[WIP/Broken]: Upgrade @tensorflow/tfjs versions to 3.x #1383
Conversation
# Conflicts: # package-lock.json # package.json
…ion` -- executes without error but cannot find face.
…nly required when using @tensorflow/tfjs-core by itself)
|
@lindapaiste , not to pressure you but is there any progress on this? Here a npm audit reportd3-color <3.1.0 node-fetch <=2.6.6 28 vulnerabilities (18 low, 10 high) To address issues that do not require attention, run: |
|
@shiffman The models which I specified as needing work are the ones that your team has been working on re-implementing. My recollection is that all of the rest seemed fine. I wonder if it makes sense to use the new repo to work out the kinks, and then once you've got those models working to copy that code back over to this repo. Then bump the dependencies and delete the 1 or 2 broken models which we didn't update. It seems like a total pain to have to copy everything over to the new repo, and makes it harder to apply changes from current PRs. This was for TFJS 3.x and they are up to 4.x now but they say that's a non-breaking version change. |
Thank you for the info @lindapaiste! I understand, it's a tricky tradeoff. I think we have changed so much about the approach to which the library is developed and maintained that keeping with the new repo as a "fresh start" is preferable. This repo would get archived and "frozen in time" and we would point back towards and highlight issues, discussion, and code contributions that are relevant. We have a meeting at NYU tomorrow and I will bring this up for discussion! |
Extends #1379
Related issue: #1337
I have installed version
^3.16.0of all@tensorflow/tfjspackages. But a lot of things broke in the process. These are some of the known errors which need to be addressed:SoundClassifierThe tfjs
speechcommandsmodel expects tensorflow version^1.1.2. Attempting to run on 3.x causes errors regarding undefined Node.js variablespathandprocess. This is the worst one as it actually breaks the whole build. I had to temporarily disable the model....But then I re-enabled and I couldn't reproduce the error 🤷
BodyPixExpects tensorflow version
~1.2.1. There is a newer version ofbodypixavailable, but the API is very different.Error stack
Facemeshfacemeshexpects tensorflow version^2.1.0and its internal dependencyblazefaceexpects1.5.2. There is a newer version available under the package name@tensorflow-models/face-landmarks-detection.I saw an error once, and then when I refreshed the page everything was fine! 🤷
Error stack
FaceApiface-api.jsdepends on tensorflow version1.7.0(not a range, that exact version). I don't know what the future of this package is. Maybe we can implement it using a different model behind the scenes. Maybe we don't need it at all since we haveFacemesh. I believe there are forks offace-api.jsout there which have updated dependencies.Error stack
SketchRNN@magenta/sketchexpects tensorflow version^1.0.2. We have a few options here:SketchRNNandCharRNNinto a genericLSTMmodelorRNNmodel. This would be cool and might make it easier to add other LSTM models in the future.Error stack
MobileNet/ImageClassifierThe code runs without error but the prediction is incorrect.
Pix2PixThe initial drawing in the example runs correctly. However I get an error when drawing on the canvas and clicking "Transfer".
It seems most likely due to some tensor getting disposed too early.
Error stack
StyleTransferI wonder if tensorflow changed the default
dtypefor certain tensors, based on the error that I get.Error stack
There might be others too, but this is plenty for me to work on!