The code can be installed with pip install -e .
In the main directory, run ./download_data.sh to get a wikipedia text dump.
Then, embeddings can be computed via embedding compute, and standard evaluation metrics can be run with embedding evaluate.
- The current release of PyTorch has a memory leak for sparse matrix multiplies and QR decomposition on the GPU. These issues are fixed in PyTorch now, and the new version can be obtained by building PyTorch from source.
- PyTorch sparse matrix multiply on a CPU is very slow. Because of this, GPU use is recommended.