Function Differences with tf.nn.ctc_loss

tf.nn.ctc_loss

tf.nn.ctc_loss(
    labels,
    logits,
    label_length,
    logit_length,
    logits_time_major=True,
    unique=None,
    blank_index=None,
    name=None
) -> Tensor

For more information, see tf.nn.ctc_loss.

mindspore.ops.CTCLoss

class mindspore.ops.CTCLoss(
    preprocess_collapse_repeated=False,
    ctc_merge_repeated=True,
    ignore_longer_outputs_than_inputs=False
)(x, labels_indices, labels_values, sequence_length) -> (Tensor, Tensor)

For more information, see mindspore.ops.CTCLoss.

Differences

TensorFlow: Calculate the loss between the continuous time sequence and the target sequence. And return only a loss that is consistent with the shape of label_length.

MindSpore: MindSpore API basically implements the same function as TensorFlow, returns loss and its gradient.

Categories	Subcategories	TensorFlow	MindSpore	Differences
Parameters	Parameter 1	labels	labels_values	The function is the same, and the parameter name is different, but in MindSpore the rank must be 1
	Parameter 2	logits	x	Same function, different parameter names
	Parameter 3	label_length	sequence_length	Same function, different parameter names
	Parameter 4	logit_length	-	MindSpore does not have this parameter
	Parameter 5	logits_time_major	-	Control how logits are arranged. MindSpore does not have this parameter
	Parameter 6	unique	-	MindSpore does not have this parameter
	Parameter 7	blank_index	-	MindSpore does not have this parameter. When it is -1, blank is represented by num_classes-1, which is consistent with MindSpore at this time
	Parameter 8	name	-	Not involved
	Parameter 9	-	preprocess_collapse_repeated	Collapse duplicate labels before CTC calculation. TensorFlow does not have this parameter
	Parameter 10	-	ctc_merge_repeated	Whether to merge non-blank labels. TensorFlow does not have this parameter
	Parameter 11	-	ignore_longer_outputs_than_inputs	Whether to ignore sequences whose output is longer than the input, TensorFlow does not have this parameter
	Parameter 12	-	labels_indices	labels_indices[i, :] = [b, t] means that labels_values[i] stores the IDs of (batch b, time t), which guarantees that the rank of labels_values is 1

Code Example

Consistent functions.

# TensorFlow
import tensorflow as tf
import numpy as np

label = tf.Variable([[0],
                     [1]])
logits = tf.constant(np.array([[[0.56352055, -0.24474338, -0.29601783],[0.8030011, -1.2187808, -0.6991761]],[[-0.81990826, -0.3598757, 0.50144005],[-1.0980303, 0.60394925, 0.3771529]]]), dtype=tf.float32)
label_length = tf.Variable(np.array([1, 1]))
logits_length = tf.Variable(np.array([2, 2]))
loss = tf.nn.ctc_loss(label, logits, label_length=label_length, logit_length=logits_length)
print(loss.numpy().shape)
# (2,)

# MindSpore
import mindspore
from mindspore import Tensor, ops
import numpy as np

x = Tensor(np.array([[[0.56352055, -0.24474338, -0.29601783], [0.8030011, -1.2187808, -0.6991761]], [[-0.81990826, -0.3598757, 0.50144005], [-1.0980303, 0.60394925, 0.3771529]]]).astype(np.float32))

labels_indices = Tensor(np.array([[0, 1], [1, 1]]), mindspore.int64)
labels_values = Tensor(np.array([0, 1]), mindspore.int32)
sequence_length = Tensor(np.array([2, 2]), mindspore.int32)

ctc_loss = ops.CTCLoss()
loss, gradient = ctc_loss(x, labels_indices, labels_values, sequence_length)
print(loss.shape)
# (2,)