New version of tensorflow breaks ``stella``

[TomWagg]

Models for stella were run with an earlier version of tensorflow but for version >= 2.16 stella breaks when trying to evaluate. E.g.

ValueError: Kernel shape must have the same length as input, but received kernel of shape (3, 1, 16) and input of shape (None, None, 200, 1).

This should be a quick fix in the requirements.txt by capping the tensorflow version

[github-actions]

Hi there! This is a test of github actions.

[benjaminpope]

I just wanted to bump this thread - I have found it pretty hard to find an environment in which stella works but would love to use it again for some student projects. I wonder does anyone have an example of a working environment they could share?

[m-schochet]

I wanted to come here because I have found a work-around solution to the issues noted by @TomWagg and @benjaminpope .

Essentially, I think enforcing the tensorflow version as suggested above is a workable solution. However, going mostly off this page on the new Keras site (https://keras.io/getting_started/), in the section on 'TensorFlow + Keras 2 backwards compatibility' it notes that after the update to Tensorflow 2.15 and onward, Keras 3 is used when calling from tensorflow import keras. Since the original stella models ran using Keras 2, we need to additionally install the maintained tf-keras package that is continuing the Keras 2 backend (https://github.com/keras-team/tf-keras). We also need to add

import os
os.environ["TF_USE_LEGACY_KERAS"] = "1"

To all our python scripts or Jupyter notebooks using stella (and they must come before calling import stella to ensure that the correct tensorflow keras backend is loaded). To make a reproducible example of how to run the quickstart tutorials currently available online (https://adina.feinste.in/stella/getting_started/shortest_demo.html), I made a new conda environment and git-cloned the current stella repo

conda create -n stellarflare python==3.13 setuptools
conda activate stellarflare
git clone https://github.com/afeinstein20/stella
cd stella
python setup.py install

(I used Python 3.13 since that was the last version I could verify that tensorflow is compatible with, and setuptools isn't autoinstalled so I needed that package too)

However, something (perhaps on my machine) causes the dependency packages to not be installed after this line. As a result, I then simply called pip install -r requirements.txt which thankfully installed all the dependencies for me. Then, after the stella dependencies were installed, I also use ipykernel to make it an environment for my Jupyter notebook for testing via

python -m ipykernel install --user --name stella

Finally, I called pip install tf-keras. After this, I opened up a jupyter notebook, and ran the tutorial code (using a TESS LC from my own research).

# extremely important commands
import os 
os.environ["TF_USE_LEGACY_KERAS"] = "1"

import numpy as np
import lightkurve as lk
import matplotlib.pyplot as plt
import astropy.units as u
import stella
from tqdm import tqdm_notebook

tess_lc = lk.read('/Users/mschochet/Desktop/tesslc/tess2025206162959-s0095-0000000441420236-0292-a_fast-lc.fits')
lc = tess_lc.PDCSAP_FLUX

ds = stella.DownloadSets()
ds.download_models()

OUT_DIR = '/Users/mschochet/Desktop/stella_results/'
cnn = stella.ConvNN(output_dir=OUT_DIR)

cnn.predict(modelname=str(ds.models[0]),
            times=lc.time.value,
            fluxes=lc.flux,
            errs=lc.flux_err)
single_pred = cnn.predictions[0]

preds = np.zeros((len(ds.models),len(cnn.predictions[0])))

for i, m in enumerate(ds.models):
    cnn.predict(modelname=m,
                times=lc.time.value,
                fluxes=lc.flux,
                errs=lc.flux_err)
    preds[i] = cnn.predictions[0]

avg_pred = np.nanmedian(preds, axis=0)

fig, (ax1, ax2) = plt.subplots(figsize=(14,8), nrows=2,
                               sharex=True, sharey=True)
im = ax1.scatter(cnn.predict_time[0], cnn.predict_flux[0],
            c=avg_pred, vmin=0, vmax=1)
ax2.scatter(cnn.predict_time[0], cnn.predict_flux[0],
            c=single_pred, vmin=0, vmax=1)
ax2.set_xlabel('Time [BJD-2457000]')
ax2.set_ylabel('Normalized Flux', y=1.2)

fig.subplots_adjust(right=0.8)
cbar_ax = fig.add_axes([0.81, 0.15, 0.02, 0.7])
fig.colorbar(im, cax=cbar_ax, label='Probability')


ax1.set_title('Averaged Predictions')
ax2.set_title('Single Model Predictions')

plt.subplots_adjust(hspace=0.4)

Which output the final tutorial figure that I was after. This solution should not require any changes to stella, but I am unsure how this relates to the solution suggested in #34