This repository includes the implementation of our ISSTA 2023 paper DyCL

Getting Started

Installation

1. Setup the Python Virtual Environment

   1. Download conda. First, run wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh It will download script for installing the latest version of conda.
   2. Install conda. Then run bash Miniconda3-latest-Linux-x86_64.sh for installation.
   3. Create a virtual environment with conda create -n DLCL python=3.9.
   4. Activate the created virtual environment with conda activate DLCL

2. Install TVM There are two ways to install TVM: (a) built from source; (b) install the pre-built binary.

   • (a) built from source. Please follow the instructions on their website to install from source code.
   • (b) install the pre-built binary. run pip install apache-tvm

3. Install the Library

   1. Running pip install -r requirements_tx2.txt to install the necessary library

File Structure

• adNN -The code repository that implements different dynamic neural networks used in our evaluation.
  • ./adNN/blockdrop -
  • ./adNN/ImgCaption
  • ./adNN/RANet
  • ./adNN/ShallowDeep
  • ./adNN/skipnet
• src the core implementation of DyCL.
• compile_onnx.py the Python script to compile DyNN with ONNXRuntime
• compile_tvm.py the Python script to compile DyNN with TVM
• evaluate_onnx.py the Python script evaluates the accuracy of compiling with OnnxRuntime.
• evaluate_tvm.py the Python script evaluates the accuracy of compiling with TVM.
• run_compile.sh the bash script to compile our experimental DyNNs.
• run_evaluate.sh the bash script to evaluate the correctness of our experimental DyNNs.

How to Run

First compilerse the dynamic neural network model bash run_compile.sh

After compilation, evaluating the compiled executable bash run_evaluate.sh

Reusability

DyCL offers a highly automated and user-friendly experience. The current implementation allows us to compile any DyNN using either OnnxRuntime or TVM.

1. To demonstrate this, we can refer to the load_model function in utils.py. In order to perform the compilation, we need to prepare the following components: the DyNN model instance ('model'), the source code of the model instance ('src'), the entry function for compilation ('compile_func'), an input demonstration of the DyNN ('example_x'), and a dataset for validating correctness ('test_loader').

2. Once these components are ready, we can customize the parameters for the compiler using the 'adnn_id' and compile it using either 'python compile_tvm.py --eval_id=adnn_id' or 'python compile_onnx.py --eval_id=adnn_id'."