this dataset was created from mnist_tiny dataset, by randomly resizing it to slighly more than 28x28 and padding with black bg - so the resulting images are somewhere between 28x28 and 35x42.
wget http://files.fast.ai/data/examples/mnist_tiny.tgz
tar -xvzf mnist_tiny.tgz
find mnist_tiny -type f -name "*png" -exec perl -le '$f=shift; $w=28+int rand 7; $h=28+int rand 14; $s="${w}x${h}"; qx[mogrify -resize $s -background black -gravity center -extent $s $f]' {} \;
mv mnist_tiny mnist_var_size_tiny
rm mnist_var_size_tiny/models/tmp.pth
tar -cvzf mnist_var_size_tiny.tgz mnist_var_size_tiny
It’s on s3 - thanks Jeremy!
‘https://s3.amazonaws.com/fast-ai-imageclas/mnist_var_size_tiny.tgz’
and it’s in fastai datasets under URLs.MNIST_VAR_SIZE_TINY
now we can start using it in the tests