search

kosta-pmf / dnn-audio-watermarking

DNN-based audio watermarking
GNU General Public License v3.0
40 stars 10 forks source link

Example File Not Running as Described in the ReadMe #2

Closed adityatb closed 1 year ago

adityatb commented 2 years ago

Hi there,

Excited to try this repo! I was trying to get this to run and found that there's seems to be some inconsistency with the pretrained model and it's expected input shape.

I tried to librosa.load(audio_file, target_sr=16000) and then run the main.embed() on it as described in the ReadMe and I get an error with Tensor Shapes not matching.

Perhaps I'm missing some extra steps to get you there. Appreciate any help you can offer!

Cheers!

I should clarify - I'm running it with the example files provided. There also seems to be a mismatch for the message shapes -> I get (1, 512) vs (16, 2, 512)

kosta-pmf commented 2 years ago

Hello, Thanks for the issue. We somehow missed it until now. We will look into it.

AAA-eng-alt commented 2 years ago

Happy to answer your question, I ran into a similar tensor mismatch when testing. What you have here is the problem of SFTF and BATCH_SIZE. You try to change the HOP_Length parameter in the config.py file, reduce the frame spacing, to match the amplitude spectrum of 64 width, try it, mine has worked.

my-yy commented 1 year ago

same issue

slavzilla commented 1 year ago

@my-yy please refer to @AAA-eng-alt comment, it is the answer to your issue.