class paddle.nn. CTCLoss ( blank=0, reduction='mean' ) [source]

An operator integrating the open source Warp-CTC library (https://github.com/baidu-research/warp-ctc) to compute Connectionist Temporal Classification (CTC) loss. It can be aliased as softmax with CTC, since a native softmax activation is interated to the Warp-CTC library to normalize values for each row of the input tensor.

  • blank (int, optional) – The blank label index of Connectionist Temporal Classification (CTC) loss, which is in the half-opened interval [0, num_classes + 1). The data type must be int32. Default is 0.

  • reduction (string, optional) – Indicate how to average the loss, the candicates are 'none' | 'mean' | 'sum'. If reduction is 'mean', the output loss will be divided by the label_lengths, and then return the mean of quotient; If reduction is 'sum', return the sum of loss; If reduction is 'none', no reduction will be applied. Default is 'mean'.

  • log_probs (Tensor): The unscaled probability sequence with padding, which is a 3-D Tensor. The tensor shape is [max_logit_length, batch_size, num_classes + 1], where max_logit_length is the longest length of input logit sequence. The data type should be float32 or float64.

  • labels (Tensor): The ground truth sequence with padding, which must be a 3-D Tensor. The tensor shape is [batch_size, max_label_length], where max_label_length is the longest length of label sequence. The data type must be int32.

  • input_lengths (Tensor): The length for each input sequence, it should have shape [batch_size] and dtype int64.

  • label_lengths (Tensor): The length for each label sequence, it should have shape [batch_size] and dtype int64.

  • norm_by_times (bool, optional): Whether to normalize the gradients by the number of time-step, which is also the sequence’s length. There is no need to normalize the gradients if reduction mode is ‘mean’. Default: False.


reduction is 'none', the shape of loss is [batch_size], otherwise, the shape of loss is []. Data type is the same as log_probs.

Return type

Tensor, The Connectionist Temporal Classification (CTC) loss between log_probs and labels. If attr


>>> # declarative mode
>>> import paddle

>>> # length of the longest logit sequence
>>> max_seq_length = 4
>>> #length of the longest label sequence
>>> max_label_length = 3
>>> # number of logit sequences
>>> batch_size = 2
>>> # class num
>>> class_num = 3

>>> log_probs = paddle.to_tensor([[[4.17021990e-01, 7.20324516e-01, 1.14374816e-04],
...                                [3.02332580e-01, 1.46755889e-01, 9.23385918e-02]],
...                               [[1.86260208e-01, 3.45560730e-01, 3.96767467e-01],
...                                [5.38816750e-01, 4.19194520e-01, 6.85219526e-01]],
...                               [[2.04452246e-01, 8.78117442e-01, 2.73875929e-02],
...                                [6.70467496e-01, 4.17304814e-01, 5.58689833e-01]],
...                               [[1.40386939e-01, 1.98101491e-01, 8.00744593e-01],
...                                [9.68261600e-01, 3.13424170e-01, 6.92322612e-01]],
...                               [[8.76389146e-01, 8.94606650e-01, 8.50442126e-02],
...                                [3.90547849e-02, 1.69830427e-01, 8.78142476e-01]]], dtype="float32")
>>> labels = paddle.to_tensor([[1, 2, 2], [1, 2, 2]], dtype="int32")
>>> input_lengths = paddle.to_tensor([5, 5], dtype="int64")
>>> label_lengths = paddle.to_tensor([3, 3], dtype="int64")

>>> loss = paddle.nn.CTCLoss(blank=0, reduction='none')(log_probs, labels, input_lengths, label_lengths)
>>> print(loss)
Tensor(shape=[2], dtype=float32, place=Place(cpu), stop_gradient=True,
[3.91798496, 2.90765214])

>>> loss = paddle.nn.CTCLoss(blank=0, reduction='mean')(log_probs, labels, input_lengths, label_lengths)
>>> print(loss)
Tensor(shape=[], dtype=float32, place=Place(cpu), stop_gradient=True,
forward ( log_probs, labels, input_lengths, label_lengths, norm_by_times=False )


Defines the computation performed at every call. Should be overridden by all subclasses.

  • *inputs (tuple) – unpacked tuple arguments

  • **kwargs (dict) – unpacked dict arguments