mnist_symmetries / README.md
ACDRepo's picture
Upload 3 files
3351716 verified
|
Raw History Blame Contribute Delete
1.21 kB
metadata
tags:
  - mnist
  - image-classification
  - computer-vision
task_categories:
  - image-classification
pretty_name: Transformed MNIST 48x48 (classes 6 and 9 removed)
size_categories:
  - 10K<n<100K

Transformed MNIST 48x48

MNIST images on a 48x48 canvas with extra margin. Classes 6 and 9 were removed. Each image received one reproducible random transform set: rotation, optional horizontal reflection, and a safe random translation that keeps the foreground away from the edge.

Splits

  • train: 48,133 examples
  • test: 8,033 examples
  • Labels: 0, 1, 2, 3, 4, 5, 7, 8

The label column uses contiguous IDs 0–7 with this mapping: 0->0, 1->1, 2->2, 3->3, 4->4, 5->5, 6->7, 7->8.

Features

  • image: grayscale 48x48 PNG bytes
  • label: contiguous class ID
  • rotation_degrees, translate_x, translate_y, reflected: transform metadata

Load

from datasets import load_dataset, Image
ds = load_dataset("parquet", data_files={
    "train": "data/train-00000-of-00001.parquet",
    "test": "data/test-00000-of-00001.parquet",
})
ds = ds.cast_column("image", Image())

The original MNIST data was downloaded from the official MNIST mirror and filtered before transformation.