PyTorch3D file io throws error - AttributeError: _evt

Question

I finally got PyTorch3D to install on my conda environment with the following configuration -- torch=1.13.0 and torchvision=0.14.0 and pytorch3d=0.7.5.

I am trying to load a mesh from an .obj file using pytorch3d.io from https://pytorch3d.readthedocs.io/en/v0.6.0/modules/io.html.

In the line pytorch3d.io.load_obj(filename), I get the below error

An exception occurred in telemetry logging.Disabling telemetry to prevent further exceptions.
Traceback (most recent call last):
  File "/home/aditya/miniconda3/envs/py3d/lib/python3.10/site-packages/iopath/common/file_io.py", line 946, in __log_tmetry_keys
    handler.log_event()
  File "/home/aditya/miniconda3/envs/py3d/lib/python3.10/site-packages/iopath/common/event_logger.py", line 97, in log_event
    del self._evt
AttributeError: _evt

I am on Nvidia RTX 4090 graphics card. Not sure what the issue is here.

Answer 1

I just encountered this problem. I looked into the for iopath files and found the culprit:

def log_event(self, topic: Optional[str] = None):
    if b_tmetry_available and self._enabled:

        # Sample the current event.
        if not self._sample_record():
            return

        if topic is None:
            topic = self.DEFAULT_TOPIC

        for writer in self._writers:
            writer.writeRecord(topic, self._evt)
    del self._evt
    self._evt = SimpleEventRecord()

The last two lines should be under the conditional (they are not):

if b_tmetry_available and self._enabled:

When I did conda intall it gave me the version of iopath with this bug, but as of writing this the latest version on GitHub has this bug corrected. https://github.com/facebookresearch/iopath/blob/main/iopath/common/event_logger.py

I noticed the import block as well:

try:
    from tmetry.simpleevent import SimpleEventRecord
    from tmetry.writer import TmetryWriter

    b_tmetry_available = True
except ImportError:
    b_tmetry_available = False

But tmetry is an internal tool used by the developers. So the fix, as per the latest version on GitHub, is that it should not execute those last two lines.

That said, I've done conda update iopath and it still gave me the broken version. I ended up fixing it manually for now.