Zalando's MNIST fashion replacement

Zalando recently released an MNIST replacement. The issues with using MNIST are known but you can read about the dataset and their motivation here.

Training

python train.py

        --model         # specify model, (FashionSimpleNet, resnet18)
        --patience      # early stopping
        --batch_size    # batch size
        --nepochs       # max epochs
        --nocuda        # no cuda
        --nworkers      # number of workers
        --seed          # random seed
        --data          # mnist, fasion

Preprocessing

Normalization, random horizontal flip, random vertical flip, random translation, random rotation.

Results

Best accuracy of the same model when run on MNIST and Fashion MNIST.

Model	MNIST	Fashion MNIST
ResNet18	0.995	0.949
SimpleNet	0.994	0.919