An experiment configuration starts with an
application type, and contains settings
The stages are configured as follows.
The application type can be
knowledge graph or
resource: gpus: [list of GPU ids] gpu_memory_limit: [limit for each GPU in bytes] cpu_per_gpu: [CPU thread per GPU] dim: [dim]
For optimal performance, modules are compiled with pre-defined dimensions in C++. As a drawback, only dimensions that are powers of 2 are supported in the library.
format: delimiters: [string of delimiter characters] comment: [prefix of comment strings]
Format section is optional. By default, delimiters are any blank character and comment is “#”, following the Python style.
graph: file_name: [file name] as_undirected: [symmetrize the graph or not]
For standard datasets, we can specify its file name by
This would make the configuration file independent of the path.
build: optimizer: type: [type] lr: [learning rate] weight_decay: [weight decay] schedule: [learning rate schedule] # and other optimizer-specific configuration num_partition: [number of partitions] num_negative: [number of negative samples] batch_size: [batch size] episode_size: [episode size]
The number of partitions determines how to deal with multi-GPU or large graph cases. The more partitions, the less GPU memory consumption and speed. The episode size controls the synchronization frequency across partitions.
See section 3.2 in GraphVite paper for a detailed illustration.
load: file_name: [file name]
Loading a model is optional.
train: model: [model] num_epoch: [number of epochs] resume: [resume training or not] log_frequency: [log frequency in batches] # and other application-specific configuration
To resume training from a loaded model, set
resume to true in
evaluate: - task: [task] # and other task-specific configuration - task: [task] ...
Evaluation is optional. There may be multiple evaluation tasks.
save: file_name: [file name] save_hyperparameter: [save hyperparameters or not]
Saving the model is optional.
For more detailed settings, we recommend to read the baseline configurations
for concrete examples. They can be found under
config/ in the Python package,
or in the GitHub repository.
We can overwrite the global settings of GraphVite in
backend: [graphvite or torch] dataset_path: [path to store downloaded datasets] float_type: [default float type] index_type: [default index type]
By default, the evaluation backend is
graphvite. The datasets are stored in
~/.graphvite/dataset. The data types are