[PWGHF] Patches for first Xic autoencoder study #10813
Conversation
macamerl
requested review from
alibuild,
vkucera,
fcolamar,
fgrosa,
fcatalan92,
mfaggin,
mmazzilli,
deepathoms,
NicoleBastid,
hahassan7,
jpxrk,
apalasciano and
zhangbiao-phy
as code owners
github-actions
bot
changed the title
Please consider the following formatting changes to AliceO2Group#10813
|
Hi @macamerl, Thanks for the development! First of all, let me ping the code owner to see if they have any comments on this implementation. @mfaggin @cterrevo and @phymanshu. Then I only have a few comments, see the review. Then could you also fix magalinter errors since they come from this PR? |
changes to fix megalinter check
changes to fix megalinter checks
Changes to fix the PR checks
|
Hi @macamerl @zhangbiao-phy |
|
Thanks @phymanshu. Yes, I locally checked that the BDT inference has no problem after the changes that I made. The flags "applyMSE" and "applyMinMax" are set false/0 by default. In any case, I will privately send you the config file that I used for the check. |
| //---- temporary MSE function | ||
| float MSE(std::vector<float> y_true, std::vector<float> y_pred, std::vector<float> ScaleMin, std::vector<float> ScaleMax, bool scaling) | ||
| { | ||
| float mse = 0.0f; | ||
| float sum = 0.0f; | ||
|
|
||
| std::vector<float> yf_true; | ||
| if (y_true.size() != y_pred.size()) { | ||
| LOG(debug) << "size of input vector =" << y_true.size(); | ||
| LOG(debug) << "size of AE output vector =" << y_pred.size(); | ||
| LOG(fatal) << "vectors of input and predictions don't have the same size"; | ||
| } else { // MSE | ||
| for (size_t j = 0; j < y_pred.size(); ++j) { // for over the features | ||
| if (scaling == true) { // MinMax scaling | ||
| LOG(debug) << "--------------> MinMax scaling Debug \t" << ScaleMin.at(j) << "\t" << ScaleMax.at(j); | ||
| yf_true.push_back((y_true.at(j) - ScaleMin.at(j)) / (ScaleMax.at(j) - ScaleMin.at(j))); | ||
| LOG(debug) << "feature = " << j << " ----> input = " << y_true.at(j) << " scaled feature = " << yf_true.at(j); | ||
| } else { | ||
| yf_true.push_back(y_true.at(j)); | ||
| } | ||
| sum += pow(((yf_true).at(j) - (y_pred).at(j)), 2); // has dimensions | ||
| LOG(debug) << "feature = " << j << " ----> input = " << yf_true.at(j) << " AE prediction = " << y_pred.at(j); | ||
| } | ||
| mse = sum / y_pred.size(); // MSE of a candidate | ||
| } |
There was a problem hiding this comment.
Since this is not HF specific, I would move this in the ML folder. You could consider to add it as a function of the https://github.com/AliceO2Group/O2Physics/blob/master/Tools/ML/MlResponse.h class, or to add a header file with utils functions.
Please also follow the naming conventions of O2Physics when you move it (camel case), thanks!
| if (applyMSE) { | ||
| hfMlXicToPKPiCandidate(outputMSEXicToPKPi, outputMSEXicToPiKP); | ||
| } else { | ||
| hfMlXicToPKPiCandidate(outputMlXicToPKPi, outputMlXicToPiKP); | ||
| } |
There was a problem hiding this comment.
Can you not avoid all the branching and just make the decision once?
Minimal changes in XicToPKPi specific files (task and selector) to evaluate the MSE between input (externally scaled) and internal scaled output of autoencoder. At moment, one can not run Autoencoder and BDT at the same time since they competitively fill candidate.mlProbXicToPKPi()[0].