tensorboardX 2.3 imports Pillow and PyTorch despite not depending on them

Describe the bug tensorboardX only has NumPy and Protobuf listed as dependencies, but if you try to import it, it tries to import PIL and torch as well.

This is a regression from 2.2.

Minimal runnable code to reproduce the behavior

$ python3 -mvenv /tmp/venv
$ /tmp/venv/bin/pip install -U pip
[...]
Successfully installed pip-21.1.3
$ /tmp/venv/bin/pip install tensorboardX==2.3
[...]
Successfully installed numpy-1.19.5 protobuf-3.17.3 six-1.16.0 tensorboardX-2.3
$ /tmp/venv/bin/python -c 'import tensorboardX'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/tmp/venv/lib/python3.6/site-packages/tensorboardX/__init__.py", line 5, in <module>
    from .torchvis import TorchVis
  File "/tmp/venv/lib/python3.6/site-packages/tensorboardX/torchvis.py", line 11, in <module>
    from .writer import SummaryWriter
  File "/tmp/venv/lib/python3.6/site-packages/tensorboardX/writer.py", line 17, in <module>
    from .comet_utils import CometLogger
  File "/tmp/venv/lib/python3.6/site-packages/tensorboardX/comet_utils.py", line 4, in <module>
    from PIL import Image
ModuleNotFoundError: No module named 'PIL'

(If you install Pillow, the same error happens with import torch in the same module.)

Expected behavior The import should succeed. All functionality that doesn’t depend on Pillow or PyTorch should work.

Screenshots N/A

Environment What is the result of pip list|grep -E "torch|proto|tensor" If the version is too old, please try to update first.

protobuf      3.17.3
tensorboardX  2.3

Python environment I’m using Python 3.6.9 from the Ubuntu 18.04 repositories.

Additional context N/A

[richardliaw]

+1, this introduces a large dependency regression for our ecosystem too!

[ujrav]

This has caused a problem/minor annoyance for my company’s project. We are now forced to install pytorch and allocate more ram on our containers despite not needing either for anything else.