Skip to content

Latest commit

 

History

History

data

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 

TorchBench input data set

This directory manages a set of minimal input data and models used by the core model set in the pytorch/benchmark repository.

How to download the dataset

The full list is stored in index.yaml. It contains two categories of inputs: INPUT_TARBALLS and MODEL_PKLS.

To download an input tarball, use the following URL:

https://ossci-datasets.s3.amazonaws.com/torchbench/data/<TARBALL_NAME>

For example, the URL to download multi30k.tar.gz is https://ossci-datasets.s3.amazonaws.com/torchbench/data/multi30k.tar.gz.

To download a model pkl, use the following URL:

https://ossci-datasets.s3.amazonaws.com/torchbench/models/<MODEL_PKL_NAME>

For example, the URL to download drq/obs.pkl is https://ossci-datasets.s3.amazonaws.com/torchbench/models/drq/obs.pkl.

How to contribute

TorchBench core model set runs on a single batch of data repeatedly to collect performance metrics. Therefore, we only accept minimal datasets which only contains less than 10 batches of input data. The total input data size should be smaller than 20 MB.

If you would like to contribute new dataset, please submit a GitHub issue.