lab-cosmo / kernel-tutorials Goto Github PK

currently under tests/ etc. working off of CSD

By default no random seed is set for the train/test partition. It would be nice to get consistent runs would be to set by default the seed of, I guess, numpy.

Sklearn has a preprocessing function called KernelCenterer that is called in kernel functions. We need to check if this centering is in line with ours and if not, write our own analogous function.

1_LinearMethods.ipynb

Error message is received when running the following cell.
The variable name should be 'n_to_select' after its change in scikit-cosmo.

var_dict = load_variables()
locals().update(var_dict)

--------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-17-82d4ac1ba4d8> in <module>
----> 1 var_dict = load_variables()
      2 locals().update(var_dict)

kernel-tutorials/utilities/general.py in load_variables(cache_name, **kwargs)
     93         X, y = load_csd_1000r(return_X_y=True)
     94         data = dict(X=X, Y=y, indices=np.array([]))
---> 95     return calculate_variables(**dict(data), **kwargs)
     96 
     97 

kernel-tutorials/utilities/general.py in calculate_variables(X, Y, indices, n_atoms, N, n_FPS, kernel_func, i_train, i_test, n_train, K_train, K_test)
    115 
    116     if n_FPS is not None and n_FPS < X.shape[1]:
--> 117         fps_idxs = FPS(n_features_to_select=n_FPS).fit(X).selected_idx_
    118         print("Taking a subsampling of ", n_FPS, "features")
    119         X = X[:, fps_idxs]

miniconda3/lib/python3.9/site-packages/skcosmo/feature_selection/_base.py in __init__(self, **kwargs)
     15 
     16     def __init__(self, **kwargs):
---> 17         super().__init__(selection_type="feature", **kwargs)
     18 
     19 

miniconda3/lib/python3.9/site-packages/skcosmo/_selection.py in __init__(self, initialize, **kwargs)
    817         self.initialize = initialize
    818 
--> 819         super().__init__(
    820             **kwargs,
    821         )

TypeError: __init__() got an unexpected keyword argument 'n_features_to_select'

Centering Class with does pre-processing and uncentering

(This issue was found by Aditi)

The tutorial requires the librascal package, so it would be helpful to have installation instructions in the foreword notebook (the most obvious attempts, pip install librascal and pip install rascal, don't work - in fact, the latter installs a completely different package that just happens to have the same name). The condensed installation instructions for librascal would be:

$ git clone https://github.com/cosmo-epfl/librascal.git
$ cd librascal
$ pip install .

And for installing scikit-cosmo, perhaps also mention that it's called skcosmo on pip?

Thanks!

Hello!

I have been looking into SparseKPCovR from the Utility Class.

1. It seems that the predicted targets are missing the average of the training set, i.e the correct prediction of "y" should be "y_pred -> y_pred + Y_train.mean(axis=0)". I can't think of why this is happening from looking at the code.

2. I noticed that in the 4th notebook, the correlation matrix C_pca is normalized. However, this is not the case in the Utility Class.

it's both faster and more stable.

In notebook 5_CUR section 5.4, the table_from_dict() functions has a T_test2 argument that is not definied before.