ODModelBuilderTF

Object detection models builder with TensorFlow.

.NET (C#) object detection models builder with TensorFlow.

ODModelBuilderTF is a .NET library which allow to train object detection models directly in .NET environment, without the needing of Python code.

• It's based on the TensorFlow framework.
  
• It can be used with all .NET languages simply including the package on your project.
  
• It's also event oriented, so it's possible to control the whole train process registering the events in the client application.
  
• Tuning parameters are provided to manage the train process directly from the client without manually writing configuration files.
  
• Can train and export TensorFlow's saved model format, frozen graph and ONNX models that can be afterwards consumed, for example, in the ML.NET framework.

Getting started with ODModelBuilderTF

Simply include the package (or the reference to the project if you include it in your solution) to used the library.
At the first initialization it will install the train environment on your device downloading it from Internet (the process can take long time).
If you prefer to have a ready to use environment (or your device is offline) you can include in your application the redistributables containing all the needed resources.

Sample apps

For a quick usage example you can take a look to the console application included in this repository, used to test the library.

Packages

LG.ODModelBuilderTF: the train library.
LG.ODModelBuilderTF-Redist-Win: the redistributable with TensorFlow object detection packages.
LG.ODModelBuilderTF-Redist-Win-TF: the redistributable with TensorFlow packages.

• LG.ODModelBuilderTF-Redist-Win-TF-A: the redistributable with TensorFlow library (part A).
  
• LG.ODModelBuilderTF-Redist-Win-TF-B: the redistributable with TensorFlow library (part B).
  

LG.ODModelBuilderTF-Redist-Win-CUDA10_1-TF: the redistributable with a special TensorFlow library for old GPUs which can work with CUDA 10.1 SM30. To use instead of the ODModelBuilderTF-Redist-Win-TF (CUDA 11) in old devices.

Simple installation with environment download at the first run:

LG.ODModelBuilderTF

Installation with full environment:

LG.ODModelBuilderTF, LG.ODModelBuilderTF-Redist-Win, LG.ODModelBuilderTF-Redist-Win-TF

Installation on old GPUs:

LG.ODModelBuilderTF, LG.ODModelBuilderTF-Redist-Win, LG.ODModelBuilderTF-Redist-Win-CUDA10_1-TF

Building ODModelBuilderTF

It's possible to build ODModelBuilderTF and the packages directly from the command line launching the build.cmd or opening the solution in Visual Studio 2019 or above.

Code examples

Here is a snippet code to show how to train a model.

using ODModelBuilderTF;
using System;
using System.Threading;
using System.Threading.Tasks;

namespace ODModelBuilderTF_Con
{
   class Program
   {
      static async Task Main(string[] args)
      {
         // Create the trainer
         var trainer = new Trainer(new Trainer.Options
         {
            BatchSize = 16,
            CheckpointEvery = null,
            CheckpointForceThreashold = null,
            EvalImagesFolder = @"Data\Images\Eval",
            ExportFolder = @"Data\Export",
            ModelType = ModelTypes.SSD_MobileNet_V2_320x320,
            NumTrainSteps = 50000,
            OnnxModelFileName = "Model.onnx",
            TensorBoardPort = 6006,
            TrainFolder = @"Data\Train",
            TrainImagesFolder = @"Data\Images\Train",
            TrainRecordsFolder = @"Data\Records",
         });
         // Step event
         var loss = default(double?);
         trainer.TrainStep += (sender, e) =>
         {
            Console.WriteLine($"Step number:{e.StepNumber}\t\tstep time: {e.StepTime:N3} secs\t\ttotal loss:{e.TotalLoss:N3}");
            if (loss == null)
               loss = e.TotalLoss;
            else if (e.TotalLoss < loss) {
               loss = e.TotalLoss;
               e.CreateCheckpoint = true;
               Console.WriteLine($"{new string('=', 40)}> Create checkpoint with total loss {e.TotalLoss}");
            }
         };
         // Evaluation event
         var mAP = default(double?);
         trainer.TrainEvaluation += (sender, e) =>
         {
            if (mAP == null)
               mAP = e.AP;
            else if (e.AP > mAP) {
               mAP = e.AP;
               e.Export = true;
            }
            Console.WriteLine($"{new string('=', 40)}> Mean average precision {e.AP}");
            if (e.Export)
               Console.WriteLine($"{new string('=', 40)}> Exporting model...");
         };
         // Export events
         trainer.ExportedSavedModel += (sender, e) =>
         {
            Console.WriteLine($"{new string('=', 40)}> SavedModel exported");
         };
         trainer.ExportedOnnx += (sender, e) =>
         {
            Console.WriteLine($"{new string('=', 40)}> Onnx model exported");
         };
         // Token for stop training
         var exitToken = new CancellationTokenSource();
         Console.CancelKeyPress += (sender, e) => exitToken.Cancel();
         // Wait the training
         await Task.Run(() => trainer.Train(exitToken.Token));
      }
   }
}

Colab notebook

It's possible to test / train directly on a google colab environment using the notebook ODModelBuilderTF.ipynb

License

ODModelBuilderTF is licensed under the MIT license and it is free to use commercially.