Has anyone tried to implement the infamous facebook prophet (fbprophet) time-series forecasting algorithm in KNIME?
Forecasting at Scale - Whitepaper
My approach was to create a local portable Python3 Installation and install fbprophet to later use the Python Scripting Nodes to handle model training and prediction. Unfortunately I did not even overcome the basic installation steps, as the installation of required packages like pystan is very troublesome (https://pystan.readthedocs.io/en/latest/windows.html) and I’m stuck with Windows on my corporate computer (won’t compile requirements, no binary packages found).
My question is if anyone ever got prophet working in KNIME and what your experiences were (and what your setup/weapons of choice were).
Many thanks in advance and best regards,
welcome to the KNIME community.
Indeed, prophet is an interesting tool for forecasting. Since we have the python (and R) integration available in KNIME, we recommend to use those to make use of such cutting edge tools. We might look into providing a dedicated node to integrate prophet in the future, but it is not on our TODO list at this stage.
As for technical difficulties installing prophet (and specifically pystan) on Windows, as long as you use conda, you can install all required packages (including relevant compilers to build pystan) as simple as
conda install -c conda-forge fbprophet from within the virtual environment that you are going to use in knime. This way i managed to install it on my work laptop on Windows 10. It is available for import in python script, but i did not rtry to train a model, as i did not have relevant data.
Let us know if you encounter problems.
thank you for your response! I regularly use several (manual) Python installations in parallel and have tried Anaconda many times before, but -tbh- found it super sluggish and almost impossible to get to work behind the corporate NTLM proxy without flaws… (pip is fine, the Anaconda Navigator and conda not so much)
Considering your encouraging comment, I think I will try my luck with a miniconda installation next and 'll see how far I get with the basic installation process, before deciding how to tackle the Python Integration of the package in KNIME (probably splitting up model-setup, training, prediction and evaluation).
I will post again with the results.
Many thanks and best regards,
yes, conda can be a pain, but it can also pleasantly surprise you
Let us know the outcome of the installation process. Using pip it indeed might be more involved, as one needs to install a compiler and build pystan and i would have no idea how to do it best on Windows, so i keep my fingers crossed
This topic was automatically closed 182 days after the last reply. New replies are no longer allowed.