Paandaman / SingleDigitMNISTDetection

A classifier trained on a single sample per class from the MNIST dataset.

SingleDigitMNISTDetection

A classifier trained on a single sample per class from the MNIST dataset.

Trained using the approach described in the paper 'Unsupervised Data Augmentation for Consistency Training" in order to utilize the 'unlabeled' samples. Evaluated on a 1k subset of the MNIST test dataset (100 samples per digit).

Requirements

python3
numpy
Pillow
pyyaml
tensorboard
tensorflow
torch
torchvision

Installation

git clone https://github.com/Paandaman/SingleDigitMNISTDetection.git
cd SingleDigitMNISTDetection
#Install required packages
conda install --file requirements.txt

Train

python main.py

Keep track of training progress

cd SingleDigitMNISTDetection
python -m tensorboard.main --logdir=.

Results

1 sample per digit, no unsupervised data used:

Average Accuracy over 5 runs: 45.25%
Highest achieved accuracy : 51.9%

1 sample per digit with unsupervised data augmentation used:

Average Accuracy over 5 runs: 57.98%
Highest achieved accuracy : 64.6%