Why does my dglke_eval not work?

[zhenshiqi1996]

When I train the TransE using the giving command and eval the model, the evaluation function dosen't work without error.

[classicsong]

Can you check cpu/gpu usage?

[vlad43210]

I have this same issue -- tried it with and without the GPU. CPU usage just quickly drops to 0 and the program hangs there with no error:

DGLBACKEND=pytorch dglke_eval \
> --data_path (data path) --dataset mydataset \
> --data_files entities.tsv relations.tsv all_ctups_2.tsv valid.tsv test.tsv --format udd_hrt \
> --model_name ComplEx \
> --hidden_dim 512 --gamma 175 \
> --num_proc 8 --num_thread 4 \
> --batch_size_eval 1024 --neg_sample_size_eval 10000 --eval_percent 5 \
> --model_path (model path)
Using backend: pytorch
Reading train triples....
Finished. Read 211968 train triples.
Reading valid triples....
Finished. Read 27026 valid triples.
Reading test triples....
Finished. Read 31796 test triples.
/opt/conda/envs/kge/lib/python3.8/site-packages/dgl/base.py:25: UserWarning: multigraph will be deprecated.DGL will treat all graphs as multigraph in the future.
  warnings.warn(msg, warn_type)
|valid|: 27026
|test|: 31796

Package versions:

- dgl==0.4.3post2
- dglke==0.1.2
- torch==1.7.1

[mtoles]

Has anyone discovered a fix? I am experiencing a similar problem.

[classicsong]

Can you provide more details?

[PoloWitty]

I meet the same issue, too. But I found it may be because the eval process is too slow.
The cmd are as followed:

dglke_eval \
    --model_name TransE \
    --data_path ~/KGE/data/BIOS \
    --dataset BIOS \
    --format raw_udd_hrt \
    --data_files train.tsv valid.tsv test.tsv \
    --hidden_dim 512 \
    --batch_size_eval 256 \
    --neg_sample_size_eval 256 \
    --model_path ~/KGE/output/TransE_BIOS_2/ \
    --gpu 0

after the process output |valid|: 8449749 |test|: 8449806, it seems that the process have died. But actually the process is still running. I don't know why the eval process is so slow as the train process can be so fast. I also tried to use --batch_size_eval 10000 --neg_sample_size_eval 10000 mentioned #106 here. But the speed is still pretty slow and the GPU util is still as slow as 10% (for A100).
I also tried to use multi-process to eval on cpu, I didn't see any improvement yet.

I wonder if there is any solution to this problem?