Converting *.h5 model with tfjs

[stewie-stew]

I am working on a project where I need to convert my *.h5 model to use it with javascript, I wanted to convert the model with tensorflowjs using the code:

import tensorflowjs as tfjs
tfjs.converters.save_keras_model(model, "tfjs_models")

an error occured saying something about deprecated np.object type. I fixed this error by replacing np.object with object in file: /venv/lib/python3.13/site-packages/tensorflowjs/write_weights.py and /venv/lib/python3.13/site-packages/tensorflowjs/read_weights.py on lines 27 and 28.
But it did not help.

A new error saying module 'tensorflow.compat.v1' has no attribute 'estimator'

my system information

• fedora linux 44
• python 3.13.14
• pip 26.1.2
• tensorflow 2.21.0
• tensorflowjs 3.18.0
• keras 3.15.0

[nileshau-afk]

Hi @stewie-stew ,

Thank you for raising this issue. We have reproduced the issue. As a workaround, you can downgrade to TensorFlow < 2.16 (which includes np.object) and NumPy < 1.24 (which includes the tf.estimator API) to ensure full compatibility with tensorflowjs. Alternatively, if you need to use newer versions of TensorFlow and NumPy, you can apply runtime monkey-patching in your Python script prior to importing tensorflowjs.

Could you please try this and let us know if it works for you. If it doesn't, please share a code snippet we can run so we can pinpoint the issue.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No

[stewie-stew]

Hi @stewie-stew ,

Thank you for raising this issue. We have reproduced the issue. As a workaround, you can downgrade to TensorFlow < 2.16 (which includes np.object) and NumPy < 1.24 (which includes the tf.estimator API) to ensure full compatibility with tensorflowjs. Alternatively, if you need to use newer versions of TensorFlow and NumPy, you can apply runtime monkey-patching in your Python script prior to importing tensorflowjs.

Could you please try this and let us know if it works for you. If it doesn't, please share a code snippet we can run so we can pinpoint the

@nileshau-afk
Thank for your reply
I tried the using older versions of tensorflow and numpy, and it worked but with some additional debugging and version management.

Is there a future fix for this issue, or other alternative that are error free or something like this.