szha commented on a change in pull request #7304: gluon bce & ctc losses
URL: https://github.com/apache/incubator-mxnet/pull/7304#discussion_r131727946
##########
File path: python/mxnet/gluon/loss.py
##########
@@ -239,3 +281,59 @@ def hybrid_forward(self, F, output, label,
sample_weight=None):
loss = label * (F.log(label+1e-8) - output)
loss = _apply_weighting(F, loss, self._weight, sample_weight)
return F.mean(loss, axis=self._batch_axis, exclude=True)
+
+class CTCLoss(Loss):
+ r"""Connectionist Temporal Classification Loss.
+
+ See `"Connectionist Temporal Classification: Labelling Unsegmented
+ Sequence Data with Recurrent Neural Networks"
+ <http://www.cs.toronto.edu/~graves/icml_2006.pdf>`_ paper for more
information.
+
+ The prediction output should be an activation vector without softmax, with
shape
+ according to the output_layout:
+ **TNC**: *(sequence_length, batch_size, alphabet_size + 1)*
+ **NTC**: *(batch_size, sequence_length, alphabet_size + 1)*
+ **out**: *(batch_size)*.
+
+ ``label`` is a tensor of integers between 1 and *alphabet_size*, with
shape according
+ to the batch_axis:
+ **batch_axis=0**: *(batch_size, label_sequence_length)*
+ **batch_axis=1**: *(label_sequence_length, batch_size)*
+
+ If a sequence of labels is shorter than *label_sequence_length*, use the
special
+ padding character 0 at the end of the sequence to conform it to the correct
+ length. For example, if *label_sequence_length* = 4, and one has two
sequences
+ of labels [2, 1] and [3, 2, 2], the resulting ```label``` tensor should be
+ padded to be::
+
+ [[2, 1, 0, 0], [3, 2, 2, 0]]
Review comment:
Yes, since 0 is special character. I will take a look at cudnn7 CTC too and
see if we can make it consistent.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
