スヴン

参考文献:

• コード
• ハグフェイス

完全な数字

次のコマンドを使用して、このデータセットを TFDS にロードします。

ds = tfds.load('huggingface:svhn/full_numbers')

• 説明：

SVHN is a real-world image dataset for developing machine learning and object recognition algorithms with minimal requirement on data preprocessing and formatting.
It can be seen as similar in flavor to MNIST (e.g., the images are of small cropped digits), but incorporates an order of magnitude more labeled data (over 600,000 digit images)
and comes from a significantly harder, unsolved, real world problem (recognizing digits and numbers in natural scene images). SVHN is obtained from house numbers in Google Street View images.

• ライセンス: カスタム (非営利)
• バージョン: 1.0.0
• 分割:

• 特徴：

{
    "image": {
        "id": null,
        "_type": "Image"
    },
    "digits": {
        "feature": {
            "bbox": {
                "feature": {
                    "dtype": "int32",
                    "id": null,
                    "_type": "Value"
                },
                "length": 4,
                "id": null,
                "_type": "Sequence"
            },
            "label": {
                "num_classes": 10,
                "names": [
                    "0",
                    "1",
                    "2",
                    "3",
                    "4",
                    "5",
                    "6",
                    "7",
                    "8",
                    "9"
                ],
                "names_file": null,
                "id": null,
                "_type": "ClassLabel"
            }
        },
        "length": -1,
        "id": null,
        "_type": "Sequence"
    }
}

切り取られた数字

次のコマンドを使用して、このデータセットを TFDS にロードします。

ds = tfds.load('huggingface:svhn/cropped_digits')

• 説明：

SVHN is a real-world image dataset for developing machine learning and object recognition algorithms with minimal requirement on data preprocessing and formatting.
It can be seen as similar in flavor to MNIST (e.g., the images are of small cropped digits), but incorporates an order of magnitude more labeled data (over 600,000 digit images)
and comes from a significantly harder, unsolved, real world problem (recognizing digits and numbers in natural scene images). SVHN is obtained from house numbers in Google Street View images.

• ライセンス: カスタム (非営利)
• バージョン: 1.0.0
• 分割:

• 特徴：

{
    "image": {
        "id": null,
        "_type": "Image"
    },
    "label": {
        "num_classes": 10,
        "names": [
            "0",
            "1",
            "2",
            "3",
            "4",
            "5",
            "6",
            "7",
            "8",
            "9"
        ],
        "names_file": null,
        "id": null,
        "_type": "ClassLabel"
    }
}