AttributeError: 'Column' object has no attribute `deferred_dtype`

[mufeili]

🐛 Bug

@ndickson-nvidia @BarclayII

The introduction of deferred_dtype in PR #3890 seems to be causing issues when loading a graph file saved with an older version of DGL. Do you have any ideas on fixing the compatibility? We might also want to add a unit test for loading a graph data file saved with an older version of DGL.

Expected behavior

Environment

• DGL Version (e.g., 1.0): 0.8.2
• Backend Library & Version (e.g., PyTorch 0.4.1, MXNet/Gluon 1.3):
• OS (e.g., Linux):
• How you installed DGL (conda, pip, source):
• Build command you used (if compiling from source):
• Python version:
• CUDA/cuDNN version (if applicable):
• GPU models and configuration (e.g. V100):
• Any other relevant information:

[nv-dlasalle]

Do we see the same issue for pinned_by_dgl as well? Or do we have an established method for handling new attributes?

[ndickson-nvidia]

Are there any plans to add a more robust system for saving/loading graph data? Relying on just serializing Python objects, like what dgl.data.utils.save_info and dgl.data.utils.load_info do, seems like a risky approach that may not be able to support any sort of compatibility, especially since there's no version information. I'd expect that a few data members of Column, for example, never need to be saved or loaded, since device, index, and deferred_dtype should be resolved and cleared before saving to disk, and there are probably other similar examples in other classes.

[LQ999666]

博主，您这个问题怎么解决的

[LQ999666]

Blogger, how did you solve this problem

[mufeili]

Blogger, how did you solve this problem

For now, downgrading to 0.8.1 should address the issue.

[LQ999666]

thank you ,my problem has solved.

[natalie-0430]

Hello, how can you downgrade to 0.8.1 from 0.8.2? I have a similar problem.

[nv-dlasalle]

@natalie-0430 That will depend on how you installed it. If you used pip for example:

pip uninstall dgl-cu113
pip install dgl-cu113==0.8.1 dglgo -f https://data.dgl.ai/wheels/repo.html

@mufeili What's your preferred fix for this? Should we add hasattr to check if the unpickled object has the new attribute, or should we override __setstate__ (https://docs.python.org/3/library/pickle.html#object.__setstate__) to explicitly populate missing attributes from old versions?

[mufeili]

@natalie-0430 That will depend on how you installed it. If you used pip for example:

pip uninstall dgl-cu113
pip install dgl-cu113==0.8.1 dglgo -f https://data.dgl.ai/wheels/repo.html

@mufeili What's your preferred fix for this? Should we add hasattr to check if the unpickled object has the new attribute, or should we override __setstate__ (https://docs.python.org/3/library/pickle.html#object.__setstate__) to explicitly populate missing attributes from old versions?

@jermainewang @BarclayII What do you think? Probably overriding __setstate__ will be more general for future changes?

[jermainewang]

I vote for overriding __setstate__ and __getstate__. Two other related questions:

• Shall we always make pickling/unpickling backward compatible? This may prevent us from making changes to DGLGraph and Frame.
• How to properly test pickling/unpickling? This is somewhat related to the question above.