This article is about a Microsoft C# 4.0 WPF implementation of a framework that allows to create, train, and test convolutional neural networks against the MNIST dataset of handwritten digits or the CIFAR-10 dataset of 10 different natural objects. There is a magnificent article by Mike O'Neill on the The Code Project about the same subject. Without his great article and C++ demo code, this project wouldn't exist. I also relied heavily on Dr. Yann LeCun's paper: Gradient-Based Learning Applied to Document Recognition to understand more about the principles of convolutional neural networks and the reason why they are so successful in the area of machine vision. Mike O'Neill uses Patrice Simard's implementation where the subsampling step is integrated in the structure of the convolutional layer itself. Dr. Yann LeCun uses in his LeNet-5 a separate subsampling step, and also uses non-fully connected layers. The framework presented allows to use all types of layers, and has an additional Max-Pooling layer that you can use instead of plain Average-Pooling. The default squashing function used is
tanh() and the value to train for is set to 0.8 because it is the value at the curvature of the second derivative of the used non-linearity so there is less saturation. The input images are all normalised [-1,1] and the input layer is at a fixed 32x32 window.
The main goal of this project was to build an enhanced and extended version of Mike O'Neill's excellent C++ project. This time written in C# 4.0 and using WPF with a simple MVVM pattern as the GUI instead of Windows Forms. I've included and used the WPF TaskDialog Wrapper from Sean A. Hanley instead of the Windows API Code Pack because the first is more compact and fit my needs perfectly. Also the Extended WPF Toolkit is used. For unzipping the CIFAR-10 dataset I used the open-source SharpDevelop SharpZipLib module. So Visual Studio 2010 and Windows Vista SP2 are the minimum requirements to use my application or just the operating system if you choose just the setup. I also made maximal use of the parallel functionality offered in C# 4.0 by letting the user at all times choose how many logical cores are used in the parallel optimised code parts with a simple manipulation of the sliderbar next to the View combobox.
Using the code
Here is the example code to construct a LeNet-5 network in my code (see the
InitializeDefaultNeuralNetwork() function in MainViewWindows.xaml.cs):
NeuralNetworks network = new NeuralNetworks("LeNet-5", 0.8D, LossFunctions.MeanSquareError, DataProviderSets.MNIST, 0.02D);
network.Layers.Add(new Layers(network, LayerTypes.Input, 1, 32, 32));
network.Layers.Add(new Layers(network, LayerTypes.Convolutional,ActivationFunctions.Tanh, 6, 28, 28, 5, 5));
network.Layers.Add(new Layers(network, LayerTypes.Subsampling, ActivationFunctions.AveragePoolingTanh, 6, 14, 14, 2, 2));
List<bool> mapCombinations = new List<bool>(16 * 6)
true, false,false,false,true, true, true, false,false,true, true, true, true, false,true, true,
true, true, false,false,false,true, true, true, false,false,true, true, true, true, false,true,
true, true, true, false,false,false,true, true, true, false,false,true, false,true, true, true,
false,true, true, true, false,false,true, true, true, true, false,false,true, false,true, true,
false,false,true, true, true, false,false,true, true, true, true, false,true, true, false,true,
false,false,false,true, true, true, false,false,true, true, true, true, false,true, true, true
network.Layers.Add(new Layers(network, LayerTypes.Convolutional, ActivationFunctions.Tanh, 16, 10, 10, 5, 5, new Mappings(network, 2, mapCombinations)));
network.Layers.Add(new Layers(network, LayerTypes.Subsampling, ActivationFunctions.AveragePoolingTanh, 16, 5, 5, 2, 2));
network.Layers.Add(new Layers(network, LayerTypes.Convolutional, ActivationFunctions.Tanh, 120, 1, 1, 5, 5));
network.Layers.Add(new Layers(network, LayerTypes.FullyConnected, ActivationFunctions.Tanh, 10));
This is Design view where you can see how the network is defined and see the weights of all the layers. When you hover with the mouse over a single weight, a tooltip shows the corresponding weight or bias value. You can always refresh the weights graphic if you have changed the block size so you can see it in the prefered size.
This is Training view where you train the network. The 'Play' button gives you the 'Select Training Parameters' dialog where you can define the basic training parameters. The 'Training Scheme Editor' button gives you the possibility to fully define your own training schemes to experiment with. At any time, the training can be easily aborted by pressing the 'Stop' button. The 'Star' button will reset (forget) all the weight values.
In Testing view, you can test your network and get a graphical confusion matrix that represents all the misses.
In Calculate view, we can test a single digit or object with the desired properties and fire it through the network and get a graphical view of all the output values in every layer.
I would love to see a DirectCompute 5.0 integration for offloading the highly parallel task of learning the neural network to a DirectX 11 compliant GPU if one is available. But I've never programmed with DirectX or any other shader based language before, so if there's anyone out there with some more experience in this area, any help is very welcome. I made an attempt to use a simple MVVM structure in this WPF application. In the Model folder, you can find the files for the neural network class and also a
DataProvider class which deals with loading and providing the necessary MNIST and CIFAR-10 training and testing samples. There is also a
NeuralNetworkDataSet class that is used by the project to load and save neural network definitions, weights, or both (full) from or to a file on disk. Then there is the View folder that contains the four different PageViews in the project and a global PageView which acts as a container for the different views (Design, Training, Testing, and Calculate). In the ViewModel folder, you will find a
PageViewModelBase class where the corresponding four ViewModels are derived from. All the rest is found in the MainViewWindows.xaml.cs class. Hope there's someone out there who can actually use this code and improve on it. Extend it with an unsupervised learning stage for example (encoder/decoder construction), or implement a better loss-function (negative log likelihood instead of MSE); extend to more test databases; make use of more advanced squashing functions, etc.
- Better garbage collection when switching between networks.
- PageViewModelBase.cs and the classes wich derive from it are cleanend from some unnecessary code.
- Refactoring & small fixes.
- Reduced memory usage for every dataset.
- Fixed: Download of MNIST dataset.
- Fixed: Training Scheme Editor works now for the CIFAR-10 dataset.
- Speed improvements in training the CNN.
- Speed improvements in creating the Design & Calculate graphic.
- Memory consumption reduced.
- Loading the CIFAR-10 dataset is now much faster.
- The performance of Design View is now better optimised for bigger networks.
- It's now possible to adjust the block size of the weight and output values graphic.
- In Design View you can refresh the weights graphic to the current block size.
- Performance improvements in training networks
- Performance improvement in displaying Design View. (still to slow for big networks)
- Minor GUI changes
- Now all the fully connected layers are displayed in Calculate View.
- Changing the background color is working properly now.
- Now you can easily reset the weights values in Training View.
- By using Max-Pooling with the CIFAR-10 dataset the results are much better. I've also horizontal flipped each training pattern to double the size of the training set.
- Some minor fixes.
- The CIFAR-10 Dataset of 10 natural objects in color is now fully supported.
- The weights in Design View are now correctly displayed. (still slow on big networks)
- The file format used to save and load weights, definitions, etc is changed and incompatible with previous versions.
- Now you can see all the weight and bias values in every layer.
- Renaming some items so they make more sense (KernelTypes.Sigmoid => ActivationFunctions.Tanh)
- As a last layer you can use LeCun's RBF layer with fixed weights.
- Now it is possible to uses ActivationFunctions.AbsTanh to have a rectified convolutional layer.
- Initial release