Why is the output.pb file almost half a gig and how wide is the input layer?

DeepSpeech
#1

For the trained model released by mozilla, why is it so darn big? Can quantization be done to it?

How wide is the input layer? What about the output layer?

Thanks, love this stuff, deepspeech is cool

#2

The model width is 2048[1] and the output is the number of elements in the target alphabet, for English around 26+1.

Unfortunately due to an old bug in TensorFlow[2] you can’t quantize. However, we’re are hoping that changes in the next release 0.2.0 will allow us to circumvent that bug.

Thanks for the encouragement!