Large int inputs in Tensorflow.

This repository contains two custom Keras layers for TensorFlow that convert between Int32 and float tensors:

• Int32ToFloatLayer: A layer that converts Int32 tensors to float tensors.
• FloatToInt32Layer: A layer that converts float tensors to Int32 tensors.

The rationale behind these layers is to retain all the bit-level information of potentially very large integer inputs in a deep learning model. While this could be achieved by simply converting the numbers to Float64 and rescaling, recent architectures tend to use lower precision. Converting an Int32 to 32 floats bit-by-bit seems excessive, so the solution is to convert an Int32 into 4 floats instead.

The method involves taking input integers (e.g., an Int32) and converting them byte-by-byte into floats.

My use case involved sorting a vector of large integers, where it was essential to verify that the exact values were returned, not just an approximation.

These layers are intended for experimental use only.

To use these layers in your TensorFlow models, simply import them from the intlayer module and add them to your model like any other Keras layer.