mindspore.ops.RNNTLoss

View Source On Gitee
class mindspore.ops.RNNTLoss(blank_label=0)[source]

Computes the RNNTLoss and its gradient with respect to the softmax outputs.

Parameters

blank_label (int) – blank label. Default: 0 .

Inputs:

  • acts (Tensor) - Tensor of shape \((B, T, U, V)\), where \(B\) is batch, \(T\) is sequence length, \(U\) is label length and \(V\) is output dim. Data type must be float16 or float32.

  • labels (Tensor) - Tensor of shape \((B, U-1)\). Data type is int32.

  • input_lengths (Tensor) - Tensor of shape \((B,)\). Data type is int32.

  • label_lengths (Tensor) - Tensor of shape \((B,)\). Data type is int32.

Outputs:

  • costs (Tensor) - Tensor of shape \((B,)\). Data type is int32.

  • grads (Tensor) - Has the same shape and dtype as acts.

Raises

  • TypeError – If acts, labels, input_lengths or label_lengths is not a Tensor.

  • TypeError – If dtype of acts is neither float16 nor float32.

  • TypeError – If dtype of labels, input_lengths or label_lengths is not int32.

Supported Platforms:

Ascend

Examples

>>> import numpy as np
>>> from mindspore import ops, Tensor
>>> B, T, U, V = 1, 2, 3, 5
>>> blank = 0
>>> acts = np.random.random((B, T, U, V)).astype(np.float32)
>>> labels = np.array([[1, 2]]).astype(np.int32)
>>> input_length = np.array([T] * B).astype(np.int32)
>>> label_length = np.array([len(l) for l in labels]).astype(np.int32)
>>> rnnt_loss = ops.RNNTLoss(blank_label=0)
>>> costs, grads = rnnt_loss(Tensor(acts), Tensor(labels), Tensor(input_length), Tensor(label_length))
>>> print(costs.shape)
(1,)
>>> print(grads.shape)
(1, 2, 3, 5)