2.5 KiB
JSON for MNIST
This repo contains two JSON files, train.json and test.json, which
are a human-readable reproduction of the training and testing data from
the MNIST database. There are
10,000 entries in train.json and another 60,000 in test.json.
Formatting
A typical line in either JSON file looks like
{"84": [[7, 6]], "185": [[7, 7]], "159": [[7, 8]], "151": [[7, 9]], "60": [[7, 10]], "36": [[7, 11]], "222": [[8, 6]], "254": [[8, 7], [8, 8], [8, 9], [8, 10], [9, 12], [9, 14], [9, 15], [9, 16], [9, 19], [9, 20], [10, 20], [13, 18], [14, 18], [15, 17], [17, 16], [19, 15], [20, 14], [21, 13], [21, 14], [22, 13], [23, 12], [23, 13], [24, 12], [24, 13], [25, 11], [25, 12], [26, 11]], "241": [[8, 11]], "198": [[8, 12], [8, 13], [8, 14], [8, 15], [8, 16], [8, 17], [8, 18], [8, 19]], "170": [[8, 20]], "52": [[8, 21], [23, 14], [24, 14]], "67": [[9, 6], [10, 14], [10, 15], [10, 16]], "114": [[9, 7], [9, 9]], "72": [[9, 8]], "163": [[9, 10]], "227": [[9, 11]], "225": [[9, 13]], "250": [[9, 17]], "229": [[9, 18]], "140": [[9, 21]], "17": [[10, 11]], "66": [[10, 12]], "14": [[10, 13]], "59": [[10, 17], [14, 16]], "21": [[10, 18]], "236": [[10, 19]], "106": [[10, 21]], "83": [[11, 18], [12, 20]], "253": [[11, 19]], "209": [[11, 20]], "18": [[11, 21], [26, 13]], "22": [[12, 17]], "233": [[12, 18]], "255": [[12, 19]], "129": [[13, 17]], "238": [[13, 19]], "44": [[13, 20]], "249": [[14, 17]], "62": [[14, 19]], "133": [[15, 16], [23, 11]], "187": [[15, 18]], "5": [[15, 19]], "9": [[16, 15]], "205": [[16, 16]], "248": [[16, 17]], "58": [[16, 18]], "126": [[17, 15]], "182": [[17, 17]], "75": [[18, 14]], "251": [[18, 15]], "240": [[18, 16]], "57": [[18, 17]], "19": [[19, 13]], "221": [[19, 14]], "166": [[19, 16]], "3": [[20, 12]], "203": [[20, 13]], "219": [[20, 15], [25, 13]], "35": [[20, 16]], "38": [[21, 12]], "77": [[21, 15]], "31": [[22, 11]], "224": [[22, 12]], "115": [[22, 14]], "1": [[22, 15]], "61": [[24, 10]], "242": [[24, 11]], "121": [[25, 10], [26, 10]], "40": [[25, 14]], "207": [[26, 12]], "label": 7}
This string is a valid dictionary in Python. Its keys are all integers which
lie between 0 and 255, except for the last key listed, which is label. Each
of the integer keys represents a greyscale value between 0 and 255. The
values for these integer keys are lists of pairs, indicating which pixels in
the image have the corresponding greyscale value. The value associated to the
label key is the number that the image is supposed to represent.