[BUG] GraphLearn doesn't work with Python 3.10 & Python 3.11

[sighingnow]

Describe the bug

GraphLearn requires a fairly low version of tensorflow (currently 2.5.2 in our CI) which is not compatible with Python 3.10 and Python 3.11.

We'll release the wheel packages for Python 3.{10,11} to pypi and that would be a concern if user tries to play with the GLE part with recently Python distributions.

For detail error message, just trying pip3 install grpahscope>=0.19.0 tensorflow in Python 3.{10,11} environments and try the demo in README.

Additional context
Add any other context about the problem here.

[goldenleaves]

I reproduced this example with ubuntu 22.04 and Python 3.10, and got the following error info:

Traceback (most recent call last):
  File "/root/workdir/codelab/test_gle.py", line 104, in <module>
    train_gcn(lg, node_type="paper", edge_type="cites",
  File "/root/workdir/codelab/test_gle.py", line 73, in train_gcn
    train_data = EgoSAGESupervisedDataLoader(
  File "/root/workdir/pip/graphscope/graphscope/learning/examples/tf/ego_sage/ego_sage_data_loader.py", line 47, in __init__
    super().__init__(graph, mask, sampler, batch_size, window)
  File "/root/workdir/pip/graphscope/graphscope/learning/examples/tf/ego_data_loader.py", line 54, in __init__
    self._dataset = tfg.Dataset(self._q, window=window)
  File "/root/workdir/pip/graphscope/graphscope/learning/graphlearn/python/nn/tf/data/dataset.py", line 64, in __init__
    self._iterator = self._make_iterator()
  File "/root/workdir/pip/graphscope/graphscope/learning/graphlearn/python/nn/tf/data/dataset.py", line 167, in _make_iterator
    output_types, output_shapes = self._raw_flatten_types_and_shapes()
  File "/root/workdir/pip/graphscope/graphscope/learning/graphlearn/python/nn/tf/data/dataset.py", line 183, in _raw_flatten_types_and_shapes
    types, shapes = self._data_types_and_shapes(node.decoder, 
  File "/root/workdir/pip/graphscope/graphscope/learning/graphlearn/python/nn/tf/data/dataset.py", line 236, in _data_types_and_shapes
    feat_shapes = np.array([tf.TensorShape([None, node_decoder.int_attr_num]),
ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (5,) + inhomogeneous part.

Is this consistent with the problem you encountered? @sighingnow

[goldenleaves]

I have located the problem and will fix it as soon as possible

[sighingnow]

After fixing the above error, new tensorflow related error would appear.

Thanks!

[goldenleaves]

@sighingnow
Another error is located in https://github.com/alibaba/graph-learn/blob/711f67796162ab6ade24cd6ebac504488150b381/graphlearn/python/nn/dataset.py#L97
and the related error info is:

Traceback (most recent call last):

  File "/root/workdir/pip/tensorflow/tensorflow/python/ops/script_ops.py", line 271, in __call__
    ret = func(*args)

  File "/root/workdir/pip/tensorflow/tensorflow/python/autograph/impl/api.py", line 642, in wrapper
    return func(*args, **kwargs)

  File "/root/workdir/pip/tensorflow/tensorflow/python/data/ops/dataset_ops.py", line 1039, in generator_py_func
    values = next(generator_state.get_iterator(iterator_id))

  File "/root/workdir/pip/graphscope/graphscope/learning/graphlearn/python/nn/tf/data/dataset.py", line 193, in _raw_flatten_generator
    yield tuple(self._rds.get_flatten_values())

  File "/root/workdir/pip/graphscope/graphscope/learning/graphlearn/python/nn/dataset.py", line 104, in get_flatten_values
    res.extend(parse_value(value, masks))

  File "/root/workdir/pip/graphscope/graphscope/learning/graphlearn/python/nn/dataset.py", line 97, in parse_value
    return list(np.array(values)[feat_masks + id_masks + sparse_masks])

ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (10,) + inhomogeneous part.

I will fix it together.

[goldenleaves]

fixed with alibaba/graph-learn#249