Tuesday, 10 June 2025

Database logging Entity Framework queries to SeriLog

Logging database queries from Entity Framework can be done using Sql Server Profiler or using Profiling tools inside VS 2022. A problem with this approach is that Sql server profiler tends to log very much other events than just the queries to the database and if you use the Profiling tool, they will truncate large queries. Note that database logging the traffic will quickly grow into large logs, of several gigabyte. Database logging can also reveal sensitive data, another challenge. But having a tailored database log can for developers be a helpful tool to both profile and check queries, especially when using an ORM like Entity Framework that can hide the actual queries being sent to the database since that part is abstracted away via IQueryable. Once you have the queries you want to inspect, you can profile them inside SQL Server Management Studio and check if the use of indexes in database are optimal or if more indexes should be added. In case you as a developer want to inspect queries that your application does, maybe activated via an app setting in config, you might be better off with a tailored way of logging the database queries that the application performs. This article will look at Entity Framework 6 for .NET Framework and Db interceptor that logs compacted, one-lined, interpolated queries that Entity Framework runs. Note that also later version of Entity Framework support db interceptors, such as EF Core 8 (for .NET 8). Let's first look at the repo for the code of this article :

https://github.com/toreaurstadboss/BulkOperationsEntityFramework

SerilogCommandInterceptor

The db interceptor looks like this, it will log to a file that uses SeriLog as a logging framework and to format the logging file(s).


using Serilog;
using System;
using System.Configuration;
using System.Data.Common;
using System.Data.Entity.Infrastructure.Interception;

namespace BulkOperationsEntityFramework.Test
{

    /// <summary>
    /// Intercepts Entity Framework database commands and logs them using Serilog.
    /// 
    /// This interceptor captures and logs SQL command text for NonQuery, Reader, and Scalar operations.
    /// Logging is configured via Serilog and can be customized using <c>App.config</c> or <c>Web.config</c> settings.
    /// 
    /// <para>
    /// <b>Configuration via AppSettings:</b>
    /// </para>
    /// <list type="table">
    ///   <item>
    ///     <term>serilog:write-to:File.path</term>
    ///     <description>Path to the log file. Default: <c>databaselogs\log.txt</c></description>
    ///   </item>
    ///   <item>
    ///     <term>serilog:write-to:File.rollingInterval</term>
    ///     <description>Rolling interval for log files (e.g., <c>Day</c>, <c>Hour</c>). Default: <c>Day</c></description>
    ///   </item>
    ///   <item>
    ///     <term>serilog:minimum-level</term>
    ///     <description>Minimum log level (e.g., <c>Information</c>, <c>Warning</c>). Default: <c>Information</c></description>
    ///   </item>
    ///   <item>
    ///     <term>serilog:write-to:File.retainedFileCountLimit</term>
    ///     <description>Number of log files to retain. Default: <c>21</c></description>
    ///   </item>
    /// </list>
    /// 
    /// <para>
    /// <b>Example App.config override:</b>
    /// </para>
    /// <code language="xml">
    /// <appSettings>
    ///   <add key="serilog:write-to:File.path" value="C:\Logs\mydb.log" />
    ///   <add key="serilog:write-to:File.rollingInterval" value="Hour" />
    ///   <add key="serilog:minimum-level" value="Warning" />
    ///   <add key="serilog:write-to:File.retainedFileCountLimit" value="10" />
    /// </appSettings>
    /// </code>
    /// </summary>
    public class SerilogCommandInterceptor : IDbCommandInterceptor
    {

        private static bool _isInitialized = false;

        public SerilogCommandInterceptor()
        {
            if (_isInitialized)
            {
                return;
            }

            var logPath = ConfigurationManager.AppSettings["serilog:write-to:File.path"] ?? "databaselogs\\log.txt";
            var logIntervalRaw = ConfigurationManager.AppSettings["serilog:write-to:File.rollingInterval"];
            var logInterval = Enum.TryParse(logIntervalRaw, true, out RollingInterval interval) ? interval : RollingInterval.Day;
            var minLevelRaw = ConfigurationManager.AppSettings["serilog:minimum-level"];
            var logLevel = Enum.TryParse(minLevelRaw, true, out Serilog.Events.LogEventLevel level) ? level : Serilog.Events.LogEventLevel.Information;
            var retainedCountRaw = ConfigurationManager.AppSettings["serilog:write-to:File.retainedFileCountLimit"];
            var retainedCount = int.TryParse(retainedCountRaw, out int count) ? count : 21;

            //Set up Serilog logging for the database logging interceptor - set up the minimum level to Information and
            //write to a file with rolling intervals. Set up the file size limit to 500 MB per file
            //in case the log file grows too large on a given day, it will roll over to a new file with the with a running number suffixed to it 
            //the logs will be stored in the "databaselogs" subfolder, or configured path in config file
            //logs will be kept or a maximum number of 21 days or specified number of days
            Log.Logger = new LoggerConfiguration()
                .MinimumLevel.Is(logLevel)
                .WriteTo.File(
                    logPath, 
                    rollingInterval: logInterval, 
                    rollOnFileSizeLimit: true,
                    fileSizeLimitBytes: 500 * 1000 * 1000,
                    outputTemplate: "[{Timestamp:yyyy-MM-dd HH:mm:ss} {Level:u3}] [SQL] {Message:lj}{NewLine}",
                    retainedFileCountLimit: retainedCount
                )
                .CreateLogger();

            _isInitialized = true;
        }

        public void NonQueryExecuted(DbCommand command, DbCommandInterceptionContext<int> interceptionContext) { }

        public void NonQueryExecuting(DbCommand command, DbCommandInterceptionContext<int> interceptionContext) =>
            Log.Information("{Tag} {Sql}", GetSqlTag(command.CommandText), CompactAndInterpolateSql(command));

        public void ReaderExecuted(DbCommand command, DbCommandInterceptionContext<DbDataReader> interceptionContext) { }

        public void ReaderExecuting(DbCommand command, DbCommandInterceptionContext<DbDataReader> interceptionContext) =>
            Log.Information("{Tag} {Sql}", GetSqlTag(command.CommandText), CompactAndInterpolateSql(command));

        public void ScalarExecuted(DbCommand command, DbCommandInterceptionContext<object> interceptionContext) { }

        public void ScalarExecuting(DbCommand command, DbCommandInterceptionContext<object> interceptionContext) =>
            Log.Information("{Tag} {Sql}", GetSqlTag(command.CommandText), CompactAndInterpolateSql(command));

        private string CompactAndInterpolateSql(DbCommand dbCommand)
        {
            string sql = InterpolateSql(dbCommand);
            return CompactSql(sql);
        }

        private string CompactSql(string sql) =>
            sql.Replace(Environment.NewLine, " ").Replace("\n", "").Replace("\r", " ").Trim();

        private string GetSqlTag(string sql)
        {
            if (sql.StartsWith("SELECT", StringComparison.OrdinalIgnoreCase)) return "[SELECT]";
            if (sql.StartsWith("INSERT", StringComparison.OrdinalIgnoreCase)) return "[INSERT]";
            if (sql.StartsWith("UPDATE", StringComparison.OrdinalIgnoreCase)) return "[UPDATE]";
            if (sql.StartsWith("DELETE", StringComparison.OrdinalIgnoreCase)) return "[DELETE]";
            return "[SQL]";
        }

        private string InterpolateSql(DbCommand dbCommand)
        {
            string sql = dbCommand.CommandText;
            foreach (DbParameter parameter in dbCommand.Parameters)
            {
                string value = FormatParameterValue(parameter.Value);
                sql = sql.Replace(parameter.ParameterName, value);
            }
            return sql;
        }

        private string FormatParameterValue(object value)
        { 
            if (value == null || value == DBNull.Value)
            {
                return "NULL";
            }
            if (value is string || value is DateTime || value is Guid)
            {
                return $"'{value}'"; // Wrap strings, DateTime, and Guid in single quotes
            }
            if (value is bool b)
            {
                return b ? "1" : "0";
            }

            return value.ToString();
        }        

    }
}


Adding a db interceptor should be preferably done once, for example in a static constructor, to ensure the db interceptor is available right away. Example of a simple db context that adds the interceptor shown above.


using BulkOperationsEntityFramework.Models;
using BulkOperationsEntityFramework.Test;
using System.ComponentModel.DataAnnotations.Schema;
using System.Data.Entity;
using System.Data.Entity.Infrastructure.Interception;

namespace BulkOperationsEntityFramework
{

    public class ApplicationDbContext : DbContext
    {

        static ApplicationDbContext()
        {
            DbInterception.Add(new SerilogCommandInterceptor());
        }

        public ApplicationDbContext() : base("name=App")
        {
        }

        public DbSet<User> Users { get; set; }

        protected override void OnModelCreating(DbModelBuilder modelBuilder)
        {
            modelBuilder.Entity<User>().HasKey(u => u.Id);
            modelBuilder.Entity<User>().Property(u => u.Id).HasDatabaseGeneratedOption(DatabaseGeneratedOption.Identity);
        }

    }

}




Since Serilog is used, a great flexibility to the logging is possible to do. For example, logging database queries can be done with rolling intervals, such as logging per hour or per day or even per minute for example. Also, a max size for logs can be set. The logging within the same time interval for the logging is split into multiple log files, when log files reaches size limit. Example app config settings are here:


<appSettings>
	<add key="serilog:write-to:File.path" value="logs\dblogv3.txt" />
	<add key="serilog:minimum-level" value="Information" />
	<add key="serilog:write-to:File.rollingInterval" value="Day" />
	<add key="serilog:file:retentionDays" value="21" />
</appSettings>


If you use a tool like TailBlazer, opening very large logs of many gigabytes is fairly fast and you can define highlighting rules to discern between different types of queries.
A screenshot of the highlighting rules of Tail Blazor for the previous screen shot shown is shown below:



Tail Blazer tool

The Tail Blazer tool can be downloaded from GitHub here:

https://github.com/RolandPheasant/TailBlazer

Sunday, 1 June 2025

Using SqlBulkCopy with EntityFramework

This article tests out variying methods of ways of doing bulk inserts in EntityFramework. The code shown in this article uses .NET Framework 4.8
and Entity Framework 6.5.0. Support is better in Entity Framework Core, so the code shown here for SqlBulkCopy must be adjusted a little bit to also work with EntityFramework Core. A Github repo has been added here:

https://github.com/toreaurstadboss/BulkOperationsEntityFramework

A benchmark has been run to test out the different approaches. Not suprisingly, SqlBulkCopy is the fastest and most economic way of performing the bulk inserts (uses the least memory). The code for handling SqlBulkCopy is provided via extension methods listed below:

BulkInsertExtensions.cs




using System.Collections.Generic;
using System.Data;
using System.Data.Entity;
using System.Data.SqlClient;
using System.Threading.Tasks;

namespace BulkOperationsEntityFramework.Lib.Extensions
{

    /// <summary>
    /// Provides extension methods for performing bulk insert operations using SqlBulkCopy.
    /// </summary>
    public static class BulkInsertExtensions
    {
        /// <summary>
        /// Performs a bulk insert of the specified entities into the database.
        /// </summary>
        /// <typeparam name="T">The type of the entity.</typeparam>
        /// <param name="context">The DbContext instance.</param>
        /// <param name="entities">The collection of entities to insert.</param>
        /// <param name="tableName">
        /// Optional: The name of the destination table. If null, the entity type name is used.
        /// </param>
        /// <param name="columnMappings">
        /// Optional: Custom column mappings (propertyName → columnName).
        /// 
        /// <example>
        /// Example 1: Rename columns
        /// <code>
        /// var mappings = new Dictionary<string, string>
        /// {
        ///     { "Name", "ProductName" },
        ///     { "Price", "UnitPrice" }
        /// };
        /// await context.BulkInsertAsync(products, "Products", mappings);
        /// </code>
        /// </example>
        /// 
        /// <example>
        /// Example 2: Ignore properties using attributes
        /// <code>
        /// public class Customer
        /// {
        ///     public int Id { get; set; }
        ///     public string FullName { get; set; }
        /// 
        ///     [BulkIgnore]
        ///     public string TempNotes { get; set; }
        /// 
        ///     [NotMapped]
        ///     public string CalculatedField { get; set; }
        /// }
        /// 
        /// var mappings = new Dictionary<string, string>
        /// {
        ///     { "FullName", "CustomerName" }
        /// };
        /// await context.BulkInsertAsync(customers, "Customers", mappings);
        /// </code>
        /// </example>
        /// 
        /// <example>
        /// Example 3: Snake_case mapping
        /// <code>
        /// var mappings = new Dictionary<string, string>
        /// {
        ///     { "FirstName", "first_name" },
        ///     { "LastName", "last_name" },
        ///     { "Email", "email_address" }
        /// };
        /// await context.BulkInsertAsync(users, "user_accounts", mappings);
        /// </code>
        /// </example>
        /// </param>
        /// <param name="bulkCopyTimout">The bulk copy timeout in seconds. Default is set to 30.</param>
        public static void BulkInsert<T>(this DbContext context, IEnumerable<T> entities, string tableName = null,
            Dictionary<string, string> columnMappings = null, int bulkCopyTimeout = 30)
            where T : class
        {
            var dataTable = entities.ToBulkDataTable(columnMappings, out var finalMappings);
            BulkCopy(context, dataTable, tableName ?? typeof(T).Name, finalMappings, bulkCopyTimeout);
        }

        /// <summary>
        /// Asynchronously performs a bulk insert of the specified entities into the database.
        /// </summary>
        /// <typeparam name="T">The type of the entity.</typeparam>
        /// <param name="context">The DbContext instance.</param>
        /// <param name="entities">The collection of entities to insert.</param>
        /// <param name="tableName">
        /// Optional: The name of the destination table. If null, the entity type name is used.
        /// </param>
        /// <param name="columnMappings">
        /// Optional: Custom column mappings (propertyName → columnName).
        /// 
        /// <example>
        /// Example 1: Rename columns
        /// <code>
        /// var mappings = new Dictionary<string, string>
        /// {
        ///     { "Name", "ProductName" },
        ///     { "Price", "UnitPrice" }
        /// };
        /// await context.BulkInsertAsync(products, "Products", mappings);
        /// </code>
        /// </example>
        /// 
        /// <example>
        /// Example 2: Ignore properties using attributes
        /// <code>
        /// public class Customer
        /// {
        ///     public int Id { get; set; }
        ///     public string FullName { get; set; }
        /// 
        ///     [BulkIgnore]
        ///     public string TempNotes { get; set; }
        /// 
        ///     [NotMapped]
        ///     public string CalculatedField { get; set; }
        /// }
        /// 
        /// var mappings = new Dictionary<string, string>
        /// {
        ///     { "FullName", "CustomerName" }
        /// };
        /// await context.BulkInsertAsync(customers, "Customers", mappings);
        /// </code>
        /// </example>
        /// 
        /// <example>
        /// Example 3: Snake_case mapping
        /// <code>
        /// var mappings = new Dictionary<string, string>
        /// {
        ///     { "FirstName", "first_name" },
        ///     { "LastName", "last_name" },
        ///     { "Email", "email_address" }
        /// };
        /// await context.BulkInsertAsync(users, "user_accounts", mappings);
        /// </code>
        /// </example>
        /// </param>
        /// <param name="bulkCopyTimout">The bulk copy timeout in seconds. Default is set to 30.</param>
        public static async Task BulkInsertAsync<T>(this DbContext context, IEnumerable<T> entities, string tableName = null,
            Dictionary<string, string> columnMappings = null, int bulkCopyTimout = 30)
            where T : class
        {
            var dataTable = entities.ToBulkDataTable(columnMappings, out var finalMappings);
            await BulkCopyAsync(context, dataTable, tableName ?? typeof(T).Name, finalMappings, 30);
        }

        private static void BulkCopy(DbContext context, DataTable table, string tableName,
            Dictionary<string, string> finalMappings, int bulkCopyTimeout = 30)
        {
            var connection = (SqlConnection)context.Database.Connection;
            var wasClosed = connection.State == ConnectionState.Closed;

            if (wasClosed)
                connection.Open();

            using (var bulkCopy = new SqlBulkCopy(connection))
            {
                bulkCopy.DestinationTableName = tableName;
                bulkCopy.BulkCopyTimeout = bulkCopyTimeout;

                foreach (var map in finalMappings)
                {
                    bulkCopy.ColumnMappings.Add(map.Key, map.Value);
                }
                bulkCopy.WriteToServer(table);
            }

            if (wasClosed)
                connection.Close(); // Ensure the connection is closed after the operation, if it was closed before
        }

        private static async Task BulkCopyAsync(DbContext context, DataTable table, string tableName, Dictionary<string, string> mappings,
            int bulkCopyTimeout = 30)
        {
            var connection = (SqlConnection)context.Database.Connection;
            var wasClosed = connection.State == ConnectionState.Closed;

            if (wasClosed)
                await connection.OpenAsync();

            using (var bulkCopy = new SqlBulkCopy(connection))
            {
                bulkCopy.DestinationTableName = tableName;
                bulkCopy.BulkCopyTimeout = bulkCopyTimeout;
                foreach (var map in mappings)
                {
                    bulkCopy.ColumnMappings.Add(map.Key, map.Value);
                }
                await bulkCopy.WriteToServerAsync(table);
            }

            if (wasClosed)
                connection.Close();  //Ensure the connection is closed after the operation, if it was closed before
        }
    }
}





As the code shows, we use the DbContext from EntityFramework and get the connection string from the Database.Connection.ConnectionString property. We also use the Database.Connection property and cast it to SqlConnection to get the connection object. As we see, SqlBulkCopy is part of ADO.Net and relies upon column mappings. The code above will ignore the properties marked by the BulkIgnore attribute and the NotMapped attribute

BulkIgnoreAttribute.cs



using System;

namespace BulkOperationsEntityFramework.Lib.Attributes
{

    /// 
    /// Indicates that a property should be ignored during bulk insert operations.
    /// 
    [AttributeUsage(AttributeTargets.Property)]
    public class BulkIgnoreAttribute : Attribute
    {
    }

}


The following helper class creates a DataTable for a collection of entities (an IEnumerable) and uses reflection to make such a mapping. Note that column mappings can also be provided in the BulkInsertExtensions methods shown above to override property to column mapping, if desired. Default sql bulk copy timeout is set to 30 seconds (it is also the default timeout)

DataTableExtensions.cs


The following code provides code for creating data table for a collection of objects (IEnumerable). The BulkInsertExtensions uses the specific method ToBulkDataTable. This data table skips properties that are marked with [NotMapped] or [BulkIgnore] attributes. Both ToDataTable methods shown here allows column mappings to customize the property to column mappings.

using BulkOperationsEntityFramework.Lib.Attributes;
using System;
using System.Collections.Generic;
using System.Data;
using System.Linq;
using System.Reflection;

namespace BulkOperationsEntityFramework.Lib.Extensions
{

    public static class DatatableExtensions
    {

        /// <summary>
        /// Converts an IEnumerable of type T to a DataTable.
        /// </summary>
        /// <typeparam name="T"></typeparam>
        /// <param name="data"></param>
        /// <param name="columnMappings"></param>
        /// <returns></returns>
        public static DataTable ToDataTable<T>(this IEnumerable<T> data, Dictionary<string, string> columnMappings = null)
        {
            var dataTable = new DataTable();
            var properties = typeof(T).GetProperties(BindingFlags.Public | BindingFlags.Instance);

            foreach (var prop in properties)
            {
                var columnName = columnMappings != null && columnMappings.ContainsKey(prop.Name) ? columnMappings[prop.Name] : prop.Name;
                dataTable.Columns.Add(columnName, Nullable.GetUnderlyingType(prop.PropertyType) ?? prop.PropertyType);
            }

            foreach (var item in data)
            {
                var values = properties.Select(p => p.GetValue(item) ?? DBNull.Value).ToArray();
                dataTable.Rows.Add(values);
            }

            return dataTable;
        }

        /// <summary>
        /// Converts an IEnumerable of type T to a DataTable with specified column mappings. Tailored for Bulk operations.
        /// </summary>
        /// <typeparam name="T"></typeparam>
        /// <param name="entities"></param>
        /// <param name="columnMappings"></param>
        /// <param name="finalMappings"></param>
        /// <returns></returns>
        public static DataTable ToBulkDataTable<T>(this IEnumerable<T> entities, Dictionary<string, string> columnMappings, out Dictionary<string, string> finalMappings)
        {
            var dataTable = new DataTable();
            var properties = typeof(T).GetProperties(BindingFlags.Public | BindingFlags.Instance)
                .Where(p =>
                    !Attribute.IsDefined(p, typeof(System.ComponentModel.DataAnnotations.Schema.NotMappedAttribute)) &&
                    !Attribute.IsDefined(p, typeof(BulkIgnoreAttribute)))
                .ToArray();

            finalMappings = new Dictionary<string, string>();

            foreach (var prop in properties)
            {
                var columnName = columnMappings != null && columnMappings.ContainsKey(prop.Name)
                    ? columnMappings[prop.Name]
                    : prop.Name;

                dataTable.Columns.Add(columnName, Nullable.GetUnderlyingType(prop.PropertyType) ?? prop.PropertyType);
                finalMappings[prop.Name] = columnName;
            }

            foreach (var entity in entities)
            {
                var values = properties.Select(p => p.GetValue(entity) ?? DBNull.Value).ToArray();
                dataTable.Rows.Add(values);
            }

            return dataTable;
        }

    }
}




BulkInsertBenchmark

The following benchmark is available in the solution for benchmarking the different approaches to bulk copy with EntityFramework.
  • EF - add one and save in a loop. Not suprisingly, performs the worst due to the many roundtrips to the database
  • EF - add one by one and save at the end. Better performance, since we have one roundtrip. Will handle poor cases where we try to add many items in a batch.
  • EF - addrange and save at the end. Similar to the one above, one roundtrip to database
  • Dapper - add as batch and save. Minor better speed than the previous two. One roundtrip to database.
  • SqlBulkCopy - clearly the fastest way to insert a batch of entities to the database



using BenchmarkDotNet.Attributes;
using Bogus;
using BulkOperationsEntityFramework.Lib.Extensions;
using BulkOperationsEntityFramework.Models;
using Dapper;
using System.Configuration;
using System.Data.SqlClient;
using System.Linq;
using System.Threading.Tasks;

namespace BulkOperationsEntityFramework.Benchmarks
{

    [MemoryDiagnoser]
    public class BulkInsertBenchmark
    {

        private static readonly Faker Faker = new Faker();

        [Params(100)]
        public int Size { get; set; }

        //First benchmark - Naive approach - add one and one entity and save changes everytime with round trip to database

        [Benchmark(Description = "EFAddOneAndSave - Add one user and then save. Then add another one. Results in many round-trips to database.")]
        public async Task EFAddOneAndSave()
        {
            using (var context = new ApplicationDbContext())
            {
                foreach (var user in GetUsers())
                {
                    context.Users.Add(user);
                    await context.SaveChangesAsync();
                }
            }
        }

        [Benchmark(Description = "EFAddOneByOneAndSave - Add one by one user, but save once after adding them. Results in one round-trip to database")]
        public async Task EFAddOneByOneAndSave()
        {
            using (var context = new ApplicationDbContext())
            {
                foreach (var user in GetUsers())
                {
                    context.Users.Add(user);

                }
                await context.SaveChangesAsync();
            }
        }

        [Benchmark(Description = "EFAddRange - Adds the users, then does the save. Results in one round-trip to database")]
        public async Task EFAddRange()
        {
            using (var context = new ApplicationDbContext())
            {
                var users = GetUsers();
                context.Users.AddRange(users);
                await context.SaveChangesAsync();
            }
        }

        [Benchmark(Description = "DapperInsertRange - Uses Dapper to insert the users. Results in one round-trip to database")]
        public async Task DapperInsertRange()
        {
            var connectionString = ConfigurationManager.ConnectionStrings["App"].ConnectionString;

            string sql = @"
                INSERT INTO Users (Email, FirstName, LastName, PhoneNumber)
                VALUES (@Email, @FirstName, @LastName, @PhoneNumber)
            ".Trim();

            using (var connection = new SqlConnection(connectionString))
            {
                var users = GetUsers().Select(u => new 
                {
                    u.Email,
                    u.FirstName,
                    u.LastName,
                    u.PhoneNumber
                }).ToArray();

                await connection.ExecuteAsync(sql, users);
            }
        }

        [Benchmark(Description = "SqlBulkCopy - Uses SqlBulkCopy to insert the users. Results in one round-trip to database")]
        public async Task SqlBulkCopy()
        {
            using (var context = new ApplicationDbContext())
            {
                await context.BulkInsertAsync(GetUsers(), "Users");
            }
        }



        private User[] GetUsers() =>
            Enumerable.Range(1, Size).Select(i => new User
            {
                Email = Faker.Internet.Email(),
                FirstName = Faker.Name.FirstName(),
                LastName = Faker.Name.LastName(),
                PhoneNumber = Faker.Phone.PhoneNumber()
            }).ToArray();

    }

}


The following screenshots shows the benchmark results after running Benchmark.NET for the mentioned approaches and comparing the performance.

Results

Benchmark for a given batch size of 100

Benchmark for a given batch size of 100

Sunday, 18 May 2025

Displaying weather data using Seaborn with Python

Using the library Seaborn, built on top of MatPlotLib, displaying weather data is convenient. Using Anaconda and Jupyter Notebook, working on the data is user friendly. First off, let's look at a data set containing some Norwegian weather data. The following data set, contains weather data from Norway from 2020-2021 for 55 meterological stations. (13,61 MB in size), available on DbCL v1.0 license (meaning, free of use , 'as-is' warranty) on the Kaggle.com website.

https://www.kaggle.com/datasets/annbengardt/noway-meteorological-data/data

First off, the following imports are done in the Jupyter Notebook. This is a free IDE part of the Anaconda distribution that is prepared for data visualization, that runs in a browser.

NorwayMeteoDemo1.pynb


import seaborn as sns
import pandas as pd
import matplotlib.pyplot as plt

Next off, using Pandas, Python's Data Analysis Library, the data is prepared from the mentioned dataset. The format is in CSV format. Also, columns are added dynamically to the dataset. A moving 14-days average of the daily maxmimum air temperature is added using the rolling method and setting window to 14. Also, a date column is added. Our dataset contains three int64 values day, month and year. We combine these to create a date column.

NorwayMeteoDemo1.pynb


df = pd.read_csv("datasets/weather/NorwayMeteoDataCompleted.csv")
df['moving_max_air_temp_avg'] = df['max(air_temperature P1D)'].rolling(window = 14, min_periods = 1).mean()
df['date'] = pd.to_datetime(df[['year', 'month', 'day']])

Note the use of min_periods set to 1 for the moving average. Or else, you will get NaN in the start of the data of your created moving average column and the way Python works, it will cause NaN for all the next periods too ! Next, choosing what data to display. The following data will be shown in the demo.
  • Station id : SN69100 (This is Værnes - Trondheim Airport weather station by the way,
  • Year 2020
The station ids can be looked up here on this user's Github GIST: https://gist.github.com/ofaltins/c1f0158f1766c8bd695b2c8d545c052c The following code set up the filtered data and saves it into a variable

NorwayMeteoDemo1.pynb


filtered_df_2020= df[ 
    (df['sourceId'] == 'SN69100')  &
    (df['year'] == 2020)
]

Next up, the data is then displayed. The demo will show two lineplots, first the maximum air temperature as a lineplot using Seaborn. In addition, another line plot with the moving 14 days average of the daily maximum air temperature. Also, a bar plot is added below these two line plots. Note that the bar plot used here is bar and not barplot. Barplot is available in Seaborn, while Bar is available in the MatPlotLib.

NorwayMeteoDemo1.pynb



sns.set_style('whitegrid') # set the style to 'whitegrid' 

# Create a 2x1 subplot

fig, (ax1, ax2) = plt.subplots(2, 1, figsize=(12, 10), sharex=True)

sns.lineplot(data = filtered_df_2020, x = filtered_df_2020['date'], y = filtered_df_2020['max(air_temperature P1D)'], label='Daily max air temperature (C)', linewidth = 1.5, color = 'pink', ax = ax1)

sns.lineplot(data = filtered_df_2020, x = filtered_df_2020['date'], y = filtered_df_2020['moving_max_air_temp_avg'], label='14-day moving average Daily max air temperature (C)', color = 'red', linewidth = 1.8, ax = ax1)

ax2.bar(filtered_df_2020['date'], filtered_df_2020['sum(precipitation_amount P1D)'], data = filtered_df_2020, color = 'blue', label = 'Daily sum precipitation (mm)')

ax1.set_title('Værnes - Weather data - 2020')

ax1.set_xlabel('Date of year')
ax1.set_ylabel('Daily max air temperature (C)')

ax2.set_xlabel('Date of year')
ax2.set_ylabel('Daily sum precipitation (mm)')



The code above shows the resulting figure consisting of a 2x1 subplots layout, the upper plot shows the 2020 daily maximum air temperature combined with a moving 14-days average as a smoothing or trending function to show the general temperature shifts every second week of the year in average. The lower plot shows a bar plot, using MatPlotLib, since Seaborn's bar plot does not handle dates as x-axis (in our dataset, datetime64 is used). Seaborn and MatPlotLib offers a ton of plotting functionality. Maybe it also could be of interest for .NET Developers to use it more often? My previous article showed how it is possible to render in the backend images using MatPlotLib and then display in a Blazor serverside app, combining both
Python and .NET. That demo used Python.Net library for the Python interop with .NET. The screenshot shows the Jupyter Notebook IDE, part of Anaconda distribution of Python tailored for data analysis and data science. It displays the plots described from the Python script above.

Monday, 5 May 2025

Using MatPlotLib from .NET

MatPlotLib is a powerful library for data visualization. It provides graphing for scientific computing. It can be used for doing both mathematical calculations and statistics. Together with additional libraries like NumPy or Numerical Python, it is clear that Python as a programming language and ecosystem provides a lot of powerful functionality that is also free to use. MatplotLib has a BSD license, which means it can be ued for personal, academic or commercial purposes without restrictions. This article will look at using MatplotLib from .NET. First off an image that displays the demo and example of using MatplotLib.

The source code shown in this article is available on Github here:

https://github.com/toreaurstadboss/SeabornBlazorVisualizer



Using MatPlotLib from .NET

First off, install Anaconda. Anaconda is a Python distribution that contains a large collection of data visualization libraries. A compatible version with the lastest version of Python.net Nuget library. The demo displayed here uses Anaconda version 2023.03.

Anaconda archived versions 2023.03 can be installed from here. Windows users can download the file: https://repo.anaconda.com/archive/Anaconda3-2023.03-1-Windows-x86_64.exe

https://repo.anaconda.com/archive/

Next up, install also Python 3.10 version. It will be used together with Anaconda. A 64-bit installer can be found here:

Python 3.10 installer (Windows 64-bits) The correct versions of NumPy and MatPlotLib can be checked against this list :

https://github.com/toreaurstadboss/SeabornBlazorVisualizer/blob/main/SeabornBlazorVisualizer/conda_list_loading_matplotlib_working_1st_May_2025.txt

Calculating the determinite integral of a function

The demo in this article show in the link at the top has got an appsettings.json file, you can adjust to your environment.

appsettings.json Application configuration file




{
  "Logging": {
    "LogLevel": {
      "Default": "Information",
      "Microsoft.AspNetCore": "Warning"
    }
  },
  "PythonConfig": {
    "PythonDllPath": "C:\\Python310\\Python310.dll",
    "PythonHome": "C:\\Programdata\\anaconda3",
    "PythonSitePackages":  "C:\\Programdata\\anaconda3\\lib\\site-packages",
    "PythonVersion": "3.10"
  },
  "AllowedHosts": "*"
}


Clone the source code and run the application. It is a Blazor server app. You can run it from VS 2022 for example. The following code shows how Python.net is set up to start using Python. Both Python 3.10 and Anaconda site libs are used here. The Python runtime and engine is set up using this helper class.

PythonInitializer.cs



using Microsoft.Extensions.Options;
using Python.Runtime;

namespace SeabornBlazorVisualizer.Data
{

    /// <summary>
    /// Helper class to initialize the Python runtime
    /// </summary>
    public static class PythonInitializer
    {

        private static bool runtime_initialized = false;

        /// <summary>
        /// Perform one-time initialization of Python runtime
        /// </summary>
        /// <param name="pythonConfig"></param>
        public static void InitializePythonRuntime(IOptions<PythonConfig> pythonConfig)
        {
            if (runtime_initialized)
                return;
            var config = pythonConfig.Value;

            // Set environment variables
            Environment.SetEnvironmentVariable("PYTHONHOME", config.PythonHome, EnvironmentVariableTarget.Process);
            Environment.SetEnvironmentVariable("PYTHONPATH", config.PythonSitePackages, EnvironmentVariableTarget.Process);
            Environment.SetEnvironmentVariable("PYTHONNET_PYDLL", config.PythonDllPath);
            Environment.SetEnvironmentVariable("PYTHONNET_PYVER", config.PythonVersion);

            PythonEngine.Initialize();

            PythonEngine.PythonHome = config.PythonHome ?? Environment.GetEnvironmentVariable("PYTHONHOME", EnvironmentVariableTarget.Process)!;
            PythonEngine.PythonPath = config.PythonDllPath ?? Environment.GetEnvironmentVariable("PYTHONNET_PYDLL", EnvironmentVariableTarget.Process)!;

            PythonEngine.BeginAllowThreads();
            AddSitePackagesToPythonPath(pythonConfig);
            runtime_initialized = true;
        }

        private static void AddSitePackagesToPythonPath(IOptions<PythonConfig> pythonConfig)
        {
            if (!runtime_initialized)
            {
                using (Py.GIL())
                {
                    dynamic sys = Py.Import("sys");
                    sys.path.append(pythonConfig.Value.PythonSitePackages);
                    Console.WriteLine(sys.path);

                    //add folders in solution this too with scripts
                    sys.path.append(@"Data/");
                }
            }
        }

    }
}



The following helper class sets up the site libraries we will use.

PythonHelper.cs



using Python.Runtime;

namespace SeabornBlazorVisualizer.Data
{

    /// <summary>
    /// Helper class to initialize the Python runtime
    /// </summary>
    public static class PythonHelper
    {

        /// <summary>
        /// Imports Python modules. Returned are the following modules:
        /// <para>np (numpy)</para>
        /// <para>os (OS module - standard library)</para>
        /// <para>scipy (scipy)</para>
        /// <para>mpl (matplotlib)</para>
        /// <para>plt (matplotlib.pyplot </para>
        /// </summary>
        /// <returns>Tuple of Python modules</returns>
        public static (dynamic np, dynamic os, dynamic scipy, dynamic mpl, dynamic plt) ImportPythonModules()
        {

            dynamic np = Py.Import("numpy");
            dynamic os = Py.Import("os");
            dynamic mpl = Py.Import("matplotlib");
            dynamic plt = Py.Import("matplotlib.pyplot");
            dynamic scipy = Py.Import("scipy");

            mpl.use("Agg");

            return (np, os, scipy, mpl, plt);
        }

    }
}



The demo is a Blazor server app. The following service will generate the plot of a determinite integral using MatPlotLib. The service saves the plot into a PNG file. This PNG file is saved into the folder wwwroot. The Blazor server app displays the image that was generated and saved.

MatPlotImageService.cs



using Microsoft.Extensions.Options;
using Python.Runtime;

namespace SeabornBlazorVisualizer.Data
{
    public class MatplotPlotImageService
    {

        private IOptions<PythonConfig>? _pythonConfig;

        private static readonly object _lock = new object();

        public MatplotPlotImageService(IOptions<PythonConfig> pythonConfig)
        {
            _pythonConfig = pythonConfig;
            PythonInitializer.InitializePythonRuntime(_pythonConfig);
        }

        public Task<string> GenerateDefiniteIntegral(string functionExpression, int lowerBound, int upperBound)
        {

            string? result = null;

            using (Py.GIL()) // Ensure thread safety for Python calls
            {
                dynamic np = Py.Import("numpy");
                dynamic plt = Py.Import("matplotlib.pyplot");

                dynamic patches = Py.Import("matplotlib.patches"); // Import patches module

                // Create a Python execution scope
                using (var scope = Py.CreateScope())
                {
                    // Define the function inside the scope
                    scope.Exec($@"
import numpy as np
def func(x):
    return {functionExpression}
");

                    // Retrieve function reference from scope
                    dynamic func = scope.Get("func");

                    // Define integration limits
                    double a = lowerBound, b = upperBound;

                    // Generate x-values
                    dynamic x = np.linspace(0, 10, 100); //generate evenly spaced values in range [0, 20], 100 values (per 0.1)
                    dynamic y = func.Invoke(x);

                    // Create plot figure
                    var fig = plt.figure();
                    var ax = fig.add_subplot(111);

                    // set title to function expression
                    plt.title(functionExpression);

                    ax.plot(x, y, "r", linewidth: 2);
                    ax.set_ylim(0, null);

                    // Select range for integral shading
                    dynamic ix = np.linspace(a, b, 100);
                    dynamic iy = func.Invoke(ix);

                    // **Fix: Separate x and y coordinates properly**
                    List<double> xCoords = new List<double> { a }; // Start at (a, 0)
                    List<double> yCoords = new List<double> { 0 };

                    int length = (int)np.size(ix);
                    for (int i = 0; i < length; i++)
                    {
                        xCoords.Add((double)ix[i]);
                        yCoords.Add((double)iy[i]);
                    }

                    xCoords.Add(b); // End at (b, 0)
                    yCoords.Add(0);

                    // Convert x and y lists to NumPy arrays
                    dynamic npVerts = np.column_stack(new object[] { np.array(xCoords), np.array(yCoords) });

                    // **Correctly Instantiate Polygon Using NumPy Array**
                    dynamic poly = patches.Polygon(npVerts, facecolor: "0.6", edgecolor: "0.2");
                    ax.add_patch(poly);

                    // Compute integral area
                    double area = np.trapezoid(iy, ix);
                    ax.text(0.5 * (a + b), 30, "$\\int_a^b f(x)\\mathrm{d}x$", ha: "center", fontsize: 20);
                    ax.text(0.5 * (a + b), 10, $"Area = {area:F2}", ha: "center", fontsize: 12);

                    plt.show();


                    result = SavePlot(plt, dpi: 150);
                }
            }
            return Task.FromResult(result);
        }

        public Task<string> GenerateHistogram(List<double> values, string title = "Provide Plot title", string xlabel = "Provide xlabel title", string ylabel = "Provide ylabel title")
        {
            string? result = null;
            using (Py.GIL()) //Python Global Interpreter Lock (GIL)
            {
                var (np, os, scipy, mpl, plt) = PythonHelper.ImportPythonModules();

                var distribution = np.array(values.ToArray());

                //// Ensure clearing the plot
                //plt.clf();

                var fig = plt.figure(); //create a new figure
                var ax1 = fig.add_subplot(1, 2, 1);
                var ax2 = fig.add_subplot(1, 2, 2);

                // Add style
                plt.style.use("ggplot");

                var counts_bins_patches = ax1.hist(distribution, edgecolor: "black");

                // Normalize counts to get colors 
                var counts = counts_bins_patches[0];
                var patches = counts_bins_patches[2];

                var norm_counts = counts / np.max(counts);

                int norm_counts_size = Convert.ToInt32(norm_counts.size.ToString());

                // Apply colors to patches based on frequency
                for (int i = 0; i < norm_counts_size; i++)
                {
                    plt.setp(patches[i], "facecolor", plt.cm.viridis(norm_counts[i])); //plt.cm is the colormap module in MatPlotlib. viridis creates color maps from normalized value 0 to 1 that is optimized for color-blind people.
                }

                // **** AX1 Histogram first - frequency counts ***** 

                ax1.set_title(title);
                ax1.set_xlabel(xlabel);
                ax1.set_ylabel(ylabel);

                string cwd = os.getcwd();

                // Calculate average and standard deviation
                var average = np.mean(distribution);
                var std_dev = np.std(distribution);
                var total_count = np.size(distribution);

                // Format average and standard deviation to two decimal places
                var average_formatted = np.round(average, 2);
                var std_dev_formatted = np.round(std_dev, 2);

                //Add legend with average and standard deviation
                ax1.legend(new string[] { $"Total count: {total_count}\n Average: {average_formatted} cm\nStd Dev: {std_dev_formatted} cm" }, framealpha: 0.5, fancybox: true);



                //***** AX2 : Set up ax2 = Percentage histogram next *******

                ax2.set_title("Percentage distribution");
                ax2.set_xlabel(xlabel);
                ax2.set_ylabel(ylabel);
                // Fix for CS1977: Cast the lambda expression to a delegate type
                ax2.yaxis.set_major_formatter((PyObject)plt.FuncFormatter(new Func<double, int, string>((y, _) => $"{y:P0}")));

                ax2.hist(distribution, edgecolor: "black", weights: np.ones(distribution.size) / distribution.size);

                // Format y-axis to show percentages
                ax2.yaxis.set_major_formatter(plt.FuncFormatter(new Func<double, int, string>((y, _) => $"{y:P0}")));

                // tight layout to prevent overlap 
                plt.tight_layout();

                // Show the plot with the two subplots at last (render to back buffer 'Agg', see method SavePlot for details)
                plt.show();

                result = SavePlot(plt, theme: "bmh", dpi: 150);
            }

            return Task.FromResult(result);
        }

        public Task<string> GeneratedCumulativeGraphFromValues(List<double> values)
        {
            string? result = null;
            using (Py.GIL()) //Python Global Interpreter Lock (GIL)
            {
                var (np, os, scipy, mpl, plt) = PythonHelper.ImportPythonModules();

                dynamic pythonValues = np.cumsum(np.array(values.ToArray()));

                // Ensure clearing the plot
                plt.clf();

                // Create a figure with increased size
                dynamic fig = plt.figure(figsize: new PyTuple(new PyObject[] { new PyFloat(6), new PyFloat(4) }));

                // Plot data
                plt.plot(values, color: "green");

                string cwd = os.getcwd();

                result = SavePlot(plt, theme: "ggplot", dpi: 200);

            }

            return Task.FromResult(result);
        }

        public Task<string> GenerateRandomizedCumulativeGraph()
        {
            string? result = null;
            using (Py.GIL()) //Python Global Interpreter Lock (GIL)
            {

                dynamic np = Py.Import("numpy");

                //TODO : Remove imports of pandas and scipy and datetime if they are not needed

                Py.Import("pandas");
                Py.Import("scipy");
                Py.Import("datetime");
                dynamic os = Py.Import("os");

                dynamic mpl = Py.Import("matplotlib");
                dynamic plt = Py.Import("matplotlib.pyplot");

                // Set dark theme
                plt.style.use("ggplot");

                mpl.use("Agg");


                // Generate data
                //dynamic x = np.arange(0, 10, 0.1);
                //dynamic y = np.multiply(2, x); // Use NumPy's multiply function

                dynamic values = np.cumsum(np.random.randn(1000, 1));


                // Ensure clearing the plot
                plt.clf();

                // Create a figure with increased size
                dynamic fig = plt.figure(figsize: new PyTuple(new PyObject[] { new PyFloat(6), new PyFloat(4) }));

                // Plot data
                plt.plot(values, color: "blue");

                string cwd = os.getcwd();

                result = SavePlot(plt, theme: "ggplot", dpi: 200);

            }

            return Task.FromResult(result);
        }

        /// <summary>
        /// Saves the plot to a PNG file with a unique name based on the current date and time
        /// </summary>
        /// <param name="plot">Plot, must be a PyPlot plot use Python.net Py.Import("matplotlib.pyplot")</param>
        /// <param name="theme"></param>
        /// <param name="dpi"></param>
        /// <returns></returns>
        public string? SavePlot(dynamic plt, string theme = "ggplot", int dpi = 200)
        {
            string? plotSavedImagePath = null;
            //using (Py.GIL()) //Python Global Interpreter Lock (GIL)
            //{
            dynamic os = Py.Import("os");
            dynamic mpl = Py.Import("matplotlib");
            // Set dark theme
            plt.style.use(theme);
            mpl.use("Agg"); //set up rendering of plot to back-buffer ('headless' mode)

            string cwd = os.getcwd();
            // Save plot to PNG file
            string imageToCreatePath = $@"GeneratedImages\{DateTime.Now.ToString("yyyyMMddHHmmss")}{Guid.NewGuid().ToString("N")}_plotimg.png";
            string imageToCreateWithFolderPath = $@"{cwd}\wwwroot\{imageToCreatePath}";
            plt.savefig(imageToCreateWithFolderPath, dpi: dpi); //save the plot to a file (use full path)
            plotSavedImagePath = imageToCreatePath;

            CleanupOldGeneratedImages(cwd);
            //}
            return plotSavedImagePath;
        }

        private static void CleanupOldGeneratedImages(string cwd)
        {
            lock (_lock)
            {

                Directory.GetFiles(cwd + @"\wwwroot\GeneratedImages", "*.png")
                 .OrderByDescending(File.GetLastWriteTime)
                 .Skip(10)
                 .ToList()
                 .ForEach(File.Delete);
            }
        }

}



The code above shows some additional examples of using MatPlotLib.
  • Histogram example
  • Line graph using cumulative sum by making use of NumPy or a helper method in .NET
These examples demonstrates also that MatPlotLib can be used for statistics, which today for .NET is mostly crunched with the help of Excel or EP Plus library for example. Since Python is considered as the home of data visualization with its vast ecosystem of data science libraries, this article and demos shows how you can get started with using this ecosystem from .NET. Note, using Python.net to create these plots in MatPlotLib is best prepared using Jupyter Notebook. When the plot displayed looks okay, it is time to integrate that Python script into .NET and C# using Python.Net library. Make note that there will be some challenges to get the Python code to work in C# of course. When passing in values to a function, sometimes you must use
for example NumPy to create compatible data types. Also note the usage of the Pystatic class here from Python.net , which offers the GIL Global Interpreter lock and a way to import Python modules.

https://jupyter.org/ A screenshot showing histogram in the demo is shown below. As we can see, MatPlotLib can be used from many different data visualizations and domains.

Tuesday, 22 April 2025

Predicting variable using Regression with ML.net

This article will look at regression with ML.net In the example, the variable "Poverty rate" measured as a percentage against amount "teenage pregnancies" per 1,000 birth. The data is fetched from a publicly available CSV file. The data is obtained from Jeff Prosise repos of ML.net here on Github:

https://github.com/jeffprosise/ML.NET/blob/master/MLN-SimpleRegression/MLN-SimpleRegression/Data/poverty.csv



In this article, Linqpad 8 will be used. First off, the following two Nuget packages are added :
  • Microsoft.ML
  • Microsoft.ML.Mkl.Components
The following method will plot a scatter graph from provided MLContext data, and add a standard linear trendline, which will work with the example in this article.

Plotutils.cs



void PlotScatterGraph<T>(MLContext mlContext, IDataView trainData, Func<T, PointItem> pointCreator, string chartTitle) where T : class, new()
{
	//Convert the IDataview to an enumerable collection
	var data = mlContext.Data.CreateEnumerable<T>(trainData, reuseRowObject: false).Select(x => pointCreator(x)).ToList();

	// Calculate trendline (simple linear regression)
	double avgX = data.Average(d => d.X);
	double avgY = data.Average(d => d.Y);
	double slope = data.Sum(d => (d.X - avgX) * (d.Y - avgY)) / data.Sum(d => (d.X - avgX) * (d.X - avgX));
	double intercept = avgY - slope * avgX;
	var trendline = data.Select(d => new { X = d.X, Y = slope * d.X + intercept }).ToList();

	//Plot the scatter graph
	var plot = data.Chart(d => d.X)
		.AddYSeries(d => d.Y, LINQPad.Util.SeriesType.Point, chartTitle)
		.AddYSeries(d => trendline.FirstOrDefault(t => t.X == d.X)?.Y ?? 0, Util.SeriesType.Line, "Trendline")
		.ToWindowsChart();
		
	plot.AntiAliasing = System.Windows.Forms.DataVisualization.Charting.AntiAliasingStyles.All;
	plot.Dump();
}



Let's look at the code for loading the CSV data and into the MLContext and then used the method TrainTestSplit to split the data into training data and testing data. Note also the classes Input and Output and the usage of LoadColumn and ColumnName

Program.cs



void Main()
{

	string inputFile = Path.Combine(Path.GetDirectoryName(Util.CurrentQueryPath)!, @"Sampledata\poverty2.csv"); //linqpad tech

	var context = new MLContext(seed: 0);

	//Train the model 
	var data = context.Data
		.LoadFromTextFile<Input>(inputFile, hasHeader: true, separatorChar: ';');
	
	// Split data into training and test sets 
	var split = context.Data.TrainTestSplit(data, testFraction: 0.2);
	var trainData = split.TrainSet;
	var testData = split.TestSet;

	var pipeline = context
		.Transforms.NormalizeMinMax("PovertyRate")
		.Append(context.Transforms.Concatenate("Features", "PovertyRate"))
		.Append(context.Regression.Trainers.Ols());

	var model = pipeline.Fit(trainData);
	// Use the model to make a prediction
	var predictor = context.Model.CreatePredictionEngine<Input, Output>(model);
	var input = new Input { PovertyRate = 8.4f };

	var actual = 36.8f;

	var prediction = predictor.Predict(input);
	Console.WriteLine($"Input poverty rate: {input.PovertyRate} . Predicted birth rate per 1000: {prediction.TeenageBirthRate:0.##}");
	Console.WriteLine($"Actual birth rate per 1000: {actual}");

	// Evaluate the regression model 
	var predictions = model.Transform(testData);
	var metrics = context.Regression.Evaluate(predictions);
	Console.WriteLine($"R-squared: {metrics.RSquared:0.##}");
	Console.WriteLine($"Root Mean Squared Error: {metrics.RootMeanSquaredError:0.##}");
	Console.WriteLine($"Mean Absolute Error: {metrics.MeanAbsoluteError:0.##}");
	Console.WriteLine($"Mean Squared Error: {metrics.MeanSquaredError:0.##}");


	PlotScatterGraph<Input>(context, trainData, (Input input) => 
		new PointItem { X = (float) Math.Round(input.PovertyRate, 2), Y = (float) Math.Round(input.TeenageBirthRate, 2) },
		"Poverty rate (%) vs Teenage Pregnancies per 1,000 birth");

}

public class PointItem {
	public float X { get; set; }
	public float Y { get; set; }
}

void PlotScatterGraph<T>(MLContext mlContext, IDataView trainData, Func<T, PointItem> pointCreator, string chartTitle) where T : class, new()
{
	//Convert the IDataview to an enumerable collection
	var data = mlContext.Data.CreateEnumerable<T>(trainData, reuseRowObject: false).Select(x => pointCreator(x)).ToList();

	// Calculate trendline (simple linear regression)
	double avgX = data.Average(d => d.X);
	double avgY = data.Average(d => d.Y);
	double slope = data.Sum(d => (d.X - avgX) * (d.Y - avgY)) / data.Sum(d => (d.X - avgX) * (d.X - avgX));
	double intercept = avgY - slope * avgX;
	var trendline = data.Select(d => new { X = d.X, Y = slope * d.X + intercept }).ToList();

	//Plot the scatter graph
	var plot = data.Chart(d => d.X)
		.AddYSeries(d => d.Y, LINQPad.Util.SeriesType.Point, chartTitle)
		.AddYSeries(d => trendline.FirstOrDefault(t => t.X == d.X)?.Y ?? 0, Util.SeriesType.Line, "Trendline")
		.ToWindowsChart();
		
	plot.AntiAliasing = System.Windows.Forms.DataVisualization.Charting.AntiAliasingStyles.All;
	plot.Dump();
}



public class Input
{

	[LoadColumn(1)]
	public float PovertyRate;

	[LoadColumn(5), ColumnName("Label")]
	public float TeenageBirthRate { get; set; }

}
public class Output
{
	[ColumnName("Score")]
	public float TeenageBirthRate;

}


A pipeline is defined for the machine learning here consisting of the following :
  • The method NormalizeMinMax will transform the poverty rate into a normalized scale between 0 and 1. The Concatenate method will be used to specify the "Features", in this case only the column Poverty rate is the feature of which we want to predict a score, this is the rate of teenage pregnancy births per 1,000 births. Note that our CSV data set contains more columns, but this is a simple regression where only one variable is taken into account.
  • The trainers used to train the machine learning algorithm is Ols, the Ordinary Least Squares.
  • The method fit will train using the training data defined from the method TrainTestSplit.
  • The resulting model is used to create a prediction engine.
  • Using the prediction engine, it is possible to predict a value value using the Predict method given one input item. Our prediction engine expects input objects of type Input and Output.
  • Using the testdata, the method Transform using the model gives us multiple predictions and it is possible to evalute the regression analysis from the predictions to check how accurate the regression model is.
  • Returning from this evaluation, we get the R-squared for example. This is a value from 0 to 1.0 where it describes how accurate the regression is in in describing the total variation of the residues of the model, the amount the data when plotted in a scatter graph where residue is the offset between the actual data and what the regression model predicts.
  • Other values such as RMSE and MSE are the root and mean squared error, which are absolute values.
  • Using the code above we got a fairly accurate regression model, but more accuracy would be achieved by taking in additional factors.


  • Output from the Linqpad 8 application shown in this article :
    
        
    Input poverty rate: 8,4 . Predicted birth rate per 1000: 35,06
    Actual birth rate per 1000: 36,8
    R-squared: 0,59
    Root Mean Squared Error: 8,99
    Mean Absolute Error: 8,01
    Mean Squared Error: 80,83
        
      
    
    Please note that there are some standard column names used for machine learning.
    
    Label: Represents the target variable (the value to predict).
    
    Features: Contains the input features used for training.
    
    Score: Stores the predicted value in regression models.
    
    PredictedLabel: Holds the predicted class in classification models.
    
    Probability: Represents the probability of a predicted class.
    
    FeatureContributions: Shows how much each feature contributes to a prediction.
    
    
    In the code above, the column names "Label", "Features" and "Score" was used to instruct the regression being calculated in the code here for ML.Net context model. The attribute ColumnName was being used here together with the Concatenate method.

Tuesday, 15 April 2025

Adding plugins to use for Semantic Kernel

With Microsoft Semantic Kernel, it is possible to consume AI services such as Azure AI and OpenAI with less code, as this framework provides simplification and standardization for consuming these services. A repo on Github with the code shown in this article is provided here :

https://github.com/toreaurstadboss/SemanticKernelPluginDemov4

The demo code is a Blazor server app. It demonstrates how to use Microsoft Semantic Kernel with plugins. I have decided to provide the Northwind database as the extra data the plugin will use. Via debugging and seeing the output, I see that the plugin is successfully called and used. It is also easy to add plugins, which provides additional data to the AI model. This is suitable for providing AI powered solutions with private data that you want to provide to the AI model. For example, when using OpenAI Chat GPT-4, providing a plugin will make it possible to specify which data are to be presented and displayed. It is a convenient way to provide a natural language interface for doing data reporting such as listing up results from this plugin. The plugin can provide kernel functions, using attributes on the method. Let's first look at the Program.cs file for wiring up the semantic kernel for a Blazor Server demo app.

Program.cs



using Microsoft.EntityFrameworkCore;
using Microsoft.SemanticKernel;
using SemanticKernelPluginDemov4.Models;
using SemanticKernelPluginDemov4.Services;


namespace SemanticKernelPluginDemov4
{
    public class Program
    {
        public static void Main(string[] args)
        {
            var builder = WebApplication.CreateBuilder(args);

            // Add services to the container.
            builder.Services.AddRazorPages();
            builder.Services.AddServerSideBlazor();

            // Add DbContext
            builder.Services.AddDbContextFactory<NorthwindContext>(options =>
                options.UseSqlServer(builder.Configuration.GetConnectionString("DefaultConnection")));

            builder.Services.AddScoped<IOpenAIChatcompletionService, OpenAIChatcompletionService>();

            builder.Services.AddScoped<NorthwindSemanticKernelPlugin>();

            builder.Services.AddScoped(sp =>
            {
                var kernelBuilder = Kernel.CreateBuilder();
                kernelBuilder.AddOpenAIChatCompletion(modelId: builder.Configuration.GetSection("OpenAI").GetValue<string>("ModelId")!,
                    apiKey: builder.Configuration.GetSection("OpenAI").GetValue<string>("ApiKey")!);

                var kernel = kernelBuilder.Build();

                var dbContextFactory = sp.GetRequiredService<IDbContextFactory<NorthwindContext>>();
                var northwindSemanticKernelPlugin = new NorthwindSemanticKernelPlugin(dbContextFactory);
                kernel.ImportPluginFromObject(northwindSemanticKernelPlugin);

                return kernel;
            });

            var app = builder.Build();

            // Configure the HTTP request pipeline.
            if (!app.Environment.IsDevelopment())
            {
                app.UseExceptionHandler("/Error");
                // The default HSTS value is 30 days. You may want to change this for production scenarios, see https://aka.ms/aspnetcore-hsts.
                app.UseHsts();
            }

            app.UseHttpsRedirection();

            app.UseStaticFiles();

            app.UseRouting();

            app.MapBlazorHub();
            app.MapFallbackToPage("/_Host");

            app.Run();
        }
    }
}



In the code above, note the following:
  • The usage of IDbContextFactory for creating a db context, injected into the plugin. This is a Blazor server app, so this service is used to create db contet, since a Blazor server will have a durable connection between client and the server over Signal-R and there needs to use this interface to create dbcontext instances as needed
  • Using the method ImportPluginFromObject to import the plugin into the semantic kernel built here. Note that we register the kernel as a scoped service here. Also the plugin is registered as a scoped service here.
The plugin looks like this.

NorthwindSemanticKernelplugin.cs



using Microsoft.EntityFrameworkCore;
using Microsoft.SemanticKernel;
using SemanticKernelPluginDemov4.Models;
using System.ComponentModel;

namespace SemanticKernelPluginDemov4.Services
{

    public class NorthwindSemanticKernelPlugin
    {
        private readonly IDbContextFactory<NorthwindContext> _northwindContext;

        public NorthwindSemanticKernelPlugin(IDbContextFactory<NorthwindContext> northwindContext)
        {
            _northwindContext = northwindContext;
        }

        [KernelFunction]
        [Description("When asked about the suppliers of Nortwind database, use this method to get all the suppliers. Inform that the data comes from the Semantic Kernel plugin called : NortwindSemanticKernelPlugin")]
        public async Task<List<string>> GetSuppliers()
        {
            using (var dbContext = _northwindContext.CreateDbContext())
            {
                return await dbContext.Suppliers.OrderBy(s => s.CompanyName).Select(s => "Kernel method 'NorthwindSemanticKernelPlugin:GetSuppliers' gave this: " + s.CompanyName).ToListAsync();
            }
        }

        [KernelFunction]
        [Description("When asked about the total sales of a given month in a year, use this method. In case asked for multiple months, call this method again multiple times, adjusting the month and year as provided. The month and year is to be in the range 1-12 for months and for year 1996-1998. Suggest for the user what the valid ranges are in case other values are provided.")]
        public async Task<decimal> GetTotalSalesInMontAndYear(int month, int year)
        {
            using (var dbContext = _northwindContext.CreateDbContext())
            {
                var sumOfOrders = await (from Order o in dbContext.Orders
                             join OrderDetail od in dbContext.OrderDetails on o.OrderId equals od.OrderId
                             where o.OrderDate.HasValue && (o.OrderDate.Value.Month == month
                             && o.OrderDate.Value.Year == year) 
                             select (od.UnitPrice * od.Quantity) * (1 - (decimal)od.Discount)).SumAsync();

                return sumOfOrders;
            }
        }

    }
}


In the code above, note the attributes used. KernelFunction tells that this is a method the Semantic kernel can use. The description attribute instructs the AI LLM model how to use the method, how to provide parameter values if any and when the method is to be called. Let's look at the OpenAI service next.

OpenAIChatcompletionService.cs



using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.ChatCompletion;
using Microsoft.SemanticKernel.Connectors.OpenAI;

namespace SemanticKernelPluginDemov4.Services
{

    public class OpenAIChatcompletionService : IOpenAIChatcompletionService
    {
        private readonly Kernel _kernel;

        private IChatCompletionService _chatCompletionService;

        public OpenAIChatcompletionService(Kernel kernel)
        {
            _kernel = kernel;
            _chatCompletionService = kernel.GetRequiredService<IChatCompletionService>();
        }

        public async IAsyncEnumerable<string?> RunQuery(string question)
        {
            var chatHistory = new ChatHistory();

            chatHistory.AddSystemMessage("You are a helpful assistant, answering only on questions about Northwind database. In case you got other questions, inform that you only can provide questions about the Northwind database. It is important that only the provided Northwind database functions added to the language model through plugin is used when answering the questions. If no answer is available, inform this.");

            chatHistory.AddUserMessage(question);

            await foreach (var chatUpdate in _chatCompletionService.GetStreamingChatMessageContentsAsync(chatHistory, CreateOpenAIExecutionSettings(), _kernel))
            {
                yield return chatUpdate.Content;
            }
        }

        private OpenAIPromptExecutionSettings? CreateOpenAIExecutionSettings()
        {
            return new OpenAIPromptExecutionSettings
            {
                ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions
            };
        }

    }
}



In the code above, the kernel is injected into this service. The kernel was registered in Program.cs as a scoped service, so it is injected here. The method GetRequiredService is similar to the method with same name of IServiceProvider used inside Program.cs. Note the use of ToolCallBehavior set to AutoInvokeKernelFunctions. Extending the AI powered functionality with plugins requires little extra code with Microsoft Semantic kernel. A screenshot of the demo is shown below.

Monday, 31 March 2025

Generating Dall-e-3 images using Microsoft Semantic Kernel

In this demo, Dall-e-3 images are generated from a console app using Microsoft Semantic Kernel. The semantic kernel is a library that offers different plugins for different AI services. It is supported for multiple languages, these are C#, Java and Python. Its goal is to ease the use of consuming AI services and building a shared infrastructure for these services and offer a way to conceptualize and abstract the consumption of these services. It can also be seen as a middleware for the services and offering a framework where consuming AI services becomes a more standardized process. A Github repo has been created with the code for this demo here:

Github repo for this demo
Dall-e-3 image generator with semantic kernel

The demo contains two steps, first building the semantic kernel itself and then the image generation. First off, the .csproj file has package references to the latest as of March 2025 nuget package of Microsoft Semantic Kernel.
DalleImageGeneratorWithSemanticKernel.csproj


<Project Sdk="Microsoft.NET.Sdk"> 

  <PropertyGroup>
    <OutputType>Exe</OutputType>
    <TargetFramework>net8.0</TargetFramework>
    <ImplicitUsings>enable</ImplicitUsings>
    <Nullable>enable</Nullable>
    <NoWarn>$(NoWarn);CS8618,IDE0009,CA1051,CA1050,CA1707,CA1054,CA2007,VSTHRD111,CS1591,RCS1110,RCS1243,CA5394,SKEXP0001,SKEXP0010,SKEXP0020,SKEXP0040,SKEXP0050,SKEXP0060,SKEXP0070,SKEXP0101,SKEXP0110</NoWarn>
  </PropertyGroup>

  <ItemGroup>
    <PackageReference Include="Microsoft.SemanticKernel" Version="1.44.0" />
    <PackageReference Include="Microsoft.Extensions.Configuration.Json" Version="8.0.1" />
  </ItemGroup>

</Project>


Note that multiple warnings are marked as no warning as semantic kernel is open for change in the future and thus flags multiple different warnings. The image generation demo is set up like this in the class ImageGeneration. Note how the Kernel object is built up here. It got a builder that offers many methods to add AI services. In this case we add an ITextToImageService. The modelName used here is "dall-e-3".
ImageGeneration.cs


using DalleImageGeneratorWithSemanticKernel;
using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.Connectors.OpenAI;
using Microsoft.SemanticKernel.TextToImage;
using OpenAI.Images;
using System;
using System.Diagnostics;

namespace UseSemanticKernelFromNET;

public class ImageGeneration
{
    public async Task GenerateBasicImage(string modelName)
    {
        Kernel kernel = Kernel
            .CreateBuilder()
            .AddOpenAITextToImage(modelId:modelName, apiKey: Environment.GetEnvironmentVariable("OPENAI_API_KEY")!).Build();

        ITextToImageService imageService = kernel.GetRequiredService<ITextToImageService>();

        Console.WriteLine("##### SEMANTIC KERNEL - IMAGE GENERATOR DALL-E-3 CONSOLE APP #####\n\n");


        string prompt =
           """
            In the humorous image, Vice President JD Vance and his wife are seen stepping out of their plane onto the icy runway of
            Thule Air Base. Just as they set foot on the frozen ground, a bunch of playful polar bears greet them enthusiastically, much like 
            overzealous fans welcoming celebrities. The surprised expressions on their faces are priceless as the couple finds 
            themselves being "chased" by these bundles of fur and excitement. JD Vance, with a mix of amusement and alarm, has one 
            shoe comically left behind in the snow, while his wife, holding onto her hat against the chilly wind, can't suppress a laugh.
            The scene is completed with members of the Air Base
            staff in the background, chuckling and capturing the moment on their phones, adding to the light-heartedness of the unexpected encounter.  
            The plane should carry the AirForce One Colors and read "United States of America". 
         """;

        Console.WriteLine($"\n ### STORY FOR THE IMAGE TO GENERATE WITH DALL-E-3 ### \n{prompt}\n\n");

        Console.WriteLine("\n\nStarting generation of dall-e-3 image...");

        var cts = new CancellationTokenSource();
        var cancellationToken = cts.Token;

        var rotationTask = Task.Run(() => ConsoleUtil.RotateDash(cancellationToken), cts.Token);

        var image = await imageService.GetOpenAIImageContentAsync(prompt,
            kernel: kernel,
            size: (1024, 1024), //for Dall-e-2 images, use: 256x256, 512x512, or 1024x1024. For dalle-3 images, use: 1024x1024, 1792x1024, 1024x1792. 
            style: "vivid",
            quality: "hd", //high
            responseFormat: "b64_json", // bytes
            cancellationToken: cancellationToken);       
        
        cts.Cancel(); //cancel to stop animating the waiting indicator

        var imageTmpFilePng = Path.ChangeExtension(Path.GetTempFileName(), "png");
        image?.FirstOrDefault()?.WriteToFile(imageTmpFilePng);

        Console.WriteLine($"Wrote image to location: {imageTmpFilePng}");

        Process.Start(new ProcessStartInfo
        {
            FileName = "explorer.exe",
            Arguments = imageTmpFilePng,
            UseShellExecute = true
        });

    }

}


A helper extension method has been added for the Open AI Dall-e-3 image creation. Please note that one should stick to not too many extension methods of semantic kernel itself as this defeats the purpose of a standardized way of using the semantic kernel. But in this case, it is just a helper method to customize the generation of particularly dall-e-3 (and dall-e-2) images from Open AI using the Semantic kernel. The code is shown below
TextToImageServiceExtensions.cs


using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.Connectors.OpenAI;
using Microsoft.SemanticKernel.Services;
using Microsoft.SemanticKernel.TextToImage;

namespace UseSemanticKernelFromNET;

public static class TextToImageServiceExtensions
{


    /// <summary>
    /// Generates OpenAI image content asynchronously based on the provided text input and settings.
    /// </summary>
    /// <param name="imageService">The image service used to generate the image content.</param>
    /// <param name="input">The text input used to generate the image.</param>
    /// <param name="kernel">An optional kernel instance for additional processing.</param>
    /// <param name="size">
    /// The desired size of the generated image. For DALL-E 2 images, use: 256x256, 512x512, or 1024x1024. 
    /// For DALL-E 3 images, use: 1024x1024, 1792x1024, or 1024x1792.
    /// </param>
    /// <param name="style">The style of the image. Must be "vivid" or "natural".</param>
    /// <param name="quality">The quality of the image. Must be "standard", "hd", or "high".</param>
    /// <param name="responseFormat">
    /// The format of the response. Must be one of the following: "url", "uri", "b64_json", or "bytes".
    /// </param>
    /// <param name="cancellationToken">A token to monitor for cancellation requests.</param>
    /// <returns>
    /// A task that represents the asynchronous operation. The task result contains a read-only list of 
    /// <see cref="ImageContent"/> objects representing the generated images.
    /// </returns>
    public static Task<IReadOnlyList<ImageContent>> GetOpenAIImageContentAsync(this ITextToImageService imageService,
        TextContent input,
        Kernel? kernel = null,
        (int width, int height) size = default((int, int)), // for Dall-e-2 images, use: 256x256, 512x512, or 1024x1024. For dalle-3 images, use: 1024x1024, 1792x1024, 1024x1792. 
        string style = "vivid",
        string quality = "hd",
        string responseFormat = "b64_json",        
        CancellationToken cancellationToken = default)
    {
        
        string? currentModelId = imageService.GetModelId();

        if (currentModelId != "dall-e-3" && currentModelId != "dall-e-2")
        {
            throw new NotSupportedException("This method is only supported for the DALL-E 2 and DALL-E 3 models.");
        }

        if (size.width == 0 || size.height == 0)
        {
            size = (1024, 1024); //defaulting here to (1024, 1024).
        }

        if (currentModelId == "dall-e-2"){
            var supportedSizes = new[]{
                (256, 256),
                (512, 512),
                (1024, 1024)
            };
            if (!supportedSizes.Contains(size))
            {
                throw new ArgumentException("For DALL-E 2, the size must be one of: 256x256, 512x512, or 1024x1024.");
            }
        }
        else if (currentModelId == "dall-e-3")
        {
            var supportedSizes = new[]{
                (1024, 1024),
                (1792, 1024),
                (1024, 1792)
            };
            if (!supportedSizes.Contains(size))
            {
                throw new ArgumentException("For DALL-E 3, the size must be one of: 256x256, 512x512, or 1024x1024.");
            }
        }

        return imageService.GetImageContentsAsync(
            input,
            new OpenAITextToImageExecutionSettings
                {
                    Size = size,
                    Style = style, //must be "vivid" or "natural"
                    Quality = quality, //must be "standard" or "hd" or "high"
                    ResponseFormat = responseFormat // url or uri or b64_json or bytes
                },
            kernel,
            cancellationToken);

    }
}


Screenshot of this demo, console app running:
The console app will generate the dall-e-3 image using OpenAI service for this and save the image as a PNG image and save it into file saved into a temporary location and then open this image using Windows default image viewer application. Example image generated :

Saturday, 22 March 2025

Image classification using ML.NET Machine Learning

I added a demo using ML.Net in a Github. The demo is available in this repository :

https://github.com/toreaurstadboss/ImageClassificationMLNetBlazorDemo

A screenshot shows the application running below :

ML.Net is Microsoft's machine learning library. It is combined with tooling inside VS 2022 an easy way to locally use machine learning models on your CPU or GPU, or hosted in Azure cloud services. The website for ML.Net is available here for more information about ML.Net and documentation:

https://dotnet.microsoft.com/en-us/apps/ai/ml-dotnet

In the demo above I have trained the model to recognize either horses or mooses. These species are both mammals and herbivores and somewhat are similar in appearance. I have trained the machine learning model in this demo only with ten images of each category, then again with ten other test images that checks if the model recognizes correctly if we see a horse or a moose. Already with just ten images, it did not miss once, and of course a better example for a real world machine learning model would have scoured over tens of thousand of images to handle all edge cases. ML.Net is very easy to run, it can be run locally on your own machine, using the CPU or GPU. The GPU must be CUDA compatible. That actually means you need a NVIDIA card with 8-series. I got such a card on a laptop of mine and have tested it. The following links points to download pages of NVIDIA for downloading the necessary software as of March 2025 to run ML.Net image classification functionality on GPUs :

Download Cuda 10.1

Cuda 10.1 can be downloaded from here: https://developer.nvidia.com/cuda-10.1-download-archive-base

CuDnn 7.6.4

CuDnn can be downloaded from here: https://developer.nvidia.com/rdp/cudnn-archive

Getting started with image classification using ML.Net

It is easiest to use VS 2022 to add a ML.Net machine learning model. Inside VS 2022, right click your project and choose Add and choose Machine Learning Model In case you do not see this option, hit the start menu and type in Visual Studio installer Now, hit the button Modify for your VS installation. Choose Individual Components Search for 'ml'. Select the ML.NET Model Builder. There are also a package called ML.NET Model Builder 2022, I also chose that.
Choosing the scenario
Now, after adding the Machine Learning model, the first page asks for a scenario. I choose Image Classification here, below Computer Vision scenario category.

Choosing the environment
Then I hit the button Local. In the next step, I select Local (CPU). Note that I have tested also Nvidia Cuda-compatible graphics card / GPU on another laptop and it also worked great and should be preferred if you have a GPU compatible and have installed Cuda 10.1 and Cdnn 7.6.4 as shown in links above.

Hit the button Next Step.
Choosing the Data
It is time to train the machine learning model with data ! I have gathered ten sample images of mooses and horses each. By pointing to a folder with images where each category of images are gathered in subfolders of this folder.

Next step is Train
Training the model
Here you can hit the button Train again. When you have trained enough here the model, you can hit the button Next step . Training the machine learning will take some time depending on you using CPU or GPU and the number of input images here. Usually it takes a few seconds, but not many minutes to churn through a couple of images as shown here, 20 images in total.

Loading up the image data and using the machine learning model

Note that ML.Net demands support to renderinteractive rendering of web apps, pure Blazor WASM apps are not supported. The following file shows how the Blazor serverside app is set up.

Program.cs

using ImageClassificationMLNetBlazorDemo.Components;

var builder = WebApplication.CreateBuilder(args);

// Add services to the container.
builder.Services.AddRazorComponents()
    .AddInteractiveServerComponents();

var app = builder.Build();

// Configure the HTTP request pipeline.
if (!app.Environment.IsDevelopment())
{
    app.UseExceptionHandler("/Error", createScopeForErrors: true);
    // The default HSTS value is 30 days. You may want to change this for production scenarios, see https://aka.ms/aspnetcore-hsts.
    app.UseHsts();
}

app.UseHttpsRedirection();

app.UseStaticFiles();
app.UseAntiforgery();

app.MapRazorComponents<App>()
    .AddInteractiveServerRenderMode();

app.Run();


InteractiveServer is set up inside the App.razor using the HeadOutlet.

App.razor


<!DOCTYPE html>
<html lang="en">

<head>
    <meta charset="utf-8" />
    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
    <base href="/" />
    <link rel="stylesheet" href="app.css" />
    <link rel="stylesheet" href="lib/bootstrap/css/bootstrap.min.css" />
    <link rel="stylesheet" href="ImageClassificationMLNetBlazorDemo.styles.css" />
    <link rel="icon" type="image/png" href="favicon.png" />

    <HeadOutlet @rendermode="InteractiveServer" />
</head>

<body>
    <Routes @rendermode="InteractiveServer" />
    <script src="_framework/blazor.web.js"></script>
</body>

</html>


The following codebehind of the razor component Home.razor in the demo repo shows how a file uploaded using the InputFile control in Blazor serverside. Home.razor.cs


@code {

    private string? _base64ImageSource = null;
    private string? _predictedLabel = "No classification";
    private IOrderedEnumerable<KeyValuePair<string, float>>? _predictedLabels = null;
    private int? _assessedPredictionQuality = null;
    private string? _errorMessage = null;

    private async Task LoadFileAsync(InputFileChangeEventArgs e)
    {
        try
        {
            ResetPrivateFields();

            if (e.File.Size <= 0 || e.File.Size >= 2 * 1024 * 1024)
            {
                _errorMessage = "Sorry, the uploaded image but be between 1 byte and 2 MB!";
                return;
            }

            byte[] imageBytes = await GetImageBytes(e.File);
            _base64ImageSource = GetBase64ImageSourceString(e.File.ContentType, imageBytes);

            PredictImageClassification(imageBytes);

        }
        catch (Exception err)
        {
            Console.WriteLine(err);
        }
    }

    private void ResetPrivateFields()
    {
        _base64ImageSource = null;
        _predictedLabel = null;
        _predictedLabels = null;
        _assessedPredictionQuality = null;
    }

    private int GetAssesPrediction()
    {
        int result = 1;
        if (_predictedLabel != null && _predictedLabels != null)
        {
            foreach (var label in _predictedLabels)
            {
                if (label.Key == _predictedLabel)
                {
                    result = label.Value switch
                    {
                        <= 0.50f => 1,
                        <= 0.70f => 2,
                        <= 0.80f => 3,
                        <= 0.85f => 4,
                        <= 0.90f => 5,
                        <= 1.0f => 6,
                        _ => 1 //default to dice we get some other score here..
                    };
                }
            }
        }

        return result;
    }

    private void PredictImageClassification(byte[] imageBytes)
    {

        var input = new ModelInput
            {
                ImageSource = imageBytes
            };
        ModelOutput output = HorseOrMooseImageClassifier.Predict(input);
        _predictedLabel = output.PredictedLabel;

        _predictedLabels = HorseOrMooseImageClassifier.PredictAllLabels(input);

        _assessedPredictionQuality = GetAssesPrediction(); //check how good the prediction is, give a score from 1-6 (dice score!)

        StateHasChanged();
    }


    private async Task<byte[]> GetImageBytes(IBrowserFile file) 
    {
        using MemoryStream memoryStream = new();
        var stream = file.OpenReadStream(2 * 1024 * 1024, CancellationToken.None);
        await stream.CopyToAsync(memoryStream);
        return memoryStream.ToArray();
    }

    private string GetBase64ImageSourceString(string contentType, byte[] bytes)
    {
        string preAmble = $"data:{contentType};base64,";
        return $"{preAmble}{(Convert.ToBase64String(bytes))}";
    }
}


As the code shows above, using the machine learning model is quite convenient, we just use the methods Predict to get the Label that is decided exists in the loaded image. This is the image classiciation that the machine learning found. Note that using the method PredictAllLabels get the confidence of the different labels show in this demo. There are no limitations on the number of categories here in the image classification labels that one could train a model to look after. A benefit with ML.Net is the option to use it on-premise servers and get fairly good result on just a few sample images. But the more sample images you obtain for a label, the more precise the machine learning model will become. It is possible to download a pre-trained model such as Inceptionv3 that is compatible with Tensorflow used here that supports up to 1000 categories. More information is available here from Microsoft about using a pre-trained model such as InceptionV3:

https://learn.microsoft.com/en-us/dotnet/machine-learning/tutorials/image-classification