Npair loss requires paired data. Npair loss has two parts: the first part is L2
regularizer on the embedding vector; the second part is cross entropy loss which
takes the similarity matrix of anchor and positive as logits.
For more information, please refer to:
`Improved Deep Metric Learning with Multi class N pair Loss Objective <http://www.nec-labs.com/uploads/images/Department-Images/MediaAnalytics/papers/nips16_npairmetriclearning.pdf>`_
Args:
anchor(Tensor): embedding vector for the anchor image. shape=[batch_size, embedding_dims],
the data type is float32 or float64.
positive(Tensor): embedding vector for the positive image. shape=[batch_size, embedding_dims],
the data type is float32 or float64.
labels(Tensor): 1-D tensor. shape=[batch_size], the data type is float32 or float64 or int64.
l2_reg(float32): L2 regularization term on embedding vector, default: 0.002.
Returns:
A Tensor representing the npair loss, the data type is the same as anchor, the shape is [1].
This op computes the edit distances, also called Levenshtein distance, between a batch of
hypothesis strings and their references. It measures how dissimilar two strings are by counting
the minimum number of operations to transform one string into another.
The operations include insertion, deletion, and substitution.
For example, given hypothesis string A = "kitten" and reference
B = "sitting", A will be transformed into B
at least after two substitutions and one insertion:
"kitten" -> "sitten" -> "sittin" -> "sitting"
So the edit distance between A and B is 3.
The input is a Tensor, the input_length and label_length should be supported.
The `batch_size` of labels should be same as `input`.
The output include the edit distance value between every pair of input and related label, and the number of sequence.
If Attr(normalized) is true,
the edit distance value will be divided by the length of label.
Parameters:
input(Tensor): The input tensor, its rank should be equal to 2 and its data type should be int64.
label(Tensor): The label tensor, its rank should be equal to 2 and its data type should be int64.
normalized(bool, default True): Indicated whether to normalize the edit distance.
ignored_tokens(list<int>, default None): Tokens that will be removed before
calculating edit distance.
input_length(Tensor): The length for each sequence in `input` if it's of Tensor type, it should have shape `(batch_size, )` and its data type should be int64.
label_length(Tensor): The length for each sequence in `label` if it's of Tensor type, it should have shape `(batch_size, )` and its data type should be int64.
NOTE: To be avoid unexpected result, the value of every elements in input_length and label_length should be equal to the value of the second dimension of input and label. For example, The input: [[1,2,3,4],[5,6,7,8],[9,10,11,12]], the shape of input is [3,4] and the input_length should be [4,4,4]
NOTE: This Api is different from fluid.metrics.EditDistance
Returns:
Tuple:
distance(Tensor): edit distance result, its data type is float32, and its shape is (batch_size, 1).
sequence_num(Tensor): sequence number, its data type is float32, and its shape is (1,).
