Keras Lambda Layers not supported for CTC Loss

Question

rmsmith OP

Created Jul ’17

Replies 1

Boosts 0

Views 1.2k

Participants 2

Hi,

I am looking to implement CTC loss function within CoreML. I tried, as a starting point, to use thekeras exampleand see if this could be converted however I recieve the following error:

File "/home/rob/py27/local/lib/python2.7/site-packages/coremltools/converters/keras/_keras2_converter.py", line 85, in _check_unsupported_layers

"Keras layer '%s' not supported. " % str(type(layer)))

ValueError: Keras layer '<class 'keras.layers.core.Lambda'>' not supported.

Okay as expected, Lambda wrapper functions aren't allowed. So what are the options here, as Keras does not have a built in CTC loss function? CTC loss is a very important for Speech Recognition and is seemingly not available. Can I build a similar graph and swap the loss function with an iOS one?. Is there support for CTC/TF in the works in the next version of CoreML?

RS

Boost

Answer 1

kerfuffle OP

Jul ’17

I have no experience with speech recognition or CTC but a loss function is usually used during training, not for inference. If this is the case with the CTC loss function too, then you can probably just remove it from the Keras model and export the model without that layer.

0