Wine Quality Detection ML Project

Author: Giuseppe Ruggeri, s292459

Folder structure

The folder structure is very simple and it is composed as following:

• classifiers - a package containing all the needed classifiers
• density_estimation - a package containing the density estimation methods
• evaluation - a package containing the methods for evaluating our models
• preproc - a package containing all the preprocess methods
• wine_project - a folder containing all the scripts, logs, and graphs for the project
  • utility - a package containing recurrent methods to be used during the analyses
  • data - contains the dataset and the cached models for fast processing
  • graphs - will contain the generated graphs
  • train_logs - will contain the generated train logs
  • scripts - it contains all the scripts for performing the different analyses

Usage

Please notice that you have to add the modules to your PYTHONPATH variable or use a sufficiently intelligent IDE (e.g. PyCharm). We briefly list, for each available analysis, the available options.
NB: For the classifiers are available all the option repeated in three variants: standard, with the prefix eval_partial_ and with the prefix eval_full_, they are omitted in the following for conciseness.
The eval_partial_ starts the same functionality but on the evaluation dataset, using 4/5 of the training dataset.
The eval_full_ starts the same functionality but on the evaluation dataset, using the full training dataset.

ds_preprocessing.py

It just splits the dataset generating the 5-folds split

• --load (default: true) - load the already 5-folds split dataset

ds_analysis.py

Performs the training dataset analysis, including raw and gaussian features, histograms, correlation matrices.

• --hist_raw (default: false) - show histograms for the raw featuers
• --hist-gau (default: false) - show histograms for the gaussianized features
• --show_correlations (defualt: false) - show correlation matrices

mvg_classifier.py

Performs the model selection, hyperparameters optimization on the train dataset and evaluation contrasting on the test dataset.

• --gridsearch (default: false) - perform a gridsearch on the train dataset
• --eval_partial_..
• --eval_full_..

lr_classifier.py

• --linear (default: false) - performs a gridsearch for linear LR on the train dataset
• --quadratic (default: false) - performs a gridsearch for quadratic LR on the train dataset
• --linear_app_specific (default: false) - embed the target priors in the objective for the best models obtained during the gridsearch
• --quadratic_app_specific (default: false) - embed the target priors in the objective for the best models obtained during the gridsearch
• --eval_partial_..
• --eval_full_..

linear_svm_classifier.py

• --gridsearch (default: false) - performs a gridsearch for linear SVM on the train dataset
• --class_balancing (default: false) - performs a small gridsearch (with different values of C and the best preproc configuration) but class-balancing with respect to the target application
• --eval_partial_..
• --eval_full_..

polynomial_svm_classifier.py

• --gridsearch (default: false) - performs a gridsearch for quadratic and cubic SVM on the train dataset
• --class_balancing (default: false) - performs a class-balancing training with the best preproc configuration and the best hyperparameters found in the gridsearch
• --eval_partial_..
• --eval_full_..

rbf_svm_classifier.py

• --gridsearch (default: false) - performs a coarse-level gridsearch for rbf svm to jointly optimize C and gamma
• --gridsearch_fine_grained (default: false) - performs a fine-grained gridsearch around the optimal values found
• --class_balancing (default: false) - performs a class-balancing training with the best preproc configuration and the best hyperparameters found in the gridseach
• --actual_dcf (default: false) - compute the actual dcf on the validation set (of the pulled scores during the K-fold cross-validation on the train set)
• --eval_partial_..
• --eval_full_..

gmm_classifier.py

• --gridsearch (default: false) - start a gridsearch cross-validation to optimize with respect to the number of components, the preprocess configuration and the model type
• --actual_dcf (default: false) - compute the actual dcf on the validation set (of the pulled scores during the K-fold cross-validation on the train set)
• --eval_partial_..
• --eval_full_..

svm_gmm_analysis.py

Performs the analysis of the two candidate models, re-calibration and fusion.

• --bayes_error_plot (default: false) - shows the bayes error plots of the two candidate models (uncalibrated) over the validation dataset (of the train dataset)
• --lr_recalibration (default: false) - recalibrate the scores using a linear LR for the best svm and gmm model and fuse them
• --eval_partial_..
• NB: eval_full_ not available