This repository provides an implementation guide for deploying PyTorch models using the Vitis AI framework on AMD KV260 evaluation boards. The design process outlined here guides users through the steps necessary to convert a Python description of a neural network model to a compiled model that can be executed on AMD KV260 hardware.
Baseline Implementation:
The baseline implementation presented here is derived from the AMD Vitis AI tutorial. While the core concepts remain similar, the implementation has been tailored to suit the AMD KV260 hardware.
The process involves several key steps:
-
Model Design: Begin by designing your neural network model in PyTorch. This includes defining the architecture, specifying layers, and configuring parameters.
-
Quantization: Perform quantization to convert the model's parameters and activations to lower precision. This step is crucial for optimizing the model to run efficiently on FPGA hardware.
-
Compilation: Utilize the Vitis AI compiler to compile the quantized model. This step transforms the model into a Xmodel format compatible with the FPGA DPU.
-
Deployment: Finally, deploy the compiled model onto the Kria KV260 evaluation board. This step involves loading the model onto the board and executing app_mt.py on target device.
app_mt.py: This Python script utilizes Xilinx's Vitis AI runtime (VART) to deploy a LeNet-5 model on Kria KV260 DPU (Deep Learning Processing Unit) for image classification tasks. It takes an image directory as input along with parameters for the number of threads and the path to the model file. The main function parses command-line arguments for these inputs, preprocesses the images, runs inference using multiple threads, calculates throughput (frames per second), performs post-processing to determine classification accuracy, and finally prints the results. The model inference is performed asynchronously using threading to improve performance.
To begin, ensure you have the following prerequisites installed:
- Python (with PyTorch)
- Vitis AI
- AMD Kria KV260 evaluation board setup
Follow these steps to run the compiled model on your evaluation board:
-
Clone Repository: Clone this repository to your local machine.
-
Model Design: Design your neural network model using PyTorch. Ensure that the model architecture is compatible with Vitis AI requirements.
-
Quantization: Perform quantization on the model using the provided scripts or customize them to suit your specific requirements.
-
Compilation: Use the Vitis AI compiler to compile the quantized model. Refer to the documentation for detailed instructions on how to compile models for your target hardware.
-
Deployment: Deploy the compiled model onto your AMD KV260 evaluation board. Follow the provided instructions to load the model onto the board and execute inference tasks.
Contributions to this repository are welcome. If you have suggestions for improvements or encounter any issues, please feel free to open an issue or submit a pull request.