transformers VS faiss

Most of this tutorial is based on Hugging Face course about Transformers and on Niels Rogge's Transformers tutorials: make sure to check their work and give them a star on GitHub, if you please ❤️

* Superconvergence + LR range finder + Fast AI's Ranger21 optimizer was the goto optimizer for CNNs, and worked fabulously well, but on transformers, the learning rate range finder sadi 1e-3 was the best, whilst 1e-5 was better. However, the 1 cycle learning rate stuck. https://github.com/huggingface/transformers/issues/16013

Thanks! :) I'm pushing them into transformers, pytorch-gemma and collabing with the Gemma team to resolve all the issues :)

The RoPE fix should already be in transformers 4.38.2: https://github.com/huggingface/transformers/pull/29285

My main PR for transformers which fixes most of the issues (some still left): https://github.com/huggingface/transformers/pull/29402

If you want to tinker with the architecture Hugging Face has a FOSS implementation in transformers: https://github.com/huggingface/transformers/blob/main/src/tr...

If you want to reproduce the training pipeline, you couldn't do that even if you wanted to because you don't have access to thousands of A100s.

I am aware that using mixed precision reduces the stability of weight and there will be little consistency but don't expect it to be this much. I have attached the graph of evaluation metrics. If someone can give me some insight into this issue, that would be great.

In transformers, they tried really hard to have a single function or method to deal with both self and cross attention mechanisms, masking, positional and relative encodings, interpolation etc. While it allows a user to use the same function/method for any model, it has led to severe parameter bloat. Just compare the original implementation of llama by FAIR with the implementation by HF to get an idea.

pip install git+https://github.com/huggingface/transformers.git@main

Or just use FAISS https://github.com/facebookresearch/faiss

You can try txtai (https://github.com/neuml/txtai) with a Faiss backend.

This Faiss wiki article might help (https://github.com/facebookresearch/faiss/wiki/Indexing-1G-v...).

For example, a partial Faiss configuration with 4-bit PQ quantization and only using 5% of the data to train an IVF index is shown below.

faiss={"components": "IVF,PQ384x4fs", "sample": 0.05}

If you want to experiment with vector stores, you can do that locally with something like faiss which has good platform support: https://github.com/facebookresearch/faiss

Doing full retrieval-augmented generation (RAG) and getting LLMs to interpret the results has more steps but you get a lot of flexibility, and there's no standard best-practice. When you use a vector DB you get the most similar texts back (or an index integer in the case of faiss), you then feed those to an LLM like a normal prompt.

The codifer for the RAG workflow is LangChain, but their demo is substantially more complex and harder-to-use than even a homegrown implementation: https://news.ycombinator.com/item?id=36725982

According to this documentation page, faiss-gpu is only supported on Linux, not on Windows.

Indexes for vector databases in high dimensions are nowhere near are effective as the 2-d indexes used in GIS or the 1-d B-tree indexes that are commonly used in databases.

Back around 2005 I was interested in similarity search and read a lot of conference proceedings on the top and was basically depressed at the state of vector database indexes and felt that at least for the systems I was prototyping I was OK with a full scan and later in 2013 I had the assignment of getting a search engine for patents using vector embeddings in front of customers and we got performance we found acceptable with full scan.

My impression today is that the scene is not too different than it was in 2005 but I can't say I haven't missed anything. That is, you have tradeoffs between faster algorithms that miss some results and slower algorithms that are more correct.

I think it's already a competitive business. You have Pinecone which had the good fortune of starting before the gold rush. Many established databases are adding vector extension. I know so many engineering managers who love postgresql and they're just going to load a vector extension and go. My RSS reader YOShInOn uses SBERT embeddings to cluster and classify text and certainly More Like This and semantic search are on the agenda, I'd expect it to take about an hour to get

https://github.com/facebookresearch/faiss

up and working, I could spend more time stuck on some "little" front end problem like getting something to look right in Bootstrap than it would take to get working.

I can totally believe somebody could make a better vector db than what's out there but will it be better enough? A startup going through YC now could spend 2-3 to get a really good product and find customers and that is forever in a world where everybody wants to build AI applications right now.

As the size of the codebase grows, storing and searching through embeddings in memory becomes inefficient. This is where vector databases come into play. Tools like Milvus, Faiss, and others are designed to handle large-scale vector data and provide efficient similarity search capabilities. I've wrtten about how to also use sqlite to store vector embeddings. By integrating a vector database, you can scale your code search tool to handle much larger codebases without compromising on search speed.

But FAISS has their own version ("FastScan") https://github.com/facebookresearch/faiss/wiki/Fast-accumula...

https://github.com/facebookresearch/faiss

Not a model, but I would use this Dense Passage Retrieval for Open Domain QA simply fine-tuning two BERT models, one for questions and one for queries, and then fine-tuning using contrastive loss between positive key/value pairs of document embeddings (the [CLS]) token. You can then use a vector database (Like Faiss, Elasticsearch, Vespa or similar) for querying the question.