggestions to make the code run as you expected.
Dataset issues:
I used the publicly available Shapenet dataset, but it seems to be missing the voxel section. The error message is as follows:
KeyError: "Unable to open object (object 'voxel' doesn't exist)"
Due to my previous work on object detection in outdoor scenes, I am not very familiar with this part of the dataset. May I ask where I can download the publicly available Shapenet dataset with voxel in HDF5 format data? Or please provide the data you used, thank you again!