Windows ds_ctcdecoder ResourceNotFound

Deep Speech
#1

I am trying to install ds_ctcdecoder on my GPU. However, util/taskcluster.py gives me the following url:

https://index.taskcluster.net/v1/task/project.deepspeech.deepspeech.native_client.v0.6.0-alpha.1.cpu-ctc/artifacts/public/ds_ctcdecoder-0.6.0a1-cp36-cp36m-windows_AMD64.whl

When pip installing the above url, I get the message that it is not supported on my wheel. After a bit of guessing, I discovered that the following link does get pass the wheel support error:

https://index.taskcluster.net/v1/task/project.deepspeech.deepspeech.native_client.v0.6.0-alpha.1.cpu-ctc/artifacts/public/ds_ctcdecoder-0.6.0a1-cp36-cp36m-win_amd64.whl

However, when I go to this resource it says ResourceNotFound:

{
“code”: “ResourceNotFound”,
“message”: “Artifact not found\n\n—\n\n* method: getLatestArtifact\n* errorCode: ResourceNotFound\n* statusCode: 404\n* time: 2019-07-10T21:24:22.279Z”,
“requestInfo”: {
“method”: “getLatestArtifact”,
“params”: {
“0”: “public/ds_ctcdecoder-0.6.0a1-cp36-cp36m-win_amd64.whl”,
“taskId”: “SeAG_HKXRB-J-o6WPPFkuw”,
“name”: “public/ds_ctcdecoder-0.6.0a1-cp36-cp36m-win_amd64.whl”
},
“payload”: {},
“time”: “2019-07-10T21:24:22.279Z”
}
}

I then looked around here: https://tools.taskcluster.net/index/project.deepspeech.deepspeech.native_client

I tried to find the corresponding resource, but I could not find it.

I am using a Windows GPU (64) with python 3.6

#2

Sorry, we don’t have ds_ctcdecoder for Windows because we don’t (yet?) support training on Windows.

#3

Oh ok. I’m confused then, how have people been using (training) deepspeech-gpu on nvidia gpus if they are all windows?

#4

You can use a trained model on Windows with deepspeech-gpu, just not train a new model from scratch. If you know people training on Windows, please tell them to send us PRs :slight_smile:

#5

Oh that is unfortunate. Thank you both for the help. Do you guys have any tips on how I can go about creating my own custom language model/trie and training my own Deep Speech model? The resources at my disposal are a MacBookPro, Windows Desktop, NVIDIA GPU.

Based on your feedback, it seems that the only way to move forward would be with my MacBookPro? But would training with 30+hours of audio be infeasible?

Or if not DeepSpeech, do you have any recommendations on other open source projects that would allow me to leverage my GPU to train a custom model with a custom language model (in python?)

Thank you all for your help!

#6

Nope, because TensorFlow upstream dropped CUDA support on macOS several releases ago.

Unfortunately, so far we only support training on Linux, because nobody on the team has Windows, not trains on Windows and getting that part to build on Windows is non-trivial even if not complicated. If you manage to get make -C native_client/ctcdecode to produce you a Python wheel, then you’re welcome to send us a PR but likely without CI for now (which means there’s a risk of regression).

Adding CI for that will also take some time (and resources).

#7

Thank you!

Do you know if it is possible for me to run a linux virtual machine on a windows computer (which has the NVIDIA GPU). Would it be possible to train on the virtual machine?

#8

Personally I don’t even think it’s worth trying to enable training on Windows. I believe nobody in the industry wants to spend any time in this. Ofcoz it must be doable in some way if you really want to investigate it for some reason. But if you’re serious, switch to Linux, please.

#9

Oh gotcha. Was just asking because I don’t have access to a linux machine. Does anyone know of good virtual linux machines that support GPU’s and DeepSpeech training?