Multi-Core-Machine-Learning-in-Python-With-Scikit-Learn.png

Multi-Core Machine Learning in Python With Scikit-Learn

Many computationally expensive tasks for machine learning can be made parallel by splitting the work across multiple CPU cores, referred to as multi-core processing.

Common machine learning tasks that can be made parallel include training models like ensembles of decision trees, evaluating models using resampling procedures like k-fold cross-validation, and tuning model hyperparameters, such as grid and random search.

Using multiple cores for common machine learning tasks can dramatically decrease the execution time as a factor of the number of cores available on your system. A common laptop and desktop computer may have 2, 4, or 8 cores. Larger server systems may have 32, 64, or more cores available, allowing machine learning tasks that take hours to be completed in minutes.

In this tutorial, you will discover how to configure scikit-learn for multi-core machine learning.

After completing this tutorial, you will know:

  • How to train machine learning models using multiple cores.
  • How to make the evaluation of machine learning models parallel.
  • How to use multiple cores to tune machine learning model hyperparameters.

Let’s get started.

Multi-Core Machine Learning in Python With Scikit-Learn

Multi-Core Machine Learning in Python With Scikit-Learn
Photo by ER Bauer, some rights reserved.

Tutorial Overview

This tutorial is divided into five parts; they are:

  • Multi-Core Scikit-Learn
  • Multi-Core Model Training
  • Multi-Core Model Evaluation
  • Multi-Core Hyperparameter Tuning
  • Recommendations
  • Multi-Core Scikit-Learn

    Machine learning can be computationally expensive.

    There are three main centers of this computational cost; they are:

    • Training machine learning models.
    • Evaluating machine learning models.
    • Hyperparameter tuning machine learning models.

    Worse, these concerns compound.

    For example, evaluating machine learning models using a resampling technique like k-fold cross-validation requires that the training process is repeated multiple times.

    • Evaluation Requires Repeated Training

    Tuning model hyperparameters compounds this further as it requires the evaluation procedure repeated for each combination of hyperparameters tested.

    • Tuning Requires Repeated Evaluation

    Most, if not all, modern computers have multi-core CPUs. This includes your workstation, your laptop, as well as larger servers.

    You can configure your machine learning models to harness multiple cores of your computer, dramatically speeding up computationally expensive operations.

    The scikit-learn Python machine learning library provides this capability via the n_jobs argument on key machine learning tasks, such as model training, model evaluation, and hyperparameter tuning.

    This configuration argument allows you to specify the number of cores to use for the task. The default is None, which will use a single core. You can also specify a number of cores as an integer, such as 1 or 2. Finally, you can specify -1, in which case the task will use all of the cores available on your system.

    • n_jobs: Specify the number of cores to use for key machine learning tasks.

    Common values are:

    • n_jobs=None: Use a single core or the default configured by your backend library.
    • n_jobs=4: Use the specified number of cores, in this case 4.
    • n_jobs=-1: Use all available cores.

    What is a core?

    A CPU may have multiple physical CPU cores, which is essentially like having multiple CPUs. Each core may also have hyper-threading, a technology that under many circumstances allows you to double the number of cores.

    For example, my workstation has four physical cores, which are doubled to eight cores due to hyper-threading. Therefore, I can experiment with 1-8 cores or specify -1 to use all cores on my workstation.

    Now that we are familiar with the scikit-learn library’s capability to support multi-core parallel processing for machine learning, let’s work through some examples.

    You will get different timings for all of the examples in this tutorial; share your results in the comments. You may also need to change the number of cores to match the number of cores on your system.

    Note: Yes, I am aware of the timeit API, but chose against it for this tutorial. We are not profiling the code examples per se; instead, I want you to focus on how and when to use the multi-core capabilities of scikit-learn and that they offer real benefits. I wanted the code examples to be clean and simple to read, even for beginners. I set it as an extension to update all examples to use the timeit API and get more accurate timings. Share your results in the comments.

    Multi-Core Model Training

    Many machine learning algorithms support multi-core training via an n_jobs argument when the model is defined.

    This affects not just the training of the model, but also the use of the model when making predictions.

    A popular example is the ensemble of decision trees, such as bagged decision trees, random forest, and gradient boosting.

    In this section we will explore accelerating the training of a RandomForestClassifier model using multiple cores. We will use a synthetic classification task for our experiments.

    In this case, we will define a random forest model with 500 trees and use a single core to train the model.


    We can record the time before and after the call to the train() function using the time() function. We can then subtract the start time from the end time and report the execution time in the number of seconds.

    The complete example of evaluating the execution time of training a random forest model with a single core is listed below.


    Running the example reports the time taken to train the model with a single core.

    In this case, we can see that it takes about 10 seconds.

    How long does it take on your system? Share your results in the comments below.


    We can now change the example to use all of the physical cores on the system, in this case, four.


    The complete example of multi-core training of the model with four cores is listed below.


    Running the example reports the time taken to train the model with a single core.

    In this case, we can see that the speed of execution more than halved to about 3.151 seconds.

    How long does it take on your system? Share your results in the comments below.


    We can now change the number of cores to eight to account for the hyper-threading supported by the four physical cores.


    We can achieve the same effect by setting n_jobs to -1 to automatically use all cores; for example:


    We will stick to manually specifying the number of cores for now.

    The complete example of multi-core training of the model with eight cores is listed below.


    Running the example reports the time taken to train the model with a single core.

    In this case, we can see that we got another drop in execution speed from about 3.151 to about 2.521 by using all cores.

    How long does it take on your system? Share your results in the comments below.


    We can make the relationship between the number of cores used during training and execution speed more concrete by comparing all values between one and eight and plotting the result.

    The complete example is listed below.


    Running the example first reports the execution speed for each number of cores used during training.

    We can see a steady decrease in execution speed from one to eight cores, although the dramatic benefits stop after four physical cores.

    How long does it take on your system? Share your results in the comments below.


    A plot is also created to show the relationship between the number of cores used during training and the execution speed, showing that we continue to see a benefit all the way to eight cores.

    Line Plot of Number of Cores Used During Training vs. Execution Speed

    Line Plot of Number of Cores Used During Training vs. Execution Speed

    Now that we are familiar with the benefit of multi-core training of machine learning models, let’s look at multi-core model evaluation.

    Multi-Core Model Evaluation

    The gold standard for model evaluation is k-fold cross-validation.

    This is a resampling procedure that requires that the model is trained and evaluated k times on different partitioned subsets of the dataset. The result is an estimate of the performance of a model when making predictions on data not used during training that can be used to compare and select a good or best model for a dataset.

    In addition, it is also a good practice to repeat this evaluation process multiple times, referred to as repeated k-fold cross-validation.

    The evaluation procedure can be configured to use multiple cores, where each model training and evaluation happens on a separate core. This can be done by setting the n_jobs argument on the call to cross_val_score() function; for example:

    We can explore the effect of multiple cores on model evaluation.

    First, let’s evaluate the model using a single core.


    We will evaluate the random forest model and use a single core in the training of the model (for now).


    The complete example is listed below.


    Running the example evaluates the model using 10-fold cross-validation with three repeats.

    In this case, we see that the evaluation of the model took about 6.412 seconds.

    How long does it take on your system? Share your results in the comments below.


    We can update the example to use all eight cores of the system and expect a large speedup.


    The complete example is listed below.


    Running the example evaluates the model using multiple cores.

    In this case, we can see the execution timing dropped from 6.412 seconds to about 2.371 seconds, giving a welcome speedup.

    How long does it take on your system? Share your results in the comments below.


    As we did in the previous section, we can time the execution speed for each number of cores from one to eight to get an idea of the relationship.

    The complete example is listed below.


    Running the example first reports the execution time in seconds for each number of cores for evaluating the model.

    We can see that there is not a dramatic improvement above four physical cores.

    We can also see a difference here when training with eight cores from the previous experiment. In this case, evaluating performance took 1.492 seconds whereas the standalone case took about 2.371 seconds.

    This highlights the limitation of the evaluation methodology we are using where we are reporting the performance of a single run rather than repeated runs. There is some spin-up time required to load classes into memory and perform any JIT optimization.

    Regardless of the accuracy of our flimsy profiling, we do see the familiar speedup of model evaluation with the increase of cores used during the process.

    How long does it take on your system? Share your results in the comments below.


    A plot of the relationship between the number of cores and the execution speed is also created.

    Line Plot of Number of Cores Used During Evaluation vs. Execution Speed

    Line Plot of Number of Cores Used During Evaluation vs. Execution Speed

    We can also make the model training process parallel during the model evaluation procedure.

    Although this is possible, should we?

    To explore this question, let’s first consider the case where model training uses all cores and model evaluation uses a single core.


    The complete example is listed below.


    Running the example evaluates the model using a single core, but each trained model uses a single core.

    In this case, we can see that the model evaluation takes more than 10 seconds, much longer than the 1 or 2 seconds when we use a single core for training and all cores for parallel model evaluation.

    How long does it take on your system? Share your results in the comments below.


    What if we split the number of cores between the training and evaluation procedures?


    The complete example is listed below.


    Running the example evaluates the model using four cores, and each model is trained using four different cores.

    We can see an improvement over training with all cores and evaluating with one core, but at least for this model on this dataset, it is more efficient to use all cores for model evaluation and a single core for model training.

    How long does it take on your system? Share your results in the comments below.


    Multi-Core Hyperparameter Tuning

    It is common to tune the hyperparameters of a machine learning model using a grid search or a random search.

    The scikit-learn library provides these capabilities via the GridSearchCV and RandomizedSearchCV classes respectively.

    Both of these search procedures can be made parallel by setting the n_jobs argument, assigning each hyperparameter configuration to a core for evaluation.

    The model evaluation itself could also be multi-core, as we saw in the previous section, and the model training for a given evaluation can also be training as we saw in the second before that. Therefore, the stack of potentially multi-core processes is starting to get challenging to configure.

    In this specific implementation, we can make the model training parallel, but we don’t have control over how each model hyperparameter and how each model evaluation is made multi-core. The documentation is not clear at the time of writing, but I would guess that each model evaluation using a single core hyperparameter configuration is split into jobs.

    Let’s explore the benefits of performing model hyperparameter tuning using multiple cores.

    First, let’s evaluate a grid of different configurations of the random forest algorithm using a single core.


    The complete example is listed below.


    Running the example tests different values of the max_features configuration for random forest, where each configuration is evaluated using repeated k-fold cross-validation.

    In this case, the grid search on a single core takes about 28.838 seconds.

    How long does it take on your system? Share your results in the comments below.


    We can now configure the grid search to use all available cores on the system, in this case, eight cores.


    We can then evaluate how long this multi-core grids search takes to execute. The complete example is listed below.


    Running the example reports execution time for the grid search.

    In this case, we see a factor of about four speed up from roughly 28.838 seconds to around 7.418 seconds.

    How long does it take on your system? Share your results in the comments below.


    Intuitively, we would expect that making the grid search multi-core should be the focus and not model training.

    Nevertheless, we can divide the number of cores between model training and the grid search to see if it offers a benefit for this model on this dataset.


    The complete example of multi-core model training and multi-core hyperparameter tuning is listed below.


    In this case, we do see a decrease in execution speed compared to a single core case, but not as much benefit as assigning all cores to the grid search process.

    How long does it take on your system? Share your results in the comments below.


    Recommendations

    This section lists some general recommendations when using multiple cores for machine learning.

    • Confirm the number of cores available on your system.
    • Consider using an AWS EC2 instance with many cores to get an immediate speed up.
    • Check the API documentation to see if the model/s you are using support multi-core training.
    • Confirm multi-core training offers a measurable benefit on your system.
    • When using k-fold cross-validation, it is probably better to assign cores to the resampling procedure and leave model training single core.
    • When using hyperparamter tuning, it is probably better to make the search multi-core and leave the model training and evaluation single core.

    Do you have any recommendations of your own?

    Further Reading

    This section provides more resources on the topic if you are looking to go deeper.

    Related Tutorials
    APIs
    Articles

    Summary

    In this tutorial, you discovered how to configure scikit-learn for multi-core machine learning.

    Specifically, you learned:

    • How to train machine learning models using multiple cores.
    • How to make the evaluation of machine learning models parallel.
    • How to use multiple cores to tune machine learning model hyperparameters.

    Do you have any questions?
    Ask your questions in the comments below and I will do my best to answer.

    Discover Fast Machine Learning in Python!

    Master Machine Learning With Python
    Develop Your Own Models in Minutes

    …with just a few lines of scikit-learn code

    Learn how in my new Ebook:
    Machine Learning Mastery With Python

    Covers self-study tutorials and end-to-end projects like:
    Loading data, visualization, modeling, tuning, and much more…

    Finally Bring Machine Learning To

    Your Own Projects

    Skip the Academics. Just Results.

    See What’s Inside

    Covid Abruzzo Basilicata Calabria Campania Emilia Romagna Friuli Venezia Giulia Lazio Liguria Lombardia Marche Molise Piemonte Puglia Sardegna Sicilia Toscana Trentino Alto Adige Umbria Valle d’Aosta Veneto Italia Agrigento Alessandria Ancona Aosta Arezzo Ascoli Piceno Asti Avellino Bari Barletta-Andria-Trani Belluno Benevento Bergamo Biella Bologna Bolzano Brescia Brindisi Cagliari Caltanissetta Campobasso Carbonia-Iglesias Caserta Catania Catanzaro Chieti Como Cosenza Cremona Crotone Cuneo Enna Fermo Ferrara Firenze Foggia Forlì-Cesena Frosinone Genova Gorizia Grosseto Imperia Isernia La Spezia L’Aquila Latina Lecce Lecco Livorno Lodi Lucca Macerata Mantova Massa-Carrara Matera Messina Milano Modena Monza e della Brianza Napoli Novara Nuoro Olbia-Tempio Oristano Padova Palermo Parma Pavia Perugia Pesaro e Urbino Pescara Piacenza Pisa Pistoia Pordenone Potenza Prato Ragusa Ravenna Reggio Calabria Reggio Emilia Rieti Rimini Roma Rovigo Salerno Medio Campidano Sassari Savona Siena Siracusa Sondrio Taranto Teramo Terni Torino Ogliastra Trapani Trento Treviso Trieste Udine Varese Venezia Verbano-Cusio-Ossola Vercelli Verona Vibo Valentia Vicenza Viterbo

    HyperOpt-for-Automated-Machine-Learning-With-Scikit-Learn.jpg

    HyperOpt for Automated Machine Learning With Scikit-Learn

    Automated Machine Learning (AutoML) refers to techniques for automatically discovering well-performing models for predictive modeling tasks with very little user involvement.

    HyperOpt is an open-source library for large scale AutoML and HyperOpt-Sklearn is a wrapper for HyperOpt that supports AutoML with HyperOpt for the popular Scikit-Learn machine learning library, including the suite of data preparation transforms and classification and regression algorithms.

    In this tutorial, you will discover how to use HyperOpt for automatic machine learning with Scikit-Learn in Python.

    After completing this tutorial, you will know:

    • Hyperopt-Sklearn is an open-source library for AutoML with scikit-learn data preparation and machine learning models.
    • How to use Hyperopt-Sklearn to automatically discover top-performing models for classification tasks.
    • How to use Hyperopt-Sklearn to automatically discover top-performing models for regression tasks.

    Let’s get started.

    HyperOpt for Automated Machine Learning With Scikit-Learn

    HyperOpt for Automated Machine Learning With Scikit-Learn
    Photo by Neil Williamson, some rights reserved.

    Tutorial Overview

    This tutorial is divided into four parts; they are:

  • HyperOpt and HyperOpt-Sklearn
  • How to Install and Use HyperOpt-Sklearn
  • HyperOpt-Sklearn for Classification
  • HyperOpt-Sklearn for Regression
  • HyperOpt and HyperOpt-Sklearn

    HyperOpt is an open-source Python library for Bayesian optimization developed by James Bergstra.

    It is designed for large-scale optimization for models with hundreds of parameters and allows the optimization procedure to be scaled across multiple cores and multiple machines.

    The library was explicitly used to optimize machine learning pipelines, including data preparation, model selection, and model hyperparameters.

    Our approach is to expose the underlying expression graph of how a performance metric (e.g. classification accuracy on validation examples) is computed from hyperparameters that govern not only how individual processing steps are applied, but even which processing steps are included.

    — Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures, 2013.

    HyperOpt is challenging to use directly, requiring the optimization procedure and search space to be carefully specified.

    An extension to HyperOpt was created called HyperOpt-Sklearn that allows the HyperOpt procedure to be applied to data preparation and machine learning models provided by the popular Scikit-Learn open-source machine learning library.

    HyperOpt-Sklearn wraps the HyperOpt library and allows for the automatic search of data preparation methods, machine learning algorithms, and model hyperparameters for classification and regression tasks.

    … we introduce Hyperopt-Sklearn: a project that brings the benefits of automatic algorithm configuration to users of Python and scikit-learn. Hyperopt-Sklearn uses Hyperopt to describe a search space over possible configurations of Scikit-Learn components, including preprocessing and classification modules.

    — Hyperopt-Sklearn: Automatic Hyperparameter Configuration for Scikit-Learn, 2014.

    Now that we are familiar with HyperOpt and HyperOpt-Sklearn, let’s look at how to use HyperOpt-Sklearn.

    How to Install and Use HyperOpt-Sklearn

    The first step is to install the HyperOpt library.

    This can be achieved using the pip package manager as follows:


    Once installed, we can confirm that the installation was successful and check the version of the library by typing the following command:


    This will summarize the installed version of HyperOpt, confirming that a modern version is being used.


    Next, we must install the HyperOpt-Sklearn library.

    This too can be installed using pip, although we must perform this operation manually by cloning the repository and running the installation from the local files, as follows:


    Again, we can confirm that the installation was successful by checking the version number with the following command:


    This will summarize the installed version of HyperOpt-Sklearn, confirming that a modern version is being used.


    Now that the required libraries are installed, we can review the HyperOpt-Sklearn API.

    Using HyperOpt-Sklearn is straightforward. The search process is defined by creating and configuring an instance of the HyperoptEstimator class.

    The algorithm used for the search can be specified via the “algo” argument, the number of evaluations performed in the search is specified via the “max_evals” argument, and a limit can be imposed on evaluating each pipeline via the “trial_timeout” argument.


    Many different optimization algorithms are available, including:

    • Random Search
    • Tree of Parzen Estimators
    • Annealing
    • Tree
    • Gaussian Process Tree

    The “Tree of Parzen Estimators” is a good default, and you can learn more about the types of algorithms in the paper “Algorithms for Hyper-Parameter Optimization. [PDF]”

    For classification tasks, the “classifier” argument specifies the search space of models, and for regression, the “regressor” argument specifies the search space of models, both of which can be set to use predefined lists of models provided by the library, e.g. “any_classifier” and “any_regressor“.

    Similarly, the search space of data preparation is specified via the “preprocessing” argument and can also use a pre-defined list of preprocessing steps via “any_preprocessing.


    For more on the other arguments to the search, you can review the source code for the class directly:

    Once the search is defined, it can be executed by calling the fit() function.


    At the end of the run, the best-performing model can be evaluated on new data by calling the score() function.


    Finally, we can retrieve the Pipeline of transforms, models, and model configurations that performed the best on the training dataset via the best_model() function.


    Now that we are familiar with the API, let’s look at some worked examples.

    HyperOpt-Sklearn for Classification

    In this section, we will use HyperOpt-Sklearn to discover a model for the sonar dataset.

    The sonar dataset is a standard machine learning dataset comprised of 208 rows of data with 60 numerical input variables and a target variable with two class values, e.g. binary classification.

    Using a test harness of repeated stratified 10-fold cross-validation with three repeats, a naive model can achieve an accuracy of about 53 percent. A top-performing model can achieve accuracy on this same test harness of about 88 percent. This provides the bounds of expected performance on this dataset.

    The dataset involves predicting whether sonar returns indicate a rock or simulated mine.

    No need to download the dataset; we will download it automatically as part of our worked examples.

    The example below downloads the dataset and summarizes its shape.


    Running the example downloads the dataset and splits it into input and output elements. As expected, we can see that there are 208 rows of data with 60 input variables.


    Next, let’s use HyperOpt-Sklearn to find a good model for the sonar dataset.

    We can perform some basic data preparation, including converting the target string to class labels, then split the dataset into train and test sets.


    Next, we can define the search procedure. We will explore all classification algorithms and all data transforms available to the library and use the TPE, or Tree of Parzen Estimators, search algorithm, described in “Algorithms for Hyper-Parameter Optimization.”

    The search will evaluate 50 pipelines and limit each evaluation to 30 seconds.


    We then start the search.


    At the end of the run, we will report the performance of the model on the holdout dataset and summarize the best performing pipeline.


    Tying this together, the complete example is listed below.


    Running the example may take a few minutes.

    The progress of the search will be reported and you will see some warnings that you can safely ignore.

    At the end of the run, the best-performing model is evaluated on the holdout dataset and the Pipeline discovered is printed for later use.

    Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

    In this case, we can see that the chosen model achieved an accuracy of about 85.5 percent on the holdout test set. The Pipeline involves a gradient boosting model with no pre-processing.


    The printed model can then be used directly, e.g. the code copy-pasted into another project.

    Next, let’s take a look at using HyperOpt-Sklearn for a regression predictive modeling problem.

    HyperOpt-Sklearn for Regression

    In this section, we will use HyperOpt-Sklearn to discover a model for the housing dataset.

    The housing dataset is a standard machine learning dataset comprised of 506 rows of data with 13 numerical input variables and a numerical target variable.

    Using a test harness of repeated stratified 10-fold cross-validation with three repeats, a naive model can achieve a mean absolute error (MAE) of about 6.6. A top-performing model can achieve a MAE on this same test harness of about 1.9. This provides the bounds of expected performance on this dataset.

    The dataset involves predicting the house price given details of the house suburb in the American city of Boston.

    No need to download the dataset; we will download it automatically as part of our worked examples.

    The example below downloads the dataset and summarizes its shape.


    Running the example downloads the dataset and splits it into input and output elements. As expected, we can see that there are 63 rows of data with one input variable.


    Next, we can use HyperOpt-Sklearn to find a good model for the auto insurance dataset.

    Using HyperOpt-Sklearn for regression is the same as using it for classification, except the “regressor” argument must be specified.

    In this case, we want to optimize the MAE, therefore, we will set the “loss_fn” argument to the mean_absolute_error() function provided by the scikit-learn library.


    Tying this together, the complete example is listed below.


    Running the example may take a few minutes.

    The progress of the search will be reported and you will see some warnings that you can safely ignore.

    At the end of the run, the best performing model is evaluated on the holdout dataset and the Pipeline discovered is printed for later use.

    Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

    In this case, we can see that the chosen model achieved a MAE of about 0.883 on the holdout test set, which appears skillful. The Pipeline involves an XGBRegressor model with no pre-processing.

    Note: for the search to use XGBoost, you must have the XGBoost library installed.


    Further Reading

    This section provides more resources on the topic if you are looking to go deeper.

    Summary

    In this tutorial, you discovered how to use HyperOpt for automatic machine learning with Scikit-Learn in Python.

    Specifically, you learned:

    • Hyperopt-Sklearn is an open-source library for AutoML with scikit-learn data preparation and machine learning models.
    • How to use Hyperopt-Sklearn to automatically discover top-performing models for classification tasks.
    • How to use Hyperopt-Sklearn to automatically discover top-performing models for regression tasks.

    Do you have any questions?
    Ask your questions in the comments below and I will do my best to answer.

    Discover Fast Machine Learning in Python!

    Master Machine Learning With Python
    Develop Your Own Models in Minutes

    …with just a few lines of scikit-learn code

    Learn how in my new Ebook:
    Machine Learning Mastery With Python

    Covers self-study tutorials and end-to-end projects like:
    Loading data, visualization, modeling, tuning, and much more…

    Finally Bring Machine Learning To

    Your Own Projects

    Skip the Academics. Just Results.

    See What’s Inside

    Covid Abruzzo Basilicata Calabria Campania Emilia Romagna Friuli Venezia Giulia Lazio Liguria Lombardia Marche Molise Piemonte Puglia Sardegna Sicilia Toscana Trentino Alto Adige Umbria Valle d’Aosta Veneto Italia Agrigento Alessandria Ancona Aosta Arezzo Ascoli Piceno Asti Avellino Bari Barletta-Andria-Trani Belluno Benevento Bergamo Biella Bologna Bolzano Brescia Brindisi Cagliari Caltanissetta Campobasso Carbonia-Iglesias Caserta Catania Catanzaro Chieti Como Cosenza Cremona Crotone Cuneo Enna Fermo Ferrara Firenze Foggia Forlì-Cesena Frosinone Genova Gorizia Grosseto Imperia Isernia La Spezia L’Aquila Latina Lecce Lecco Livorno Lodi Lucca Macerata Mantova Massa-Carrara Matera Messina Milano Modena Monza e della Brianza Napoli Novara Nuoro Olbia-Tempio Oristano Padova Palermo Parma Pavia Perugia Pesaro e Urbino Pescara Piacenza Pisa Pistoia Pordenone Potenza Prato Ragusa Ravenna Reggio Calabria Reggio Emilia Rieti Rimini Roma Rovigo Salerno Medio Campidano Sassari Savona Siena Siracusa Sondrio Taranto Teramo Terni Torino Ogliastra Trapani Trento Treviso Trieste Udine Varese Venezia Verbano-Cusio-Ossola Vercelli Verona Vibo Valentia Vicenza Viterbo

    Recent Posts

    Archives