Polish analysis of a multiclosure test #1950
Conversation
|
If you can do this as a branch off the new data reader that would be immensely helpful :__ and it should also work for everything other than ttbar and jet/dijets ... |
I would do it but we actually need jets and ttbar |
| import matplotlib.pyplot as plt | ||
| import matplotlib.colors |
There was a problem hiding this comment.
I think we should use validphys.subplots instead
|
|
||
| import numpy as np | ||
| from sklearn.decomposition import PCA | ||
| from sklearn import preprocessing |
There was a problem hiding this comment.
| from sklearn import preprocessing |
Seems to be unused
| # fits_dataset_predictions = [ | ||
| # ThPredictionsResult.from_convolution(pdf, dataset) for pdf in fits_pdf | ||
| # ] | ||
|
|
||
| # dimensions here are (Nfits, Ndat, Nrep) | ||
| # reps = np.asarray([th.error_members for th in fits_dataset_predictions]) | ||
|
|
||
| # reshape so as to get PCs from all the samples | ||
| # reps = reps.reshape(reps.shape[1],-1) | ||
|
|
There was a problem hiding this comment.
| # fits_dataset_predictions = [ | |
| # ThPredictionsResult.from_convolution(pdf, dataset) for pdf in fits_pdf | |
| # ] | |
| # dimensions here are (Nfits, Ndat, Nrep) | |
| # reps = np.asarray([th.error_members for th in fits_dataset_predictions]) | |
| # reshape so as to get PCs from all the samples | |
| # reps = reps.reshape(reps.shape[1],-1) |
| # rescale feature matrix | ||
| reps_scaled = reps # preprocessing.scale(reps) | ||
|
|
||
| # choose number of principal components (PCs) based on explained variance ratio | ||
| n_comp = 1 | ||
| for _ in range(reps.shape[0]): | ||
| pca = PCA(n_comp).fit(reps_scaled.T) | ||
| if np.sum(pca.explained_variance_ratio_) >= explained_variance_ratio: | ||
| break | ||
| n_comp += 1 |
There was a problem hiding this comment.
| # rescale feature matrix | |
| reps_scaled = reps # preprocessing.scale(reps) | |
| # choose number of principal components (PCs) based on explained variance ratio | |
| n_comp = 1 | |
| for _ in range(reps.shape[0]): | |
| pca = PCA(n_comp).fit(reps_scaled.T) | |
| if np.sum(pca.explained_variance_ratio_) >= explained_variance_ratio: | |
| break | |
| n_comp += 1 | |
| # choose number of principal components (PCs) based on explained variance ratio | |
| n_comp = 1 | |
| for _ in range(reps.shape[0]): | |
| pca = PCA(n_comp).fit(reps.T) | |
| if np.sum(pca.explained_variance_ratio_) >= explained_variance_ratio: | |
| break | |
| n_comp += 1 |
I think that we can remove this. Or otherwise set preprocessing as an option. The idea was to rescale the feature matrix so as to have it unit free. I am not sure however, whether we really want to have it as we compute the PCA for each dataset separately
| if n_comp <=1: | ||
| return None, None, n_comp |
There was a problem hiding this comment.
This is probably a bit ugly, but it's the way I had found to deal either with datasets that have one datapoint only or with those that for the given explained variance ratio shrink to 1.
I would be happy to discuss on this
|
Hi @andreab1997, is it fine for you if I commit to this branch? |
Yes fell free. Maybe first we want to rebase this on the new commondata reader, as @scarlehoff was suggesting |
|
@andreab1997, @scarlehoff I am trying to rebase on master but I get a series of conflict that I don't understand too much. |
@scarlehoff Is this still an issue? @andreab1997 If the one above is not an issue please feel free to close this PR and open one for this new branch. |
No, @andreab1997 fixed it. And on the process of fixing it we found out that some of the values were wrong... (I think for dijet, the value of Q had an extra square or an extra square root, I don't remember right now). It should be now fixed in master. |
Ok, great, thanks for that! |
Here we collect new functions and template that allow the analysis of multiclosure tests.
@comane @giovannidecrescenzo @mariaubiali