mnist_hls

Lenet for MNIST handwritten digit recognition using Vivado hls tool.

• Accuracy : 97.5938% (8-bit Mode).
• Speed : 71.43 FPS (100MHz, Zedboard)
• The complete project is provided, please download it from Baidu net disk, key:1bv4

requirement

• Vivado hls 2019.2
• petalinux 2019.2 (optional)
• Vitis 2019.2
• python 3+

usage

• Download MNIST dataset from MNIST and cpoy the dataset to '~/Vivado_hls/MNIST_DATA/'

• build hls project

# When you are in the project root directory
cd Vivado_hls
vivado_hls -f script.tcl

• Generate 'tanh' table (default: ap_fixed<8,3>)

# When you are in the project root directory
cd Vivado_hls/tools
python ./tanh_table.py

• build vivado project

# When you are in the project root directory
cd Vivado
vivado -source source/tcl

• Generate output product && export hardware.

• Open Vitis && create platform from *.xsa generated by vivado.

• Create an application project based on the 'hello world' template.

• Use this file './Vitis/helloworld.c' replace the original file and add file './Vitis/mnist_data.h'.

• Run and see the results

TODO

• Simplify lookup table of 'tanh'.
• Complete PS design based on embedded linux.