https://github.com/citiususc/ho-mnist

Hardware Oriented CNN for MNIST recognition
https://github.com/citiususc/ho-mnist

Last synced: about 1 year ago
JSON representation

Hardware Oriented CNN for MNIST recognition

• Host: GitHub
• URL: https://github.com/citiususc/ho-mnist
• Owner: citiususc
• Created: 2023-04-05T11:55:34.000Z (about 3 years ago)
• Default Branch: main
• Last Pushed: 2023-04-05T16:25:22.000Z (about 3 years ago)
• Last Synced: 2025-01-29T19:23:56.088Z (over 1 year ago)
• Language: Jupyter Notebook
• Size: 682 KB
• Stars: 2
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# HO-MNIST

Hardware Oriented CNN for MNIST recognition.

![HO-MNIST experiment setup](img/setup.jpg)

## Online demo

You can try an online simulation of our CIM macro! [![Binder](https://mybinder.org/badge_logo.svg)](https://mybinder.org/v2/gh/citiususc/ho-mnist/HEAD?labpath=Demo.ipynb) [![Google Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/citiususc/ho-mnist/blob/master/Demo.ipynb) 

## Creation of the Hardware Oriented model

### Train model

```shell

python3 ./train.py --output-path=./models/mnist_base.pth

```

### Quantize model

```shell

python3 ./quantize.py --model-path=./models/mnist_base.pth --output-path=./models/mnist_quantized.pth

```

### Convert model to int8 and store scales

```shell

python3 ./convert.py --model-path=./models/mnist_quantized.pth --output-path=./models/mnist_quantized_converted.pth

```

## Visualization of the 5-bit kernels

```shell

python3 ./display_weights.py --model-path=./models/mnist_quantized_converted.pth --chip-format

```

## Testing of the model

```shell

python3 ./test.py --test-base --model-path=./models/mnist_base.pth  # Accuracy: 97.56%

python3 ./test.py --model-path=./models/mnist_quantized_converted.pth  # Accuracy: 90.57%

python3 ./test.py --test-hardware --model-path=./models/mnist_hacc --serial-port=/dev/ttyUSB2  # Accuracy: 91.89%

```

## Live demo

```shell

python3 ./demo.py --model-path=./models/mnist_quantized_converted.pth --live

python3 ./demo.py --hardware --model-path=./models/mnist_hacc --serial-port=/dev/ttyUSB2 --live

```