My Code Challenge Submission for my GSoC Proposal for the University of Alaska Anchorage [2] Support for Logarithmic Number Systems in a Deep-Learning Framework for GSoC (Code Challenge)
Challenge 1: Demonstrated the inability to convert directly from xln (Logarithmic Number System) format to PyTorch tensor or Tensorflow tensor and perform basic calculations. Note: This can still be achieved in NumPy
Challenge 2: Plan to implement a multilayer perceptron in PyTorch with the flow: 28*28 input -> 100 hidden nodes -> ReLU -> 10 output nodes -> softmax and compare performance with arn_generic.py
Challenge 3: Use PyTorch tensors to create a toy-example that performs the same way as the overloaded addition operation does for xlnsnp arrays.
Other things I tried:
- Creating a version of arn_generic.py using only numpy and then inserting xlns weights into the system and running the code