6 ms·
A TensorFlow Implementation of DeepMind's WaveNet Paper
- alexbeloi 10y agoI had plans to do the same, I'm glad somebody beat me to it.
- Kenji 10y agoThat is an admirable mindset - I cannot help but be a bit frustrated when someone independently implements my ideas first. On the one hand, it validates the idea. On the other hand, building it would be fun, but now that someone else did it, doing it again would be akin to reinventing the wheel and it is more productive to turn your attention to something new (unless you think you can execute the project much better).
- sherjilozair 10y agoThis project (and others in different frameworks) are far from complete. The prize for reproducing WaveNet in open source remains unclaimed.
- vintermann 10y agoIn this case, it's DeepMind's idea anyway :) I get more disappointed when the opposite happens. I think something like, "Yeah, I'm totally going to add support in torch for noisy activation functions like in this paper!" (https://arxiv.org/pdf/1603.00391.pdf https://arxiv.org/pdf/1603.00391.pdf). Then I procrastinate and put it off. Then I think, "No matter, someone else has surely done it by now". Then they haven't.
- jonnycowboy 10y agoDo you know how long it took to train using that dataset and with what hardware configuration?
- Dowwie 10y agoI wonder whether accents could be layered on top of the trained data?
- aab0 10y agoProbably. You can add in 'speaker' as a bit of metadata to the samples (this is what is meant by 'conditioning on') and teach it to speak like different people, so if you have a diverse sample of speakers and you add in 'accent' as another variable, it might well learn to disentangle individual speakers from their accents and then you can control generated accents by changing the metadata.
- teddyknox 10y agoDoing a forward pass for every sample sounds like it would be prohibitive for real-time applications.
- nicklo 10y agoIt absolutely is. DeepMind reported that 1 second of audio generation takes about 90 minutes to generate.
- throwawaymsft 10y agoAssuming it's computation bound, it's a factor of 5400 (~13 doublings in CPU power required to get to real-time, assuming no algorithmic improvements).
- mattnewton 10y agoDo they mention it was CPU trained? I assumed GPU. If it was CPU trained, I wonder what the operations keeping it off the GPU were?
- Houshalter 10y agoGoogle has special neural net ASICs now.
- ogrisel 10y agoGoogle never stated they use those to train models as far as I know. It seems that they are primarily used to spare energy when deploying trained models at scale.
- Houshalter 10y agoTheres no reason they couldn't use them to train, as long as they can account for the lower precision operations. I think it would be much cheaper to train on them, at that scale anyway.
- blaurence5 10y agoHere's a Theano implementation: https://github.com/huyouare/WaveNet-Theano https://github.com/huyouare/WaveNet-Theano
- sargun 10y agoI'm really interested in building an RNN and training it against something like movies. I'd love to then take it and input a song and translate it to a music video that's a composite. I'm also interested in the legal ramifications of doing such a thing... Does anyone know of prior art?