tf.raw_ops.Fingerprint

Generates fingerprint values.

tf.raw_ops.Fingerprint(
    data, method, name=None
)

Generates fingerprint values of data.

Fingerprint op considers the first dimension of data as the batch dimension, and output[i] contains the fingerprint value generated from contents in data[i, ...] for all i.

Fingerprint op writes fingerprint values as byte arrays. For example, the default method farmhash64 generates a 64-bit fingerprint value at a time. This 8-byte value is written out as an uint8 array of size 8, in little-endian order.

For example, suppose that data has data type DT_INT32 and shape (2, 3, 4), and that the fingerprint method is farmhash64. In this case, the output shape is (2, 8), where 2 is the batch dimension size of data, and 8 is the size of each fingerprint value in bytes. output[0, :] is generated from 12 integers in data[0, :, :] and similarly output[1, :] is generated from other 12 integers in data[1, :, :].

Note that this op fingerprints the raw underlying buffer, and it does not fingerprint Tensor's metadata such as data type and/or shape. For example, the fingerprint values are invariant under reshapes and bitcasts as long as the batch dimension remain the same:

Fingerprint(data) == Fingerprint(Reshape(data, ...))
Fingerprint(data) == Fingerprint(Bitcast(data, ...))

For string data, one should expect Fingerprint(data) != Fingerprint(ReduceJoin(data)) in general.