Keras Integration in KNIME 5.9.0 – TensorFlow version conflict / best practice?

KNIME Analytics Platform
, ,
1

Hello KNIME team,

I’m facing a persistent issue with the Keras integration in KNIME and would appreciate guidance on best practice.

My setup:

  • OS: Windows 10 (64-bit)
  • KNIME Analytics Platform: 5.9.0
  • Anaconda: 2025.12.1 (64-bit)
  • Python: 3.9
  • Conda integration configured correctly and detected by KNIME

Problem:

  • Keras nodes fail to execute due to TensorFlow / Keras version incompatibility
  • Errors such as:
    “Unable to convert function return value to a Python type”
  • TensorFlow is detected but KNIME reports incompatible versions for Keras nodes
  • I prefer using Keras-style nodes (high-level DL workflow) and not raw TensorFlow scripting

What I already tried:

  • Creating fresh Conda environments
  • Installing different TensorFlow / Keras versions
  • Verifying Python Deep Learning preferences
  • Restarting KNIME after each change

Questions:

  1. What is the recommended TensorFlow + Keras version combination for KNIME 5.9.0?
  2. Is TensorFlow strictly required under the hood even when using Keras nodes?
  3. Is there an officially supported Conda environment recipe for this setup?
  4. Should Keras nodes be avoided in favor of TensorFlow nodes going forward?

Any official guidance or reference documentation would be greatly appreciated.

Thank you!

This issue blocks production workflows and reproducibility across KNIME environments.

Screenshot (3417)
Screenshot (3417)1280×1024 97.2 KB

2

@medoteto I have written about the challenges of setting up knime and deep learning in this slightly older article:

https://medium.com/low-code-for-advanced-data-science/knime-and-python-setting-up-deep-learning-environments-for-keras-and-tensorflow-4b66003858f4

1 Like