Fix DMRL save() failing when log_metrics=True

[Shxiao101]

Description

DMRL(log_metrics=True) cannot be saved, as reported in #656.

DMRL.__init__ created a tensorboard SummaryWriter and stored it as self.tb_writer. Recommender.save() calls copy.deepcopy(self), and Recommender.__deepcopy__ deep-copies every attribute except ignored_attrs. The writer holds thread locks and an open event file, so saving fails:

• TypeError: cannot pickle '_thread.lock' object (plain PyTorch/tensorboard)
• cannot pickle 'tensorflow.python.lib.io._pywrap_file_io.WritableFile', the variant reported in [BUG] i am not able to do save dmrl.save() in Disentangled Multimodal Representation Learning for Recommendation (DMRL) #656 (on Google Colab, where TensorFlow is installed; I only reproduced the plain PyTorch/tensorboard error above)

A side effect: just constructing DMRL(log_metrics=True) created temp/tb_data/run_1 in the working directory, even if the model was never trained.

The writer is only used inside _fit_dmrl. This PR creates it there as a local variable, right before logging the hyperparameters, and closes it when training finishes. The log directory and the logged scalars/histograms are unchanged.

I considered adding tb_writer to ignored_attrs. It would fix save(), but a loaded model trained again with log_metrics=True would then have no writer, and __init__ would still create the directory.

How I tested

Per review, the save/load test that was in the first version of this PR has been removed, so the PR now only changes cornac/models/dmrl/recom_dmrl.py.

I verified the fix with a local script (Windows, Python 3.12, torch 2.14.0+cpu, sentence-transformers 6.0.1, cornac built from source). It trains a tiny DMRL on random feedback plus pre-encoded image features, with log_metrics=True and no model download, then calls save() and Recommender.load():

check	master	this PR
temp/ created by DMRL(log_metrics=True) before fit	yes	no (created during fit)
save() after fit	TypeError: cannot pickle '_thread.lock' object	succeeds
parameters after Recommender.load()	—	identical to the trained model
pytest tests/cornac/models	10 passed, 3 failed	10 passed, 3 failed

The 3 failures are the tests/cornac/models/dmrl/test_transformertext.py tests. They need to download a Hugging Face model and I ran offline, so they fail the same way on master.

check	master	this PR
flake8 . --select=E9,F63,F7,F82 (as in CI)	—	0

Related Issues

Fixes #656

Checklist:

• I have added tests. (a save/load test was dropped from this PR per review)
• I have updated the documentation accordingly. (N/A, no API change)
• I have updated README.md (if you are adding a new model). (N/A)
• I have updated examples/README.md (if you are adding a new example). (N/A)
• I have updated datasets/README.md (if you are adding a new dataset). (N/A)

🤖 Generated with Claude Code

[qtuantruong]

The fix looks good to me. Though we should remove the test file from this PR. Thanks

[Shxiao101]

Thanks for the review! Done. I pushed a commit that removes tests/cornac/models/dmrl/test_recom_dmrl.py, so the PR now only changes cornac/models/dmrl/recom_dmrl.py.

I updated the description. The fix is still verified by a local script (fit with log_metrics=True, then save() and Recommender.load()): it fails on master with cannot pickle '_thread.lock' object and passes with this change.