This project aims to make a compiler with optimization and code generation that can take neural network inference requests and divide the load between classical hardware (CPU/ GPU) and novel photonic hardware. We take advantage of the TVM compiler to translate Tensorflow and Pytorch models into their internal Relay IR, at which point this compiler conducts the next layer of scheduling and translation.
-
inference_pratice: recreating popular modles from scrach using numpy
- submodule for GPT2 Debugging huggingface_transformers
- LeNet
- numpy LeNet lenet.ipynb
- GPT2
- numpy GPT2 np_gpt2.py
- numpy to Transformer model GPT2 comparison gpt2_compairson.py
- Trace function calls of numpy GPT2 model_trace.py
- Tensor product broken to dotproducts opps.ipynb
-
json_parse: IN PROGRESS. Everything related to code generation and parsing of TVM Relay IR .json files
- parser.py: script for instruction generation and file structure
- simple_LeNet_parsed.txt: generated instructions
-
ONNX-ResNet: model files/parameters for ResNet ML model in the ONNX format
-
Pytorch-LeNet: model files/parameters and code for loading LeNet from pytorch
-
Transformer-GPT2 Compile GPT2 to Relay IR
- Relay Graph.json tooling GPT2_tooling.ipynb
- Compule GPT2 to Relay IR transformer_GPT2.py