The doc_enc library is devoted to the computation of cross-lingual vector representations of long texts' embeddings applicable in information retrieval and classification tasks.

Full documentation in Russian

Quick start:

First of all, you should download a small dataset of documents and a pre-trained model:

curl -O dn11.isa.ru:8080/doc-enc-data/datasets.docs.mini.v1.tar.gz
tar xf datasets.docs.mini.v1.tar.gz
find docs-mini/texts/ -name "*.txt" > files.txt
curl -O http://dn11.isa.ru:8080/doc-enc-data/models.def.pt

The data is available for downloading by the link, in case the above links are not working. To start the conversion process, it is convenient to use a pre-built Docker image. Before doing so, ensure that you have installed the NVIDIA Container Toolkit by following the instructions provided in the NVIDIA Container Toolkit installation guide.

docker run  --gpus=1  --rm  -v $(pwd):/temp/ -w /temp  \
  semvectors/doc_enc:0.1.2 \
  docenccli docs -i /temp/files.txt -o /temp/vecs -m /temp/models.def.pt

Vectors will be stored in the vecs directory alongside their corresponding file names using the numpy.savez function. Below is an example of how to load the vectors from these files:

import numpy as np

obj = np.load('vecs/0000.npz')
print(obj['ids'][:2])
print(obj['embs'][:2])