New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Rapsberry Pi #15
Comments
Hi, It is actually on our todo list to migrate models from Pytorch to Mozilla's implementation of DeepSpeech in TensorFlow and rely on their framework for increased stability in production settings. Not completely sure how TensorFlow Lite works but I think this will allow us to export a trained model to TensorFlow Lite and support this as well. We will be looking specifically at this problem next week to cook up a plan. If we are able to simply convert the weights of DanSpeech models to TensorFlow, then we might have something running within the DanSpeech library sometime in July. My previous similar experiments have however shown that this is a lot harder than it sounds. Hence, if we need to train new models to support Mozilla's DeepSpeech implementation, then perhaps mid august / start september. You're welcome to try converting a DanSpeech Pytorch Model to Mozilla's TensorFlow implementation and share the results! :) Best, |
Sounds good. I'm looking forward to seeing your results. I tried looking into converting the Pytorch model into TFlite (throught ONNX) but I'm not too knowledgeable about this. According to what I could find (like https://michhar.github.io/convert-pytorch-onnx/) I need a class of the model I'm converting. I'm not sure how to create that or find it in your library, so if you have any ideas I would love to know. |
Hey I didn't have much luck in converting your model to TensorFlow Lite, but I'm still very much interested if it is still on your list to look into. Best, |
Hey Risskov, Sorry for the very slow answer - we did look into converting to the Mozilla framework briefly, but unfortunately it wasn't an entirely trivial task and as such we have parked the conversion to tensorflow for now while we work on other todo's. We will start work on a lightweight version eventually, but I can't say how long it might be - we are more than happy to welcome a PR or provide a link to a new repo if you manage to solve the issue before we get to it. Best, Martin |
Hi
I've been trying out danspeech on a Raspberry but found that it runs very very slowly compared to on a laptop. It seems that DeepSpeech has a version that has been optimized for mobile devices by using TensorFlow Lite, but is it possible to also use this for danspeech?
The text was updated successfully, but these errors were encountered: