This project investigate Machine Learning techniques to Particle Track reconstruction problems to HEP, it is part of SPRACE sponsored by Serrapilheira. This is a work flow of proposal.
We use different Machine Learning tecnhiques to resolve this big problem in the physics community. If you want to reproduce our results, we have written some general steps. You are welcome, if you have some ideas our suggestations, please let us know.
You need to install miniconda before on a linux system
- Configure your conda environtment with env.yml file.
$ conda env create -f env.yml
$ conda activate trackmlTo run:
- Clone the repository
$ git clone https://github.com/SPRACE/track-ml.git- go to
track-mldirectory created
you will need to have a GPU or some descent CPU.
We transformed the detector into three kinematical regions to train our models with different datasets.
- The first region is formed by the internal barrel with
$\eta$coordinate from($-1.0$, $1.0$). - The second region is the intermediary barrel, (overlap) with
$\eta$coordinate from($-2.0$ to $-1.0$)or($2.0$ to $1.0$). - The last region is external with
$\eta$values between($-3.0$ to $-2.0$)or($3.0$ to $2.0$).
Considering the mentioned regions and the symmetry of the detector, each dataset was filtered to contain only high energy particles pT > 1.0 GeV with $\phi$ values between ($-0.5$, $0.5$), in order to obtain tracks with larger curvature radius, facilitating initial training.
Short datasets are in dataset directory. We split a region in three parts (this is too big to be here).
There are some predefined config files to train diferents models (MLP, CNN, LSTM, CNN-parallel and others). If you need to change the parameters then change the config_*.json file. We used internal barrel as dataset, this dataset is previously transformed and linked in json file:
$ python main_train.py --config configs/config_lstm_parallel_internal.jsonThere are other configurations for example a CNN model:
$ python main_train.py --config configs/config_cnn_parallel_internal.jsonIf you want to use a gaussian model
$ python main_train.py --config configs/config_lstm_gaussian.jsonIf you want to see the training process when ajust any parameters of .json file. Run the notebook:
$ main_train.ipynbYou can inference the previous compiled model with the same config file:
$ python main_inference.py --config configs/config_lstm_parallel_internal.json --load True This will produce an report at results/encrypt_name/results-test.txt file. With the number of reconstructed tracks and quantity of hits correctly predicted by layer.
If you want to inference with a short dataset:
$ python main_inference.py --config configs/config_lstm_parallel_internal.json --load True --samples 500or
$ python main_inference.py --config configs/config_lstm_gaussian.json --load TrueYou can automatizate trainings and testings with scripts. there are some default configurations:
$ ./run_trains.sh
$ ./run_tests.shWe are using regressions metrics for accuracy of models. We show 2 groups of metrics.
-
The principal metrics is a scoring. Scoring counts how many correct hits were found per layer and comparates with original truth hits. Finally we count the quantity of tracks reconstructed.
-
The other metrics are regression metrics, we measure the error between real and predicted hits per layer.
For example, to see the accuracy of training algorithm, go to results/encrypt_name/results-train.txt file and the scoring of correct and tracks reconstructed go to results/encrypt_name/results-test.txt file.
Output test file:
[Output] Results
---Parameters---
Model Name : gaussian-lstm
Dataset : eta_n0.5-0.5_phi_n0.5-0.5_20200518171238_tracks.csv
Tracks : 3000
Model saved : compiled/model-gaussian-lstm-MEnGHa5DCmJGE9C9BpAg4i-coord-cylin-normalise-true-epochs-30-batch-30.h5
Test date : 02/10/2020 20:23:55
Coordenates : cylin
Coordenate 3D : True
Model Scaled : True
Model Optimizer : adam
Prediction Opt : gaussian
Distance metric : ['polar', 'euclidean', 'mahalanobis', 'cosine']
Remove hit : False
Correct hits per layer (polar) ['191.0(6.37%)', '135.0(4.5%)', '182.0(6.07%)', '167.0(5.57%)', '121.0(4.03%)', '127.0(4.23%)'] of 3000 tracks:
Correct hits per layer (euclidean) ['61.0(2.03%)', '54.0(1.8%)', '64.0(2.13%)', '62.0(2.07%)', '49.0(1.63%)', '36.0(1.2%)'] of 3000 tracks:
Correct hits per layer (mahalanobis) ['191.0(6.37%)', '193.0(6.43%)', '178.0(5.93%)', '193.0(6.43%)', '165.0(5.5%)', '133.0(4.43%)'] of 3000 tracks:
Correct hits per layer (cosine) ['2323.0(77.43%)', '1980.0(66.0%)', '1792.0(59.73%)', '1702.0(56.73%)', '1487.0(49.57%)', '1351.0(45.03%)'] of 3000 tracks:
Reconstructed tracks: 887 of 3000 tracks (29.566666666666666)
Above output shows scoring per layer for example 48% with 256 hits were matched at the first layer, results are 74 tracks reconstructed of 528 tracks(it is a short dataset just). We also write other info like what kind of coordinate, if we use the nearest optimization, epochs, batchs, optimazer used, model name etc.
Regression metrics per layer are:
---Regression Scores---
R_2 statistics (R2) = 0.992
Mean Square Error (MSE) = 882.525
Root Mean Square Error(RMSE) = 29.707
Mean Absolute Error (MAE) = 9.858
layer 5
---Regression Scores---
R_2 statistics (R2) = 1.0
Mean Square Error (MSE) = 6.818
Root Mean Square Error(RMSE) = 2.611
Mean Absolute Error (MAE) = 1.325
layer 6
---Regression Scores---
R_2 statistics (R2) = 0.999
Mean Square Error (MSE) = 27.603
Root Mean Square Error(RMSE) = 5.254
Mean Absolute Error (MAE) = 2.541
layer 7
---Regression Scores---
R_2 statistics (R2) = 0.998
Mean Square Error (MSE) = 141.074
Root Mean Square Error(RMSE) = 11.877
Mean Absolute Error (MAE) = 5.285
The last output shows one geral metric for all hits and four (R^2, MSE, RMSE, MAE) metrics per layer.
If you want to see the results with plots, go to the plot_prediction.ipynb file at notebooks directory. This example plots cartesian coordinates.
 |
The next plot shows all hits.
 |
There is other notebook for plot the gaussian distributions plot_distributions.ipynb
 |
The pull of distributions for
 |
 |