Make datapipe.py compatible with torchdata==0.11 or later

[drivanov]

With torch 2.7/torchdata 0.11 releases, the datapype.py example fails

python3 /workspace/examples/datapipe.py 
Traceback (most recent call last):
  File "/workspace/examples/datapipe.py", line 17, in <module>
    from torchdata.datapipes.iter import FileLister, FileOpener, IterDataPipe
ModuleNotFoundError: No module named 'torchdata.datapipes'

fails because some classes have been moved from torchdata to torch nd a few methods, including in_memory_cache, have been removed.

The current PR resolves this issue in that example.

NOTES:
Running this example with the option --task mesh was failing even with previous versions of torch/torchdata packages

• because 3991 out of 7981 are binary, not text files in the required "OFF" format;
• if the meshio and torch_claster packages are not installed.

Proposed changes address both these issues.

[akihironitta]

How is it different from this?

Suggested change

	_, cached_datapipe = tee(datapipe)
	cached_datapipe = tee(datapipe, 1)

Also, can you remind me of why we need tee?

[drivanov]

1. With the change you proposed, this example fails:

<itertools._tee object at 0x7f3512666b80>
Iterating over all data...
Traceback (most recent call last):
  File "/workspace/examples/datapipe.py", line 131, in <module>
    for batch in datapipe:
  File "/usr/local/lib/python3.12/dist-packages/torch/utils/data/datapipes/_hook_iterator.py", line 204, in wrap_generator
    response = gen.send(None)
               ^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch/utils/data/datapipes/iter/grouping.py", line 97, in __iter__
    yield self.wrapper_class(batch)
          ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch_geometric/data/batch.py", line 97, in from_data_list
    batch, slice_dict, inc_dict = collate(
                                  ^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch_geometric/data/collate.py", line 56, in collate
    out = cls(_base_cls=data_list[0].__class__)  # type: ignore
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.12/dist-packages/torch_geometric/data/batch.py", line 36, in __call__
    globals()[name] = MetaResolver(name, (cls, base_cls), {})
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: type 'itertools._tee' is not an acceptable base type
This exception is thrown by __iter__ of Batcher(batch_size=32, drop_last=False)

2. I tested both approaches:

(a) With tee:

    _, cached_datapipe = tee(datapipe)
    datapipe = IterableWrapper(cached_datapipe)

(b) Without tee:

    datapipe = IterableWrapper(datapipe)

for both molecule and mesh values of the --task option.

• For molecule, there is no significant difference.
• For mesh, using version (a) results in:

Iterating over all data...
Done! [68.57s]
Iterating over all data a second time...
Done! [0.00s]

Whereas version (b) results in:

Iterating over all data...
Done! [68.50s]
Iterating over all data a second time...
Done! [67.46s]

Since version (a) significantly improves performance on the second iteration for mesh, using it for both cases would be more appropriate for consistency.

[akihironitta]

With the change you proposed, this example fails:

Sorry for the typo! What I wanted to ask is, what the difference is between:

    _, cached_datapipe = tee(datapipe)

and

    cached_datapipe, = tee(datapipe, 1)

To me, these are exactly the same, but your way just creates an extra unnecesary object.

[akihironitta]

https://docs.python.org/3/library/itertools.html#itertools.tee

[drivanov]

Yes, it works. Thank you for the great suggestion!