CosmoPower

[cmbant]

For fun I tried to see if I could get cosmopower working following instructions, but did not get far- looks quite out of date? e.g. presumably data files should be linked to https://github.com/cosmopower-organization/jense_2024_emulators rather than original file, and unclear how one is supposed to use a TT+EE+TE likelihood with all the separate network files.

(As previously discussed, I guess ideally it does not need to be in SOLikeT at all if the cobaya component is included in the cosmopower package)

When I try to run the example on the CosmoPower home page I also fail with this

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
Cell In[2], line 7
      4 import cosmopower as cp
      6 # load pre-trained NN model: maps cosmological parameters to CMB TT log-C_ell
----> 7 cp_nn = cp.cosmopower_NN(restore=True, 
      8                          restore_filename=r'C:\Work\Dist\git\SO\SOLikeT\soliket\cosmopower\data\CP_paper\cmb_TT_NN')
     10 # create a dict of cosmological parameters
     11 params = {'omega_b': [0.0225],
     12           'omega_cdm': [0.113],
     13           'h': [0.7],
   (...)
     16           'ln10^{10}A_s': [3.07],
     17           }

File ~\anaconda3\envs\py12\Lib\site-packages\cosmopower\cosmopower_NN.py:68, in cosmopower_NN.__init__(self, parameters, modes, parameters_mean, parameters_std, features_mean, features_std, n_hidden, restore, restore_filename, trainable, optimizer, verbose)
     66 # restore
     67 if restore is True:
---> 68     self.restore(restore_filename)
     70 # else set variables from input arguments
     71 else:
     72     # attributes
     73     self.parameters = parameters

File ~\anaconda3\envs\py12\Lib\site-packages\cosmopower\cosmopower_NN.py:285, in cosmopower_NN.restore(self, filename)
    278 # load attributes
    279 with open(filename + ".pkl", 'rb') as f:
    280     self.W_, self.b_, self.alphas_, self.betas_, \
    281     self.parameters_mean_, self.parameters_std_, \
    282     self.features_mean_, self.features_std_, \
    283     self.n_parameters, self.parameters, \
    284     self.n_modes, self.modes, \
--> 285     self.n_hidden, self.n_layers, self.architecture = pickle.load(f)

ModuleNotFoundError: No module named 'tensorflow.python.training.tracking'

[borisbolliet]

Which tensorflow version are you running with?

Currently, cosmopower requires:
tensorflow<2.14
tensorflow_probability<0.22

[borisbolliet]

see also: alessiospuriomancini/cosmopower#22

[cmbant]

I do have tensorflow 2.17. But I did do a clean install. Specifically I did

conda create -n cosmopower python=3.12 scipy matplotlib
activate cosmopower
pip install cosmopower
pip install ipykernel jupyterlab
python -m ipykernel install --user --name=cosmopower

However, running cosmopower now gives

Failed to import TF-Keras. Please note that TF-Keras is not installed by default when you install TensorFlow Probability. This is so that JAX-only users do not have to install TensorFlow or TF-Keras. To use TensorFlow Probability with TensorFlow, please install the tf-keras or tf-keras-nightly package.
This can be be done through installing the tensorflow-probability[tf] extra.

So I did
pip install tensorflow-probability[tf]
Perhaps a version-tagged version of tensorflow-probability[tf] needs to be in requirements?

[cmbant]

Converting all pkl to numpy does seem like a good idea (dpiras/cosmopower-jax#5)

[alessiospuriomancini]

This has now been fixed in commit 65fa9d6 and incorporated in the new v0.2.0 release. Note that python versions up to 3.11 are supported. python=3.12 is currently in bugfix and support for it will likely be added in alessiospuriomancini/cosmopower#28.

[cmbant]

Thanks! (btw, Anaconda's default download is now python 12)