CODEWITHSUNDEEP
pythontensorflowtensorflow.jstensorflowjs-converter
OrderAndChaos
OrderAndChaos

Reputation: 3860

Load Python Tensorflow saved_model with Tensorflow.js

I've saved a model I created with Python Tensorflow. Is it possible to load this with Tensorflow.js?

Save

tf.saved_model.save(translator, 'ipa_translator',
                    signatures={'serving_default': translator.tf_translate})

Load

reloaded = tf.saved_model.load('ipa_translator')
result = reloaded.tf_translate(input_text)

I've found Tensorflow import_saved_model which looks like it'll handle the conversion, but I can't figure out how it's meant to be used.

I also tried following this guide, but got errors about mismatching model types.

There appears to be a wizard script pip3 install tensorflowjs[wizard], when I try to install it I get no matches found: tensorflowjs[wizard]

This is the model that I want to convert if that helps: https://colab.research.google.com/drive/16ge-HE2RZ6TmG9_zS2_L4dLKOO00mVRB?usp=sharing

I would appreciate any guidance at all on getting this running.

Upvotes: 1

Views: 165

Answers (1)

OrderAndChaos
OrderAndChaos

Reputation: 3860

Looks like the some of the issues I was having were to do with incompatibilities between some of the libraries.

I set up a new Poetry environment. While installing tensorflow and tensorflowjs it flagged up some version issues with dependencies.

After I downgraded tensorflow to 2.11 I was able to install tensorflowjs and the wizard. I had to wrap the wizard lib name in quotes so that it was recognised poetry add -D "tensorflowjs[wizard]".

I could then run tensorflowjs_wizard in the terminal and follow the steps to complete the conversion.

Here's the poetry set up that worked for me.

pyproject.toml

[tool.poetry]
name = "eng_to_ipa_model"
version = "0.1.0"

[tool.poetry.dependencies]
python = ">=3.9,<3.11"
tensorflow = "2.11"


[tool.poetry.group.dev.dependencies]
tensorflowjs = {extras = ["wizard"], version = "^4.4.0"}

[build-system]
requires = ["poetry-core"]
build-backend = "poetry.core.masonry.api"

The final piece of the puzzle was about using Tensorflow v1 api, I had to migrate the script to v2, which was painful to say the least.

Upvotes: 0

Related Questions