tfjs: Cannot import @tensorflow/tfjs-node-gpu on Windows

@MarcDelp @adwellj A PR fixing this issue has been merged and the next release (most likely after thankgiving) will support the latest version of node.

Right, turns out I fixed it? Some how???

Idk how I did but it had nothing to do with NodeJS? I was looking at Voice Mimicing with AI, installed some Python 3.7 libraries, one of which being Tensorflow GPU & PyTorch, now it’s suddenly working

I legit have 0 clue as to what got it working 😕

I have the same error running tfjs-node@1.3.1 with node 12.13.0. Using node 10.16.3 seems to fix it.

I too had problems running newer versions of tfjs on newer versions of Node on Windows 10.

Node v10.17.0 and tfjs 1.2.11 would produce a ‘module not found’ error when trying to require("@tensorflow/tfjs-node"). Node v10.16.3 and tfjs 1.2.11 will work though, as JoshNewham1 mentioned above.

@Mobles can you try go to the @tensorflow/tfjs-node-gpu directory, like cd node_modules\@tensorflow\tfjs-node-gpu, then run node-pre-gyp rebuild (you might need to npm install -g node-pre-gyp). Then see if you can import tfjs-node-gpu?

I had the same issue and I ran node-pre-gyp rebuild and got this output: Along with a whole bunch of more above.

I keep getting the same error: My path does not contain any spaces.

I would like to underline the fact that this problem is not resolved. I’ve tried everything mentioned above without any success to get tjfs-node-gpu working on windows 10. It always yields UnhandledPromiseRejectionWarning: Error: The Node.js native addon module (tfjs_binding.node) can not be found at path: C:\...\node_modules\@tensorflow\tfjs-node-gpu\lib\napi-v4\tfjs_binding.node..

It feels like finding right versions of node.js + tfjs + tfjs-node-gpu is purely experimental task which is a little frustrating. Also I have to mention that tfjs works well on CPU.

UPDATE: So it turns out adding tensorflow.dll as mentioned in comment below actually did the job:

If this problem is still affecting anyone, I think I might’ve found a temporary fix. I transplanted the tensorflow.dll file from my Node v10.16.3 modules (node_modules@tensorflow\tfjs-node\lib\napi-v4) to my Node 12 modules which didn’t seem to have one. Then, I got past the “specified module could not be found” error. This is quite a bad bodge and definitely not a solution for production environments but seems to work fine for development until this issue is fixed. The DLL file is hosted here, if you can’t find it.

The final set of configurations working for me seem to be following: Node.js: v10.16.3 @tensorflow/tfjs: v1.2.11 @tensorflow/tfjs-node: v1.2.9 @tensorflow/tfjs-node-gpu: v1.2.9 CUDA toolkit: v10.0.0 cuDNN sdk: any version compatible with CUDA toolkit v10.0.0 - so for example v7.4.2

And don’t forget adding required CUDA paths into env PATH.

I’m having the exact same issue on a new Windows machine under Node v12.13.0. I’ve installed Python 2 and the dev tools with absolutely no problems. I can build the bindings and can confirm that they exist, but I get the same import error. This is with TFJS-Node version 1.3.1.

On my other system, I managed to get it to work perfectly with Node v10.16.3, TFJS v1.2.11 and TFJS-Node v1.2.9. I’ve no idea why this worked, and trying to update any packages results in the “module not found” issue for “tfjs_binding.node”. Both of these systems are on Windows 10 1903.

Thanks @JoshNewham1, save my project =D

I’ve recently found that starting with node 12.15, you need to be on tfjs-node-gpu 1.5.x. Below 12.15 requires 1.4.0 (or below).

Also, tfjs-node-gpu 1.5.x requires CuDNN 7.6, whereas earlier versions used 7.4.

I hope this helps you guys. The poor / incomplete documentation regarding version compatibility is frustrating.

Yes, I got it working on node 10.16.3 with tfjs 1.3.1 but I’d really like to make it on node 12 as we are starting a new project from scratch. Do you have any idea (even if not a precise one) on when the fix may be available?

hi @ealionel

I can reproduce the error on a new windows machine (though it works on the windows machine we use to develop). I’ll try to figure out the reason.

@Mobles can you try with CUDA 10.0 and cuDNN >= 7.4.1? These are the official supported version: https://www.tensorflow.org/install/gpu

@Mobles unsure if this is relevant, but here are some additional things I had done previously to this new PC. They’re steps that I used to have to do for version 0.3.2 when tfjs-node-gpu was using node-gyp to build on the client.

npm i -g node-gyp Then from an elevated console (Run as Administrator) npm i -g --production windows-build-tools Then create a copy of python.exe in C:\Users<user>.windows-build-tools\python27 and rename it to python2.exe. Without this step node-gyp configure --verbose would complain about not being able to find python2.

My understanding is that node-gyp is no longer needed, so this is likely not relevant, but these are some other configurations that I’ve made to this new machine.