Updating DECIMER to be compatible with Python 3.12

[mam4000]

Hi DECIMER team,

I have run DECIMER on your web site, and it worked well for a small portion of a document. Now I want to install it on my system so I can process full documents.

I’m trying to run DECIMER/DECIMER-Segmentation in Google Colab, which currently provides Python 3.12 by default. The current packages appear to require the TF/Keras 2.15 stack, which isn’t available as wheels for Python 3.12, and also conflicts with Keras 3. I consistently hit:

• AttributeError: module 'keras.callbacks' has no attribute 'experimental' (Keras 3 removal)
• ImportError: numpy.core.umath failed to import (NumPy/ABI mismatch when newer NumPy is pulled with Py3.12)
• ModuleNotFoundError: decimer_segmentation (when falling back to module entry points without compatible installs)

Environment:

• Python: 3.12 (Colab default)
• Desired: wheels for Linux x86_64 (Colab), CPU is fine
• Current Colab pulls Keras 3 / NumPy ≥ 2.0 by default

Request: Publish DECIMER and DECIMER-Segmentation versions that:

1. Support Python 3.12 (publish wheels on PyPI),
2. Run on a TensorFlow/Keras version with official Py3.12 wheels (e.g., TF ≥ 2.16/2.17/2.18),
3. Are updated for Keras 3 (remove keras.callbacks.experimental.* and any other deprecated APIs),
4. Work with NumPy ≥ 2.0 (or pin a compatible NumPy in install_requires and ship matching wheels).

Why this matters: Colab users increasingly have no option to select Python 3.10/3.11. A Py3.12-compatible release would let us run DECIMER without fragile environment pinning or downgrades.

I can provide a minimal Colab repro notebook if helpful. Thanks!

[Kohulan]

Hi @mam4000 ,

Currently, I have no plan towards supporting this for now. It needs some work, and I have to make sure Segmentation and Transformer work fine on Python 3.12. I will look into this when I have time.

The collab notebook would be useful. Thanks.

• Kohulan