shannonai / fast-knn-nmt Goto Github PK

Why is this done here?

In the repo we have fairseq files in fast-knn-nmt/fast_knn_nmt/custom_fairseq/ as well as fast-knn-nmt/thirdparty/fairseq/ the files/changes to, for example, the sequence_scorer.py differ in these two. Which one is the correct one to use and why do the other ones exist?

I am encountering following issue during k nearest neighbors script run, can you please help

The transformer_layer.py also has changes regarding the ret_ffn_inp flag that is introduced here which aren't listed in the readme.

This isn't really related to your code but: Since the WMT dataset for the vanilla knn-MT has a pretty big datastore (I think in the ballpark of 900M keys), the corresponding FAISS index is quite big and loading it to a single GPU together with the model is a pretty big issue. Naturally, you'd want to shard the index over multiple GPUs but FAISS has some problems searching over a sharded index with GPU tensors, see: facebookresearch/faiss#2074 or facebookresearch/faiss#1997. Also loading it to a separate GPU, transferring tensors there for searching, and moving the results back, runs into some memory deallocation issues inside FAISS.

I wonder how you ran the experiments on that or does your codebase has some neat trick that I didn't catch yet?

From my understanding you don't shard as well, according to this line and then you also transfer the tensors you search on to CPU here (you can pass GPU tensors btw if you import faiss.contrib.torch_utils that changes the function signatures to accept GPU tensors instead of np arrays). Does your index for WMT with vanilla knn-MT just fit in GPU memory?

For the domain dataset, the datastores are relatively small so this isn't a big issue and I got them set-up just fine!