metadata

task_categories:
  - image-classification
  - unconditional-image-generation
pretty_name: Easy MNIST
size_categories:
  - 10K<n<100K

Easy MNIST

MNIST processed into three easy to use formats. Each .zip file contains a labels_and_paths.csv file and a data directory.

mnist_png.zip

MNIST in the png format.

       label            path
0          5      data/0.png
1          0      data/1.png
2          4      data/2.png
3          1      data/3.png
4          9      data/4.png
...      ...             ...
69995      2  data/69995.png
69996      3  data/69996.png
69997      4  data/69997.png
69998      5  data/69998.png
69999      6  data/69999.png

mnist_numpy.zip

MNIST in the npy format.

       label            path
0          5      data/0.npy
1          0      data/1.npy
2          4      data/2.npy
3          1      data/3.npy
4          9      data/4.npy
...      ...             ...
69995      2  data/69995.npy
69996      3  data/69996.npy
69997      4  data/69997.npy
69998      5  data/69998.npy
69999      6  data/69999.npy

mnist_numpy_flat.zip

MNIST in the npy format, flattened to 784 dimensional vectors.

       label            path
0          5      data/0.npy
1          0      data/1.npy
2          4      data/2.npy
3          1      data/3.npy
4          9      data/4.npy
...      ...             ...
69995      2  data/69995.npy
69996      3  data/69996.npy
69997      4  data/69997.npy
69998      5  data/69998.npy
69999      6  data/69999.npy

Acknowledgements

Yann LeCun, Courant Institute, NYU
Corinna Cortes, Google Labs, New York
Christopher J.C. Burges, Microsoft Research, Redmond