Training knowledge graph with one GPU

[gustavopuma]

I’m using one virtual machine with 6 CPU / 1 GPU (11GB) with 56 GB RAM . Concerning the training the dataset Wikidata5m, I tried to train the same model using the same dataset with num_partitions = auto , eipisode_size = auto , gpu_memory_limit = auto , cpu_per_gpu = auto and I able to start training only with dimensions 128 and still training after 24 hours , its expected, given the benchmark with 24 CPU 4 GPU ?

[KiddoZhu]

Could you please provide more details? Like hyperparameters output

[gustavopuma]

This is my complete python code trainning : import graphvite as gv import graphvite.application as gap app1 = gap.KnowledgeGraphApplication(dim=128,gpu_memory_limit=gv.auto,cpu_per_gpu=gv.auto) app1.load(file_name=gv.dataset.wikidata5m.train) app1.build(optimizer=gv.auto,num_partition=6,episode_size=gv.auto) app1.train(model='RotatE',margin=9) app1.save_model("rotate_wikidata5m.pkl")

I tried several option of hyperparameter and with embedded dimition 512 I'm not able even to start the trainning because of memory allocation even if I put several partitions its too long or not possible to load the and execute the trainning

[gustavopuma]

I tried as well the config file in : https://graphvite.io/docs/latest/_downloads/d86d91f69cdc0bdcdd521a92b6294306/rotate_wikidata5m.yaml

[KiddoZhu]

For 1 GPU, the best setting for acceleration is to use num_partition=1 and episode size=auto.

For embedding performance, I guess you can follow most hyperparameters in the config file above, except changing dim to 128 and changing adversarial temperature to some mild value like 2.