sentence-transformers VS hummingbird

txtai is an open-source first system. Given it's own open-source roots, like-minded projects such as sentence-transformers are prioritized during development. But that doesn't mean txtai can't work with Embeddings API services.

Ok great. My use case is not very specific, but rather general. I am looking for a model that can perform asymmetric semantic search for the languages I mentioned earlier (Urdu, Persian, Arabic etc.). I have also looked into the sentence-transformer training documentation. Do you think it would be a good idea to use the XNLI dataset for fine-tuning? Or maybe you can suggest much better dataset. Furthermore, I am not sure if fine-tuning is suitable for my task. Because my use case is general so I can use already trained model.

The task you are looking for is semantic textual similarity. There are a few models and datasets out there that can do this. I'd probably start with the SemEval2017 Task 1 task description and competition entries here and then work outward from there (using something like SemanticScholar or Papers With Code to find newer state of the art works that cite these models if needed). For what it's worth you might find that Sentence Bert (SBERT) gives good vectors for cosine similarity comparison out of the box for this task.

Check out the sentence-transformers implementation. If I don't miss anything they don't exclude CLS when the pooling strategy is set to 'mean'

Break up transcript into shorter segments and convert segments to a 768 vector array. Use a process known as embedding using our second ML model, UKP Labs BERT’s sentence transformer model.

Not sure what kind of texts you have, but these models have a max sequence limit of 512 (approx 350 words or so). If you're texts are longer than that, consider splitting them up into chunks or creating a summary and taking an embedding of that. Some clustering algorithm may be the way to go here. Here's a bunch of examples. I use agglomerative for my use case.

https://github.com/UKPLab/sentence-transformers/tree/master/examples/training/data_augmentation (Augmented Encoding)

For vector representation you can take the mean and then pca to get the size that you want, but if you have time then use sentence transformers to train a vector representation instead.

If are interested in just predictions you can try Hummingbird. It is part of the PyTorch ecosystem. We get already trained scikit-learn models and translate them into PyTorch models. From them you can run your model on any hardware support by PyTorch, export it into TVM, ONNX, etc. Performance on hardware acceleration is quite good (orders of magnitude better than scikit-learn is some cases)

I think Rapids AI's cuML tried to go into this direction (essentially scikit-learn on the GPU): https://docs.rapids.ai/api/cuml/stable/api.html#logistic-reg.... For some reason it never took really off though.

Btw., going on a tangent, you might like Hummingbird (https://github.com/microsoft/hummingbird). It allows you trained scikit-learn tree-based models to PyTorch. I watched the SciPy talk last year, and it's a super smart & elegant idea.

ONNX opens an avenue for direct inference using a number of languages and platforms. For example, a model could be run directly on Android to limit data sent to a third party service. ONNX is an exciting development with a lot of promise. Microsoft has also released Hummingbird which enables exporting traditional models (sklearn, decision trees, logistical regression..) to ONNX.

And Python.

For inference, you can convert your models to other formats that support GPU acceleration. See Hummingbird https://github.com/microsoft/hummingbird

The surprising thing is that Hummingbird can be faster than the GPU implementation of LightGBM (and XGBoost) if you use tensor compilers such as TVM. [The paper](https://www.usenix.org/conference/osdi20/presentation/nakandala) describes our findings. We have also open sourced the [benchmark code](https://github.com/microsoft/hummingbird/tree/main/benchmarks) so you try yourself!

I took their sample code from Github and tweaked it to spit out times for each model's prediction, as well as increase the number of rows to 5 million. I used Google's Colab and selected GPU for my hardware accelerator. This gives an option to run code on GPU, not that all computations will happen on the GPU.