checkpoints.md

Checkpoints for the experiments in the paper

Checkpoints (trained on 10M vectors) are available for various models presented in the paper (see links below). 8x8 denotes 8 bytes models and 16x8 denotes 16 bytes encoding. L indicates the number of residual blocks, i.e. the capapacity of the model (see Fig. 4 in the paper). To use the checkpoints for encoding and decoding, the code assumes that they are stored in a models/ subdirectory.

8 bytes:

• bigann: bigann_8x8_L2 bigann_8x8_L4 bigann_8x8_L16

• deep: deep_8x8_L2 deep_8x8_L4 deep_8x8_L16

• fb-ssnpp: FB_ssnpp_8x8_L2 FB_ssnpp_8x8_L4 FB_ssnpp_8x8_L16

• Contriever : Contriever_8x8_L2 Contriever_8x8_L12

16 bytes:

• bigann: bigann_16x8_L2 bigann_16x8_L4 bigann_16x8_L16
• deep: deep_16x8_L2 deep_16x8_L4 deep_16x8_L16
• fb-ssnpp: FB_ssnpp_16x8_L2 FB_ssnpp_16x8_L4 FB_ssnpp_16x8_L16
• Contriever: Contriever_16x8_L12

The examples below assume there is a GPU supported by Pytorch present on the machine (see installation) However, it is possible to run it on cpu as well by adding --device cpu to the command lines. It also assumes that you have the dataset to be tested installed, see [downloading data][downloading_data.md].

Encoding and decoding

To encode a set of vectors using one of the downloaded checkpoints (in this case bigann_8x8_L4.pt):

python codec_qinco.py \
    --model models/bigann_8x8_L4.pt --encode \
    --i data/bigann/bigann_query.bvecs \
    --o tmp/codes.npy

This outputs this gist.

This example loads the queries of BigANN and encodes them to an array of codes. The script can also take .fvecs or .npy files as input.

This also reports the MSE on the set of encoded vectors.

To decode:

python codec_qinco.py \
    --model models/bigann_8x8_L4.pt --decode \
    --i tmp/codes.npy \
    --o tmp/decoded.npy

Which yields this result, where we also verify the MSE (note that we compute MSE here on the queries (for speeding up computation), rather than on the large database as done in the paper).

The compressed representation is still represented as an int32 array so the .npy file size is much larger than expected. To get a file without overheads, add option --raw at encoding and decoding time.

For example:

python codec_qinco.py \
    --model models/bigann_8x8_L4.pt --encode \
    --i data/bigann/bigann_query.bvecs \
    --o tmp/codes.raw --raw

produces a codes.raw file of exactly 80000 bytes (see here)

Evaluating on one of the paper's datasets

python -u eval_qinco.py --db bigann1M --model models/bigann_8x8_L16.pt

This outputs this text, which reproduces the entry for BigANN1M 8 bytes in Table 1 of the paper.