Papers such as Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference have interested me in exploring integer-based data science. In particular, I'm thinking of econometric analysis and deep learning for time series data on modern CPUs (with AVX vectorization, etc), using mainly integer-based math for feature engineering (including scaling/normalization), weights, training, and inference. My main areas of application would be classification, anomaly detection, and forecasting. There would likely be a tradeoff of accuracy for runtime performance, which is specifically the area I want to explore.

I am aware of practices such as training with floating-point values then exporting to integer-based inference (as mentioned in the paper), but I am looking rather for avenues which would allow me to utilize common frameworks to perform the entire ML lifecycle (or most of it) with integers. That said, I don't mind dropping down to python or c++ to override some pluggable steps in the process, if 80-90% of the job is still handled by the framework.

I understand that TensorFlow utilizes the eigen library for linear algebra and that (on its own) this lib supports integer-based arrays & structs. However, I couldn't find any resources online showing (for example) the creation and training of a RNN or a 1d CNN on tabular data that didn't assume the use of floating-point numbers. Perhaps it is as easy as declaring integer-based primary structures and then proceeding from there with framework-provided methods? Or, maybe the community is aware of other frameworks with better support for this venture. Which frameworks/libs best support integer-based data science?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.