is v0.2 available now? about colbert

[okhat]

Hi @DreamsofGg !

You can use this:

CUDA_VISIBLE_DEVICES="0,1,2,3" python -m torch.distributed.launch --nproc_per_node=4 -m \
colbert.train --amp --doc_maxlen 180 --mask-punctuation --bsize 32 --accum 1 \
--triples /path/to/triples.tsv --similarity l2 \
--root /root/to/log/experiments/data/   --experiment ExperimentName

You can modify the number of GPUs (e.g., CUDA_VISIBLE_DEVICES="0" and --nproc_per_node=1) as needed.

Let me know if you face any issues!

[DreamsofGg]

Wow! your command just need half time to finish training! (Also with a single gpu) Great thanks!

[DreamsofGg]

@okhat sorry to bother again. I found that the similarity used in the paper is cosine_sim, so how l2_sim influence the results?

[okhat]

Actually both are mentioned in the paper, see Table 1 vs. Table 2.

There shouldn't be a major difference in performance; feel free to use either one.