FashionMNIST-CNN-Accelerator

An FPGA accelerator for CNN on the Fashion-MNIST dataset using HLS. The aim of this project is to accelerate the classification task for image recognition on the FashionMNIST dataset. Due to the time limitation of the project duration, our model will only include:

1. CONV
2. RELU
3. POOL
4. FC

This is a course project for COMP4601 @UNSW.

Project Planning

This section is bound to be removed after completion of project. https://www.notion.so/FasionMNIST-CNN-Accelerator-978da2b835f0482b9ddab8841dcc0221

Repo structure:

• cnn_accelerator/ - Includes the C++ source code for FashionMNIST-CNN and Vivado HLS tcl script and directives.
• cnn_training/ - Includes the CNN training code that generates the parameters used in the layers in FashionMNIST-CNN
• data/ - Includes the training and test data sets for MNIST and FashionMNIST

Quick start

Weight training and extraction

Since we only want to accelerate the classification process, we have a seperate program for training the model and saving the trained parameters to a weights.txt file. The getParameters.py script will scrap the trained weights and produce a weights.h header file saved in the cnn_accelerator/include. This will be compiled to the classification program.

cd path/to/FashionMNIST-CNN-Accelerator/
cd cnn_training
./train_parameters.sh

CNN Classification Accelerator

Resources

• Fashion-MNIST repository: link
• Tensorflow keras tutorial on classifying FasionMNIST: link
• simple-cnn repository: link